Docs

1 - Getting Started

Welcome to the Gardener Getting Started section! Here you will be able to get accustomed to the way Gardener functions and learn how its components work together in order to seamlessly run Kubernetes clusters on various hyperscalers.

The following topics aim to be useful to both complete beginners and those already somewhat familiar with Gardener. While the content is structured, with Introduction serving as the starting point, if you’re feeling confident in your knowledge, feel free to skip to a topic you’re more interested in.

1.1 - Introduction to Gardener

Problem Space

Let’s discuss the problem space first. Why does anyone need something like Gardener?

Running Software

The starting point is this rather simple question: Why would you want to run some software?

Typically, software is run with a purpose and not just for the sake of running it. Whether it is a digital ledger, a company’s inventory or a blog - software provides a service to its user.

Which brings us to the way this software is being consumed. Traditionally, software has been shipped on physical / digital media to the customer or end user. There, someone had to install, configure, and operate it. In recent times, the pattern has shifted. More and more solutions are operated by the vendor or a hosting partner and sold as a service ready to be used.

But still, someone needs to install, configure, and maintain it - regardless of where it is installed. And of course, it will run forever once started and is generally resilient to any kind of failures.

For smaller installations things like maintenance, scaling, debugging or configuration can be done in a semi-automatic way. It’s probably no fun and most importantly, only a limited amount of instances can be taken care of - similar to how one would take care of a pet.

But when hosting services at scale, there is no way someone can do all this manually at acceptable costs. So we need some vehicle to easily spin up new instances, do lifecycle operations, get some basic failure resilience, and more. How can we achieve that?

Solution Space 1 - Kubernetes

Let’s start solving some of the problems described earlier with Container technology and Kubernetes.

Containers

Container technology is at the core of the solution space. A container forms a vehicle that is shippable, can easily run in any supported environment and generally adds a powerful abstraction layer to the infrastructure.

However, plain containers do not help with resilience or scaling. Therefore, we need another system for orchestration.

Orchestration

“Classical” orchestration that just follows the “notes” and moves from state A to state B doesn’t solve all of our problems. We need something else.

Kubernetes operates on the principle of “desired state”. With it, you write a construction plan, then have controllers cycle through “observe -> analyze -> act” and transition the actual to the desired state. Those reconciliations ensure that whatever breaks there is a path back to a healthy state.

Summary

Containers (famously brought to the mainstream as “Docker”) and Kubernetes are the ingredients of a fundamental shift in IT. Similar to how the Operating System layer enabled the decoupling of software and hardware, container-related technologies provide an abstract interface to any kind of infrastructure platform for the next-generation of applications.

Solution Space 2 - Gardener

So, Kubernetes solves a lot of problems. But how do you get a Kubernetes cluster?

Either:

• Buy a cluster as a service from an external vendor
• Run a Gardener instance and host yourself a cluster with its help

Essentially, it was a “make or buy” decision that led to the founding of Gardener.

The Reason Why We Choose to “Make It”

Gardener allows to run Kubernetes clusters on various hyperscalers. It offers the same set of basic configuration options independent of the chosen infrastructure. This kind of harmonization supports any multi-vendor strategy while reducing adoption costs for the individual teams. Just imagine having to deal with multiple vendors all offering vastly different Kubernetes clusters.

Of course, there are plenty more reasons - from acquiring operational knowledge to having influence on the developed features - that made the pendulum swing towards “make it”.

What exactly is Gardener?

Gardener is a system to manage Kubernetes clusters. It is driven by the same “desired state” pattern as Kubernetes itself. In fact, it is using Kubernetes to run Kubernetes.

A user may “desire” clusters with specific configuration on infrastructures such as GCP, AWS, Azure, Alicloud, Openstack, vsphere, … and Gardener will make sure to create such a cluster and keep it running.

If you take this rather simplistic principle of reconciliation and add the feature-richness of Gardener to it, you end up with universal Kubernetes at scale.

Whether you need fleet management at minimal TCO or to look for a highly customizable control plane - we have it all.

On top of that, Gardener-managed Kubernetes clusters fulfill the conformance standard set out by the CNCF and we submit our test results for certification.

Have a look at the CNCF map for more information or dive into the testgrid directly.

Gardener itself is open-source. Under the umbrella of github.com/gardener we develop the core functionalities as well as the extensions and you are welcome to contribute (by opening issues, feature requests or submitting code).

Last time we counted, there were already 131 projects. That’s actually more projects than members of the organization.

As of today, Gardener is mainly developed by SAP employees and SAP is an “adopter” as well, among STACKIT, Telekom, Finanz Informatik Technologie Services GmbH and others. For a full list of adopters, see the Adopters page.

1.2 - Architecture

Kubeception

Kubeception - Kubernetes in Kubernetes in Kubernetes

In the classic setup, there is a dedicated host / VM to host the master components / control plane of a Kubernetes cluster. However, these are just normal programs that can easily be put into containers. Once in containers, Kubernetes Deployments and StatefulSets (for the etcd) can be made to watch over them. And by putting all that into a separate, dedicated Kubernetes cluster you get Kubernetes on Kubernetes, aka Kubeception (named after the famous movie Inception with Leonardo DiCaprio).

But what are the advantages of running Kubernetes on Kubernetes? For one, it makes use of resources more reasonably. Instead of providing a dedicated computer or virtual machine for the control plane of a Kubernetes cluster - which will probably never be the right size but either too small or too big - you can dynamically scale the individual control plane components based on demand and maximize resource usage by combining the control planes of multiple Kubernetes clusters.

In addition to that, it helps introducing a first layer of high availability. What happens if the API server suddenly stops responding to requests? In a traditional setup, someone would have to find out and manually restart the API server. In the Kubeception model, the API server is a Kubernetes Deployment and of course, it has sophisticated liveness- and readiness-probes. Should the API server fail, its liveness-probe will fail too and the pod in question simply gets restarted automatically - sometimes even before anybody would have noticed about the API server being unresponsive.

In Gardener’s terminology, the cluster hosting the control plane components is called a seed cluster. The cluster that end users actually use (and whose control plane is hosted in the seed) is called a shoot cluster.

The worker nodes of a shoot cluster are plain, simple virtual machines in a hyperscaler (EC2 instances in AWS, GCE instances in GCP or ECS instances in Alibaba Cloud). They run an operating system, a container runtime (e.g., containerd), and the kubelet that gets configured during node bootstrap to connect to the shoot’s API server. The API server in turn runs in the seed cluster and is exposed through an ingress. This connection happens over public internet and is - of course - TLS encrypted.

In other terms: you use Kubernetes to run Kubernetes.

Cluster Hierarchy in Gardener

Gardener uses many Kubernetes clusters to eventually provide you with your very own shoot cluster.

At the heart of Gardener’s cluster hierarchy is the garden cluster. Since Gardener is 100% Kubernetes native, a Kubernetes cluster is needed to store all Gardener related resources. The garden cluster is actually nodeless - it only consists of a control plane, an API server (actually two), an etcd, and a bunch of controllers. The garden cluster is the central brain of a Gardener landscape and the one you connect to in order to create, modify or delete shoot clusters - either with kubectl and a dedicated kubeconfig or through the Gardener dashboard.

The seed clusters are next in the hierarchy - they are the clusters which will host the “kubeceptioned” control planes of the shoot clusters. For every hyperscaler supported in a Gardener landscape, there would be at least one seed cluster. However, to reduce latencies as well as for scaling, Gardener landscapes have several different seeds in different regions across the globe to keep the distance between control planes and actual worker nodes small.

Finally, there are the shoot clusters - what Gardener is all about. Shoot clusters are the clusters which you create through Gardener and which your workload gets deployed to.

Gardener Components Overview

From a very high level point of view, the important components of Gardener are:

The Gardener API Endpoint

You can connect to the Gardener API Endpoint (i.e., the API server in the garden cluster) either through the dashboard or with kubectl, given that you have a proper kubeconfig for it.

The Seeds Running the Shoot Cluster Control Planes

Inside each seed is one of the most important controllers in Gardener - the gardenlet. It spawns many other controllers, which will eventually create all resources for a shoot cluster, including all resources on the cloud providers such as virtual networks, security groups, and virtual machines.

Gardener’s API Endpoint

Kubernetes’ API can be extended - either by CRDs or by API aggregation.

API aggregation involves setting up a so called extension-API-server and registering it with the main Kubernetes API server. The extension API server will then serve resources of custom-defined API groups on its own. While the main Kubernetes API server is still used to handle RBAC, authorization, namespacing, quotas, limits, etc., all custom resources will be delegated to the extension-API-server. This is done through an APIService resource in the main API server - it specifies that, e.g., the API group core.gardener.cloud is served by a dedicated extension-API-server and all requests concerning this API group should be forwarded the specified IP address or Kubernetes service name. Extension API servers can persist their resources in their very own etcd but they do not have to - instead, they can use the main API servers etcd as well.

Gardener uses its very own extension API server for its resources like Shoot, Seed, CloudProfile, SecretBinding, etc… However, Gardener does not set up a dedicated etcd for its own extension API server - instead, it reuses the existing etcd of the main Kubernetes API server. This is absolutely possible since the resources of Gardener’s API are part of the API group gardener.cloud and thus will not interfere with any resources of the main Kubernetes API in etcd.

In case you are interested, you can read more on:

• API Extension
• API Aggregation
• APIService Resource

Gardener API Resources

Since Gardener’s API endpoint is a regular Kubernetes cluster, it would theoretically serve all resources from the Kubernetes core API, including Pods, Deployments, etc. However, Gardener implements RBAC rules and disables certain controllers that make these resources inaccessible. Objects like Secrets, Namespaces, and ResourceQuotas are still available, though, as they play a vital role in Gardener.

In addition, through Gardener’s extension API server, the API endpoint also serves Gardener’s custom resources like Projects, Shoots, CloudProfiles, Seeds, SecretBindings (those are relevant for users), ControllerRegistrations, ControllerDeployments, BackupBuckets, BackupEntries (those are relevant to an operator), etc.

1.3 - Gardener Projects

Overview

Gardener is all about Kubernetes clusters, which we call shoots. However, Gardener also does user management, delicate permission management and offers technical accounts to integrate its services into other infrastructures. It allows you to create several quotas and it needs credentials to connect to cloud providers. All of these are arranged in multiple fully contained projects, each of which belongs to a dedicated user and / or group.

Projects on YAML Level

Projects are a Kubernetes resource which can be expressed by YAML. The resource specification can be found in the API reference documentation.

A project’s specification defines a name, a description (which is a free-text field), a purpose (again, a free-text field), an owner, and members. In Gardener, user management is done on a project level. Therefore, projects can have different members with certain roles.

In Gardener, a user can have one of five different roles: owner, admin, viewer, UAM, and service account manager. A member with the viewer role can see and list all clusters but cannot create, delete or modify them. For that, a member would need the admin role. Another important role would be the uam role - members with that role are allowed to manage members and technical users for a project. The owner of a project is allowed to do all of that, regardless of what other roles might be assigned to him.

Projects are getting reconciled by Gardener’s project-controller, a component of Gardener’s controller manager. The status of the last reconcilation, along with any potential failures, will be recorded in the project’s status field.

For more information, see Projects.

In case you are interested, you can also view the source code for:

• The structure of a project API object
• Reconciling a project

Gardener Projects and Kubernetes Namespaces

Even though projects are a dedicated Kubernetes resource, every project also corresponds to a dedicated namespace in the garden cluster. All project resources - including shoots - are placed into this namespace.

You can ask Gardener to use a specific namespace name in the project manifest but usually, this field should be left empty. The namespace then gets created automatically by Gardener’s project-controller, with its name getting generated from the project’s name, prefixed by “garden-”.

ResourceQuotas - if any - will be enforced on the project namespace.

Infrastructure Secrets

For Gardener to create all relevant infrastructure that a shoot cluster needs inside a cloud provider, it needs to know how to authenticate to the cloud provider’s API. This is done through regular secrets.

Through the Gardener dashboard, secrets can be created for each supported cloud provider (using the dashboard is the preferred way, as it provides interactive help on what information needs to be placed into the secret and how the corresponding user account on the cloud provider should be configured). All of that is stored in a standard, opaque Kubernetes secret.

Inside of a shoot manifest, a reference to that secret is given so that Gardener knows which secret to use for a given shoot. Consequently, different shoots, even though they are in the same project, can be created on multiple different cloud provider accounts. However, instead of referring to the secret directly, Gardener introduces another layer of indirection called a SecretBinding.

In the shoot manifest, we refer to a SecretBinding and the SecretBinding in turn refers to the actual secret.

SecretBindings

With SecretBindings, it is possible to reference the same infrastructure secret in different projects across namespaces. This has the following advantages:​

• Infrastructure secrets can be kept in one project (and thus namespace) with limited access. Through SecretsBindings, the secrets can be used in other projects (and thus namespaces) without being able to read their contents.​
• Infrastructure secrets can be kept at one central place (a dedicated project) and be used by many other projects. This way, if a credential rotation is required, they only need to be changed in the secrets at that central place and not in all projects that reference them.

Service Accounts

Since Gardener is 100% Kubernetes, it can be easily used in a programmatic way - by just sending the resource manifest of a Gardener resource to its API server. To do so, a kubeconfig file and a (technical) user that the kubeconfig maps to are required.

Next to project members, a project can have several service accounts - simple Kubernetes service accounts that are created in a project’s namespace. Consequently, every service account will also have its own, dedicated kubeconfig and they can be granted different roles through RoleBindings.

To integrate Gardener with other infrastructure or CI/CD platforms, one can create a service account, obtain its kubeconfig and then automatically send shoot manifests to the Gardener API server. With that, Kubernetes clusters can be created, modified or deleted on the fly whenever they are needed.

1.4 - Gardener Shoots

Overview

A Kubernetes cluster consists of a control plane and a data plane. The data plane runs the actual containers on worker nodes (which translate to physical or virtual machines). For the control and data plane to work together properly, lots of components need matching configuration.

Some configurations are standardized but some are also very specific to the needs of a cluster’s user / workload. Ideally, you want a properly configured cluster with the possibility to fine-tune some settings.

Concept of a “Shoot”

In Gardener, Kubernetes clusters (with their control plane and their data plane) are called shoot clusters or simply shoots. For Gardener, a shoot is just another Kubernetes resource. Gardener components watch it and act upon changes (e.g., creation). It comes with reasonable default settings but also allows fine-tuned configuration. And on top of it, you get a status providing health information, information about ongoing operations, and so on.

Luckily there is a dashboard to get started.

Basic Configuration Options

Every cluster needs a name - after all, it is a Kubernetes resource and therefore unique within a namespace.

The Kubernetes version will be used as a starting point. Once a newer version is available, you can always update your existing clusters (but not downgrade, as this is not supported by Kubernetes in general).

The “purpose” affects some configuration (like automatic deployment of a monitoring stack or setting up certain alerting rules) and generally indicates the importance of a cluster.

Start by selecting the infrastructure you want to use. The choice will be mapped to a cloud profile that contains provider specific information like the available (actual) OS images, zones and regions or machine types.

Each data plane runs in an infrastructure account owned by the end user. By selecting the infrastructure secret containing the accounts credentials, you are granting Gardener access to the respective account to create / manage resources.

As part of the infrastructure you chose, the region for data plane has to be chosen as well. The Gardener scheduler will try to place the control plane on a seed cluster based on a minimal distance strategy. See Gardener Scheduler for more details.

Up next, the networking provider (CNI) for the cluster has to be selected. At the point of writing, it is possible to choose between Calico and Cilium. If not specified in the shoot’s manifest, default CIDR ranges for nodes, services, and pods will be used.

In order to run any workloads in your cluster, you need nodes. The worker section lets you specify the most important configuration options. For beginners, the machine type is probably the most relevant field, together with the machine image (operating system).

The machine type is provider-specific and configured in the cloud profile. Check your respective cloud profile if you’re missing a machine type. Maybe it is available in general but unavailable in your selected region.

The operating system your machines will run is the next thing to choose. Debian-based GardenLinux is the best choice for most use cases.

Other specifications for the workers include the volume type and size. These settings affect the root disk of each node. Therefore we would always recommend to use an SSD-based type to avoid i/o issues.

The autoscaler parameter defines the initial elasticity / scalability of your cluster. The cluster-autoscaler will add more nodes up to the maximum defined here when your workload grows and remove nodes in case your workload shrinks. The minimum number of nodes should be equal to or higher than the number of zones. You can distribute the nodes of a worker pool among all zones available to your cluster. This is the first step in running HA workloads.

Once per day, all clusters reconcile. This means all controllers will check if there are any updates they have to apply (e.g., new image version for ETCD). The maintenance window defines when this daily operation will be triggered. It is important to understand that there is no opt-out for reconciliation.

It is also possible to confine updates to the shoot spec to be applied only during this time. This can come in handy when you want to bundle changes or prevent changes to be applied outside a well-known time window.

You can allow Gardener to automatically update your cluster’s Kubernetes patch version and/or OS version (of the nodes). Take this decision consciously! Whenever a new Kubernetes patch version or OS version is set to supported in the respective cloud profile, auto update will upgrade your cluster during the next maintenance window. If you fail to (manually) upgrade the Kubernetes or OS version before they expire, force-upgrades will take place during the maintenance window.

Result

The result of your provided inputs and a set of conscious default values is a shoot resource that, once applied, will be acted upon by various Gardener components. The status section represents the intermediate steps / results of these operations. A typical shoot creation flow would look like this:

1. Assign control plane to a seed.
2. Create infrastructure resources in the data plane account (e.g., VPC, gateways, …)
3. Deploy control plane incl. DNS records.
4. Create nodes (VMs) and bootstrap kubelets.
5. Deploy kube-system components to nodes.

How to Access a Shoot

Static credentials for shoots were discontinued in Gardener with Kubernetes v1.27. Short lived credentials need to be used instead. You can create/request tokens directly via Gardener or delegate authentication to an identity provider.

A short-lived admin kubeconfig can be requested by using kubectl. If this is something you do frequently, consider switching to gardenlogin, which helps you with it.

An alternative is to use an identity provider and issue OIDC tokens.

What can you configure?

With the basic configuration options having been introduced, it is time to discuss more possibilities. Gardener offers a variety of options to tweak the control plane’s behavior - like defining an event TTL (default 1h), adding an OIDC configuration or activating some feature gates. You could alter the scheduling profile and define an audit logging policy. In addition, the control plane can be configured to run in HA mode (applied on a node or zone level), but keep in mind that once you enable HA, you cannot go back.

In case you have specific requirements for the cluster internal DNS, Gardener offers a plugin mechanism for custom core DNS rules or optimization with node-local DNS. For more information, see Custom DNS Configuration and NodeLocalDNS Configuration.

Another category of configuration options is dedicated to the nodes and the infrastructure they are running on. Every provider has their own perks and some of them are exposed. Check the detailed documentation of the relevant extension for your infrastructure provider.

You can fine-tune the cluster-autoscaler or help the kubelet to cope better with your workload.

Worker Pools

There are a couple of ways to configure a worker pool. One of them is to set everything in the Gardener dashboard. However, only a subset of options is presented there.

A slightly more complex way is to set the configuration through the yaml file itself.

This allows you to configure much more properties of a worker pool, like the timeout after which an unhealthy machine is getting replaced. For more options, see the Worker API reference.

How to Change Things

Since a shoot is just another Kubernetes resource, changes can be applied via kubectl. For convenience, the basic settings are configurable via the dashboard’s UI. It also has a “yaml” tab where you can alter all of the shoot’s specification in your browser. Once applied, the cluster will reconcile eventually and your changes become active (or cause an error).

Immutability in a Shoot

While Gardener allows you to modify existing shoot clusters, it is important to remember that not all properties of a shoot can be changed after it is created.

For example, it is not possible to move a shoot to a different infrastructure account. This is mainly rooted in the fact that discs and network resources are bound to your account.

Another set of options that become immutable are most of the network aspects of a cluster. On an infrastructure level the VPC cannot be changed and on a cluster level things like the pod / service cidr ranges, together with the nodeCIDRmask, are set for the lifetime of the cluster.

Some other things can be changed, but not reverted. While it is possible to add more zones to a cluster on an infrastructure level (assuming that an appropriate CIDR range is available), removing zones is not supported. Similarly, upgrading Kubernetes versions is comparable to a one-way ticket. As of now, Kubernetes does not support downgrading. Lastly, the HA setting of the control plane is immutable once specified.

Crazy Botany

Since remembering all these options can be quite challenging, here is very helpful resource - an example shoot with all the latest options 🎉

1.5 - Control Plane Components

Overview

A cluster has a data plane and a control plane. The data plane is like a space station. It has certain components which keep everyone / everything alive and can operate autonomously to a certain extent. However, without mission control (and the occasional delivery of supplies) it cannot share information or receive new instructions.

So let’s see what the mission control (control plane) of a Kubernetes cluster looks like.

Kubeception

Kubeception - Kubernetes in Kubernetes in Kubernetes

In the classic setup, there is a dedicated host / VM to host the master components / control plane of a Kubernetes cluster. However, these are just normal programs that can easily be put into containers. Once in containers, we can make Kubernetes Deployments and StatefulSets (for the etcd) watch over them. And now we put all that into a separate, dedicated Kubernetes cluster - et voilà, we have Kubernetes in Kubernetes, aka Kubeception (named after the famous movie Inception with Leonardo DiCaprio).

In Gardener’s terminology, the cluster hosting the control plane components is called a seed cluster. The cluster that end users actually use (and whose control plane is hosted in the seed) is called a shoot cluster.

Control Plane Components on the Seed

All control-plane components of a shoot cluster run in a dedicated namespace on the seed.

A control plane has lots of components:

• Everything needed to run vanilla Kubernetes
• etcd main & events (split for performance reasons)
• Kube-.*-manager
• CSI driver

Additionally, we deploy components needed to manage the cluster:

• Gardener Resource Manager (GRM)
• Machine Controller Manager (MCM)
• DNS Management
• VPN

There is also a set of components making our life easier (logging, monitoring) or adding additional features (cert manager).

Core Components

Let’s take a close look at the API server as well as etcd.

Secrets are encrypted at rest. When asking etcd for the data, the reply is still encrypted. Decryption is done by the API server which knows the necessary key.

For non-HA clusters etcd has only 1 replica, while for HA clusters there are 3 replicas.

One special remark is needed for Gardener’s deployment of etcd. The pods coming from the etcd-main StatefulSet contain two containers - one runs etcd, the other runs a program that periodically backs up etcd’s contents to an object store that is set up per seed cluster to make sure no data is lost. After all, etcd is the Achilles heel of all Kubernetes clusters. The backup container is also capable of performing a restore from the object store as well as defragment and compact the etcd datastore. For performance reasons, Gardener stores Kubernetes events in a separate etcd instance. By default, events are retained for 1h but can be kept longer if defined in the shoot.spec.

The kube API server (often called “kapi”) scales both horizontally and vertically.

The kube API server is not directly exposed / reachable via its public hostname. Instead, Gardener runs a single LoadBalancer service backed by an istio gateway / envoy, which uses SNI to forward traffic.

The kube-controller-manager (aka KCM) is the component that contains all the controllers for the core Kubernetes objects such as Deployments, Services, PVCs, etc.

The Kubernetes scheduler will assign pods to nodes.

The Cloud Controller Manager (aka CCM) is the component that contains all functionality to talk to Cloud environments (e.g., create LoadBalancer services).

The CSI driver is the storage subsystem of Kubernetes. It provisions and manages anything related to persistence.

Without the cluster autoscaler, nodes could not be added or removed based on current pressure on the cluster resources. Without the VPA, pods would have fixed resource limits that could not change on demand.

Gardener-Specific Components

Shoot DNS service: External DNS management for resources within the cluster.

Machine Controller Manager: Responsible for managing VMs which will become nodes in the cluster.

Virtual Private Network deployments (aka VPN): Almost every communication between Kubernetes controllers and the API server is unidirectional - the controllers are given a kubeconfig and will establish a connection to the API server, which is exposed to all nodes of the cluster through a LoadBalancer. However, there are a few operations that require the API server to connect to the kubelet instead (e.g., for every webhook, when using kubectl exec or kubectl logs). Since every good Kubernetes cluster will have its worker nodes shielded behind firewalls to reduce the attack surface, Gardener establishes a VPN connection from the shoot’s internal network to the API server in the seed. For that, every shoot, as well as every control plane namespace in the seed, have openVPN pods in them that connect to each other (with the connection being established from the shoot to the seed).

Gardener Resource Manager: Tooling to deploy and manage Kubernetes resources required for cluster functionality.

Machines

Machine Controller Manager (aka MCM):

The machine controller manager, which lives on the seed in a shoot’s control plane namespace, is the key component responsible for provisioning and removing worker nodes for a Kubernetes cluster. It acts on MachineClass, MachineDeployment, and MachineSet resources in the seed (think of them as the equivalent of Deployments and ReplicaSets) and controls the lifecycle of machine objects. Through a system of plugins, the MCM is the component that phones to the cloud provider’s API and bootstraps virtual machines.

For more information, see MCM and Cluster-autoscaler.

ManagedResources

Gardener Resource Manager (aka GRM):

Gardener not only deploys components into the control plane namespace of the seed but also to the shoot (e.g., the counterpart of the VPN). Together with the components in the seed, Gardener needs to have a way to reconcile them.

Enter the GRM - it reconciles on ManagedResources objects, which are descriptions of Kubernetes resources which are deployed into the seed or shoot by GRM. If any of these resources are modified or deleted by accident, the usual observe-analyze-act cycle will revert these potentially malicious changes back to the values that Gardener envisioned. In fact, all the components found in a shoot’s kube-system namespace are ManagedResources governed by the GRM. The actual resource definition is contained in secrets (as they may contain “secret” data), while the ManagedResources contain a reference to the secret containing the actual resource to be deployed and reconciled.

DNS Records - “Internal” and “External”

The internal domain name is used by all Gardener components to talk to the API server. Even though it is called “internal”, it is still publicly routable.

But most importantly, it is pre-defined and not configurable by the end user.

Therefore, the “external” domain name exists. It is either a user owned domain or can be pre-defined for a Gardener landscape. It is used by any end user accessing the cluster’s API server.

For more information, see Contract: DNSRecord Resources.

Features and Observability

Gardener runs various health checks to ensure that the cluster works properly. The Network Problem Detector gives information about connectivity within the cluster and to the API server.

Certificate Management: allows to request certificates via the ACME protocol (e.g., issued by Let’s Encrypt) from within the cluster. For detailed information, have a look at the cert-manager project.

Observability stack: Gardener deploys observability components and gathers logs and metrics for the control-plane & kube-system namespace. Also provided out-of-the-box is a UI based on Plutono (fork of Grafana) with pre-defined dashboards to access and query the monitoring data. For more information, see Observability.

HA Control Plane

As the title indicates, the HA control plane feature is only about the control plane. Setting up the data plane to span multiple zones is part of the worker spec of a shoot.

HA control planes can be configured as part of the shoot’s spec. The available types are:

• Node
• Zone

Both work similarly and just differ in the failure domain the concepts are applied to.

For detailed guidance and more information, see the High Availability Guides.

Zonal HA Control Planes

Zonal HA is the most likely setup for shoots with purpose: production.

The starting point is a regular (non-HA) control plane. etcd and most controllers are singletons and the kube-apiserver might have been scaled up to several replicas.

To get to an HA setup we need:

• A minimum of 3 replicas of the API server
• 3 replicas for etcd (both main and events)
• A second instance for each controller (e.g., controller manager, csi-driver, scheduler, etc.) that can take over in case of failure (active / passive).

To distribute those pods across zones, well-known concepts like PodTopologySpreadConstraints or Affinities are applied.

kube-system Namespace

For a fully functional cluster, a few components need to run on the data plane side of the diagram. They all exist in the kube-system namespace. Let’s have a closer look at them.

Networking

On each node we need a CNI (container network interface) plugin. Gardener offers Calico or Cilium as network provider for a shoot. When using Calico, a kube-proxy is deployed. Cilium does not need a kube-proxy, as it takes care of its tasks as well.

The CNI plugin ensures pod-to-pod communication within the cluster. As part of it, it assigns cluster-internal IP addresses to the pods and manages the network devices associated with them. When an overlay network is enabled, calico will also manage the routing of pod traffic between different nodes.

On the other hand, kube-proxy implements the actual service routing (cilium can do this as well and no kube-proxy is needed). Whenever packets go to a service’s IP address, they are re-routed based on IPtables rules maintained by kube-proxy to reach the actual pods backing the service. kube-proxy operates on endpoint-slices and manages IPtables on EVERY node. In addition, kube-proxy provides a health check endpoint for services with externalTrafficPolicy=local, where traffic only gets to nodes that run a pod matching the selector of the service.

The egress filter implements basic filtering of outgoing traffic to be compliant with SAP’s policies.

And what happens if the pods crashloop, are missing or otherwise broken?

Well, in case kube-proxy is broken, service traffic will degrade over time (depending on the pod churn rate and how many kube-proxy pods are broken).

When calico is failing on a node, no new pods can start there as they don’t get any IP address assigned. It might also fail to add routes to newly added nodes. Depending on the error, deleting the pod might help.

DNS System

For a normal service in Kubernetes, a cluster-internal DNS record that resolves to the service’s ClusterIP address is being created. In Gardener (similar to most other Kubernetes offerings) CoreDNS takes care of this aspect. To reduce the load when it comes to upstream DNS queries, Gardener deploys a DNS cache to each node by default. It will also forward queries outside the cluster’s search domain directly to the upstream DNS server. For more information, see NodeLocalDNS Configuration and DNS autoscaling.

In addition to this optimization, Gardener allows custom DNS configuration to be added to CoreDNS via a dedicated ConfigMap.

In case this customization is related to non-Kubernetes entities, you may configure the shoot’s NodeLocalDNS to forward to CoreDNS instead of upstream (disableForwardToUpstreamDNS: true).

A broken DNS system on any level will cause disruption / service degradation for applications within the cluster.

Health Checks and Metrics

Gardener deploys probes checking the health of individual nodes. In a similar fashion, a network health check probes connectivity within the cluster (node to node, pod to pod, pod to api-server, …).

They provide the data foundation for Gardener’s monitoring stack together with the metrics collecting / exporting components.

Connectivity Components

From the perspective of the data plane, the shoot’s API server is reachable via the cluster-internal service kubernetes.default.svc.cluster.local. The apiserver-proxy intercepts connections to this destination and changes it so that the traffic is forwarded to the kube-apiserver service in the seed cluster. For more information, see kube-apiserver via apiserver-proxy.

The second component here is the VPN shoot. It initiates a VPN connection to its counterpart in the seed. This way, there is no open port / Loadbalancer needed on the data plane. The VPN connection is used for any traffic flowing from the control plane to the data plane. If the VPN connection is broken, port-forwarding or log querying with kubectl will not work. In addition, webhooks will stop functioning properly.

csi-driver

The last component to mention here is the csi-driver that is deployed as a Daemonset to all nodes. It registers with the kubelet and takes care of the mounting of volume types it is responsible for.

1.6 - Shoot Lifecycle

Reconciliation in Kubernetes and Gardener

The starting point of all reconciliation cycles is the constant observation of both the desired and actual state. A component would analyze any differences between the two states and try to converge the actual towards the desired state using appropriate actions. Typically, a component is responsible for a single resource type but it also watches others that have an implication on it.

As an example, the Kubernetes controller for ReplicaSets will watch pods belonging to it in order to ensure that the specified replica count is fulfilled. If one pod gets deleted, the controller will create a new pod to enforce the desired over the actual state.

This is all standard behaviour, as Gardener is following the native Kubernetes approach. All elements of a shoot cluster have a representation in Kubernetes resources and controllers are watching / acting upon them.

If we pick up the example of the ReplicaSet - a user typically creates a deployment resource and the ReplicaSet is implicitly generated on the way to create the pods. Similarly, Gardener takes the user’s intent (shoot) and creates lots of domain specific resources on the way. They all reconcile and make sure their actual and desired states match.

Updating the Desired State of a Shoot

Based on the shoot’s specifications, Gardener will create network resources on a hyperscaler, backup resources for the ETCD, credentials, and other resources, but also representations of the worker pools. Eventually, this process will result in a fully functional Kubernetes cluster.

If you change the desired state, Gardener will reconcile the shoot and run through the same cycle to ensure the actual state matches the desired state.

For example, the (infrastructure-specific) machine type can be changed within the shoot resource. The following reconciliation will pick up the change and initiate the creation of new nodes with a different machine type and the removal of the old nodes.

Maintenance Window and Daily Reconciliation

EVERY shoot cluster reconciles once per day during the so-called “maintenance window”. You can confine the rollout of spec changes to this window.

Additionally, the daily reconciliation will help pick up all kind of version changes. When a new Gardener version was rolled out to the landscape, shoot clusters will pick up any changes during their next reconciliation. For example, if a new Calico version is introduced to fix some bug, it will automatically reach all shoots.

Impact of a Change

It is important to be aware of the impacts that a change can have on a cluster and the workloads within it.

An operator pushing a new Gardener version with a new calico image to a landscape will cause all calico pods to be re-created. Another example would be the rollout of a new etcd backup-restore image. This would cause etcd pods to be re-created, rendering a non-HA control plane unavailable until etcd is up and running again.

When you change the shoot spec, it can also have significant impact on the cluster. Imagine that you have changes the machine type of a worker pool. This will cause new machines to be created and old machines to be deleted. Or in other words: all nodes will be drained, the pods will be evicted and then re-created on newly created nodes.

Kubernetes Version Update (Minor + Patch)

Some operations are rather common and have to be performed on a regular basis. Updating the Kubernetes version is one them. Patch updates cause relatively little disruption, as only the control-plane pods will be re-created with new images and the kubelets on all nodes will restart.

A minor version update is more impactful - it will cause all nodes to be recreated and rolls components of the control plane.

OS Version Update

The OS version is defined for each worker pool and can be changed per worker pool. You can freely switch back and forth. However, as there is no in-place update, each change will cause the entire worker pool to roll and nodes will be replaced. For OS versions different update strategies can be configured. Please check the documentation for details.

Available Versions​

Gardener has a dedicated resource to maintain a list of available versions – the so-called cloudProfile.

A cloudProfile provides information about supported​:

• Kubernetes versions​
• OS versions (and where to find those images)​
• Regions (and their zones)​
• Machine types​

Each shoot references a cloudProfile in order to obtain information about available / possible versions and configurations.

Version Classifications

Gardener has the following classifications for Kubernetes and OS image versions:

• preview: still in testing phase (several versions can be in preview at the same time)

• supported: recommended version

• deprecated: a new version has been set to “supported”, updating is recommended (might have an expiration date)

• expired: cannot be used anymore, clusters using this version will be force-upgraded

Version information is maintained in the relevant cloud profile resource. There might be circumstances where a version will never become supported but instead move to deprecated directly. Similarly, a version might be directly introduced as supported.

AutoUpdate / Forced Updates

AutoUpdate for a machine image version will update all node pools to the latest supported version based on the defined update strategy. Whenever a new version is set to supported, the cluster will pick it up during its next maintenance window.

For Kubernetes versions the mechanism is the same, but only applied to patch version. This means that the cluster will be kept on the latest supported patch version of a specific minor version.

In case a version used in a cluster expires, there is a force update during the next maintenance window. In a worst case scenario, 2 minor versions expire simultaneously. Then there will be two consecutive minor updates enforced.

For more information, see Shoot Kubernetes and Operating System Versioning in Gardener.

Applying Changes to a Seed

It is important to keep in mind that a seed is just another Kubernetes cluster. As such, it has its own lifecycle (daily reconciliation, maintenance, etc.) and is also a subject to change.

From time to time changes need to be applied to the seed as well. Some (like updating the OS version) cause the node pool to roll. In turn, this will cause the eviction of ALL pods running on the affected node. If your etcd is evicted and you don’t have a highly available control plane, it will cause downtime for your cluster. Your workloads will continue to run ,of course, but your cluster’s API server will not function until the etcd is up and running again.

1.7 - Observability

Overview

Gardener offers out-of-the-box observability for the control plane, Gardener managed system-components, and the nodes of a shoot cluster.

Having your workload survive on day 2 can be a challenge. The goal of this topic is to give you the tools with which to observe, analyze, and alert when the control plane or system components of your cluster become unhealthy. This will let you guide your containers through the storm of operating in a production environment.

1.7.1 - Components

Core Components

The core Observability components which Gardener offers out-of-the-box are:

• Prometheus - for Metrics and Alerting
• Vali - a Loki fork for Logging
• Plutono - a Grafana fork for Dashboard visualization

Both forks are done from the last version with an Apache license.

Control Plane Components on the Seed

Prometheus, Plutono, and Vali are all located in the seed cluster. They run next to the control plane of your cluster.

The next sections will explore those components in detail.

Logging into Plutono

Let us start by giving some visual hints on how to access Plutono. Plutono allows us to query logs and metrics and visualise those in form of dashboards. Plutono is shipped ready-to-use with a Gardener shoot cluster.

In order to access the Gardener provided dashboards, open the Plutono link provided in the Gardener dashboard and use the username and password provided next to it.

The password you can use to log in can be retrieved as shown below:

Accessing the Dashboards

After logging in, you will be greeted with a Plutono welcome screen. Navigate to General/Home, as depicted with the red arrow in the next picture:

Then you will be able to select the dashboards. Some interesting ones to look at are:

• The Kubernetes Control Plane Status dashboard allows you to check control plane availability during a certain time frame.
• The API Server dashboard gives you an overview on which requests are done towards your apiserver and how long they take.
• With the Node Details dashboard you can analyze CPU/Network pressure or memory usage for nodes.
• The Network Problem Detector dashboard illustrates the results of periodic networking checks between nodes and to the APIServer.

Here is a picture with the Kubernetes Control Plane Status dashboard.

Prometheus

Prometheus is a monitoring system and a time series database. It can be queried using PromQL, the so called Prometheus Querying Language.

This example query describes the current uptime status of the kube apiserver.

Prometheus and Plutono

Time series data from Prometheus can be made visible with Plutono. Here we see how the query above which describes the uptime of a Kubernetes cluster is visualized with a Plutono dashboard.

Vali Logs via Plutono

Vali is our logging solution. In order to access the logs provided by Vali, you need to:

1. Log into Plutono.

2. Choose Explore, which is depicted as the little compass symbol:

1. Select Vali at the top left, as shown here:

There you can browse logs or events of the control plane components.

Here are some examples of helpful queries:

• {container_name="cluster-autoscaler" } to get cluster-autoscaler logs and see why certain node groups were scaled up.
• {container_name="kube-apiserver"} |~ "error" to get the logs of the kube-apiserver container and filter for errors.
• {unit="kubelet.service", nodename="ip-123"} to get the kubelet logs of a specific node.
• {unit="containerd.service", nodename="ip-123"} to retrieve the containerd logs for a specific node.

Choose Help > in order to see what options exist to filter the results.

For more information on how to retrieve K8s events from the past, see How to Access Logs.

Detailed View

Data Flow

Our monitoring and logging solutions Vali and Prometheus both run next to the control plane of the shoot cluster.

Data Flow - Logging

The following diagram allows a more detailed look at Vali and the data flow.

On the very left, we see Plutono as it displays the logs. Vali is aggregating the logs from different sources.

Valitail and Fluentbit send the logs to Vali, which in turn stores them.

Valitail

Valitail is a systemd service that runs on each node. It scrapes kubelet, containerd, kernel logs, and the logs of the pods in the kube-system namespace.

Fluentbit

Fluentbit runs as a daemonset on each seed node. It scrapes logs of the kubernetes control plane components, like apiserver or etcd.

It also scrapes logs of the Gardener deployed components which run next to the control plane of the cluster, like the machine-controller-manager or the cluster autoscaler. Debugging those components, for example, would be helpful when finding out why certain worker groups got scaled up or why nodes were replaced.

Data Flow - Monitoring

Next to each shoot’s control plane, we deploy an instance of Prometheus in the seed.

Gardener uses Prometheus for storing and accessing shoot-related metrics and alerting.

The diagram below shows the data flow of metrics. Plutono uses PromQL queries to query data from Prometheus. It then visualises those metrics in dashboards. Prometheus itself scrapes various targets for metrics, as seen in the diagram below by the arrows pointing to the Prometheus instance.

Let us have a look what metrics we scrape for debugging purposes:

Container performance metrics

cAdvisor is an open-source agent integrated into the kubelet binary that monitors resource usage and analyzes the performance of containers. It collects statistics about the CPU, memory, file, and network usage for all containers running on a given node. We use it to scrape data for all pods running in the kube-system namespace in the shoot cluster.

Hardware and kernel-related metrics

The Prometheus Node Exporter runs as a daemonset in the kube-system namespace of your shoot cluster. It exposes a wide variety of hardware and kernel-related metrics. Some of the metrics we scrape are, for example, the current usage of the filesystem (node_filesystem_free_bytes) or current CPU usage (node_cpu_seconds_total). Both can help you identify if nodes are running out of hardware resources, which could lead to your workload experiencing downtimes.

Control plane component specific metrics

The different control plane pods (for example, etcd, API server, and kube-controller-manager) emit metrics over the /metrics endpoint. This includes metrics like how long webhooks take, the request count of the apiserver and storage information, like how many and what kind of objects are stored in etcd.

Metrics about the state of Kubernetes objects

kube-state-metrics is a simple service that listens to the Kubernetes API server and generates metrics about the state of the objects. It is not concerned with metrics about the Kubernetes components, but rather it exposes metrics calculated from the status of Kubernetes objects (for example, resource requests or health of pods).

In the following image a few example metrics, which are exposed by the various components, are listed:

We only store metrics for Gardener deployed components. Those include the Kubernetes control plane, Gardener managed system components (e.g., pods) in the kube-system namespace of the shoot cluster or systemd units on the nodes. We do not gather metrics for workload deployed in the shoot cluster. This is also shown in the picture below.

This means that for any workload you deploy into your shoot cluster, you need to deploy monitoring and logging yourself.

Logs or metrics are kept up to 14 days or when a configured space limit is reached.

1.7.2 - Alerts

Overview

In this overview, we want to present two ways to receive alerts for control plane and Gardener managed system-components:

• Predefined Gardener alerts
• Custom alerts

Predefined Control Plane Alerts

In the shoot spec it is possible to configure emailReceivers. On this email address you will automatically receive email notifications for predefined alerts of your control plane. Such alerts are deployed in the shoot Prometheus and have visibility owner or all. For more alert details, shoot owners can use this visibility to find these alerts in their shoot Prometheus UI.

spec:
  monitoring:
    alerting:
      emailReceivers:
      - john.doe@example.com

For more information, see Alerting.

Custom Alerts - Federation

If you need more customization for alerts for control plane metrics, you have the option to deploy your own Prometheus into your shoot control plane.

Then you can use federation, which is a Prometheus feature, to forward the metrics from the Gardener managed Prometheus to your custom deployed Prometheus. Since as a shoot owner you do not have access to the control plane pods, this is the only way to get those metrics.

The credentials and endpoint for the Gardener managed Prometheus are exposed over the Gardener dashboard or programmatically in the garden project as a secret (<shoot-name>.monitoring).

1.7.3 - Shoot Status

Overview

In this topic you can see various shoot statuses and how you can use them to monitor your shoot cluster.

Shoot Status - Conditions

You can retrieve the shoot status by using kubectl get shoot -oyaml

It contains conditions, which give you information about the healthiness of your cluster. Those conditions are also forwarded to the Gardener dashboard and show your cluster as healthy or unhealthy.

Shoot Status - Constraints

The shoot status also contains constraints. If these constraints are met, your cluster operations are impaired and the cluster is likely to fail at some point. Please watch them and act accordingly.

Shoot Status - Last Operation

The lastOperation, lastErrors, and lastMaintenance give you information on what was last happening in your clusters. This is especially useful when you are facing an error.

In this example, nodes are being recreated and not all machines have reached the desired state yet.

Shoot Status - Credentials Rotation

You can also see the status of the last credentials rotation. Here you can also programmatically derive when the last rotation was down in order to trigger the next rotation.

1.8 - Features

1.8.1 - Hibernation

Hibernation

Some clusters need to be up all the time - typically, they would be hosting some kind of production workload. Others might be used for development purposes or testing during business hours only. Keeping them up and running all the time is a waste of money. Gardener can help you here with its “hibernation” feature. Essentially, hibernation means to shut down all components of a cluster.

How Hibernation Works

The hibernation flow for a shoot attempts to reduce the resources consumed as much as possible. Hence everything not state-related is being decommissioned.

Data Plane

All nodes will be drained and the VMs will be deleted. As a result, all pods will be “stuck” in a Pending state since no new nodes are added. Of course, PVC / PV holding data is not deleted.

Services of type LoadBalancer will keep their external IP addresses.

Control Plane

All components will be scaled down and no pods will remain running. ETCD data is kept safe on the disk.

The DNS records routing traffic for the API server are also destroyed. Trying to connect to a hibernated cluster via kubectl will result in a DNS lookup failure / no-such-host message.

When waking up a cluster, all control plane components will be scaled up again and the DNS records will be re-created. Nodes will be created again and pods scheduled to run on them.

How to Configure / Trigger Hibernation

The easiest way to configure hibernation schedules is via the dashboard. Of course, this is reflected in the shoot’s spec and can also be maintained there. Before a cluster is hibernated, constraints in the shoot’s status will be evaluated. There might be conditions (mostly revolving around mutating / validating webhooks) that would block a successful wake-up. In such a case, the constraint will block hibernation in the first place.

To wake-up or hibernate a shoot immediately, the dashboard can be used or a patch to the shoot’s spec can be applied directly.

1.8.2 - Workerless Shoots

Controlplane as a Service

Sometimes, there may be use cases for Kubernetes clusters that don’t require pods but only features of the control plane. Gardener can create the so-called “workerless” shoots, which are exactly that. A Kubernetes cluster without nodes (and without any controller related to them).

In a scenario where you already have multiple clusters, you can use it for orchestration (leases) or factor out components that require many CRDs.

As part of the control plane, the following components are deployed in the seed cluster for workerless shoot:

• etcds
• kube-apiserver
• kube-controller-manager
• gardener-resource-manager
• Logging and monitoring components
• Extension components (to find out if they support workerless shoots, see the Extensions documentation)

1.8.3 - Credential Rotation

Keys

There are plenty of keys in Gardener. The ETCD needs one to store resources like secrets encrypted at rest. Gardener generates certificate authorities (CAs) to ensure secured communication between the various components and actors and service account tokens are signed with a dedicated key. There is also an SSH key pair to allow debugging of nodes and the observability stack has its own passwords too.

All of these keys share a common property: they are managed by Gardener. Rotating them, however, is potentially very disruptive. Hence, Gardener does not do it automatically, but offers you means to perform these tasks easily. For a single cluster, you may conveniently use the dashboard.

Where possible, the rotation happens in two phases - Preparing and Completing.

Prepare Rotation of All Credentials

The Preparing phase introduces new keys while the old ones are still valid. Users can safely exchange keys / CA bundles wherever they are used. It is possible to start the preparation by annotating the shoot resource accordingly:

kubectl -n <shoot-namespace> annotate shoot <shoot-name> gardener.cloud/operation=rotate-credentials-start

Complete Rotation of All Credentials

Afterward, the Completing phase will invalidate the old keys / CA bundles. Annotate the shoot resource accordingly:

kubectl -n <shoot-namespace> annotate shoot <shoot-name> gardener.cloud/operation=rotate-credentials-complete

Rotation Phases

At the beginning, only the old set of credentials exists. By triggering the rotation, new credentials are created in the Preparing phase and both sets are valid. Now, all clients have to update and start using the new credentials. Only afterward it is safe to trigger the Completing phase, which invalidates the old credentials.

The shoot’s status will always show the current status / phase of the rotation.

For more information, see Credentials Rotation for Shoot Clusters.

User-Provided Credentials

You grant Gardener permissions to create resources by handing over cloud provider keys. These keys are stored in a secret and referenced to a shoot via a SecretBinding. Gardener uses the keys to create the network for the cluster resources, routes, VMs, disks, and IP addresses.

When you rotate credentials, the new keys have to be stored in the same secret and the shoot needs to reconcile successfully to ensure the replication to every controller. Afterward, the old keys can be deleted safely from Gardener’s perspective.

While the reconciliation can be triggered manually, there is no need for it (if you’re not in a hurry). Each shoot reconciles once within 24h and the new keys will be picked up during the next maintenance window.

1.8.4 - External DNS Management

External DNS Management

When you deploy to Kubernetes, there is no native management of external DNS. Instead, the cloud-controller-manager requests (mostly IPv4) addresses for every service of type LoadBalancer. Of course, the Ingress resource helps here, but how is the external DNS entry for the ingress controller managed?

Essentially, some sort of automation for DNS management is missing.

Automating DNS Management

From a user’s perspective, it is desirable to work with already known resources and concepts. Hence, the DNS management offered by Gardener plugs seamlessly into Kubernetes resources and you do not need to “leave” the context of the shoot cluster.

To request a DNS record creation / update, a Service or Ingress resource is annotated accordingly. The shoot-dns-service extension will (if configured) will pick up the request and create a DNSEntry resource + reconcile it to have an actual DNS record created at a configured DNS provider. Gardener supports the following providers:

• aws-route53
• azure-dns
• azure-private-dns
• google-clouddns
• openstack-designate
• alicloud-dns
• cloudflare-dns

For more information, see DNS Names.

DNS Provider

For the above to work, we need some ingredients. Primarily, this is implemented via a so-called DNSProvider. Every shoot has a default provider that is used to set up the API server’s public DNS record. It can be used to request sub-domains as well.

In addition, a shoot can reference credentials to a DNS provider. Those can be used to manage custom domains.

Please have a look at the documentation for further details.

1.8.5 - Certificate Management

Certificate Management

For proper consumption, any service should present a TLS certificate to its consumers. However, self-signed certificates are not fit for this purpose - the certificate should be signed by a CA trusted by an application’s userbase. Luckily, Issuers like Let’s Encrypt and others help here by offering a signing service that issues certificates based on the ACME challenge (Automatic Certificate Management Environment).

There are plenty of tools you can use to perform the challenge. For Kubernetes, cert-manager certainly is the most common, however its configuration is rather cumbersome and error prone. So let’s see how a Gardener extension can help here.

Manage Certificates with Gardener

You may annotate a Service or Ingress resource to trigger the cert-manager to request a certificate from the any configured issuer (e.g. Let’s Encrypt) and perform the challenge. A Gardener operator can add a default issuer for convenience. With the DNS extension discussed previously, setting up the DNS TXT record for the ACME challenge is fairly easy. The requested certificate can be customized by the means of several other annotations known to the controller. Most notably, it is possible to specify SANs via cert.gardener.cloud/dnsnames to accommodate domain names that have more than 64 characters (the limit for the CN field).

The user’s request for a certificate manifests as a certificate resource. The status, issuer, and other properties can be checked there.

Once successful, the resulting certificate will be stored in a secret and is ready for usage.

With additional configuration, it is also possible to define custom issuers of certificates.

For more information, see the Manage certificates with Gardener for public domain topic and the cert-management repository.

1.8.6 - Vertical Pod Autoscaler

Vertical Pod Autoscaler

When a pod’s resource CPU or memory grows, it will hit a limit eventually. Either the pod has resource limits specified or the node will run short of resources. In both cases, the workload might be throttled or even terminated. When this happens, it is often desirable to increase the request or limits. To do this autonomously within certain boundaries is the goal of the Vertical Pod Autoscaler project.

Since it is not part of the standard Kubernetes API, you have to install the CRDs and controller manually. With Gardener, you can simply flip the switch in the shoot’s spec and start creating your VPA objects.

Please be aware that VPA and HPA operate in similar domains and might interfere.

A controller & CRDs for vertical pod auto-scaling can be activated via the shoot’s spec.

1.8.7 - Cluster Autoscaler

Obtaining Aditional Nodes

The scheduler will assign pods to nodes, as long as they have capacity (CPU, memory, Pod limit, # attachable disks, …). But what happens when all nodes are fully utilized and the scheduler does not find any suitable target?

Option 1: Evict other pods based on priority. However, this has the downside that other workloads with lower priority might become unschedulable.

Option 2: Add more nodes. There is an upstream Cluster Autoscaler project that does exactly this. It simulates the scheduling and reacts to pods not being schedulable events. Gardener has forked it to make it work with machine-controller-manager abstraction of how node (groups) are defined in Gardener. The cluster autoscaler respects the limits (min / max) of any worker pool in a shoot’s spec. It can also scale down nodes based on utilization thresholds. For more details, see the autoscaler documentation.

Scaling by Priority

For clusters with more than one node pool, the cluster autoscaler has to decide which group to scale up. By default, it randomly picks from the available / applicable. However, this behavior is customizable by the use of so-called expanders.

This section will focus on the priority based expander.

Each worker pool gets a priority and the cluster autoscaler will scale up the one with the highest priority until it reaches its limit.

To get more information on the current status of the autoscaler, you can check a “status” configmap in the kube-system namespace with the following command:

kubectl get cm -n kube-system cluster-autoscaler-status -oyaml

To obtain information about the decision making, you can check the logs of the cluster-autoscaler pod by using the shoot’s monitoring stack.

For more information, see the cluster-autoscaler FAQ and the Priority based expander for cluster-autoscaler topic.

1.9 - Common Pitfalls

Architecture

Containers will NOT fix a broken architecture!

Running a highly distributed system has advantages, but of course, those come at a cost. In order to succeed, one would need:

• Logging
• Tracing
• No singleton
• Tolerance to failure of individual instances
• Automated config / change management
• Kubernetes knowledge

Scalability

Most scalability dimensions are interconnected with others. If a cluster grows beyond reasonable defaults, it can still function very well. But tuning it comes at the cost of time and can influence stability negatively.

Take the number of nodes and pods, for example. Both are connected and you cannot grow both towards their individual limits, as you would face issues way before reaching any theoretical limits.

Reading the Scalability of Gardener Managed Kubernetes Clusters guide is strongly recommended in order to understand the topic of scalability within Kubernetes and Gardener.

A Small Sample of Things That Can Grow Beyond Reasonable Limits

When scaling a cluster, there are plenty of resources that can be exhausted or reach a limit:

• The API server will be scaled horizontally and vertically by Gardener. However, it can still consume too much resources to fit onto a single node on the seed. In this case, you can only reduce the load on the API server. This should not happen with regular usage patterns though.
• ETCD disk space: 8GB is the limit. If you have too many resources or a high churn rate, a cluster can run out of ETCD capacity. In such a scenario it will stop working until defragmented, compacted, and cleaned up.
• The number of nodes is limited by the network configuration (pod cidr range & node cidr mask). Also, there is a reasonable number of nodes (300) that most workloads should not exceed. It is possible to go beyond but doing so requires careful tuning and consideration of connected scaling dimensions (like the number of pods per node).

The availability of your cluster is directly impacted by the way you use it.

Infrastructure Capacity and Quotas

Sometimes requests cannot be fulfilled due to shortages on the infrastructure side. For example, a certain instance type might not be available and new Kubernetes nodes of this type cannot be added. It is a good practice to use the cluster-autoscaler’s priority expander and have a secondary node pool.

Sometimes, it is not the physical capacity but exhausted quotas within an infrastructure account that result in limits. Obviously, there should be sufficient quota to create as many VMs as needed. But there are also other resources that are created in the infrastructure that need proper quotas:

• Loadbalancers
• VPC
• Disks
• Routes (often forgotten, but very important for clusters without overlay network; typically defaults to around 50 routes, meaning that 50 nodes is the maximum a cluster can have)
• …

NodeCIDRMaskSize

Upon cluster creation, there are several settings that are network related. For example, the address space for Pods has to be defined. In this case, it is a /16 subnet that includes a total of 65.536 hosts. However, that does not imply that you can easily use all addresses at the same point in time.

As part of the Kubernetes network setup, the /16 network is divided into smaller subnets and each node gets a distinct subnet. The size of this subnet defaults to /24. It can also be specified (but not changed later).

Now, as you create more nodes, you have a total of 256 subnets that can be assigned to nodes, thus limiting the total number of nodes of this cluster to 256.

For more information, see Shoot Networking.

Overlapping VPCs

Avoid Overlapping CIDR Ranges in VPCs

Gardener can create shoot cluster resources in an existing / user-created VPC. However, you have to make sure that the CIDR ranges used by the shoots nodes or subnets for zones do not overlap with other shoots deployed to the same VPC.

In case of an overlap, there might be strange routing effects, and packets ending up at a wrong location.

Expired Credentials

Credentials expire or get revoked. When this happens to the actively used infrastructure credentials of a shoot, the cluster will stop working after a while. New nodes cannot be added, LoadBalancers cannot be created, and so on.

You can update the credentials stored in the project namespace and reconcile the cluster to replicate the new keys to all relevant controllers. Similarly, when doing a planned rotation one should wait until the shoot reconciled successfully before invalidating the old credentials.

AutoUpdate Breaking Clusters

Gardener can automatically update a shoot’s Kubernetes patch version, when a new patch version is labeled as “supported”. Automatically updating of the OS images works in a similar way. Both are triggered by the “supported” classification in the respective cloud profile and can be enabled / disabled as part a shoot’s spec.

Additionally, when a minor Kubernetes / OS version expires, Gardener will force-update the shoot to the next supported version.

Turning on AutoUpdate for a shoot may be convenient but comes at the risk of potentially unwanted changes. While it is possible to switch to another OS version, updates to the Kubernetes version are a one way operation and cannot be reverted.

Node Draining

Node Draining and Pod Disruption Budget

Typically, nodes are drained when:

• There is a update of the OS / Kubernetes minor version
• An Operator cordons & drains a node
• The cluster-autoscaler wants to scale down

Without a PodDistruptionBudget, pods will be terminated as fast as possible. If an application has 2 out of 2 replicas running on the drained node, this will probably cause availability issues.

Node Draining with PDB

PodDisruptionBudgets can help to manage a graceful node drain. However, if no disruptions are allowed there, the node drain will be blocked until it reaches a timeout. Only then will the nodes be terminated but without respecting PDB thresholds.

Pod Resource Requests and Limits

Resource Consumption

Pods consume resources and, of course, there are only so many resources available on a single node. Setting requests will make the scheduling much better, as the scheduler has more information available.

Specifying limits can help, but can also limit an application in unintended ways. A recommendation to start with:

• Do not set CPU limits (CPU is compressible and throttling is really hard to detect)
• Set memory limits and monitor OOM kills / restarts of workload (typically detectable by container status exit code 137 and corresponding events). This will decrease the likelihood of OOM situations on the node itself. However, for critical workloads it might be better to have uncapped growth and rather risk a node going OOM.

Next, consider if assigning the workload to quality of service class guaranteed is needed. Again - this can help or be counterproductive. It is important to be aware of its implications. For more information, see Pod Quality of Service Classes.

Tune shoot.spec.Kubernetes.kubeReserved to protect the node (kubelet) in case of a workload pod consuming too much resources. It is very helpful to ensure a high level of stability.

If the usage profile changes over time, the VPA can help a lot to adapt the resource requests / limits automatically.

Webhooks

User-Deployed Webhooks in Kubernetes

By default, any request to the API server will go through a chain of checks. Let’s take the example of creating a pod.

When the resource is submitted to the API server, it will be checked against the following validations:

• Is the user authorized to perform this action?
• Is the pod definitionactually valid?
• Are the specified values allowed?

Additionally, there is the defaulting - like the injection of the default service account’s name, if nothing else is specified.

This chain of admission control and mutation can be enhanced by the user. Read about dynamic admission control for more details.

ValidatingWebhookConfiguration: allow or deny requests based on custom rules

MutatingWebhookConfiguration: change а resource before it is actually stored in etcd (that is, before any other controller acts upon)

Both ValidatingWebhookConfiguration as well as MutatingWebhookConfiguration resources:

• specify for which resources and operations these checks should be executed.
• specify how to reach the webhook server (typically a service running on the data plane of a cluster)
• rely on a webhook server performing a review and reply to the admissionReview request

What could possibly go wrong? Due to the separation of control plane and data plane in Gardener’s architecture, webhooks have the potential to break a cluster. If the webhook server is not responding in time with a valid answer, the request should timeout and the failure policy is invoked. Depending on the scope of the webhook, frequent failures may cause downtime for applications. Common causes for failure are:

• The call to the webhook is made through the VPN tunnel. VPN / connection issues can happen both on the side of the seed as well as the shoot and would render the webhook unavailable from the perspective of the control plane.
• The traffic cannot reach the pod (network issue, pod not available)
• The pod is processing too slow (e.g., because there are too many requests)

Timeout

Webhooks are a very helpful feature of Kubernetes. However, they can easily be configured to break a shoot cluster. Take the timeout, for example. High timeouts (>15s) can lead to blocking requests of control plane components. That’s because most control-plane API calls are made with a client-side timeout of 30s, so if a webhook has timeoutSeconds=30, the overall request might still fail as there is overhead in communication with the API server and other potential webhooks.

Recommendations

Problematic webhooks are reported as part of a shoot’s status. In addition to timeouts, it is crucial to exclude the kube-system namespace and (potentially non-namespaced) resources that are necessary for the cluster to function properly. Those should not be subject to a user-defined webhook.

In particular, a webhook should not operate on:

• the kube-system namespace
• Endpoints or EndpointSlices
• Nodes
• PodSecurityPolicies
• ClusterRoles
• ClusterRoleBindings
• CustomResourceDefinitions
• ApiServices
• CertificateSigningRequests
• PriorityClasses

Example:

A webhook checks node objects upon creation and has a failurePolicy: fail. If the webhook does not answer in time (either due to latency or because there is no pod serving it), new nodes cannot join the cluster.

For more information, see Shoot Status.

Conversion Webhooks

Who installs a conversion webhook?

If you have written your own CustomResourceDefinition (CRD) and made a version upgrade, you will also have consciously written & deployed the conversion webhook.

However, sometimes, you simply use helm or kustomize to install a (third-party) dependency that contains CRDs. Of course, those can contain conversion webhooks as well. As a user of a cluster, please make sure to be aware what you deploy.

CRD with a Conversion Webhook

Conversion webhooks are tricky. Similarly to regular webhooks, they should have a low timeout. However, they cannot be remediated automatically and can cause errors in the control plane. For example, if a webhook is invoked but not available, it can block the garbage collection run by the kube-controller-manager.

In turn, when deleting something like a deployment, dependent resources like pods will not be deleted automatically.

For more information, see the Webhook Conversion, Upgrade Existing Objects to a New Stored Version, and Version Priority topics in the Kubernetes documentation.

2 - Guides

2.1 - Set Up Client Tools

2.1.1 - Fun with kubectl Aliases

Speed up Your Terminal Workflow

Use the Kubernetes command-line tool, kubectl, to deploy and manage applications on Kubernetes. Using kubectl, you can inspect cluster resources, as well as create, delete, and update components.

You will probably run more than a hundred kubectl commands on some days and you should speed up your terminal workflow with with some shortcuts. Of course, there are good shortcuts and bad shortcuts (lazy coding, lack of security review, etc.), but let’s stick with the positives and talk about a good shortcut: bash aliases in your .profile.

What are those mysterious .profile and .bash_profile files you’ve heard about?

What’s the .bash_profile then? It’s exactly the same, but under a different name. The unix shell you are logging into, in this case OS X, looks for etc/profile and loads it if it exists. Then it looks for ~/.bash_profile, ~/.bash_login and finally ~/.profile, and loads the first one of these it finds.

Populating the .profile File

Here is the fantastic time saver that needs to be in your shell profile:

# time save number one. shortcut for kubectl
#
alias k="kubectl"

# Start a shell in a pod AND kill them after leaving
#
alias ksh="kubectl run busybox -i --tty --image=busybox --restart=Never --rm -- sh"

# opens a bash
#
alias kbash="kubectl run busybox -i --tty --image=busybox --restart=Never --rm -- ash"

# activate/exports the kuberconfig.yaml in the current working directory
#
alias kexport="export KUBECONFIG=`pwd`/kubeconfig.yaml"


# usage: kurl http://your-svc.namespace.cluster.local
#
# we need for this our very own image...never trust an unknown image..
alias kurl="docker run --rm byrnedo/alpine-curl"

All the kubectl tab completions still work fine with these aliases, so you’re not losing that speed.

# time save number one. shortcut for kubectl
#
alias k="kubectl"

# Enable kubectl completion
source <(k completion bash | sed s/kubectl/k/g)

2.1.2 - Kubeconfig Context as bash Prompt

Overview

Use the Kubernetes command-line tool, kubectl, to deploy and manage applications on Kubernetes. Using kubectl, you can inspect cluster resources, as well as create, delete, and update components.

By default, the kubectl configuration is located at ~/.kube/config.

Let us suppose that you have two clusters, one for development work and one for scratch work.

How to handle this easily without copying the used configuration always to the right place?

Export the KUBECONFIG Environment Variable

bash$ export KUBECONFIG=<PATH-TO-M>-CONFIG>/kubeconfig-dev.yaml

How to determine which cluster is used by the kubectl command?

Determine Active Cluster

bash$ kubectl cluster-info
Kubernetes master is running at https://api.dev.garden.shoot.canary.k8s-hana.ondemand.com
KubeDNS is running at https://api.dev.garden.shoot.canary.k8s-hana.ondemand.com/api/v1/proxy/namespaces/kube-system/services/kube-dns

To further debug and diagnose cluster problems, use 'kubectl cluster-info dump'.
bash$ 

Display Cluster in the bash - Linux and Alike

I found this tip on Stackoverflow and find it worth to be added here.

Edit your ~/.bash_profile and add the following code snippet to show the current K8s context in the shell’s prompt:

prompt_k8s(){
  k8s_current_context=$(kubectl config current-context 2> /dev/null)
  if [[ $? -eq 0 ]] ; then echo -e "(${k8s_current_context}) "; fi
}
 
 
PS1+='$(prompt_k8s)'

After this, your bash command prompt contains the active KUBECONFIG context and you always know which cluster is active - develop or production.

For example:

bash$ export KUBECONFIG=/Users/d023280/Documents/workspace/gardener-ui/kubeconfig_gardendev.yaml 
bash (garden_dev)$ 

Note the (garden_dev) prefix in the bash command prompt.

This helps immensely to avoid thoughtless mistakes.

Display Cluster in the PowerShell - Windows

Display the current K8s cluster in the title of PowerShell window.

Create a profile file for your shell under %UserProfile%\Documents\Windows­PowerShell\Microsoft.PowerShell_profile.ps1

Copy following code to Microsoft.PowerShell_profile.ps1

 function prompt_k8s {
     $k8s_current_context = (kubectl config current-context) | Out-String
     if($?) {
         return $k8s_current_context
     }else {
         return "No K8S contenxt found"
     }
 }

 $host.ui.rawui.WindowTitle = prompt_k8s

If you want to switch to different cluster, you can set KUBECONFIG to new value, and re-run the file Microsoft.PowerShell_profile.ps1

2.1.3 - Organizing Access Using kubeconfig Files

Overview

The kubectl command-line tool uses kubeconfig files to find the information it needs to choose a cluster and communicate with the API server of a cluster.

Problem

If you’ve become aware of a security breach that affects you, you may want to revoke or cycle credentials in case anything was leaked. However, this is not possible with the initial or master kubeconfig from your cluster.

Pitfall

Never distribute the kubeconfig, which you can download directly within the Gardener dashboard, for a productive cluster.

Create a Custom kubeconfig File for Each User

Create a separate kubeconfig for each user. One of the big advantages of this approach is that you can revoke them and control the permissions better. A limitation to single namespaces is also possible here.

The script creates a new ServiceAccount with read privileges in the whole cluster (Secrets are excluded). To run the script, Deno, a secure TypeScript runtime, must be installed.

#!/usr/bin/env -S deno run --allow-run

/*
* This script create Kubernetes ServiceAccount and other required resource and print KUBECONFIG to console.
* Depending on your requirements you might want change clusterRoleBindingTemplate() function
*
* In order to execute this script it's required to install Deno.js https://deno.land/ (TypeScript & JavaScript runtime).
* It's single executable binary for the major OSs from the original author of the Node.js
* example: deno run --allow-run kubeconfig-for-custom-user.ts d00001
* example: deno run --allow-run kubeconfig-for-custom-user.ts d00001 --delete
*
* known issue: shebang does works under the Linux but not for Windows Linux Subsystem
*/

const KUBECTL = "/usr/local/bin/kubectl" //or
// const KUBECTL = "C:\\Program Files\\Docker\\Docker\\resources\\bin\\kubectl.exe"

const serviceAccName = Deno.args[0]
const deleteIt = Deno.args[1]
if (serviceAccName == undefined || serviceAccName == "--delete" ) {
    console.log("please provide username as an argument, for example: deno run --allow-run kubeconfig-for-custom-user.ts USER_NAME [--delete]")
    Deno.exit(1)
}

if (deleteIt == "--delete") {
    exec([KUBECTL, "delete", "serviceaccount", serviceAccName])
    exec([KUBECTL, "delete", "secret", `${serviceAccName}-secret`])
    exec([KUBECTL, "delete", "clusterrolebinding", `view-${serviceAccName}-global`])
    Deno.exit(0)
}

await exec([KUBECTL, "create", "serviceaccount", serviceAccName, "-o", "json"])

await exec([KUBECTL, "create", "-o", "json", "-f", "-"], secretYamlTemplate())
let secret = await exec([KUBECTL, "get", "secret", `${serviceAccName}-secret`, "-o", "json"])
let caCRT = secret.data["ca.crt"];
let userToken = atob(secret.data["token"]); //decode base64

let kubeConfig = await exec([KUBECTL, "config", "view", "--minify", "-o", "json"]);
let clusterApi = kubeConfig.clusters[0].cluster.server
let clusterName = kubeConfig.clusters[0].name

await exec([KUBECTL, "create", "-o", "json", "-f", "-"], clusterRoleBindingTemplate())

console.log(kubeConfigTemplate(caCRT, userToken, clusterApi, clusterName, serviceAccName + "-" + clusterName))

async function exec(args: string[], stdInput?: string): Promise<Object> {
    console.log("# "+args.join(" "))
    let opt: Deno.RunOptions = {
        cmd: args,
        stdout: "piped",
        stderr: "piped",
        stdin: "piped",
    };

    const p = Deno.run(opt);

    if (stdInput != undefined) {
        await p.stdin.write(new TextEncoder().encode(stdInput));
        await p.stdin.close();
    }

    const status = await p.status()
    const output = await p.output()
    const stderrOutput = await p.stderrOutput()
    if (status.code === 0) {
        return JSON.parse(new TextDecoder().decode(output))
    } else {
        let error = new TextDecoder().decode(stderrOutput);
        return ""
    }
}

function clusterRoleBindingTemplate() {
    return `
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: view-${serviceAccName}-global
subjects:
- kind: ServiceAccount
  name: ${serviceAccName}
  namespace: default
roleRef:
  kind: ClusterRole
  name: view
  apiGroup: rbac.authorization.k8s.io    
`
}

function secretYamlTemplate() {
    return `
apiVersion: v1
kind: Secret
metadata:
  name: ${serviceAccName}-secret
  annotations:
    kubernetes.io/service-account.name: ${serviceAccName}
type: kubernetes.io/service-account-token`
}

function kubeConfigTemplate(certificateAuthority: string, token: string, clusterApi: string, clusterName: string, username: string) {
    return `
## KUBECONFIG generated on ${new Date()}
apiVersion: v1
clusters:
- cluster:
    certificate-authority-data: ${certificateAuthority}
    server: ${clusterApi}
  name: ${clusterName}
contexts:
- context:
    cluster: ${clusterName}
    user: ${username}
  name: ${clusterName}
current-context: ${clusterName}
kind: Config
preferences: {}
users:
- name: ${username}
  user:
    token: ${token}
`
}

If edit or admin rights are to be assigned, the ClusterRoleBinding must be adapted in the roleRef section with the roles listed below.

Furthermore, you can restrict this to a single namespace by not creating a ClusterRoleBinding but only a RoleBinding within the desired namespace.

2.2 - High Availability

2.2.1 - Best Practices

Implementing High Availability and Tolerating Zone Outages

Developing highly available workload that can tolerate a zone outage is no trivial task. You will find here various recommendations to get closer to that goal. While many recommendations are general enough, the examples are specific in how to achieve this in a Gardener-managed cluster and where/how to tweak the different control plane components. If you do not use Gardener, it may be still a worthwhile read.

First however, what is a zone outage? It sounds like a clear-cut “thing”, but it isn’t. There are many things that can go haywire. Here are some examples:

• Elevated cloud provider API error rates for individual or multiple services
• Network bandwidth reduced or latency increased, usually also effecting storage sub systems as they are network attached
• No networking at all, no DNS, machines shutting down or restarting, …
• Functional issues, of either the entire service (e.g. all block device operations) or only parts of it (e.g. LB listener registration)
• All services down, temporarily or permanently (the proverbial burning down data center 🔥)

This and everything in between make it hard to prepare for such events, but you can still do a lot. The most important recommendation is to not target specific issues exclusively - tomorrow another service will fail in an unanticipated way. Also, focus more on meaningful availability than on internal signals (useful, but not as relevant as the former). Always prefer automation over manual intervention (e.g. leader election is a pretty robust mechanism, auto-scaling may be required as well, etc.).

Also remember that HA is costly - you need to balance it against the cost of an outage as silly as this may sound, e.g. running all this excess capacity “just in case” vs. “going down” vs. a risk-based approach in between where you have means that will kick in, but they are not guaranteed to work (e.g. if the cloud provider is out of resource capacity). Maybe some of your components must run at the highest possible availability level, but others not - that’s a decision only you can make.

Control Plane

The Kubernetes cluster control plane is managed by Gardener (as pods in separate infrastructure clusters to which you have no direct access) and can be set up with no failure tolerance (control plane pods will be recreated best-effort when resources are available) or one of the failure tolerance types node or zone.

Strictly speaking, static workload does not depend on the (high) availability of the control plane, but static workload doesn’t rhyme with Cloud and Kubernetes and also means, that when you possibly need it the most, e.g. during a zone outage, critical self-healing or auto-scaling functionality won’t be available to you and your workload, if your control plane is down as well. That’s why, even though the resource consumption is significantly higher, we generally recommend to use the failure tolerance type zone for the control planes of productive clusters, at least in all regions that have 3+ zones. Regions that have only 1 or 2 zones don’t support the failure tolerance type zone and then your second best option is the failure tolerance type node, which means a zone outage can still take down your control plane, but individual node outages won’t.

In the shoot resource it’s merely only this what you need to add:

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  controlPlane:
    highAvailability:
      failureTolerance:
        type: zone # valid values are `node` and `zone` (only available if your control plane resides in a region with 3+ zones)

This setting will scale out all control plane components for a Gardener cluster as necessary, so that no single zone outage can take down the control plane for longer than just a few seconds for the fail-over to take place (e.g. lease expiration and new leader election or readiness probe failure and endpoint removal). Components run highly available in either active-active (servers) or active-passive (controllers) mode at all times, the persistence (ETCD), which is consensus-based, will tolerate the loss of one zone and still maintain quorum and therefore remain operational. These are all patterns that we will revisit down below also for your own workload.

Worker Pools

Now that you have configured your Kubernetes cluster control plane in HA, i.e. spread it across multiple zones, you need to do the same for your own workload, but in order to do so, you need to spread your nodes across multiple zones first.

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  provider:
    workers:
    - name: ...
      minimum: 6
      maximum: 60
      zones:
      - ...

Prefer regions with at least 2, better 3+ zones and list the zones in the zones section for each of your worker pools. Whether you need 2 or 3 zones at a minimum depends on your fail-over concept:

• Consensus-based software components (like ETCD) depend on maintaining a quorum of (n/2)+1, so you need at least 3 zones to tolerate the outage of 1 zone.
• Primary/Secondary-based software components need just 2 zones to tolerate the outage of 1 zone.
• Then there are software components that can scale out horizontally. They are probably fine with 2 zones, but you also need to think about the load-shift and that the remaining zone must then pick up the work of the unhealthy zone. With 2 zones, the remaining zone must cope with an increase of 100% load. With 3 zones, the remaining zones must only cope with an increase of 50% load (per zone).

In general, the question is also whether you have the fail-over capacity already up and running or not. If not, i.e. you depend on re-scheduling to a healthy zone or auto-scaling, be aware that during a zone outage, you will see a resource crunch in the healthy zones. If you have no automation, i.e. only human operators (a.k.a. “red button approach”), you probably will not get the machines you need and even with automation, it may be tricky. But holding the capacity available at all times is costly. In the end, that’s a decision only you can make. If you made that decision, please adapt the minimum, maximum, maxSurge and maxUnavailable settings for your worker pools accordingly (visit this section for more information).

Also, consider fall-back worker pools (with different/alternative machine types) and cluster autoscaler expanders using a priority-based strategy.

Gardener-managed clusters deploy the cluster autoscaler or CA for short and you can tweak the general CA knobs for Gardener-managed clusters like this:

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  kubernetes:
    clusterAutoscaler:
      expander: "least-waste"
      scanInterval: 10s
      scaleDownDelayAfterAdd: 60m
      scaleDownDelayAfterDelete: 0s
      scaleDownDelayAfterFailure: 3m
      scaleDownUnneededTime: 30m
      scaleDownUtilizationThreshold: 0.5

If you want to be ready for a sudden spike or have some buffer in general, over-provision nodes by means of “placeholder” pods with low priority and appropriate resource requests. This way, they will demand nodes to be provisioned for them, but if any pod comes up with a regular/higher priority, the low priority pods will be evicted to make space for the more important ones. Strictly speaking, this is not related to HA, but it may be important to keep this in mind as you generally want critical components to be rescheduled as fast as possible and if there is no node available, it may take 3 minutes or longer to do so (depending on the cloud provider). Besides, not only zones can fail, but also individual nodes.

Replicas (Horizontal Scaling)

Now let’s talk about your workload. In most cases, this will mean to run multiple replicas. If you cannot do that (a.k.a. you have a singleton), that’s a bad situation to be in. Maybe you can run a spare (secondary) as backup? If you cannot, you depend on quick detection and rescheduling of your singleton (more on that below).

Obviously, things get messier with persistence. If you have persistence, you should ideally replicate your data, i.e. let your spare (secondary) “follow” your main (primary). If your software doesn’t support that, you have to deploy other means, e.g. volume snapshotting or side-backups (specific to the software you deploy; keep the backups regional, so that you can switch to another zone at all times). If you have to do those, your HA scenario becomes more a DR scenario and terms like RPO and RTO become relevant to you:

• Recovery Point Objective (RPO): Potential data loss, i.e. how much data will you lose at most (time between backups)
• Recovery Time Objective (RTO): Time until recovery, i.e. how long does it take you to be operational again (time to restore)

Also, keep in mind that your persistent volumes are usually zonal, i.e. once you have a volume in one zone, it’s bound to that zone and you cannot get up your pod in another zone w/o first recreating the volume yourself (Kubernetes won’t help you here directly).

Anyway, best avoid that, if you can (from technical and cost perspective). The best solution (and also the most costly one) is to run multiple replicas in multiple zones and keep your data replicated at all times, so that your RPO is always 0 (best). That’s what we do for Gardener-managed cluster HA control planes (ETCD) as any data loss may be disastrous and lead to orphaned resources (in addition, we deploy side cars that do side-backups for disaster recovery, with full and incremental snapshots with an RPO of 5m).

So, how to run with multiple replicas? That’s the easiest part in Kubernetes and the two most important resources, Deployments and StatefulSet, support that out of the box:

apiVersion: apps/v1
kind: Deployment | StatefulSet
spec:
  replicas: ...

The problem comes with the number of replicas. It’s easy only if the number is static, e.g. 2 for active-active/passive or 3 for consensus-based software components, but what with software components that can scale out horizontally? Here you usually do not set the number of replicas statically, but make use of the horizontal pod autoscaler or HPA for short (built-in; part of the kube-controller-manager). There are also other options like the cluster proportional autoscaler, but while the former works based on metrics, the latter is more a guestimate approach that derives the number of replicas from the number of nodes/cores in a cluster. Sometimes useful, but often blind to the actual demand.

So, HPA it is then for most of the cases. However, what is the resource (e.g. CPU or memory) that drives the number of desired replicas? Again, this is up to you, but not always are CPU or memory the best choices. In some cases, custom metrics may be more appropriate, e.g. requests per second (it was also for us).

You will have to create specific HorizontalPodAutoscaler resources for your scale target and can tweak the general HPA knobs for Gardener-managed clusters like this:

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  kubernetes:
    kubeControllerManager:
      horizontalPodAutoscaler:
        syncPeriod: 15s
        tolerance: 0.1
        downscaleStabilization: 5m0s
        initialReadinessDelay: 30s
        cpuInitializationPeriod: 5m0s

Resources (Vertical Scaling)

While it is important to set a sufficient number of replicas, it is also important to give the pods sufficient resources (CPU and memory). This is especially true when you think about HA. When a zone goes down, you might need to get up replacement pods, if you don’t have them running already to take over the load from the impacted zone. Likewise, e.g. with active-active software components, you can expect the remaining pods to receive more load. If you cannot scale them out horizontally to serve the load, you will probably need to scale them out (or rather up) vertically. This is done by the vertical pod autoscaler or VPA for short (not built-in; part of the kubernetes/autoscaler repository).

A few caveats though:

• You cannot use HPA and VPA on the same metrics as they would influence each other, which would lead to pod trashing (more replicas require fewer resources; fewer resources require more replicas)
• Scaling horizontally doesn’t cause downtimes (at least not when out-scaling and only one replica is affected when in-scaling), but scaling vertically does (if the pod runs OOM anyway, but also when new recommendations are applied, resource requests for existing pods may be changed, which causes the pods to be rescheduled). Although the discussion is going on for a very long time now, that is still not supported in-place yet (see KEP 1287, implementation in Kubernetes, implementation in VPA).

VPA is a useful tool and Gardener-managed clusters deploy a VPA by default for you (HPA is supported anyway as it’s built into the kube-controller-manager). You will have to create specific VerticalPodAutoscaler resources for your scale target and can tweak the general VPA knobs for Gardener-managed clusters like this:

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  kubernetes:
    verticalPodAutoscaler:
      enabled: true
      evictAfterOOMThreshold: 10m0s
      evictionRateBurst: 1
      evictionRateLimit: -1
      evictionTolerance: 0.5
      recommendationMarginFraction: 0.15
      updaterInterval: 1m0s
      recommenderInterval: 1m0s

While horizontal pod autoscaling is relatively straight-forward, it takes a long time to master vertical pod autoscaling. We saw performance issues, hard-coded behavior (on OOM, memory is bumped by +20% and it may take a few iterations to reach a good level), unintended pod disruptions by applying new resource requests (after 12h all targeted pods will receive new requests even though individually they would be fine without, which also drives active-passive resource consumption up), difficulties to deal with spiky workload in general (due to the algorithmic approach it takes), recommended requests may exceed node capacity, limit scaling is proportional and therefore often questionable, and more. VPA is a double-edged sword: useful and necessary, but not easy to handle.

For the Gardener-managed components, we mostly removed limits. Why?

• CPU limits have almost always only downsides. They cause needless CPU throttling, which is not even easily visible. CPU requests turn into cpu shares, so if the node has capacity, the pod may consume the freely available CPU, but not if you have set limits, which curtail the pod by means of cpu quota. There are only certain scenarios in which they may make sense, e.g. if you set requests=limits and thereby define a pod with guaranteed QoS, which influences your cgroup placement. However, that is difficult to do for the components you implement yourself and practically impossible for the components you just consume, because what’s the correct value for requests/limits and will it hold true also if the load increases and what happens if a zone goes down or with the next update/version of this component? If anything, CPU limits caused outages, not helped prevent them.
• As for memory limits, they are slightly more useful, because CPU is compressible and memory is not, so if one pod runs berserk, it may take others down (with CPU, cpu shares make it as fair as possible), depending on which OOM killer strikes (a complicated topic by itself). You don’t want the operating system OOM killer to strike as the result is unpredictable. Better, it’s the cgroup OOM killer or even the kubelet’s eviction, if the consumption is slow enough as it takes priorities into consideration even. If your component is critical and a singleton (e.g. node daemon set pods), you are better off also without memory limits, because letting the pod go OOM because of artificial/wrong memory limits can mean that the node becomes unusable. Hence, such components also better run only with no or a very high memory limit, so that you can catch the occasional memory leak (bug) eventually, but under normal operation, if you cannot decide about a true upper limit, rather not have limits and cause endless outages through them or when you need the pods the most (during a zone outage) where all your assumptions went out of the window.

The downside of having poor or no limits and poor and no requests is that nodes may “die” more often. Contrary to the expectation, even for managed services, the managed service is not responsible or cannot guarantee the health of a node under all circumstances, since the end user defines what is run on the nodes (shared responsibility). If the workload exhausts any resource, it will be the end of the node, e.g. by compressing the CPU too much (so that the kubelet fails to do its work), exhausting the main memory too fast, disk space, file handles, or any other resource.

The kubelet allows for explicit reservation of resources for operating system daemons (system-reserved) and Kubernetes daemons (kube-reserved) that are subtracted from the actual node resources and become the allocatable node resources for your workload/pods. All managed services configure these settings “by rule of thumb” (a balancing act), but cannot guarantee that the values won’t waste resources or always will be sufficient. You will have to fine-tune them eventually and adapt them to your needs. In addition, you can configure soft and hard eviction thresholds to give the kubelet some headroom to evict “greedy” pods in a controlled way. These settings can be configured for Gardener-managed clusters like this:

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  kubernetes:
    kubelet:
      kubeReserved:                            # explicit resource reservation for Kubernetes daemons
        cpu: 100m
        memory: 1Gi
        ephemeralStorage: 1Gi
        pid: 1000
      evictionSoft:                            # soft, i.e. graceful eviction (used if the node is about to run out of resources, avoiding hard evictions)
        memoryAvailable: 200Mi
        imageFSAvailable: 10%
        imageFSInodesFree: 10%
        nodeFSAvailable: 10%
        nodeFSInodesFree: 10%
      evictionSoftGracePeriod:                 # caps pod's `terminationGracePeriodSeconds` value during soft evictions (specific grace periods)
        memoryAvailable: 1m30s
        imageFSAvailable: 1m30s
        imageFSInodesFree: 1m30s
        nodeFSAvailable: 1m30s
        nodeFSInodesFree: 1m30s
      evictionHard:                            # hard, i.e. immediate eviction (used if the node is out of resources, avoiding the OS generally run out of resources fail processes indiscriminately)
        memoryAvailable: 100Mi
        imageFSAvailable: 5%
        imageFSInodesFree: 5%
        nodeFSAvailable: 5%
        nodeFSInodesFree: 5%
      evictionMinimumReclaim:                  # additional resources to reclaim after hitting the hard eviction thresholds to not hit the same thresholds soon after again
        memoryAvailable: 0Mi
        imageFSAvailable: 0Mi
        imageFSInodesFree: 0Mi
        nodeFSAvailable: 0Mi
        nodeFSInodesFree: 0Mi
      evictionMaxPodGracePeriod: 90            # caps pod's `terminationGracePeriodSeconds` value during soft evictions (general grace periods)
      evictionPressureTransitionPeriod: 5m0s   # stabilization time window to avoid flapping of node eviction state

You can tweak these settings also individually per worker pool (spec.provider.workers.kubernetes.kubelet...), which makes sense especially with different machine types (and also workload that you may want to schedule there).

Physical memory is not compressible, but you can overcome this issue to some degree (alpha since Kubernetes v1.22 in combination with the feature gate NodeSwap on the kubelet) with swap memory. You can read more in this introductory blog and the docs. If you chose to use it (still only alpha at the time of this writing) you may want to consider also the risks associated with swap memory:

• Reduced performance predictability
• Reduced performance up to page trashing
• Reduced security as secrets, normally held only in memory, could be swapped out to disk

That said, the various options mentioned above are only remotely related to HA and will not be further explored throughout this document, but just to remind you: if a zone goes down, load patterns will shift, existing pods will probably receive more load and will require more resources (especially because it is often practically impossible to set “proper” resource requests, which drive node allocation - limits are always ignored by the scheduler) or more pods will/must be placed on the existing and/or new nodes and then these settings, which are generally critical (especially if you switch on bin-packing for Gardener-managed clusters as a cost saving measure), will become even more critical during a zone outage.

Probes

Before we go down the rabbit hole even further and talk about how to spread your replicas, we need to talk about probes first, as they will become relevant later. Kubernetes supports three kinds of probes: startup, liveness, and readiness probes. If you are a visual thinker, also check out this slide deck by Tim Hockin (Kubernetes networking SIG chair).

Basically, the startupProbe and the livenessProbe help you restart the container, if it’s unhealthy for whatever reason, by letting the kubelet that orchestrates your containers on a node know, that it’s unhealthy. The former is a special case of the latter and only applied at the startup of your container, if you need to handle the startup phase differently (e.g. with very slow starting containers) from the rest of the lifetime of the container.

Now, the readinessProbe helps you manage the ready status of your container and thereby pod (any container that is not ready turns the pod not ready). This again has impact on endpoints and pod disruption budgets:

• If the pod is not ready, the endpoint will be removed and the pod will not receive traffic anymore
• If the pod is not ready, the pod counts into the pod disruption budget and if the budget is exceeded, no further voluntary pod disruptions will be permitted for the remaining ready pods (e.g. no eviction, no voluntary horizontal or vertical scaling, if the pod runs on a node that is about to be drained or in draining, draining will be paused until the max drain timeout passes)

As you can see, all of these probes are (also) related to HA (mostly the readinessProbe, but depending on your workload, you can also leverage livenessProbe and startupProbe into your HA strategy). If Kubernetes doesn’t know about the individual status of your container/pod, it won’t do anything for you (right away). That said, later/indirectly something might/will happen via the node status that can also be ready or not ready, which influences the pods and load balancer listener registration (a not ready node will not receive cluster traffic anymore), but this process is worker pool global and reacts delayed and also doesn’t discriminate between the containers/pods on a node.

In addition, Kubernetes also offers pod readiness gates to amend your pod readiness with additional custom conditions (normally, only the sum of the container readiness matters, but pod readiness gates additionally count into the overall pod readiness). This may be useful if you want to block (by means of pod disruption budgets that we will talk about next) the roll-out of your workload/nodes in case some (possibly external) condition fails.

Pod Disruption Budgets

One of the most important resources that help you on your way to HA are pod disruption budgets or PDB for short. They tell Kubernetes how to deal with voluntary pod disruptions, e.g. during the deployment of your workload, when the nodes are rolled, or just in general when a pod shall be evicted/terminated. Basically, if the budget is reached, they block all voluntary pod disruptions (at least for a while until possibly other timeouts act or things happen that leave Kubernetes no choice anymore, e.g. the node is forcefully terminated). You should always define them for your workload.

Very important to note is that they are based on the readinessProbe, i.e. even if all of your replicas are lively, but not enough of them are ready, this blocks voluntary pod disruptions, so they are very critical and useful. Here an example (you can specify either minAvailable or maxUnavailable in absolute numbers or as percentage):

apiVersion: policy/v1
kind: PodDisruptionBudget
spec:
  maxUnavailable: 1
  selector:
    matchLabels:
      ...

And please do not specify a PDB of maxUnavailable being 0 or similar. That’s pointless, even detrimental, as it blocks then even useful operations, forces always the hard timeouts that are less graceful and it doesn’t make sense in the context of HA. You cannot “force” HA by preventing voluntary pod disruptions, you must work with the pod disruptions in a resilient way. Besides, PDBs are really only about voluntary pod disruptions - something bad can happen to a node/pod at any time and PDBs won’t make this reality go away for you.

PDBs will not always work as expected and can also get in your way, e.g. if the PDB is violated or would be violated, it may possibly block whatever you are trying to do to salvage the situation, e.g. drain a node or deploy a patch version (if the PDB is or would be violated, not even unhealthy pods would be evicted as they could theoretically become healthy again, which Kubernetes doesn’t know). In order to overcome this issue, it is now possible (alpha since Kubernetes v1.26 in combination with the feature gate PDBUnhealthyPodEvictionPolicy on the API server, beta and enabled by default since Kubernetes v1.27) to configure the so-called unhealthy pod eviction policy. The default is still IfHealthyBudget as a change in default would have changed the behavior (as described above), but you can now also set AlwaysAllow at the PDB (spec.unhealthyPodEvictionPolicy). For more information, please check out this discussion, the PR and this document and balance the pros and cons for yourself. In short, the new AlwaysAllow option is probably the better choice in most of the cases while IfHealthyBudget is useful only if you have frequent temporary transitions or for special cases where you have already implemented controllers that depend on the old behavior.

Pod Topology Spread Constraints

Pod topology spread constraints or PTSC for short (no official abbreviation exists, but we will use this in the following) are enormously helpful to distribute your replicas across multiple zones, nodes, or any other user-defined topology domain. They complement and improve on pod (anti-)affinities that still exist and can be used in combination.

PTSCs are an improvement, because they allow for maxSkew and minDomains. You can steer the “level of tolerated imbalance” with maxSkew, e.g. you probably want that to be at least 1, so that you can perform a rolling update, but this all depends on your deployment (maxUnavailable and maxSurge), etc. Stateful sets are a bit different (maxUnavailable) as they are bound to volumes and depend on them, so there usually cannot be 2 pods requiring the same volume. minDomains is a hint to tell the scheduler how far to spread, e.g. if all nodes in one zone disappeared because of a zone outage, it may “appear” as if there are only 2 zones in a 3 zones cluster and the scheduling decisions may end up wrong, so a minDomains of 3 will tell the scheduler to spread to 3 zones before adding another replica in one zone. Be careful with this setting as it also means, if one zone is down the “spread” is already at least 1, if pods run in the other zones. This is useful where you have exactly as many replicas as you have zones and you do not want any imbalance. Imbalance is critical as if you end up with one, nobody is going to do the (active) re-balancing for you (unless you deploy and configure additional non-standard components such as the descheduler). So, for instance, if you have something like a DBMS that you want to spread across 2 zones (active-passive) or 3 zones (consensus-based), you better specify minDomains of 2 respectively 3 to force your replicas into at least that many zones before adding more replicas to another zone (if supported).

Anyway, PTSCs are critical to have, but not perfect, so we saw (unsurprisingly, because that’s how the scheduler works), that the scheduler may block the deployment of new pods because it takes the decision pod-by-pod (see for instance #109364).

Pod Affinities and Anti-Affinities

As said, you can combine PTSCs with pod affinities and/or anti-affinities. Especially inter-pod (anti-)affinities may be helpful to place pods apart, e.g. because they are fall-backs for each other or you do not want multiple potentially resource-hungry “best-effort” or “burstable” pods side-by-side (noisy neighbor problem), or together, e.g. because they form a unit and you want to reduce the failure domain, reduce the network latency, and reduce the costs.

Topology Aware Hints

While topology aware hints are not directly related to HA, they are very relevant in the HA context. Spreading your workload across multiple zones may increase network latency and cost significantly, if the traffic is not shaped. Topology aware hints (beta since Kubernetes v1.23, replacing the now deprecated topology aware traffic routing with topology keys) help to route the traffic within the originating zone, if possible. Basically, they tell kube-proxy how to setup your routing information, so that clients can talk to endpoints that are located within the same zone.

Be aware however, that there are some limitations. Those are called safeguards and if they strike, the hints are off and traffic is routed again randomly. Especially controversial is the balancing limitation as there is the assumption, that the load that hits an endpoint is determined by the allocatable CPUs in that topology zone, but that’s not always, if even often, the case (see for instance #113731 and #110714). So, this limitation hits far too often and your hints are off, but then again, it’s about network latency and cost optimization first, so it’s better than nothing.

Networking

We have talked about networking only to some small degree so far (readiness probes, pod disruption budgets, topology aware hints). The most important component is probably your ingress load balancer - everything else is managed by Kubernetes. AWS, Azure, GCP, and also OpenStack offer multi-zonal load balancers, so make use of them. In Azure and GCP, LBs are regional whereas in AWS and OpenStack, they need to be bound to a zone, which the cloud-controller-manager does by observing the zone labels at the nodes (please note that this behavior is not always working as expected, see #570 where the AWS cloud-controller-manager is not readjusting to newly observed zones).

Please be reminded that even if you use a service mesh like Istio, the off-the-shelf installation/configuration usually never comes with productive settings (to simplify first-time installation and improve first-time user experience) and you will have to fine-tune your installation/configuration, much like the rest of your workload.

Relevant Cluster Settings

Following now a summary/list of the more relevant settings you may like to tune for Gardener-managed clusters:

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  controlPlane:
    highAvailability:
      failureTolerance:
        type: zone # valid values are `node` and `zone` (only available if your control plane resides in a region with 3+ zones)
  kubernetes:
    kubeAPIServer:
      defaultNotReadyTolerationSeconds: 300
      defaultUnreachableTolerationSeconds: 300
    kubelet:
      ...
    kubeScheduler:
      featureGates:
        MinDomainsInPodTopologySpread: true
    kubeControllerManager:
      nodeMonitorGracePeriod: 40s
      horizontalPodAutoscaler:
        syncPeriod: 15s
        tolerance: 0.1
        downscaleStabilization: 5m0s
        initialReadinessDelay: 30s
        cpuInitializationPeriod: 5m0s
    verticalPodAutoscaler:
      enabled: true
      evictAfterOOMThreshold: 10m0s
      evictionRateBurst: 1
      evictionRateLimit: -1
      evictionTolerance: 0.5
      recommendationMarginFraction: 0.15
      updaterInterval: 1m0s
      recommenderInterval: 1m0s
    clusterAutoscaler:
      expander: "least-waste"
      scanInterval: 10s
      scaleDownDelayAfterAdd: 60m
      scaleDownDelayAfterDelete: 0s
      scaleDownDelayAfterFailure: 3m
      scaleDownUnneededTime: 30m
      scaleDownUtilizationThreshold: 0.5
  provider:
    workers:
    - name: ...
      minimum: 6
      maximum: 60
      maxSurge: 3
      maxUnavailable: 0
      zones:
      - ... # list of zones you want your worker pool nodes to be spread across, see above
      kubernetes:
        kubelet:
          ... # similar to `kubelet` above (cluster-wide settings), but here per worker pool (pool-specific settings), see above
      machineControllerManager: # optional, it allows to configure the machine-controller settings.
        machineCreationTimeout: 20m
        machineHealthTimeout: 10m
        machineDrainTimeout: 60h
  systemComponents:
    coreDNS:
      autoscaling:
        mode: horizontal # valid values are `horizontal` (driven by CPU load) and `cluster-proportional` (driven by number of nodes/cores)

On spec.controlPlane.highAvailability.failureTolerance.type

If set, determines the degree of failure tolerance for your control plane. zone is preferred, but only available if your control plane resides in a region with 3+ zones. See above and the docs.

On spec.kubernetes.kubeAPIServer.defaultUnreachableTolerationSeconds and defaultNotReadyTolerationSeconds

This is a very interesting API server setting that lets Kubernetes decide how fast to evict pods from nodes whose status condition of type Ready is either Unknown (node status unknown, a.k.a unreachable) or False (kubelet not ready) (see node status conditions; please note that kubectl shows both values as NotReady which is a somewhat “simplified” visualization).

You can also override the cluster-wide API server settings individually per pod:

spec:
  tolerations:
  - key: "node.kubernetes.io/unreachable"
    operator: "Exists"
    effect: "NoExecute"
    tolerationSeconds: 0
  - key: "node.kubernetes.io/not-ready"
    operator: "Exists"
    effect: "NoExecute"
    tolerationSeconds: 0

This will evict pods on unreachable or not-ready nodes immediately, but be cautious: 0 is very aggressive and may lead to unnecessary disruptions. Again, you must decide for your own workload and balance out the pros and cons (e.g. long startup time).

Please note, these settings replace spec.kubernetes.kubeControllerManager.podEvictionTimeout that was deprecated with Kubernetes v1.26 (and acted as an upper bound).

On spec.kubernetes.kubeScheduler.featureGates.MinDomainsInPodTopologySpread

Required to be enabled for minDomains to work with PTSCs (beta since Kubernetes v1.25, but off by default). See above and the docs. This tells the scheduler, how many topology domains to expect (=zones in the context of this document).

On spec.kubernetes.kubeControllerManager.nodeMonitorGracePeriod

This is another very interesting kube-controller-manager setting that can help you speed up or slow down how fast a node shall be considered Unknown (node status unknown, a.k.a unreachable) when the kubelet is not updating its status anymore (see node status conditions), which effects eviction (see spec.kubernetes.kubeAPIServer.defaultUnreachableTolerationSeconds and defaultNotReadyTolerationSeconds above). The shorter the time window, the faster Kubernetes will act, but the higher the chance of flapping behavior and pod trashing, so you may want to balance that out according to your needs, otherwise stick to the default which is a reasonable compromise.

On spec.kubernetes.kubeControllerManager.horizontalPodAutoscaler...

This configures horizontal pod autoscaling in Gardener-managed clusters. See above and the docs for the detailed fields.

On spec.kubernetes.verticalPodAutoscaler...

This configures vertical pod autoscaling in Gardener-managed clusters. See above and the docs for the detailed fields.

On spec.kubernetes.clusterAutoscaler...

This configures node auto-scaling in Gardener-managed clusters. See above and the docs for the detailed fields, especially about expanders, which may become life-saving in case of a zone outage when a resource crunch is setting in and everybody rushes to get machines in the healthy zones.

In case of a zone outage, it is critical to understand how the cluster autoscaler will put a worker pool in one zone into “back-off” and what the consequences for your workload will be. Unfortunately, the official cluster autoscaler documentation does not explain these details, but you can find hints in the source code:

If a node fails to come up, the node group (worker pool in that zone) will go into “back-off”, at first 5m, then exponentially longer until the maximum of 30m is reached. The “back-off” is reset after 3 hours. This in turn means, that nodes must be first considered Unknown, which happens when spec.kubernetes.kubeControllerManager.nodeMonitorGracePeriod lapses (e.g. at the beginning of a zone outage). Then they must either remain in this state until spec.provider.workers.machineControllerManager.machineHealthTimeout lapses for them to be recreated, which will fail in the unhealthy zone, or spec.kubernetes.kubeAPIServer.defaultUnreachableTolerationSeconds lapses for the pods to be evicted (usually faster than node replacements, depending on your configuration), which will trigger the cluster autoscaler to create more capacity, but very likely in the same zone as it tries to balance its node groups at first, which will fail in the unhealthy zone. It will be considered failed only when maxNodeProvisionTime lapses (usually close to spec.provider.workers.machineControllerManager.machineCreationTimeout) and only then put the node group into “back-off” and not retry for 5m (at first and then exponentially longer). Only then you can expect new node capacity to be brought up somewhere else.

During the time of ongoing node provisioning (before a node group goes into “back-off”), the cluster autoscaler may have “virtually scheduled” pending pods onto those new upcoming nodes and will not reevaluate these pods anymore unless the node provisioning fails (which will fail during a zone outage, but the cluster autoscaler cannot know that and will therefore reevaluate its decision only after it has given up on the new nodes).

It’s critical to keep that in mind and accommodate for it. If you have already capacity up and running, the reaction time is usually much faster with leases (whatever you set) or endpoints (spec.kubernetes.kubeControllerManager.nodeMonitorGracePeriod), but if you depend on new/fresh capacity, the above should inform you how long you will have to wait for it and for how long pods might be pending (because capacity is generally missing and pending pods may have been “virtually scheduled” to new nodes that won’t come up until the node group goes eventually into “back-off” and nodes in the healthy zones come up).

On spec.provider.workers.minimum, maximum, maxSurge, maxUnavailable, zones, and machineControllerManager

Each worker pool in Gardener may be configured differently. Among many other settings like machine type, root disk, Kubernetes version, kubelet settings, and many more you can also specify the lower and upper bound for the number of machines (minimum and maximum), how many machines may be added additionally during a rolling update (maxSurge) and how many machines may be in termination/recreation during a rolling update (maxUnavailable), and of course across how many zones the nodes shall be spread (zones).

Gardener divides minimum, maximum, maxSurge, maxUnavailable values by the number of zones specified for this worker pool. This fact must be considered when you plan the sizing of your worker pools.

Example:

  provider:
    workers:
    - name: ...
      minimum: 6
      maximum: 60
      maxSurge: 3
      maxUnavailable: 0
      zones: ["a", "b", "c"]

• The resulting MachineDeployments per zone will get minimum: 2, maximum: 20, maxSurge: 1, maxUnavailable: 0.
• If another zone is added all values will be divided by 4, resulting in:
  • Less workers per zone.
  • ⚠️ One MachineDeployment with maxSurge: 0, i.e. there will be a replacement of nodes without rolling updates.

Interesting is also the configuration for Gardener’s machine-controller-manager or MCM for short that provisions, monitors, terminates, replaces, or updates machines that back your nodes:

• The shorter machineCreationTimeout is, the faster MCM will retry to create a machine/node, if the process is stuck on cloud provider side. It is set to useful/practical timeouts for the different cloud providers and you probably don’t want to change those (in the context of HA at least). Please align with the cluster autoscaler’s maxNodeProvisionTime.
• The shorter machineHealthTimeout is, the faster MCM will replace machines/nodes in case the kubelet isn’t reporting back, which translates to Unknown, or reports back with NotReady, or the node-problem-detector that Gardener deploys for you reports a non-recoverable issue/condition (e.g. read-only file system). If it is too short however, you risk node and pod trashing, so be careful.
• The shorter machineDrainTimeout is, the faster you can get rid of machines/nodes that MCM decided to remove, but this puts a cap on the grace periods and PDBs. They are respected up until the drain timeout lapses - then the machine/node will be forcefully terminated, whether or not the pods are still in termination or not even terminated because of PDBs. Those PDBs will then be violated, so be careful here as well. Please align with the cluster autoscaler’s maxGracefulTerminationSeconds.

Especially the last two settings may help you recover faster from cloud provider issues.

On spec.systemComponents.coreDNS.autoscaling

DNS is critical, in general and also within a Kubernetes cluster. Gardener-managed clusters deploy CoreDNS, a graduated CNCF project. Gardener supports 2 auto-scaling modes for it, horizontal (using HPA based on CPU) and cluster-proportional (using cluster proportional autoscaler that scales the number of pods based on the number of nodes/cores, not to be confused with the cluster autoscaler that scales nodes based on their utilization). Check out the docs, especially the trade-offs why you would chose one over the other (cluster-proportional gives you more configuration options, if CPU-based horizontal scaling is insufficient to your needs). Consider also Gardener’s feature node-local DNS to decouple you further from the DNS pods and stabilize DNS. Again, that’s not strictly related to HA, but may become important during a zone outage, when load patterns shift and pods start to initialize/resolve DNS records more frequently in bulk.

More Caveats

Unfortunately, there are a few more things of note when it comes to HA in a Kubernetes cluster that may be “surprising” and hard to mitigate:

• If the kubelet restarts, it will report all pods as NotReady on startup until it reruns its probes (#100277), which leads to temporary endpoint and load balancer target removal (#102367). This topic is somewhat controversial. Gardener uses rolling updates and a jitter to spread necessary kubelet restarts as good as possible.
• If a kube-proxy pod on a node turns NotReady, all load balancer traffic to all pods (on this node) under services with externalTrafficPolicy local will cease as the load balancer will then take this node out of serving. This topic is somewhat controversial as well. So, please remember that externalTrafficPolicy local not only has the disadvantage of imbalanced traffic spreading, but also a dependency to the kube-proxy pod that may and will be unavailable during updates. Gardener uses rolling updates to spread necessary kube-proxy updates as good as possible.

These are just a few additional considerations. They may or may not affect you, but other intricacies may. It’s a reminder to be watchful as Kubernetes may have one or two relevant quirks that you need to consider (and will probably only find out over time and with extensive testing).

Meaningful Availability

Finally, let’s go back to where we started. We recommended to measure meaningful availability. For instance, in Gardener, we do not trust only internal signals, but track also whether Gardener or the control planes that it manages are externally available through the external DNS records and load balancers, SNI-routing Istio gateways, etc. (the same path all users must take). It’s a huge difference whether the API server’s internal readiness probe passes or the user can actually reach the API server and it does what it’s supposed to do. Most likely, you will be in a similar spot and can do the same.

What you do with these signals is another matter. Maybe there are some actionable metrics and you can trigger some active fail-over, maybe you can only use it to improve your HA setup altogether. In our case, we also use it to deploy mitigations, e.g. via our dependency-watchdog that watches, for instance, Gardener-managed API servers and shuts down components like the controller managers to avert cascading knock-off effects (e.g. melt-down if the kubelets cannot reach the API server, but the controller managers can and start taking down nodes and pods).

Either way, understanding how users perceive your service is key to the improvement process as a whole. Even if you are not struck by a zone outage, the measures above and tracking the meaningful availability will help you improve your service.

Thank you for your interest.

2.2.2 - Chaos Engineering

Overview

Gardener provides chaostoolkit modules to simulate compute and network outages for various cloud providers such as AWS, Azure, GCP, OpenStack/Converged Cloud, and VMware vSphere, as well as pod disruptions for any Kubernetes cluster.

The API, parameterization, and implementation is as homogeneous as possible across the different cloud providers, so that you have only minimal effort. As a Gardener user, you benefit from an additional garden module that leverages the generic modules, but exposes their functionality in the most simple, homogeneous, and secure way (no need to specify cloud provider credentials, cluster credentials, or filters explicitly; retrieves credentials and stores them in memory only).

Installation

The name of the package is chaosgarden and it was developed and tested with Python 3.9+. It’s being published to PyPI, so that you can comfortably install it via Python’s package installer pip (you may want to create a virtual environment before installing it):

pip install chaosgarden

ℹ️ If you want to use the VMware vSphere module, please note the remarks in requirements.txt for vSphere. Those are not contained in the published PyPI package.

The package can be used directly from Python scripts and supports this usage scenario with additional convenience that helps launch actions and probes in background (more on actions and probes later), so that you can compose also complex scenarios with ease.

If this technology is new to you, you will probably prefer the chaostoolkit CLI in combination with experiment files, so we need to install the CLI next:

pip install chaostoolkit

Please verify that it was installed properly by running:

chaos --help

Usage

ℹ️ We assume you are using Gardener and run Gardener-managed shoot clusters. You can also use the generic cloud provider and Kubernetes chaosgarden modules, but configuration and secrets will then differ. Please see the module docs for details.

A Simple Experiment

The most important command is the run command, but before we can use it, we need to compile an experiment file first. Let’s start with a simple one, invoking only a read-only 📖 action from chaosgarden that lists cloud provider machines and networks (depends on cloud provider) for the “first” zone of one of your shoot clusters.

Let’s assume, your project is called my-project and your shoot is called my-shoot, then we need to create the following experiment:

{
    "title": "assess-filters-impact",
    "description": "assess-filters-impact",
    "method": [
        {
            "type": "action",
            "name": "assess-filters-impact",
            "provider": {
                "type": "python",
                "module": "chaosgarden.garden.actions",
                "func": "assess_cloud_provider_filters_impact",
                "arguments": {
                    "zone": 0
                }
            }
        }
    ],
    "configuration": {
        "garden_project": "my-project",
        "garden_shoot": "my-shoot"
    }
}

We are not yet there and need one more thing to do before we can run it: We need to “target” the Gardener landscape resp. Gardener API server where you have created your shoot cluster (not to be confused with your shoot cluster API server). If you do not know what this is or how to download the Gardener API server kubeconfig, please follow these instructions. You can either download your personal credentials or project credentials (see creation of a serviceaccount) to interact with Gardener. For now (fastest and most convenient way, but generally not recommended), let’s use your personal credentials, but if you later plan to automate your experiments, please use proper project credentials (a serviceaccount is not bound to your person, but to the project, and can be restricted using RBAC roles and role bindings, which is why we recommend this for production).

To download your personal credentials, open the Gardener Dashboard and click on your avatar in the upper right corner of the page. Click “My Account”, then look for the “Access” pane, then “Kubeconfig”, then press the “Download” button and save the kubeconfig to disk. Run the following command next:

export KUBECONFIG=path/to/kubeconfig

We are now set and you can run your first experiment:

chaos run path/to/experiment

You should see output like this (depends on cloud provider):

[INFO] Validating the experiment's syntax
[INFO] Installing signal handlers to terminate all active background threads on involuntary signals (note that SIGKILL cannot be handled).
[INFO] Experiment looks valid
[INFO] Running experiment: assess-filters-impact
[INFO] Steady-state strategy: default
[INFO] Rollbacks strategy: default
[INFO] No steady state hypothesis defined. That's ok, just exploring.
[INFO] Playing your experiment's method now...
[INFO] Action: assess-filters-impact
[INFO] Validating client credentials and listing probably impacted instances and/or networks with the given arguments zone='world-1a' and filters={'instances': [{'Name': 'tag-key', 'Values': ['kubernetes.io/cluster/shoot--my-project--my-shoot']}], 'vpcs': [{'Name': 'tag-key', 'Values': ['kubernetes.io/cluster/shoot--my-project--my-shoot']}]}:
[INFO] 1 instance(s) would be impacted:
[INFO] - i-aabbccddeeff0000
[INFO] 1 VPC(s) would be impacted:
[INFO] - vpc-aabbccddeeff0000
[INFO] Let's rollback...
[INFO] No declared rollbacks, let's move on.
[INFO] Experiment ended with status: completed

🎉 Congratulations! You successfully ran your first chaosgarden experiment.

A Destructive Experiment

Now let’s break 🪓 your cluster. Be advised that this experiment will be destructive in the sense that we will temporarily network-partition all nodes in one availability zone (machine termination or restart is available with chaosgarden as well). That means, these nodes and their pods won’t be able to “talk” to other nodes, pods, and services. Also, the API server will become unreachable for them and the API server will report them as unreachable (confusingly shown as NotReady when you run kubectl get nodes and Unknown in the status Ready condition when you run kubectl get nodes --output yaml).

Being unreachable will trigger service endpoint and load balancer de-registration (when the node’s grace period lapses) as well as eventually pod eviction and machine replacement (which will continue to fail under test). We won’t run the experiment long enough for all of these effects to materialize, but the longer you run it, the more will happen, up to temporarily giving up/going into “back-off” for the affected worker pool in that zone. You will also see that the Kubernetes cluster autoscaler will try to create a new machine almost immediately, if pods are pending for the affected zone (which will initially fail under test, but may succeed later, which again depends on the runtime of the experiment and whether or not the cluster autoscaler goes into “back-off” or not).

But for now, all of this doesn’t matter as we want to start “small”. You can later read up more on the various settings and effects in our best practices guide on high availability.

Please create a new experiment file, this time with this content:

{
    "title": "run-network-failure-simulation",
    "description": "run-network-failure-simulation",
    "method": [
        {
            "type": "action",
            "name": "run-network-failure-simulation",
            "provider": {
                "type": "python",
                "module": "chaosgarden.garden.actions",
                "func": "run_cloud_provider_network_failure_simulation",
                "arguments": {
                    "mode": "total",
                    "zone": 0,
                    "duration": 60
                }
            }
        }
    ],
    "rollbacks": [
        {
            "type": "action",
            "name": "rollback-network-failure-simulation",
            "provider": {
                "type": "python",
                "module": "chaosgarden.garden.actions",
                "func": "rollback_cloud_provider_network_failure_simulation",
                "arguments": {
                    "mode": "total",
                    "zone": 0
                }
            }
        }
    ],
    "configuration": {
        "garden_project": {
            "type": "env",
            "key": "GARDEN_PROJECT"
        },
        "garden_shoot": {
            "type": "env",
            "key": "GARDEN_SHOOT"
        }
    }
}

ℹ️ There is an even more destructive action that terminates or alternatively restarts machines in a given zone 🔥 (immediately or delayed with some randomness/chaos for maximum inconvenience for the nodes and pods). You can find links to all these examples at the end of this tutorial.

This experiment is very similar, but this time we will break 🪓 your cluster - for 60s. If that’s too short to even see a node or pod transition from Ready to NotReady (actually Unknown), then increase the duration. Depending on the workload that your cluster runs, you may already see effects of the network partitioning, because it is effective immediately. It’s just that Kubernetes cannot know immediately and rather assumes that something is failing only after the node’s grace period lapses, but the actual workload is impacted immediately.

Most notably, this experiment also has a rollbacks section, which is invoked even if you abort the experiment or it fails unexpectedly, but only if you run the CLI with the option --rollback-strategy always which we will do soon. Any chaosgarden action that can undo its activity, will do that implicitly when the duration lapses, but it is a best practice to always configure a rollbacks section in case something unexpected happens. Should you be in panic and just want to run the rollbacks section, you can remove all other actions and the CLI will execute the rollbacks section immediately.

One other thing is different in the second experiment as well. We now read the name of the project and the shoot from the environment, i.e. a configuration section can automatically expand environment variables. Also useful to know (not shown here), chaostoolkit supports variable substitution too, so that you have to define variables only once. Please note that you can also add a secrets section that can also automatically expand environment variables. For instance, instead of targeting the Gardener API server via $KUBECONFIG, which is supported by our chaosgarden package natively, you can also explicitly refer to it in a secrets section (for brevity reasons not shown here either).

Let’s now run your second experiment (please watch your nodes and pods in parallel, e.g. by running watch kubectl get nodes,pods --output wide in another terminal):

export GARDEN_PROJECT=my-project
export GARDEN_SHOOT=my-shoot
chaos run --rollback-strategy always path/to/experiment

The output of the run command will be similar to the one above, but longer. It will mention either machines or networks that were network-partitioned (depends on cloud provider), but should revert everything back to normal.

Normally, you would not only run actions in the method section, but also probes as part of a steady state hypothesis. Such steady state hypothesis probes are run before and after the actions to validate that the “system” was in a healthy state before and gets back to a healthy state after the actions ran, hence show that the “system” is in a steady state when not under test. Eventually, you will write your own probes that don’t even have to be part of a steady state hypothesis. We at Gardener run multi-zone (multiple zones at once) and rolling-zone (strike each zone once) outages with continuous custom probes all within the method section to validate our KPIs continuously under test (e.g. how long do the individual fail-overs take/how long is the actual outage). The most complex scenarios are even run via Python scripts as all actions and probes can also be invoked directly (which is what the CLI does).

High Availability

Developing highly available workload that can tolerate a zone outage is no trivial task. You can find more information on how to achieve this goal in our best practices guide on high availability.

Thank you for your interest in Gardener chaos engineering and making your workload more resilient.

Further Reading

Here some links for further reading:

• Examples: Experiments, Scripts
• Gardener Chaos Engineering: GitHub, PyPI, Module Docs for Gardener Users
• Chaos Toolkit Core: Home Page, Installation, Concepts, GitHub

2.2.3 - Control Plane

Highly Available Shoot Control Plane

Shoot resource offers a way to request for a highly available control plane.

Failure Tolerance Types

A highly available shoot control plane can be setup with either a failure tolerance of zone or node.

Node Failure Tolerance

The failure tolerance of a node will have the following characteristics:

• Control plane components will be spread across different nodes within a single availability zone. There will not be more than one replica per node for each control plane component which has more than one replica.
• Worker pool should have a minimum of 3 nodes.
• A multi-node etcd (quorum size of 3) will be provisioned, offering zero-downtime capabilities with each member in a different node within a single availability zone.

Zone Failure Tolerance

The failure tolerance of a zone will have the following characteristics:

• Control plane components will be spread across different availability zones. There will be at least one replica per zone for each control plane component which has more than one replica.
• Gardener scheduler will automatically select a seed which has a minimum of 3 zones to host the shoot control plane.
• A multi-node etcd (quorum size of 3) will be provisioned, offering zero-downtime capabilities with each member in a different zone.

Shoot Spec

To request for a highly available shoot control plane Gardener provides the following configuration in the shoot spec:

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
spec:
  controlPlane:
    highAvailability:
      failureTolerance:
        type: <node | zone>

Allowed Transitions

If you already have a shoot cluster with non-HA control plane, then the following upgrades are possible:

• Upgrade of non-HA shoot control plane to HA shoot control plane with node failure tolerance.
• Upgrade of non-HA shoot control plane to HA shoot control plane with zone failure tolerance. However, it is essential that the seed which is currently hosting the shoot control plane should be multi-zonal. If it is not, then the request to upgrade will be rejected.

Note: There will be a small downtime during the upgrade, especially for etcd, which will transition from a single node etcd cluster to a multi-node etcd cluster.

Disallowed Transitions

If you already have a shoot cluster with HA control plane, then the following transitions are not possible:

• Upgrade of HA shoot control plane from node failure tolerance to zone failure tolerance is currently not supported, mainly because already existing volumes are bound to the zone they were created in originally.
• Downgrade of HA shoot control plane with zone failure tolerance to node failure tolerance is currently not supported, mainly because of the same reason as above, that already existing volumes are bound to the respective zones they were created in originally.
• Downgrade of HA shoot control plane with either node or zone failure tolerance, to a non-HA shoot control plane is currently not supported, mainly because etcd-druid does not currently support scaling down of a multi-node etcd cluster to a single-node etcd cluster.

Zone Outage Situation

Implementing highly available software that can tolerate even a zone outage unscathed is no trivial task. You may find our HA Best Practices helpful to get closer to that goal. In this document, we collected many options and settings for you that also Gardener internally uses to provide a highly available service.

During a zone outage, you may be forced to change your cluster setup on short notice in order to compensate for failures and shortages resulting from the outage. For instance, if the shoot cluster has worker nodes across three zones where one zone goes down, the computing power from these nodes is also gone during that time. Changing the worker pool (shoot.spec.provider.workers[]) and infrastructure (shoot.spec.provider.infrastructureConfig) configuration can eliminate this disbalance, having enough machines in healthy availability zones that can cope with the requests of your applications.

Gardener relies on a sophisticated reconciliation flow with several dependencies for which various flow steps wait for the readiness of prior ones. During a zone outage, this can block the entire flow, e.g., because all three etcd replicas can never be ready when a zone is down, and required changes mentioned above can never be accomplished. For this, a special one-off annotation shoot.gardener.cloud/skip-readiness helps to skip any readiness checks in the flow.

The shoot.gardener.cloud/skip-readiness annotation serves as a last resort if reconciliation is stuck because of important changes during an AZ outage. Use it with caution, only in exceptional cases and after a case-by-case evaluation with your Gardener landscape administrator. If used together with other operations like Kubernetes version upgrades or credential rotation, the annotation may lead to a severe outage of your shoot control plane.

2.3 - Networking

2.3.1 - Enable IPv4/IPv6 (dual-stack) Ingress on AWS

Using IPv4/IPv6 (dual-stack) Ingress in an IPv4 single-stack cluster

Motivation

IPv6 adoption is continuously growing, already overtaking IPv4 in certain regions, e.g. India, or scenarios, e.g. mobile. Even though most IPv6 installations deploy means to reach IPv4, it might still be beneficial to expose services natively via IPv4 and IPv6 instead of just relying on IPv4.

Disadvantages of full IPv4/IPv6 (dual-stack) Deployments

Enabling full IPv4/IPv6 (dual-stack) support in a kubernetes cluster is a major endeavor. It requires a lot of changes and restarts of all pods so that all pods get addresses for both IP families. A side-effect of dual-stack networking is that failures may be hidden as network traffic may take the other protocol to reach the target. For this reason and also due to reduced operational complexity, service teams might lean towards staying in a single-stack environment as much as possible. Luckily, this is possible with Gardener and IPv4/IPv6 (dual-stack) ingress on AWS.

Simplifying IPv4/IPv6 (dual-stack) Ingress with Protocol Translation on AWS

Fortunately, the network load balancer on AWS supports automatic protocol translation, i.e. it can expose both IPv4 and IPv6 endpoints while communicating with just one protocol to the backends. Under the hood, automatic protocol translation takes place. Client IP address preservation can be achieved by using proxy protocol.

This approach enables users to expose IPv4 workload to IPv6-only clients without having to change the workload/service. Without requiring invasive changes, it allows a fairly simple first step into the IPv6 world for services just requiring ingress (incoming) communication.

Necessary Shoot Cluster Configuration Changes for IPv4/IPv6 (dual-stack) Ingress

To be able to utilize IPv4/IPv6 (dual-stack) Ingress in an IPv4 shoot cluster, the cluster needs to meet two preconditions:

1. dualStack.enabled needs to be set to true to configure VPC/subnet for IPv6 and add a routing rule for IPv6. (This does not add IPv6 addresses to kubernetes nodes.)
2. loadBalancerController.enabled needs to be set to true as well to use the load balancer controller, which supports dual-stack ingress.

apiVersion: core.gardener.cloud/v1beta1
kind: Shoot
...
spec:
  provider:
    type: aws
    infrastructureConfig:
      apiVersion: aws.provider.extensions.gardener.cloud/v1alpha1
      kind: InfrastructureConfig
      dualStack:
        enabled: true
    controlPlaneConfig:
      apiVersion: aws.provider.extensions.gardener.cloud/v1alpha1
      kind: ControlPlaneConfig
      loadBalancerController:
        enabled: true
...

When infrastructureConfig.networks.vpc.id is set to the ID of an existing VPC, please make sure that your VPC has an Amazon-provided IPv6 CIDR block added.

After adapting the shoot specification and reconciling the cluster, dual-stack load balancers can be created using kubernetes services objects.

Creating an IPv4/IPv6 (dual-stack) Ingress

With the preconditions set, creating an IPv4/IPv6 load balancer is as easy as annotating a service with the correct annotations:

apiVersion: v1
kind: Service
metadata:
  annotations:
    service.beta.kubernetes.io/aws-load-balancer-ip-address-type: dualstack
    service.beta.kubernetes.io/aws-load-balancer-scheme: internet-facing
    service.beta.kubernetes.io/aws-load-balancer-nlb-target-type: instance
    service.beta.kubernetes.io/aws-load-balancer-type: external
  name: ...
  namespace: ...
spec:
  ...
  type: LoadBalancer

In case the client IP address should be preserved, the following annotation can be used to enable proxy protocol. (The pod receiving the traffic needs to be configured for proxy protocol as well.)

    service.beta.kubernetes.io/aws-load-balancer-proxy-protocol: "*"

Please note that changing an existing Service to dual-stack may cause the creation of a new load balancer without deletion of the old AWS load balancer resource. While this helps in a seamless migration by not cutting existing connections it may lead to wasted/forgotten resources. Therefore, the (manual) cleanup needs to be taken into account when migrating an existing Service instance.

For more details see AWS Load Balancer Documentation - Network Load Balancer.

DNS Considerations to Prevent Downtime During a Dual-Stack Migration

In case the migration of an existing service is desired, please check if there are DNS entries directly linked to the corresponding load balancer. The migrated load balancer will have a new domain name immediately, which will not be ready in the beginning. Therefore, a direct migration of the domain name entries is not desired as it may cause a short downtime, i.e. domain name entries without backing IP addresses.

If there are DNS entries directly linked to the corresponding load balancer and they are managed by the shoot-dns-service, you can identify this via annotations with the prefix dns.gardener.cloud/. Those annotations can be linked to a Service, Ingress or Gateway resources. Alternatively, they may also use DNSEntry or DNSAnnotation resources.

For a seamless migration without downtime use the following three step approach:

1. Temporarily prevent direct DNS updates
2. Migrate the load balancer and wait until it is operational
3. Allow DNS updates again

To prevent direct updates of the DNS entries when the load balancer is migrated add the annotation dns.gardener.cloud/ignore: 'true' to all affected resources next to the other dns.gardener.cloud/... annotations before starting the migration. For example, in case of a Service ensure that the service looks like the following:

kind: Service
metadata:
  annotations:
    dns.gardener.cloud/ignore: 'true'
    dns.gardener.cloud/class: garden
    dns.gardener.cloud/dnsnames: '...'
    ...

Next, migrate the load balancer to be dual-stack enabled by adding/changing the corresponding annotations.

You have multiple options how to check that the load balancer has been provisioned successfully. It might be useful to peek into status.loadBalancer.ingress of the corresponding Service to identify the load balancer:

• Check in the AWS console for the corresponding load balancer provisioning state
• Perform domain name lookups with nslookup/dig to check whether the name resolves to an IP address.
• Call your workload via the new load balancer, e.g. using curl --resolve <my-domain-name>:<port>:<IP-address> https://<my-domain-name>:<port>, which allows you to call your service with the “correct” domain name without using actual name resolution.
• Wait a fixed period of time as load balancer creation is usually finished within 15 minutes

Once the load balancer has been provisioned, you can remove the annotation dns.gardener.cloud/ignore: 'true' again from the affected resources. It may take some additional time until the domain name change finally propagates (up to one hour).

2.3.2 - Support for IPv6 on AWS

Support for IPv6

Overview

Gardener supports different levels of IPv6 support in shoot clusters. This document describes the differences between them and what to consider when using them.

In IPv6 Ingress for IPv4 Shoot Clusters, the focus is on how an existing IPv4-only shoot cluster can provide dual-stack services to clients. Section IPv6-only Shoot Clusters describes how to create a shoot cluster that only supports IPv6. Finally, Dual-Stack Shoot Clusters explains how to create a shoot cluster that supports both IPv4 and IPv6.

IPv6 Ingress for IPv4 Shoot Clusters

Per default, Gardener shoot clusters use only IPv4. Therefore, they also expose their services only via load balancers with IPv4 addresses. To allow external clients to also use IPv6 to access services in an IPv4 shoot cluster, the cluster needs to be configured to support dual-stack ingress.

It is possible to configure a shoot cluster to support dual-stack ingress, see Using IPv4/IPv6 (dual-stack) Ingress in an IPv4 single-stack cluster for more information.

The main benefit of this approach is that the existing cluster stays almost as is without major changes, keeping the operational simplicity. It works very well for services that only require incoming communication, e.g. pure web services.

The main drawback is that certain scenarios, especially related to IPv6 callbacks, are not possible. This means that services, which actively call to their clients via web hooks, will not be able to do so over IPv6. Hence, those services will not be able to allow full-usage via IPv6.

IPv6-only Shoot Clusters

Motivation

IPv6-only shoot clusters are the best option to verify that services are fully IPv6-compatible. While Dual-Stack Shoot Clusters may fall back on using IPv4 transparently, IPv6-only shoot clusters enforce the usage of IPv6 inside the cluster. Therefore, it is recommended to check with IPv6-only shoot clusters if a workload is fully IPv6-compatible.

In addition to being a good testbed for IPv6 compatibility, IPv6-only shoot clusters may also be a desirable eventual target in the IPv6 migration as they allow to support both IPv4 and IPv6 clients while having a single-stack with the cluster.

Creating an IPv6-only Shoot Cluster

To create an IPv6-only shoot cluster, the following needs to be specified in the Shoot resource (see also here):

kind: Shoot
apiVersion: core.gardener.cloud/v1beta1
metadata:
  ...
spec:
  ...
  networking:
    type: ...
    ipFamilies:
      - IPv6
  ...
  provider:
    type: aws
    infrastructureConfig:
      apiVersion: aws.provider.extensions.gardener.cloud/v1alpha1
      kind: InfrastructureConfig
      networks:
        vpc:
          cidr: 192.168.0.0/16
        zones:
          - name: ...
            public: 192.168.32.0/20
            internal: 192.168.48.0/20

Warning

Please note that nodes, pods and services should not be specified in .spec.networking resource.

In contrast to that, it is still required to specify IPv4 ranges for the VPC and the public/internal subnets. This is mainly due to the fact that public/internal load balancers still require IPv4 addresses as there are no pure IPv6-only load balancers as of now. The ranges can be sized according to the expected amount of load balancers per zone/type.

The IPv6 address ranges are provided by AWS. It is ensured that the IPv6 ranges are globally unique und internet routable.

Load Balancer Configuration

The AWS Load Balancer Controller is automatically deployed when using an IPv6-only shoot cluster. When creating a load balancer, the corresponding annotations need to be configured, see AWS Load Balancer Documentation - Network Load Balancer for details.

The AWS Load Balancer Controller allows dual-stack ingress so that an IPv6-only shoot cluster can serve IPv4 and IPv6 clients. You can find an example here.

Warning

When accessing Network Load Balancers (NLB) from within the same IPv6-only cluster, it is crucial to add the annotation service.beta.kubernetes.io/aws-load-balancer-target-group-attributes: preserve_client_ip.enabled=false. Without this annotation, if a request is routed by the NLB to the same target instance from which it originated, the client IP and destination IP will be identical. This situation, known as the hair-pinning effect, will prevent the request from being processed. (This also happens for internal load balancers in IPv4 clusters, but is mitigated by the NAT gateway for external IPv4 load balancers.)

Connectivity to IPv4-only Services

The IPv6-only shoot cluster can connect to IPv4-only services via DNS64/NAT64. The cluster is configured to use the DNS64/NAT64 service of the underlying cloud provider. This allows the cluster to resolve IPv4-only DNS names and to connect to IPv4-only services.

Please note that traffic going through NAT64 incurs the same cost as ordinary NAT traffic in an IPv4-only cluster. Therefore, it might be beneficial to prefer IPv6 for services, which provide IPv4 and IPv6.

Dual-Stack Shoot Clusters

Motivation

Dual-stack shoot clusters support IPv4 and IPv6 out-of-the-box. They can be the intermediate step on the way towards IPv6 for any existing (IPv4-only) clusters.

Creating a Dual-Stack Shoot Cluster

To create a dual-stack shoot cluster, the following needs to be specified in the Shoot resource:

kind: Shoot
apiVersion: core.gardener.cloud/v1beta1
metadata:
  ...
spec:
  ...
  networking:
    type: ...
    pods: 192.168.128.0/17
    nodes: 192.168.0.0/18
    services: 192.168.64.0/18
    ipFamilies:
      - IPv4
      - IPv6
  ...
  provider:
    type: aws
    infrastructureConfig:
      apiVersion: aws.provider.extensions.gardener.cloud/v1alpha1
      kind: InfrastructureConfig
      networks:
        vpc:
          cidr: 192.168.0.0/18
        zones:
          - name: ...
            workers: 192.168.0.0/19
            public: 192.168.32.0/20
            internal: 192.168.48.0/20

Please note that the only change compared to an IPv4-only shoot cluster is the addition of IPv6 to the .spec.networking.ipFamilies field. The order of the IP families defines the preference of the IP family. In this case, IPv4 is preferred over IPv6, e.g. services specifying no IP family will get only an IPv4 address.

Migration of IPv4-only Shoot Clusters to Dual-Stack

Eventually, migration should be as easy as changing the .spec.networking.ipFamilies field in the Shoot resource from IPv4 to IPv4, IPv6. However, as of now, this is not supported.

It is worth recognizing that the migration from an IPv4-only shoot cluster to a dual-stack shoot cluster involves rolling of the nodes/workload as well. Nodes will not get a new IPv6 address assigned automatically. The same is true for pods as well. Once the migration is supported, the detailed caveats will be documented here.

Load Balancer Configuration

The AWS Load Balancer Controller is automatically deployed when using a dual-stack shoot cluster. When creating a load balancer, the corresponding annotations need to be configured, see AWS Load Balancer Documentation - Network Load Balancer for details.

Warning

Please note that load balancer services without any special annotations will default to IPv4-only regardless how .spec.ipFamilies is set.

The AWS Load Balancer Controller allows dual-stack ingress so that a dual-stack shoot cluster can serve IPv4 and IPv6 clients. You can find an example here.

Warning

When accessing external Network Load Balancers (NLB) from within the same cluster via IPv6 or internal NLBs via IPv4, it is crucial to add the annotation service.beta.kubernetes.io/aws-load-balancer-target-group-attributes: preserve_client_ip.enabled=false. Without this annotation, if a request is routed by the NLB to the same target instance from which it originated, the client IP and destination IP will be identical. This situation, known as the hair-pinning effect, will prevent the request from being processed.

2.3.3 - Manage Certificates with Gardener

Manage certificates with Gardener for public domain

Introduction

Dealing with applications on Kubernetes which offer a secure service endpoints (e.g. HTTPS) also require you to enable a secured communication via SSL/TLS. With the certificate extension enabled, Gardener can manage commonly trusted X.509 certificate for your application endpoint. From initially requesting certificate, it also handeles their renewal in time using the free Let’s Encrypt API.

There are two senarios with which you can use the certificate extension

• You want to use a certificate for a subdomain the shoot’s default DNS (see .spec.dns.domain of your shoot resource, e.g. short.ingress.shoot.project.default-domain.gardener.cloud). If this is your case, please see Manage certificates with Gardener for default domain
• You want to use a certificate for a custom domain. If this is your case, please keep reading this article.

Prerequisites

Before you start this guide there are a few requirements you need to fulfill:

• You have an existing shoot cluster
• Your custom domain is under a public top level domain (e.g. .com)
• Your custom zone is resolvable with a public resolver via the internet (e.g. 8.8.8.8)
• You have a custom DNS provider configured and working (see “DNS Providers”)

As part of the Let’s Encrypt ACME challenge validation process, Gardener sets a DNS TXT entry and Let’s Encrypt checks if it can both resolve and authenticate it. Therefore, it’s important that your DNS-entries are publicly resolvable. You can check this by querying e.g. Googles public DNS server and if it returns an entry your DNS is publicly visible:

# returns the A record for cert-example.example.com using Googles DNS server (8.8.8.8)
dig cert-example.example.com @8.8.8.8 A

DNS provider

In order to issue certificates for a custom domain you need to specify a DNS provider which is permitted to create DNS records for subdomains of your requested domain in the certificate. For example, if you request a certificate for host.example.com your DNS provider must be capable of managing subdomains of host.example.com.

DNS providers are normally specified in the shoot manifest. To learn more on how to configure one, please see the DNS provider documentation.

Issue a certificate

Every X.509 certificate is represented by a Kubernetes custom resource certificate.cert.gardener.cloud in your cluster. A Certificate resource may be used to initiate a new certificate request as well as to manage its lifecycle. Gardener’s certificate service regularly checks the expiration timestamp of Certificates, triggers a renewal process if necessary and replaces the existing X.509 certificate with a new one.

Your application should be able to reload replaced certificates in a timely manner to avoid service disruptions.

Certificates can be requested via 3 resources type

• Ingress
• Service (type LoadBalancer)
• Gateways (both Istio gateways and from the Gateway API)
• Certificate (Gardener CRD)

If either of the first 2 are used, a corresponding Certificate resource will be created automatically.

Using an Ingress Resource

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: amazing-ingress
  annotations:
    cert.gardener.cloud/purpose: managed
    # Optional but recommended, this is going to create the DNS entry at the same time
    dns.gardener.cloud/class: garden
    dns.gardener.cloud/ttl: "600"
    #cert.gardener.cloud/commonname: "*.example.com"              # optional, if not specified the first name from spec.tls[].hosts is used as common name
    #cert.gardener.cloud/dnsnames: ""                             # optional, if not specified the names from spec.tls[].hosts are used
    #cert.gardener.cloud/follow-cname: "true"                     # optional, same as spec.followCNAME in certificates
    #cert.gardener.cloud/secret-labels: "key1=value1,key2=value2" # optional labels for the certificate secret
    #cert.gardener.cloud/issuer: custom-issuer                    # optional to specify custom issuer (use namespace/name for shoot issuers)
    #cert.gardener.cloud/preferred-chain: "chain name"            # optional to specify preferred-chain (value is the Subject Common Name of the root issuer)
    #cert.gardener.cloud/private-key-algorithm: ECDSA             # optional to specify algorithm for private key, allowed values are 'RSA' or 'ECDSA'
    #cert.gardener.cloud/private-key-size: "384"                  # optional to specify size of private key, allowed values for RSA are "2048", "3072", "4096" and for ECDSA "256" and "384"

spec:
  tls:
  - hosts:
    # Must not exceed 64 characters.
    - amazing.example.com
    # Certificate and private key reside in this secret.
    secretName: tls-secret
  rules:
  - host: amazing.example.com
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: amazing-svc
            port:
              number: 8080

Replace the hosts and rules[].host value again with your own domain and adjust the remaining Ingress attributes in accordance with your deployment (e.g. the above is for an istio Ingress controller and forwards traffic to a service1 on port 80).

Using a Service of type LoadBalancer

apiVersion: v1
kind: Service
metadata:
  annotations:
    cert.gardener.cloud/secretname: tls-secret
    dns.gardener.cloud/dnsnames: example.example.com
    dns.gardener.cloud/class: garden
    # Optional
    dns.gardener.cloud/ttl: "600"
    cert.gardener.cloud/commonname: "*.example.example.com"
    cert.gardener.cloud/dnsnames: ""
    #cert.gardener.cloud/follow-cname: "true"                     # optional, same as spec.followCNAME in certificates
    #cert.gardener.cloud/secret-labels: "key1=value1,key2=value2" # optional labels for the certificate secret
    #cert.gardener.cloud/issuer: custom-issuer                    # optional to specify custom issuer (use namespace/name for shoot issuers)
    #cert.gardener.cloud/preferred-chain: "chain name"            # optional to specify preferred-chain (value is the Subject Common Name of the root issuer)
    #cert.gardener.cloud/private-key-algorithm: ECDSA             # optional to specify algorithm for private key, allowed values are 'RSA' or 'ECDSA'
    #cert.gardener.cloud/private-key-size: "384"                  # optional to specify size of private key, allowed values for RSA are "2048", "3072", "4096" and for ECDSA "256" and "384"
    
  name: test-service
  namespace: default
spec:
  ports:
    - name: http
      port: 80
      protocol: TCP
      targetPort: 8080
  type: LoadBalancer

Using a Gateway resource

Please see Istio Gateways or Gateway API for details.

Using the custom Certificate resource

apiVersion: cert.gardener.cloud/v1alpha1
kind: Certificate
metadata:
  name: cert-example
  namespace: default
spec:
  commonName: amazing.example.com
  secretRef:
    name: tls-secret
    namespace: default
  # Optionnal if using the default issuer
  issuerRef:
    name: garden

  # If delegated domain for DNS01 challenge should be used. This has only an effect if a CNAME record is set for
  # '_acme-challenge.amazing.example.com'.
  # For example: If a CNAME record exists '_acme-challenge.amazing.example.com' => '_acme-challenge.writable.domain.com',
  # the DNS challenge will be written to '_acme-challenge.writable.domain.com'.
  #followCNAME: true

  # optionally set labels for the secret
  #secretLabels:
  #  key1: value1
  #  key2: value2

  # Optionally specify the preferred certificate chain: if the CA offers multiple certificate chains, prefer the chain with an issuer matching this Subject Common Name. If no match, the default offered chain will be used.
  #preferredChain: "ISRG Root X1"

  # Optionally specify algorithm and key size for private key. Allowed algorithms: "RSA" (allowed sizes: 2048, 3072, 4096) and "ECDSA" (allowed sizes: 256, 384)
  # If not specified, RSA with 2048 is used.
  #privateKey:
  #  algorithm: ECDSA
  #  size: 384

Supported attributes

Here is a list of all supported annotations regarding the certificate extension:

Request a wildcard certificate

In order to avoid the creation of multiples certificates for every single endpoints, you may want to create a wildcard certificate for your shoot’s default cluster.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: amazing-ingress
  annotations:
    cert.gardener.cloud/purpose: managed
    cert.gardener.cloud/commonName: "*.example.com"
spec:
  tls:
  - hosts:
    - amazing.example.com
    secretName: tls-secret
  rules:
  - host: amazing.example.com
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: amazing-svc
            port:
              number: 8080

Please note that this can also be achived by directly adding an annotation to a Service type LoadBalancer. You could also create a Certificate object with a wildcard domain.

Using a custom Issuer

Most Gardener deployment with the certification extension enabled have a preconfigured garden issuer. It is also usually configured to use Let’s Encrypt as the certificate provider.

If you need a custom issuer for a specific cluster, please see Using a custom Issuer

Quotas

For security reasons there may be a default quota on the certificate requests per day set globally in the controller registration of the shoot-cert-service.

The default quota only applies if there is no explicit quota defined for the issuer itself with the field requestsPerDayQuota, e.g.:

kind: Shoot
...
spec:
  extensions:
  - type: shoot-cert-service
    providerConfig:
      apiVersion: service.cert.extensions.gardener.cloud/v1alpha1
      kind: CertConfig
      issuers:
        - email: your-email@example.com
          name: custom-issuer # issuer name must be specified in every custom issuer request, must not be "garden"
          server: 'https://acme-v02.api.letsencrypt.org/directory'
          requestsPerDayQuota: 10

DNS Propagation

As stated before, cert-manager uses the ACME challenge protocol to authenticate that you are the DNS owner for the domain’s certificate you are requesting. This works by creating a DNS TXT record in your DNS provider under _acme-challenge.example.example.com containing a token to compare with. The TXT record is only applied during the domain validation. Typically, the record is propagated within a few minutes. But if the record is not visible to the ACME server for any reasons, the certificate request is retried again after several minutes. This means you may have to wait up to one hour after the propagation problem has been resolved before the certificate request is retried. Take a look in the events with kubectl describe ingress example for troubleshooting.

Character Restrictions

Due to restriction of the common name to 64 characters, you may to leave the common name unset in such cases.

For example, the following request is invalid:

apiVersion: cert.gardener.cloud/v1alpha1
kind: Certificate
metadata:
  name: cert-invalid
  namespace: default
spec:
  commonName: morethan64characters.ingress.shoot.project.default-domain.gardener.cloud

But it is valid to request a certificate for this domain if you have left the common name unset:

apiVersion: cert.gardener.cloud/v1alpha1
kind: Certificate
metadata:
  name: cert-example
  namespace: default
spec:
  dnsNames:
  - morethan64characters.ingress.shoot.project.default-domain.gardener.cloud

References

• Gardener cert-management
• Managing DNS with Gardener

2.3.4 - Manage Certificates with Gardener for Default Domain

Manage certificates with Gardener for default domain

Introduction

Dealing with applications on Kubernetes which offer a secure service endpoints (e.g. HTTPS) also require you to enable a secured communication via SSL/TLS. With the certificate extension enabled, Gardener can manage commonly trusted X.509 certificate for your application endpoint. From initially requesting certificate, it also handeles their renewal in time using the free Let’s Encrypt API.

There are two senarios with which you can use the certificate extension

• You want to use a certificate for a subdomain the shoot’s default DNS (see .spec.dns.domain of your shoot resource, e.g. short.ingress.shoot.project.default-domain.gardener.cloud). If this is your case, please keep reading this article.
• You want to use a certificate for a custom domain. If this is your case, please see Manage certificates with Gardener for public domain

Prerequisites

Before you start this guide there are a few requirements you need to fulfill:

• You have an existing shoot cluster

Since you are using the default DNS name, all DNS configuration should already be done and ready.

Issue a certificate

Every X.509 certificate is represented by a Kubernetes custom resource certificate.cert.gardener.cloud in your cluster. A Certificate resource may be used to initiate a new certificate request as well as to manage its lifecycle. Gardener’s certificate service regularly checks the expiration timestamp of Certificates, triggers a renewal process if necessary and replaces the existing X.509 certificate with a new one.

Your application should be able to reload replaced certificates in a timely manner to avoid service disruptions.

Certificates can be requested via 3 resources type

• Ingress
• Service (type LoadBalancer)
• certificate (Gardener CRD)

If either of the first 2 are used, a corresponding Certificate resource will automatically be created.

Using an ingress Resource

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: amazing-ingress
  annotations:
    cert.gardener.cloud/purpose: managed
    #cert.gardener.cloud/issuer: custom-issuer                    # optional to specify custom issuer (use namespace/name for shoot issuers)
    #cert.gardener.cloud/follow-cname: "true"                     # optional, same as spec.followCNAME in certificates
    #cert.gardener.cloud/secret-labels: "key1=value1,key2=value2" # optional labels for the certificate secret
    #cert.gardener.cloud/preferred-chain: "chain name"            # optional to specify preferred-chain (value is the Subject Common Name of the root issuer)
    #cert.gardener.cloud/private-key-algorithm: ECDSA             # optional to specify algorithm for private key, allowed values are 'RSA' or 'ECDSA'
    #cert.gardener.cloud/private-key-size: "384"                  # optional to specify size of private key, allowed values for RSA are "2048", "3072", "4096" and for ECDSA "256" and "384"spec:
  tls:
  - hosts:
    # Must not exceed 64 characters.
    - short.ingress.shoot.project.default-domain.gardener.cloud
    # Certificate and private key reside in this secret.
    secretName: tls-secret
  rules:
  - host: short.ingress.shoot.project.default-domain.gardener.cloud
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: amazing-svc
            port:
              number: 8080

Using a service type LoadBalancer

apiVersion: v1
kind: Service
metadata:
  annotations:
    cert.gardener.cloud/purpose: managed
    # Certificate and private key reside in this secret.
    cert.gardener.cloud/secretname: tls-secret
    # You may add more domains separated by commas (e.g. "service.shoot.project.default-domain.gardener.cloud, amazing.shoot.project.default-domain.gardener.cloud")
    dns.gardener.cloud/dnsnames: "service.shoot.project.default-domain.gardener.cloud" 
    dns.gardener.cloud/ttl: "600"
    #cert.gardener.cloud/issuer: custom-issuer                    # optional to specify custom issuer (use namespace/name for shoot issuers)
    #cert.gardener.cloud/follow-cname: "true"                     # optional, same as spec.followCNAME in certificates
    #cert.gardener.cloud/secret-labels: "key1=value1,key2=value2" # optional labels for the certificate secret
    #cert.gardener.cloud/preferred-chain: "chain name"            # optional to specify preferred-chain (value is the Subject Common Name of the root issuer)
    #cert.gardener.cloud/private-key-algorithm: ECDSA             # optional to specify algorithm for private key, allowed values are 'RSA' or 'ECDSA'
    #cert.gardener.cloud/private-key-size: "384"                  # optional to specify size of private key, allowed values for RSA are "2048", "3072", "4096" and for ECDSA "256" and "384"  name: test-service
  namespace: default
spec:
  ports:
    - name: http
      port: 80
      protocol: TCP
      targetPort: 8080
  type: LoadBalancer

Using the custom Certificate resource

apiVersion: cert.gardener.cloud/v1alpha1
kind: Certificate
metadata:
  name: cert-example
  namespace: default
spec:
  commonName: short.ingress.shoot.project.default-domain.gardener.cloud
  secretRef:
    name: tls-secret
    namespace: default
  # Optionnal if using the default issuer
  issuerRef:
    name: garden

If you’re interested in the current progress of your request, you’re advised to consult the description, more specifically the status attribute in case the issuance failed.

Request a wildcard certificate

In order to avoid the creation of multiples certificates for every single endpoints, you may want to create a wildcard certificate for your shoot’s default cluster.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: amazing-ingress
  annotations:
    cert.gardener.cloud/purpose: managed
    cert.gardener.cloud/commonName: "*.ingress.shoot.project.default-domain.gardener.cloud"
spec:
  tls:
  - hosts:
    - amazing.ingress.shoot.project.default-domain.gardener.cloud
    secretName: tls-secret
  rules:
  - host: amazing.ingress.shoot.project.default-domain.gardener.cloud
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: amazing-svc
            port:
              number: 8080

Please note that this can also be achived by directly adding an annotation to a Service type LoadBalancer. You could also create a Certificate object with a wildcard domain.

More information

For more information and more examples about using the certificate extension, please see Manage certificates with Gardener for public domain

2.3.5 - Managing DNS with Gardener

Request DNS Names in Shoot Clusters

Introduction

Within a shoot cluster, it is possible to request DNS records via the following resource types:

• Ingress
• Service
• DNSEntry

It is necessary that the Gardener installation your shoot cluster runs in is equipped with a shoot-dns-service extension. This extension uses the seed’s dns management infrastructure to maintain DNS names for shoot clusters. Please ask your Gardener operator if the extension is available in your environment.

Shoot Feature Gate

In some Gardener setups the shoot-dns-service extension is not enabled globally and thus must be configured per shoot cluster. Please adapt the shoot specification by the configuration shown below to activate the extension individually.

kind: Shoot
...
spec:
  extensions:
    - type: shoot-dns-service
...

Before you start

You should :

• Have created a shoot cluster
• Have created and correctly configured a DNS Provider (Please consult this page for more information)
• Have a basic understanding of DNS (see link under References)

There are 2 types of DNS that you can use within Kubernetes :

• internal (usually managed by coreDNS)
• external (managed by a public DNS provider).

This page, and the extension, exclusively works for external DNS handling.

Gardener allows 2 way of managing your external DNS:

• Manually, which means you are in charge of creating / maintaining your Kubernetes related DNS entries
• Via the Gardener DNS extension

Gardener DNS extension

The managed external DNS records feature of the Gardener clusters makes all this easier. You do not need DNS service provider specific knowledge, and in fact you do not need to leave your cluster at all to achieve that. You simply annotate the Ingress / Service that needs its DNS records managed and it will be automatically created / managed by Gardener.

Managed external DNS records are supported with the following DNS provider types:

• aws-route53
• azure-dns
• azure-private-dns
• google-clouddns
• openstack-designate
• alicloud-dns
• cloudflare-dns

Request DNS records for Ingress resources

To request a DNS name for Ingress, Service or Gateway (Istio or Gateway API) objects in the shoot cluster it must be annotated with the DNS class garden and an annotation denoting the desired DNS names.

Example for an annotated Ingress resource:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: amazing-ingress
  annotations:
    # Let Gardener manage external DNS records for this Ingress.
    dns.gardener.cloud/dnsnames: special.example.com # Use "*" to collects domains names from .spec.rules[].host
    dns.gardener.cloud/ttl: "600"
    dns.gardener.cloud/class: garden
    # If you are delegating the certificate management to Gardener, uncomment the following line
    #cert.gardener.cloud/purpose: managed
spec:
  rules:
  - host: special.example.com
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: amazing-svc
            port:
              number: 8080
  # Uncomment the following part if you are delegating the certificate management to Gardener
  #tls:
  #  - hosts:
  #      - special.example.com
  #    secretName: my-cert-secret-name

For an Ingress, the DNS names are already declared in the specification. Nevertheless the dnsnames annotation must be present. Here a subset of the DNS names of the ingress can be specified. If DNS names for all names are desired, the value all can be used.

Keep in mind that ingress resources are ignored unless an ingress controller is set up. Gardener does not provide an ingress controller by default. For more details, see Ingress Controllers and Service in the Kubernetes documentation.

Request DNS records for service type LoadBalancer

Example for an annotated Service (it must have the type LoadBalancer) resource:

apiVersion: v1
kind: Service
metadata:
  name: amazing-svc
  annotations:
    # Let Gardener manage external DNS records for this Service.
    dns.gardener.cloud/dnsnames: special.example.com
    dns.gardener.cloud/ttl: "600"
    dns.gardener.cloud/class: garden
spec:
  selector:
    app: amazing-app
  ports:
    - protocol: TCP
      port: 80
      targetPort: 8080
  type: LoadBalancer

Request DNS records for Gateway resources

Please see Istio Gateways or Gateway API for details.

Creating a DNSEntry resource explicitly

It is also possible to create a DNS entry via the Kubernetes resource called DNSEntry:

apiVersion: dns.gardener.cloud/v1alpha1
kind: DNSEntry
metadata:
  annotations:
    # Let Gardener manage this DNS entry.
    dns.gardener.cloud/class: garden
  name: special-dnsentry
  namespace: default
spec:
  dnsName: special.example.com
  ttl: 600
  targets:
  - 1.2.3.4

If one of the accepted DNS names is a direct subname of the shoot’s ingress domain, this is already handled by the standard wildcard entry for the ingress domain. Therefore this name should be excluded from the dnsnames list in the annotation. If only this DNS name is configured in the ingress, no explicit DNS entry is required, and the DNS annotations should be omitted at all.

You can check the status of the DNSEntry with

$ kubectl get dnsentry
NAME          DNS                                                            TYPE          PROVIDER      STATUS    AGE
mydnsentry    special.example.com     aws-route53   default/aws   Ready     24s

As soon as the status of the entry is Ready, the provider has accepted the new DNS record. Depending on the provider and your DNS settings and cache, it may take up to 24 hours for the new entry to be propagated over all internet.

More examples can be found here

Request DNS records for Service/Ingress resources using a DNSAnnotation resource

In rare cases it may not be possible to add annotations to a Service or Ingress resource object.

E.g.: the helm chart used to deploy the resource may not be adaptable for some reasons or some automation is used, which always restores the original content of the resource object by dropping any additional annotations.

In these cases, it is recommended to use an additional DNSAnnotation resource in order to have more flexibility that DNSentry resources. The DNSAnnotation resource makes the DNS shoot service behave as if annotations have been added to the referenced resource.

For the Ingress example shown above, you can create a DNSAnnotation resource alternatively to provide the annotations.

apiVersion: dns.gardener.cloud/v1alpha1
kind: DNSAnnotation
metadata:
  annotations:
    dns.gardener.cloud/class: garden
  name: test-ingress-annotation
  namespace: default
spec:
  resourceRef:
    kind: Ingress
    apiVersion: networking.k8s.io/v1
    name: test-ingress
    namespace: default
  annotations:
    dns.gardener.cloud/dnsnames: '*'
    dns.gardener.cloud/class: garden    

Note that the DNSAnnotation resource itself needs the dns.gardener.cloud/class=garden annotation. This also only works for annotations known to the DNS shoot service (see Accepted External DNS Records Annotations).

For more details, see also DNSAnnotation objects

Accepted External DNS Records Annotations

Here are all of the accepted annotation related to the DNS extension:

If one of the accepted DNS names is a direct subdomain of the shoot’s ingress domain, this is already handled by the standard wildcard entry for the ingress domain. Therefore, this name should be excluded from the dnsnames list in the annotation. If only this DNS name is configured in the ingress, no explicit DNS entry is required, and the DNS annotations should be omitted at all.

Troubleshooting

General DNS tools

To check the DNS resolution, use the nslookup or dig command.

$ nslookup special.your-domain.com

or with dig

$ dig +short special.example.com
Depending on your network settings, you may get a successful response faster using a public DNS server (e.g. 8.8.8.8, 8.8.4.4, or 1.1.1.1)

dig @8.8.8.8 +short special.example.com

DNS record events

The DNS controller publishes Kubernetes events for the resource which requested the DNS record (Ingress, Service, DNSEntry). These events reveal more information about the DNS requests being processed and are especially useful to check any kind of misconfiguration, e.g. requests for a domain you don’t own.

Events for a successfully created DNS record:

$ kubectl describe service my-service

Events:
  Type    Reason          Age                From                    Message
  ----    ------          ----               ----                    -------
  Normal  dns-annotation  19s                dns-controller-manager  special.example.com: dns entry is pending
  Normal  dns-annotation  19s (x3 over 19s)  dns-controller-manager  special.example.com: dns entry pending: waiting for dns reconciliation
  Normal  dns-annotation  9s (x3 over 10s)   dns-controller-manager  special.example.com: dns entry active

Please note, events vanish after their retention period (usually 1h).

DNSEntry status

DNSEntry resources offer a .status sub-resource which can be used to check the current state of the object.

Status of a erroneous DNSEntry.

  status:
    message: No responsible provider found
    observedGeneration: 3
    provider: remote
    state: Error

References

• Understanding DNS
• Kubernetes Internal DNS
• DNSEntry API (Golang)
• Managing Certificates with Gardener

2.4 - Administer Client (Shoot) Clusters

2.4.1 - Scalability of Gardener Managed Kubernetes Clusters

Have you ever wondered how much more your Kubernetes cluster can scale before it breaks down?

Of course, the answer is heavily dependent on your workloads. But be assured, any cluster will break eventually. Therefore, the best mitigation is to plan for sharding early and run multiple clusters instead of trying to optimize everything hoping to survive with a single cluster. Still, it is helpful to know when the time has come to scale out. This document aims at giving you the basic knowledge to keep a Gardener-managed Kubernetes cluster up and running while it scales according to your needs.

Welcome to Planet Scale, Please Mind the Gap!

For a complex, distributed system like Kubernetes it is impossible to give absolute thresholds for its scalability. Instead, the limit of a cluster’s scalability is a combination of various, interconnected dimensions.

Let’s take a rather simple example of two dimensions - the number of Pods per Node and number of Nodes in a cluster. According to the scalability thresholds documentation, Kubernetes can scale up to 5000 Nodes and with default settings accommodate a maximum of 110 Pods on a single Node. Pushing only a single dimension towards its limit will likely harm the cluster. But if both are pushed simultaneously, any cluster will break way before reaching one dimension’s limit.

What sounds rather straightforward in theory can be a bit trickier in reality. While 110 Pods is the default limit, we successfully pushed beyond that and in certain cases run up to 200 Pods per Node without breaking the cluster. This is possible in an environment where one knows and controls all workloads and cluster configurations. It still requires careful testing, though, and comes at the cost of limiting the scalability of other dimensions, like the number of Nodes.

Of course, a Kubernetes cluster has a plethora of dimensions. Thus, when looking at a simple questions like “How many resources can I store in ETCD?”, the only meaningful answer must be: “it depends”

The following sections will help you to identify relevant dimensions and how they affect a Gardener-managed Kubernetes cluster’s scalability.

“Official” Kubernetes Thresholds and Scalability Considerations

To get started with the topic, please check the basic guidance provided by the Kubernetes community (specifically SIG Scalability):

• How we define scalability?
• Kubernetes Scalability Thresholds

Furthermore, the problem space has been discussed in a KubeCon talk, the slides for which can be found here. You should at least read the slides before continuing.

Essentially, it comes down to this:

If you promise to:

• correctly configure your cluster
• use extensibility features “reasonably”
• keep the load in the cluster within recommended limits

Then we promise that your cluster will function properly.

With that knowledge in mind, let’s look at Gardener and eventually pick up the question about the number of objects in ETCD raised above.

Gardener-Specific Considerations

The following considerations are based on experience with various large clusters that scaled in different dimensions. Just as explained above, pushing beyond even one of the limits is likely to cause issues at some point in time (but not guaranteed). Depending on the setup of your workloads however, it might work unexpectedly well. Nevertheless, we urge you take conscious decisions and rather think about sharding your workloads. Please keep in mind - your workload affects the overall stability and scalability of a cluster significantly.

ETCD

The following section is based on a setup where ETCD Pods run on a dedicated Node pool and each Node has 8 vCPU and 32GB memory at least.

ETCD has a practical space limit of 8 GB. It caps the number of objects one can technically have in a Kubernetes cluster.

Of course, the number is heavily influenced by each object’s size, especially when considering that secrets and configmaps may store up to 1MB of data. Another dimension is a cluster’s churn rate. Since ETCD stores a history of the keyspace, a higher churn rate reduces the number of objects. Gardener runs compaction every 30min and defragmentation once per day during a cluster’s maintenance window to ensure proper ETCD operations. However, it is still possible to overload ETCD. If the space limit is reached, ETCD will only accept READ or DELETE requests and manual interaction by a Gardener operator is needed to disarm the alarm, once you got below the threshold.

To avoid such a situation, you can monitor the current ETCD usage via the “ETCD” dashboard of the monitoring stack. It gives you the current DB size, as well as historical data for the past 2 weeks. While there are improvements planned to trigger compaction and defragmentation based on DB size, an ETCD should not grow up to this threshold. A typical, healthy DB size is less than 3 GB.

Furthermore, the dashboard has a panel called “Memory”, which indicates the memory usage of the etcd pod(s). Using more than 16GB memory is a clear red flag, and you should reduce the load on ETCD.

Another dimension you should be aware of is the object count in ETCD. You can check it via the “API Server” dashboard, which features a “ETCD Object Counts By Resource” panel. The overall number of objects (excluding events, as they are stored in a different etcd instance) should not exceed 100k for most use cases.

Kube API Server

The following section is based on a setup where kube-apiserver run as Pods and are scheduled to Nodes with at least 8 vCPU and 32GB memory.

Gardener can scale the Deployment of a kube-apiserver horizontally and vertically. Horizontal scaling is limited to a certain number of replicas and should not concern a stakeholder much. However, the CPU / memory consumption of an individual kube-apiserver pod poses a potential threat to the overall availability of your cluster. The vertical scaling of any kube-apiserver is limited by the amount of resources available on a single Node. Outgrowing the resources of a Node will cause a downtime and render the cluster unavailable.

In general, continuous CPU usage of up to 3 cores and 16 GB memory per kube-apiserver pod is considered to be safe. This gives some room to absorb spikes, for example when the caches are initialized. You can check the resource consumption by selecting kube-apiserver Pods in the “Kubernetes Pods” dashboard. If these boundaries are exceeded constantly, you need to investigate and derive measures to lower the load.

Further information is also recorded and made available through the monitoring stack. The dashboard “API Server Request Duration and Response Size” provides insights into the request processing time of kube-apiserver Pods. Related information like request rates, dropped requests or termination codes (e.g., 429 for too many requests) can be obtained from the dashboards “API Server” and “Kubernetes API Server Details”. They provide a good indicator for how well the system is dealing with its current load.

Reducing the load on the API servers can become a challenge. To get started, you may try to:

• Use immutable secrets and configmaps where possible to save watches. This pays off, especially when you have a high number of Nodes or just lots of secrets in general.
• Applications interacting with the K8s API: If you know an object by its name, use it. Using label selector queries is expensive, as the filtering happens only within the kube-apiserver and not etcd, hence all resources must first pass completely from etcd to kube-apiserver.
• Use (single object) caches within your controllers. Check the “Use cache for ShootStates in Gardenlet” issue for an example.

Nodes

When talking about the scalability of a Kubernetes cluster, Nodes are probably mentioned in the first place… well, obviously not in this guide. While vanilla Kubernetes lists 5000 Nodes as its upper limit, pushing that dimension is not feasible. Most clusters should run with fewer than 300 Nodes. But of course, the actual limit depends on the workloads deployed and can be lower or higher. As you scale your cluster, be extra careful and closely monitor ETCD and kube-apiserver.

The scalability of Nodes is subject to a range of limiting factors. Some of them can only be defined upon cluster creation and remain immutable during a cluster lifetime. So let’s discuss the most important dimensions.

CIDR:

Upon cluster creation, you have to specify or use the default values for several network segments. There are dedicated CIDRs for services, Pods, and Nodes. Each defines a range of IP addresses available for the individual resource type. Obviously, the maximum of possible Nodes is capped by the CIDR for Nodes. However, there is a second limiting factor, which is the pod CIDR combined with the nodeCIDRMaskSize. This mask is used to divide the pod CIDR into smaller subnets, where each blocks gets assigned to a node. With a /16 pod network and a /24 nodeCIDRMaskSize, a cluster can scale up to 256 Nodes. Please check Shoot Networking for details.

Even though a /24 nodeCIDRMaskSize translates to a theoretical 256 pod IP addresses per Node, the maxPods setting should be less than 1/2 of this value. This gives the system some breathing room for churn and minimizes the risk for strange effects like mis-routed packages caused by immediate re-use of IPs.

Cloud provider capacity:

Most of the time, Nodes in Kubernetes translate to virtual machines on a hyperscaler. An attempt to add more Nodes to a cluster might fail due to capacity issues resulting in an error message like this:

Cloud provider message - machine codes error: code = [Internal] message = [InsufficientInstanceCapacity: We currently do not have sufficient <instance type> capacity in the Availability Zone you requested. Our system will be working on provisioning additional capacity. 

In heavily utilized regions, individual clusters are competing for scarce resources. So before choosing a region / zone, try to ensure that the hyperscaler supports your anticipated growth. This might be done through quota requests or by contacting the respective support teams. To mitigate such a situation, you may configure a worker pool with a different Node type and a corresponding priority expander as part of a shoot’s autoscaler section. Please consult the Autoscaler FAQ for more details.

Rolling of Node pools:

The overall number of Nodes is affecting the duration of a cluster’s maintenance. When upgrading a Node pool to a new OS image or Kubernetes version, all machines will be drained and deleted, and replaced with new ones. The more Nodes a cluster has, the longer this process will take, given that workloads are typically protected by PodDisruptionBudgets. Check Shoot Updates and Upgrades for details. Be sure to take this into consideration when planning maintenance.

Root disk:

You should be aware that the Node configuration impacts your workload’s performance too. Take the root disk of a Node, for example. While most hyperscalers offer the usage of HDD and SSD disks, it is strongly recommended to use SSD volumes as root disks. When there are lots of Pods on a Node or workloads making extensive use of emptyDir volumes, disk throttling becomes an issue. When a disk hits its IOPS limits, processes are stuck in IO-wait and slow down significantly. This can lead to a slow-down in the kubelet’s heartbeat mechanism and result in Nodes being replaced automatically, as they appear to be unhealthy. To analyze such a situation, you might have to run tools like iostat, sar or top directly on a Node.

Switching to an I/O optimized instance type (if offered for your infrastructure) can help to resolve issue. Please keep in mind that disks used via PersistentVolumeClaims have I/O limits as well. Sometimes these limits are related to the size and/or can be increased for individual disks.

Cloud Provider (Infrastructure) Limits

In addition to the already mentioned capacity restrictions, a cloud provider may impose other limitations to a Kubernetes cluster’s scalability. One category is the account quota defining the number of resources allowed globally or per region. Make sure to request appropriate values that suit your needs and contain a buffer, for example for having more Nodes during a rolling update.

Another dimension is the network throughput per VM or network interface. While you may be able to choose a network-optimized Node type for your workload to mitigate issues, you cannot influence the available bandwidth for control plane components. Therefore, please ensure that the traffic on the ETCD does not exceed 100MB/s. The ETCD dashboard provides data for monitoring this metric.

In some environments the upstream DNS might become an issue too and make your workloads subject to rate limiting. Given the heterogeneity of cloud providers incl. private data centers, it is not possible to give any thresholds. Still, the “CoreDNS” and “NodeLocalDNS” dashboards can help to derive a workload’s usage pattern. Check the DNS autoscaling and NodeLocalDNS documentations for available configuration options.

Webhooks

While webhooks provide powerful means to manage a cluster, they are equally powerful in breaking a cluster upon a malfunction or unavailability. Imagine using a policy enforcing system like Kyverno or Open Policy Agent Gatekeeper. As part of the stack, both will deploy webhooks which are invoked for almost everything that happens in a cluster. Now, if this webhook gets either overloaded or is simply not available, the cluster will stop functioning properly.

Hence, you have to ensure proper sizing, quick processing time, and availability of the webhook serving Pods when deploying webhooks. Please consult Dynamic Admission Control (Availability and Timeouts sections) for details. You should also be aware of the time added to any request that has to go through a webhook, as the kube-apiserver sends the request for mutation / validation to another pod and waits for the response. The more resources being subject to an external webhook, the more likely this will become a bottleneck when having a high churn rate on resources. Within the Gardener monitoring stack, you can check the extra time per webhook via the “API Server (Admission Details)” dashboard, which has a panel for “Duration per Webhook”.

In Gardener, any webhook timeout should be less than 15 seconds. Due to the separation of Kubernetes data-plane (shoot) and control-plane (seed) in Gardener, the extra hop from kube-apiserver (control-plane) to webhook (data-plane) is more expensive. Please check Shoot Status for more details.

Custom Resource Definitions

Using Custom Resource Definitions (CRD) to extend a cluster’s API is a common Kubernetes pattern and so is writing an operator to act upon custom resources. Writing an efficient controller reduces the load on the kube-apiserver and allows for better scaling. As a starting point, you might want to read Gardener’s Kubernetes Clients Guide.

Another problematic dimension is the usage of conversion webhooks when having resources stored in different versions. Not only do they add latency (see Webhooks) but can also block the kube-controllermanager’s garbage collection. If a conversion webhook is unavailable, the garbage collector fails to list all resources and does not perform any cleanup. In order to avoid such a situation, it is highly recommended to use conversion webhooks only when necessary and complete the migration to a new version as soon as possible.

Conclusion

As outlined by SIG Scalability, it is quite impossible to give limits or even recommendations fitting every individual use case. Instead, this guide outlines relevant dimensions and gives rather conservative recommendations based on usage patterns observed. By combining this information, it is possible to operate and scale a cluster in stable manner.

While going beyond is certainly possible for some dimensions, it significantly increases the risk of instability. Typically, limits on the control-plane are introduced by the availability of resources like CPU or memory on a single machine and can hardly be influenced by any user. Therefore, utilizing the existing resources efficiently is key. Other parameters are controlled by a user. In these cases, careful testing may reveal actual limits for a specific use case.

Please keep in mind that all aspects of a workload greatly influence the stability and scalability of a Kubernetes cluster.

2.4.2 - Authenticating with an Identity Provider

Prerequisites

Please read the following background material on Authenticating.

Overview

Kubernetes on its own doesn’t provide any user management. In other words, users aren’t managed through Kubernetes resources. Whenever you refer to a human user it’s sufficient to use a unique ID, for example, an email address. Nevertheless, Gardener project owners can use an identity provider to authenticate user access for shoot clusters in the following way:

1. Configure an Identity Provider using OpenID Connect (OIDC).
2. Configure a local kubectl oidc-login to enable oidc-login.
3. Configure the shoot cluster to share details of the OIDC-compliant identity provider with the Kubernetes API Server.
4. Authorize an authenticated user using role-based access control (RBAC).
5. Verify the result

Configure an Identity Provider

Create a tenant in an OIDC compatible Identity Provider. For simplicity, we use Auth0, which has a free plan.

1. In your tenant, create a client application to use authentication with kubectl:

   

2. Provide a Name, choose Native as application type, and choose CREATE.

   

3. In the tab Settings, copy the following parameters to a local text file:

   • Domain

     Corresponds to the issuer in OIDC. It must be an https-secured endpoint (Auth0 requires a trailing / at the end). For more information, see Issuer Identifier.

   • Client ID

   • Client Secret

     

4. Configure the client to have a callback url of http://localhost:8000. This callback connects to your local kubectl oidc-login plugin:

   

5. Save your changes.

6. Verify that https://<Auth0 Domain>/.well-known/openid-configuration is reachable.

7. Choose Users & Roles > Users > CREATE USERS to create a user with a user and password:

   

Configure a Local kubectl oidc-login

1. Install the kubectl plugin oidc-login. We highly recommend the krew installation tool, which also makes other plugins easily available.

   kubectl krew install oidc-login
   

   The response looks like this:

   Updated the local copy of plugin index.
   Installing plugin: oidc-login
   CAVEATS:
   \
   |  You need to setup the OIDC provider, Kubernetes API server, role binding and kubeconfig.
   |  See https://github.com/int128/kubelogin for more.
   /
   Installed plugin: oidc-login
   

2. Prepare a kubeconfig for later use:

   cp ~/.kube/config ~/.kube/config-oidc
   

3. Modify the configuration of ~/.kube/config-oidc as follows:

   apiVersion: v1
   kind: Config
   
   ...
   
   contexts:
   - context:
       cluster: shoot--project--mycluster
       user: my-oidc
     name: shoot--project--mycluster
   
   ...
   
   users:
   - name: my-oidc
     user:
       exec:
         apiVersion: client.authentication.k8s.io/v1beta1
         command: kubectl
         args:
         - oidc-login
         - get-token
         - --oidc-issuer-url=https://<Issuer>/ 
         - --oidc-client-id=<Client ID>
         - --oidc-client-secret=<Client Secret>
         - --oidc-extra-scope=email,offline_access,profile
   

To test our OIDC-based authentication, the context shoot--project--mycluster of ~/.kube/config-oidc is used in a later step. For now, continue to use the configuration ~/.kube/config with administration rights for your cluster.

Configure the Shoot Cluster

Modify the shoot cluster YAML as follows, using the client ID and the domain (as issuer) from the settings of the client application you created in Auth0:

kind: Shoot
apiVersion: garden.sapcloud.io/v1beta1
metadata:
  name: mycluster
  namespace: garden-project
...
spec:
  kubernetes:
    kubeAPIServer:
      oidcConfig:
        clientID: <Client ID>
        issuerURL: "https://<Issuer>/"
        usernameClaim: email

This change of the Shoot manifest triggers a reconciliation. Once the reconciliation is finished, your OIDC configuration is applied. It doesn’t invalidate other certificate-based authentication methods. Wait for Gardener to reconcile the change. It can take up to 5 minutes.

Authorize an Authenticated User

In Auth0, you created a user with a verified email address, test@test.com in our example. For simplicity, we authorize a single user identified by this email address with the cluster role view:

apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: viewer-test
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: view
subjects:
- apiGroup: rbac.authorization.k8s.io
  kind: User
  name: test@test.com

As administrator, apply the cluster role binding in your shoot cluster.

Verify the Result

1. To step into the shoes of your user, use the prepared kubeconfig file ~/.kube/config-oidc, and switch to the context that uses oidc-login:

   cd ~/.kube
   export KUBECONFIG=$(pwd)/config-oidc
   kubectl config use-context `shoot--project--mycluster`
   

2. kubectl delegates the authentication to plugin oidc-login the first time the user uses kubectl to contact the API server, for example:

   kubectl get all
   

   The plugin opens a browser for an interactive authentication session with Auth0, and in parallel serves a local webserver for the configured callback.

3. Enter your login credentials.

   

   You should get a successful response from the API server:

   Opening in existing browser session.
   NAME                 TYPE        CLUSTER-IP   EXTERNAL-IP   PORT(S)   AGE
   service/kubernetes   ClusterIP   100.64.0.1   <none>        443/TCP   86m
   

1. To see if your user uses the cluster role view, do some checks with kubectl auth can-i.

   • The response for the following commands should be no:

     kubectl auth can-i create clusterrolebindings
     

     kubectl auth can-i get secrets
     

     kubectl auth can-i describe secrets
     

   • The response for the following commands should be yes:

     kubectl auth can-i list pods
     

     kubectl auth can-i get pods
     

If the last step is successful, you’ve configured your cluster to authenticate against an identity provider using OIDC.

Related Links

• Auth0 Pricing

2.4.3 - Backup and Restore of Kubernetes Objects

TL;DR

In general, Backup and Restore (BR) covers activities enabling an organization to bring a system back in a consistent state, e.g., after a disaster or to setup a new system. These activities vary in a very broad way depending on the applications and its persistency.

Kubernetes objects like Pods, Deployments, NetworkPolicies, etc. configure Kubernetes internal components and might as well include external components like load balancer and persistent volumes of the cloud provider. The BR of external components and their configurations might be difficult to handle in case manual configurations were needed to prepare these components.

To set the expectations right from the beginning, this tutorial covers the BR of Kubernetes deployments which might use persistent volumes. The BR of any manual configuration of external components, e.g., via the cloud providers console, is not covered here, as well as the BR of a whole Kubernetes system.

This tutorial puts the focus on the open source tool Velero (formerly Heptio Ark) and its functionality to explain the BR process.

Basically, Velero allows you to:

• backup and restore your Kubernetes cluster resources and persistent volumes (on-demand or scheduled)
• backup or restore all objects in your cluster, or filter resources by type, namespace, and/or label
• by default, all persistent volumes are backed up (configurable)
• replicate your production environment for development and testing environments
• define an expiration date per backup
• execute pre- and post-activities in a container of a pod when a backup is created (see Hooks)
• extend Velero by Plugins, e.g., for Object and Block store (see Plugins)

Velero consists of a server side component and a client tool. The server components consists of Custom Resource Definitions (CRD) and controllers to perform the activities. The client tool communicates with the K8s API server to, e.g., create objects like a Backup object.

The diagram below explains the backup process. When creating a backup, Velero client makes a call to the Kubernetes API server to create a Backup object (1). The BackupController notices the new Backup object, validates the object (2) and begins the backup process (3). Based on the filter settings provided by the Velero client it collects the resources in question (3). The BackupController creates a tar ball with the Kubernetes objects and stores it in the backup location, e.g., AWS S3 (4) as well as snapshots of persistent volumes (5).

The size of the backup tar ball corresponds to the number of objects in etcd. The gzipped archive contains the Json representations of the objects.

Getting Started

At first, clone the Velero GitHub repository and get the Velero client from the releases or build it from source via make all in the main directory of the cloned GitHub repository.

To use an AWS S3 bucket as storage for the backup files and the persistent volumes, you need to:

• create a S3 bucket as the backup target
• create an AWS IAM user for Velero
• configure the Velero server
• create a secret for your AWS credentials

For details about this setup, check the Set Permissions for Velero documentation. Moreover, it is possible to use other supported storage providers.

Velero offers a wide range of filter possibilities for Kubernetes resources, e.g filter by namespaces, labels or resource types. The filter settings can be combined and used as include or exclude, which gives a great flexibility for selecting resources.

Exemplary Use Cases

Below are some use cases which could give you an idea on how to use Velero. You can also check Velero’s documentation for other introductory examples.

Helm Based Deployments

To be able to use Helm charts in your Kubernetes cluster, you need to install the Helm client helm and the server component tiller. Per default the server component is installed in the namespace kube-system. Even if it is possible to select single deployments via the filter settings of Velero, you should consider to install tiller in a separate namespace via helm init --tiller-namespace <your namespace>. This approach applies as well for all Helm charts to be deployed - consider separate namespaces for your deployments as well by using the parameter --namespace.

To backup a Helm based deployment, you need to backup both Tiller and the deployment. Only then the deployments could be managed via Helm. As mentioned above, the selection of resources would be easier in case they are separated in namespaces.

Separate Backup Locations

In case you run all your Kubernetes clusters on a single cloud provider, there is probably no need to store the backups in a bucket of a different cloud provider. However, if you run Kubernetes clusters on different cloud provider, you might consider to use a bucket on just one cloud provider as the target for the backups, e.g., to benefit from a lower price tag for the storage.

Per default, Velero assumes that both the persistent volumes and the backup location are on the same cloud provider. During the setup of Velero, a secret is created using the credentials for a cloud provider user who has access to both objects (see the policies, e.g., for the AWS configuration).

Now, since the backup location is different from the volume location, you need to follow these steps (described here for AWS):

• configure as documented the volume storage location in examples/aws/06-volumesnapshotlocation.yaml and provide the user credentials. In this case, the S3 related settings like the policies can be omitted

• create the bucket for the backup in the cloud provider in question and a user with the appropriate credentials and store them in a separate file similar to credentials-ark

• create a secret which contains two credentials, one for the volumes and one for the backup target, e.g., by using the command kubectl create secret generic cloud-credentials --namespace heptio-ark --from-file cloud=credentials-ark --from-file backup-target=backup-ark

• configure in the deployment manifest examples/aws/10-deployment.yaml the entries in volumeMounts, env and volumes accordingly, e.g., for a cluster running on AWS and the backup target bucket on GCP a configuration could look similar to:

  Note

  Some links might get broken in the near future since Heptio was taken over by VMWare which might result in different GitHub repositories or other changes. Please don’t hesitate to inform us in case you encounter any issues.

  Example Velero deployment

  # Copyright 2017 the Heptio Ark contributors.
  #
  # Licensed under the Apache License, Version 2.0 (the "License");
  # you may not use this file except in compliance with the License.
  # You may obtain a copy of the License at
  #
  #     http://www.apache.org/licenses/LICENSE-2.0
  #
  # Unless required by applicable law or agreed to in writing, software
  # distributed under the License is distributed on an "AS IS" BASIS,
  # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  # See the License for the specific language governing permissions and
  # limitations under the License.
  
  ---
  apiVersion: apps/v1beta1
  kind: Deployment
  metadata:
    namespace: velero
    name: velero
  spec:
    replicas: 1
    template:
      metadata:
        labels:
          component: velero
        annotations:
          prometheus.io/scrape: "true"
          prometheus.io/port: "8085"
          prometheus.io/path: "/metrics"
      spec:
        restartPolicy: Always
        serviceAccountName: velero
        containers:
          - name: velero
            image: gcr.io/heptio-images/velero:latest
            command:
              - /velero
            args:
              - server
            volumeMounts:
              - name: cloud-credentials
                mountPath: /credentials
              - name: plugins
                mountPath: /plugins
              - name: scratch
                mountPath: /scratch
            env:
              - name: AWS_SHARED_CREDENTIALS_FILE
                value: /credentials/cloud
              - name: GOOGLE_APPLICATION_CREDENTIALS
                value: /credentials/backup-target
              - name: VELERO_SCRATCH_DIR
                value: /scratch
        volumes:
          - name: cloud-credentials
            secret:
              secretName: cloud-credentials
          - name: plugins
            emptyDir: {}
          - name: scratch
            emptyDir: {}
  

• finally, configure the backup storage location in examples/aws/05-backupstoragelocation.yaml to use, in this case, a GCP bucket

Limitations

Below is a potentially incomplete list of limitations. You can also consult Velero’s documentation to get up to date information.

• Only full backups of selected resources are supported. Incremental backups are not (yet) supported. However, by using filters it is possible to restrict the backup to specific resources
• Inconsistencies might occur in case of changes during the creation of the backup
• Application specific actions are not considered by default. However, they might be handled by using Velero’s Hooks or Plugins

2.4.4 - Create / Delete a Shoot Cluster

Create a Shoot Cluster

As you have already prepared an example Shoot manifest in the steps described in the development documentation, please open another Terminal pane/window with the KUBECONFIG environment variable pointing to the Garden development cluster and send the manifest to the Kubernetes API server:

kubectl apply -f your-shoot-aws.yaml

You should see that Gardener has immediately picked up your manifest and has started to deploy the Shoot cluster.

In order to investigate what is happening in the Seed cluster, please download its proper Kubeconfig yourself (see next paragraph). The namespace of the Shoot cluster in the Seed cluster will look like that: shoot-johndoe-johndoe-1, whereas the first johndoe is your namespace in the Garden cluster (also called “project”) and the johndoe-1 suffix is the actual name of the Shoot cluster.

To connect to the newly created Shoot cluster, you must download its Kubeconfig as well. Please connect to the proper Seed cluster, navigate to the Shoot namespace, and download the Kubeconfig from the kubecfg secret in that namespace.

Delete a Shoot Cluster

In order to delete your cluster, you have to set an annotation confirming the deletion first, and trigger the deletion after that. You can use the prepared delete shoot script which takes the Shoot name as first parameter. The namespace can be specified by the second parameter, but it is optional. If you don’t state it, it defaults to your namespace (the username you are logged in with to your machine).

./hack/usage/delete shoot johndoe-1 johndoe

(the hack bash script can be found at GitHub)

Configure a Shoot Cluster Aalert Receiver

The receiver of the Shoot alerts can be configured from the .spec.monitoring.alerting.emailReceivers section in the Shoot specification. The value of the field has to be a list of valid mail addresses.

The alerting for the Shoot clusters is handled by the Prometheus Alertmanager. The Alertmanager will be deployed next to the control plane when the Shoot resource specifies .spec.monitoring.alerting.emailReceivers and if a SMTP secret exists.

If the field gets removed then the Alertmanager will be also removed during the next reconcilation of the cluster. The opposite is also valid if the field is added to an existing cluster.

2.4.5 - Create a Shoot Cluster Into an Existing AWS VPC

Overview

Gardener can create a new VPC, or use an existing one for your shoot cluster. Depending on your needs, you may want to create shoot(s) into an already created VPC. The tutorial describes how to create a shoot cluster into an existing AWS VPC. The steps are identical for Alicloud, Azure, and GCP. Please note that the existing VPC must be in the same region like the shoot cluster that you want to deploy into the VPC.

TL;DR

If .spec.provider.infrastructureConfig.networks.vpc.cidr is specified, Gardener will create a new VPC with the given CIDR block and respectively will delete it on shoot deletion.
If .spec.provider.infrastructureConfig.networks.vpc.id is specified, Gardener will use the existing VPC and respectively won’t delete it on shoot deletion.

1. Configure the AWS CLI

The aws configure command is a convenient way to setup your AWS CLI. It will prompt you for your credentials and settings which will be used in the following AWS CLI invocations:

aws configure
AWS Access Key ID [None]: <ACCESS_KEY_ID>
AWS Secret Access Key [None]: <SECRET_ACCESS_KEY>
Default region name [None]: <DEFAULT_REGION>
Default output format [None]: <DEFAULT_OUTPUT_FORMAT>

2. Create a VPC

Create the VPC by running the following command:

aws ec2 create-vpc --cidr-block <cidr-block>
{
  "Vpc": {
      "VpcId": "vpc-ff7bbf86",
      "InstanceTenancy": "default",
      "Tags": [],
      "CidrBlockAssociations": [
          {
              "AssociationId": "vpc-cidr-assoc-6e42b505",
              "CidrBlock": "10.0.0.0/16",
              "CidrBlockState": {
                  "State": "associated"
              }
          }
      ],
      "Ipv6CidrBlockAssociationSet": [],
      "State": "pending",
      "DhcpOptionsId": "dopt-38f7a057",
      "CidrBlock": "10.0.0.0/16",
      "IsDefault": false
  }
}

Gardener requires the VPC to have enabled DNS support, i.e the attributes enableDnsSupport and enableDnsHostnames must be set to true. enableDnsSupport attribute is enabled by default, enableDnsHostnames - not. Set the enableDnsHostnames attribute to true:

aws ec2 modify-vpc-attribute --vpc-id vpc-ff7bbf86 --enable-dns-hostnames

3. Create an Internet Gateway

Gardener also requires that an internet gateway is attached to the VPC. You can create one by using:

aws ec2 create-internet-gateway
{
    "InternetGateway": {
        "Tags": [],
        "InternetGatewayId": "igw-c0a643a9",
        "Attachments": []
    }
}

and attach it to the VPC using:

aws ec2 attach-internet-gateway --internet-gateway-id igw-c0a643a9 --vpc-id vpc-ff7bbf86

4. Create the Shoot

Prepare your shoot manifest (you could check the example manifests). Please make sure that you choose the region in which you had created the VPC earlier (step 2). Also, put your VPC ID in the .spec.provider.infrastructureConfig.networks.vpc.id field:

spec:
  region: <aws-region-of-vpc>
  provider:
    type: aws
    infrastructureConfig:
      apiVersion: aws.provider.extensions.gardener.cloud/v1alpha1
      kind: InfrastructureConfig
      networks:
        vpc:
          id: vpc-ff7bbf86
    # ...

Apply your shoot manifest:

kubectl apply -f your-shoot-aws.yaml

Ensure that the shoot cluster is properly created:

kubectl get shoot $SHOOT_NAME -n $SHOOT_NAMESPACE
NAME           CLOUDPROFILE   VERSION   SEED   DOMAIN           OPERATION   PROGRESS   APISERVER   CONTROL   NODES   SYSTEM   AGE
<SHOOT_NAME>   aws            1.15.0    aws    <SHOOT_DOMAIN>   Succeeded   100        True        True      True    True     20m

2.4.6 - Fix Problematic Conversion Webhooks

Reasoning

Custom Resource Definition (CRD) is what you use to define a Custom Resource. This is a powerful way to extend Kubernetes capabilities beyond the default installation, adding any kind of API objects useful for your application.

The CustomResourceDefinition API provides a workflow for introducing and upgrading to new versions of a CustomResourceDefinition. In a scenario where a CRD adds support for a new version and switches its spec.versions.storage field to it (i.e., from v1beta1 to v1), existing objects are not migrated in etcd. For more information, see Versions in CustomResourceDefinitions.

This creates a mismatch between the requested and stored version for all clients (kubectl, KCM, etc.). When the CRD also declares the usage of a conversion webhook, it gets called whenever a client requests information about a resource that still exists in the old version. If the CRD is created by the end-user, the webhook runs on the shoot side, whereas controllers / kapi-servers run separated, as part of the control-plane. For the webhook to be reachable, a working VPN connection seed -> shoot is essential. In scenarios where the VPN connection is broken, the kube-controller-manager eventually stops its garbage collection, as that requires it to list v1.PartialObjectMetadata for everything to build a dependency graph. Without the kube-controller-manager’s garbage collector, managed resources get stuck during update/rollout.

Breaking Situations

When a user upgrades to failureTolerance: node|zone, that will cause the VPN deployments to be replaced by statefulsets. However, as the VPN connection is broken upon teardown of the deployment, garbage collection will fail, leading to a situation that is stuck until an operator manually tackles it.

Such a situation can be avoided if the end-user has correctly configured CRDs containing conversion webhooks.

Checking Problematic CRDs

In order to make sure there are no version problematic CRDs, please run the script below in your shoot. It will return the name of the CRDs in case they have one of the 2 problems:

• the returned version of the CR is different than what is maintained in the status.storedVersions field of the CRD.
• the status.storedVersions field of the CRD has more than 1 version defined.

#!/bin/bash

set -e -o pipefail

echo "Checking all CRDs in the cluster..."
for p in $(kubectl get crd | awk 'NR>1' | awk '{print $1}'); do
  strategy=$(kubectl get crd "$p" -o json | jq -r .spec.conversion.strategy)

  if [ "$strategy" == "Webhook" ]; then
     crd_name=$(kubectl get crd "$p" -o json | jq -r .metadata.name)

     number_of_stored_versions=$(kubectl get crd "$crd_name" -o json  | jq '.status.storedVersions | length')

      if [[ "$number_of_stored_versions" == 1 ]]; then
         returned_cr_version=$(kubectl get "$crd_name" -A -o json |  jq -r '.items[] | .apiVersion'  | sed 's:.*/::')
         if [ -z "$returned_cr_version" ]; then
           continue
         else
           variable=$(echo "$returned_cr_version" | xargs -n1 | sort -u | xargs)
           present_version=$(kubectl get crd "$crd_name" -o json  |  jq -cr '.status.storedVersions |.[]')
           if [[ $variable != "$present_version" ]]; then
             echo "ERROR: Stored version differs from the version that CRs are being returned. $crd_name with conversion webhook needs to be fixed"
           fi
         fi
      fi

      if [[ "$number_of_stored_versions" -gt 1 ]]; then
         returned_cr_version=$(kubectl get "$crd_name" -A -o json |  jq -r '.items[] | .apiVersion'  | sed 's:.*/::')
         if [ -z "$returned_cr_version" ]; then
           continue
         else
           echo "ERROR: Too many stored versions defined. $crd_name with conversion webhook needs to be fixed"
         fi
      fi
  fi
done
echo "Problematic CRDs are reported above."

Resolve CRDs

Below we give the steps needed to be taken in order to fix the CRDs reported by the script above.

Inspect all your CRDs that have conversion webhooks in place. If you have more than 1 version defined in its spec.status.storedVersions field, then initiate migration as described in Option 2 in the Upgrade existing objects to a new stored version guide.

For convenience, we have provided the necessary steps below.

1. Please check/set the old CR version to storage:false and set the new CR version to storage:true.

   For the sake of an example, let’s consider the two versions v1beta1 (old) and v1 (new).

   Before:

   spec:
   versions:
   - name: v1beta1
   ......
   storage: true
   
   - name: v1
   ......
   storage: false
   

   After:

   spec:
   versions:
   - name: v1beta1
   ......
   storage: false
   
   - name: v1
   ......
   storage: true
   

2. Convert custom-resources to the newest version.

   kubectl get <custom-resource-name> -A -ojson | k apply -f -
   

3. Patch the CRD to keep only the latest version under storedVersions.

   kubectl patch customresourcedefinitions <crd-name> --subresource='status' --type='merge' -p '{"status":{"storedVersions":["your-latest-cr-version"]}}'
   

2.4.7 - GPU Enabled Cluster

Disclaimer

Be aware, that the following sections might be opinionated. Kubernetes, and the GPU support in particular, are rapidly evolving, which means that this guide is likely to be outdated sometime soon. For this reason, contributions are highly appreciated to update this guide.

Create a Cluster

First thing first, let’s create a Kubernetes (K8s) cluster with GPU accelerated nodes. In this example we will use an AWS p2.xlarge EC2 instance because it’s the cheapest available option at the moment. Use such cheap instances for learning to limit your resource costs. This costs around 1€/hour per GPU

Install NVidia Driver as Daemonset

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: nvidia-driver-installer
  namespace: kube-system
  labels:
    k8s-app: nvidia-driver-installer
spec:
  selector:
    matchLabels:
      name: nvidia-driver-installer
      k8s-app: nvidia-driver-installer
  template:
    metadata:
      labels:
        name: nvidia-driver-installer
        k8s-app: nvidia-driver-installer
    spec:
      hostPID: true
      initContainers:
      - image: squat/modulus:4a1799e7aa0143bcbb70d354bab3e419b1f54972
        name: modulus
        args:
        - compile
        - nvidia
        - "410.104"
        securityContext:
          privileged: true
        env:
        - name: MODULUS_CHROOT
          value: "true"
        - name: MODULUS_INSTALL
          value: "true"
        - name: MODULUS_INSTALL_DIR
          value: /opt/drivers
        - name: MODULUS_CACHE_DIR
          value: /opt/modulus/cache
        - name: MODULUS_LD_ROOT
          value: /root
        - name: IGNORE_MISSING_MODULE_SYMVERS
          value: "1"          
        volumeMounts:
        - name: etc-coreos
          mountPath: /etc/coreos
          readOnly: true
        - name: usr-share-coreos
          mountPath: /usr/share/coreos
          readOnly: true
        - name: ld-root
          mountPath: /root
        - name: module-cache
          mountPath: /opt/modulus/cache
        - name: module-install-dir-base
          mountPath: /opt/drivers
        - name: dev
          mountPath: /dev
      containers:
      - image: "gcr.io/google-containers/pause:3.1"
        name: pause
      tolerations:
      - key: "nvidia.com/gpu"
        effect: "NoSchedule"
        operator: "Exists"
      volumes:
      - name: etc-coreos
        hostPath:
          path: /etc/coreos
      - name: usr-share-coreos
        hostPath:
          path: /usr/share/coreos
      - name: ld-root
        hostPath:
          path: /
      - name: module-cache
        hostPath:
          path: /opt/modulus/cache
      - name: dev
        hostPath:
          path: /dev
      - name: module-install-dir-base
        hostPath:
          path: /opt/drivers

Install Device Plugin

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: nvidia-gpu-device-plugin
  namespace: kube-system
  labels:
    k8s-app: nvidia-gpu-device-plugin
    #addonmanager.kubernetes.io/mode: Reconcile
spec:
  selector:
    matchLabels:
      k8s-app: nvidia-gpu-device-plugin
  template:
    metadata:
      labels:
        k8s-app: nvidia-gpu-device-plugin
      annotations:
        scheduler.alpha.kubernetes.io/critical-pod: ''
    spec:
      priorityClassName: system-node-critical
      volumes:
      - name: device-plugin
        hostPath:
          path: /var/lib/kubelet/device-plugins
      - name: dev
        hostPath:
          path: /dev
      containers:
      - image: "k8s.gcr.io/nvidia-gpu-device-plugin@sha256:08509a36233c5096bb273a492251a9a5ca28558ab36d74007ca2a9d3f0b61e1d"
        command: ["/usr/bin/nvidia-gpu-device-plugin", "-logtostderr", "-host-path=/opt/drivers/nvidia"]
        name: nvidia-gpu-device-plugin
        resources:
          requests:
            cpu: 50m
            memory: 10Mi
          limits:
            cpu: 50m
            memory: 10Mi
        securityContext:
          privileged: true
        volumeMounts:
        - name: device-plugin
          mountPath: /device-plugin
        - name: dev
          mountPath: /dev
  updateStrategy:
    type: RollingUpdate

Test

To run an example training on a GPU node, first start a base image with Tensorflow with GPU support & Keras:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: deeplearning-workbench
  namespace: default
spec:
  replicas: 1
  selector:
    matchLabels:
      app: deeplearning-workbench
  template:
    metadata:
      labels:
        app: deeplearning-workbench
    spec:
      containers:
      - name: deeplearning-workbench
        image: afritzler/deeplearning-workbench
        resources:
          limits:
            nvidia.com/gpu: 1
      tolerations:
      - key: "nvidia.com/gpu"
        effect: "NoSchedule"
        operator: "Exists"

  kubernetes:
    kubeAPIServer:
      admissionPlugins:
      - name: ExtendedResourceToleration

Now exec into the container and start an example Keras training:

kubectl exec -it deeplearning-workbench-8676458f5d-p4d2v -- /bin/bash
cd /keras/example
python imdb_cnn.py

Related Links

• Andreas Fritzler from the Gardener Core team for the R&D, who has provided this setup.
• Build and install NVIDIA driver on CoreOS

2.4.8 - Shoot Cluster Maintenance

Overview

Day two operations for shoot clusters are related to:

• The Kubernetes version of the control plane and the worker nodes
• The operating system version of the worker nodes

The following table summarizes what options Gardener offers to maintain these versions:

Allowed Target Versions in the CloudProfile

Administrators maintain the allowed target versions that you can update to in the CloudProfile for each IaaS-Provider. Users with access to a Gardener project can check supported target versions with:

kubectl get cloudprofile [IAAS-SPECIFIC-PROFILE] -o yaml

Both the Kubernetes version and the operating system version follow semantic versioning that allows Gardener to handle updates automatically.

For more information, see Semantic Versioning.

Impact of Version Classifications on Updates

Gardener allows to classify versions in the CloudProfile as preview, supported, deprecated, or expired. During maintenance operations, preview versions are excluded from updates, because they’re often recently released versions that haven’t yet undergone thorough testing and may contain bugs or security issues.

For more information, see Version Classifications.

Let Gardener Manage Your Updates

The Maintenance Window

Gardener can manage updates for you automatically. It offers users to specify a maintenance window during which updates are scheduled:

• The time interval of the maintenance window can’t be less than 30 minutes or more than 6 hours.
• If there’s no maintenance window specified during the creation of a shoot cluster, Gardener chooses a maintenance window randomly to spread the load.

You can either specify the maintenance window in the shoot cluster specification (.spec.maintenance.timeWindow) or the start time of the maintenance window using the Gardener dashboard (CLUSTERS > [YOUR-CLUSTER] > OVERVIEW > Lifecycle > Maintenance).

Auto-Update and Forceful Updates

To trigger updates during the maintenance window automatically, Gardener offers the following methods:

• Auto-update:
  Gardener starts an update during the next maintenance window whenever there’s a version available in the CloudProfile that is higher than the one of your shoot cluster specification, and that isn’t classified as preview version. For Kubernetes versions, auto-update only updates to higher patch levels.

  You can either activate auto-update on the Gardener dashboard (CLUSTERS > [YOUR-CLUSTER] > OVERVIEW > Lifecycle > Maintenance) or in the shoot cluster specification:

  • .spec.maintenance.autoUpdate.kubernetesVersion: true
  • .spec.maintenance.autoUpdate.machineImageVersion: true

• Forceful updates:
  In the maintenance window, Gardener compares the current version given in the shoot cluster specification with the version list in the CloudProfile. If the version has an expiration date and if the date is before the start of the maintenance window, Gardener starts an update to the highest version available in the CloudProfile that isn’t classified as preview version. The highest version in CloudProfile can’t have an expiration date. For Kubernetes versions, Gardener only updates to higher patch levels or consecutive minor versions.

If you don’t want to wait for the next maintenance window, you can annotate the shoot cluster specification with shoot.gardener.cloud/operation: maintain. Gardener then checks immediately if there’s an auto-update or a forceful update needed.

With expiration dates, administrators can give shoot cluster owners more time for testing before the actual version update happens, which allows for smoother transitions to new versions.

Kubernetes Update Paths

The bigger the delta of the Kubernetes source version and the Kubernetes target version, the better it must be planned and executed by operators. Gardener only provides automatic support for updates that can be applied safely to the cluster workload:

Gardener doesn’t support automatic updates of nonconsecutive minor versions, because Kubernetes doesn’t guarantee updateability in this case. However, multiple minor version updates are possible if not only the minor source version is expired, but also the minor target version is expired. Gardener then updates the Kubernetes version first to the expired target version, and waits for the next maintenance window to update this version to the next minor target version.

Manual Updates

To update the Kubernetes version or the node operating system manually, change the .spec.kubernetes.version field or the .spec.provider.workers.machine.image.version field correspondingly.

Manual updates are required if you would like to do a minor update of the Kubernetes version. Gardener doesn’t do such updates automatically, as they can have breaking changes that could impact the cluster workload.

Manual updates are either executed immediately (default) or can be confined to the maintenance time window.
Choosing the latter option causes changes to the cluster (for example, node pool rolling-updates) and the subsequent reconciliation to only predictably happen during a defined time window (available since Gardener version 1.4).

For more information, see Confine Specification Changes/Update Roll Out.

Examples

In the examples for the CloudProfile and the shoot cluster specification, only the fields relevant for the example are shown.

Auto-Update of Kubernetes Version

Let’s assume that the Kubernetes versions 1.10.5 and 1.11.0 were added in the following CloudProfile:

spec:
  kubernetes:
    versions:
    - version: 1.11.0
    - version: 1.10.5
    - version: 1.10.0

Before this change, the shoot cluster specification looked like this:

spec:
  kubernetes:
    version: 1.10.0
  maintenance:
    timeWindow:
      begin: 220000+0000
      end: 230000+0000
    autoUpdate:
      kubernetesVersion: true

As a consequence, the shoot cluster is updated to Kubernetes version 1.10.5 between 22:00-23:00 UTC. Your shoot cluster isn’t updated automatically to 1.11.0, even though it’s the highest Kubernetes version in the CloudProfile, because Gardener only does automatic updates of the Kubernetes patch level.

Forceful Update Due to Expired Kubernetes Version

Let’s assume the following CloudProfile exists on the cluster:

spec:
  kubernetes:
    versions:
    - version: 1.12.8
    - version: 1.11.10
    - version: 1.10.13
    - version: 1.10.12
      expirationDate: "2019-04-13T08:00:00Z"

Let’s assume the shoot cluster has the following specification:

spec:
  kubernetes:
    version: 1.10.12
  maintenance:
    timeWindow:
      begin: 220000+0100
      end: 230000+0100
    autoUpdate:
      kubernetesVersion: false

The shoot cluster specification refers to a Kubernetes version that has an expirationDate. In the maintenance window on 2019-04-12, the Kubernetes version stays the same as it’s still not expired. But in the maintenance window on 2019-04-14, the Kubernetes version of the shoot cluster is updated to 1.10.13 (independently of the value of .spec.maintenance.autoUpdate.kubernetesVersion).

Forceful Update to New Minor Kubernetes Version

Let’s assume the following CloudProfile exists on the cluster:

spec:
  kubernetes:
    versions:
    - version: 1.12.8
    - version: 1.11.10
    - version: 1.11.09
    - version: 1.10.12
      expirationDate: "2019-04-13T08:00:00Z"

Let’s assume the shoot cluster has the following specification:

spec:
  kubernetes:
    version: 1.10.12
  maintenance:
    timeWindow:
      begin: 220000+0100
      end: 230000+0100
    autoUpdate:
      kubernetesVersion: false

The shoot cluster specification refers a Kubernetes version that has an expirationDate. In the maintenance window on 2019-04-14, the Kubernetes version of the shoot cluster is updated to 1.11.10, which is the highest patch version of minor target version 1.11 that follows the source version 1.10.

Automatic Update from Expired Machine Image Version

Let’s assume the following CloudProfile exists on the cluster:

spec:
  machineImages:
  - name: coreos
    versions:
    - version: 2191.5.0
    - version: 2191.4.1
    - version: 2135.6.0
      expirationDate: "2019-04-13T08:00:00Z"

Let’s assume the shoot cluster has the following specification:

spec:
  provider:
    type: aws
    workers:
    - name: name
      maximum: 1
      minimum: 1
      maxSurge: 1
      maxUnavailable: 0
      image:
        name: coreos
        version: 2135.6.0
        type: m5.large
      volume:
        type: gp2
        size: 20Gi
  maintenance:
    timeWindow:
      begin: 220000+0100
      end: 230000+0100
    autoUpdate:
      machineImageVersion: false

The shoot cluster specification refers a machine image version that has an expirationDate. In the maintenance window on 2019-04-12, the machine image version stays the same as it’s still not expired. But in the maintenance window on 2019-04-14, the machine image version of the shoot cluster is updated to 2191.5.0 (independently of the value of .spec.maintenance.autoUpdate.machineImageVersion) as version 2135.6.0 is expired.

2.4.9 - Tailscale

Access the Kubernetes apiserver from your tailnet

Overview

If you would like to strengthen the security of your Kubernetes cluster even further, this guide post explains how this can be achieved.

The most common way to secure a Kubernetes cluster which was created with Gardener is to apply the ACLs described in the Gardener ACL Extension repository or to use ExposureClass, which exposes the Kubernetes apiserver in a corporate network not exposed to the public internet.

However, those solutions are not without their drawbacks. Managing the ACL extension becomes fairly difficult with the growing number of participants, especially in a dynamic environment and work from home scenarios, and using ExposureClass requires you to first have a corporate network suitable for this purpose.

But there is a solution which bridges the gap between these two approaches by the use of a mesh VPN based on WireGuard

Tailscale

Tailscale is a mesh VPN network which uses Wireguard under the hood, but automates the key exchange procedure. Please consult the official tailscale documentation for a detailed explanation.

Target Architecture

Installation

In order to be able to access the Kubernetes apiserver only from a tailscale VPN, you need this steps:

1. Create a tailscale account and ensure MagicDNS is enabled.
2. Create an OAuth ClientID and Secret OAuth ClientID and Secret. Don’t forget to create the required tags.
3. Install the tailscale operator tailscale operator.

If all went well after the operator installation, you should be able to see the tailscale operator by running tailscale status:

# tailscale status
...
100.83.240.121  tailscale-operator   tagged-devices linux   -
...

Expose the Kubernetes apiserver

Now you are ready to expose the Kubernetes apiserver in the tailnet by annotating the service which was created by Gardener:

kubectl annotate -n default kubernetes tailscale.com/expose=true tailscale.com/hostname=kubernetes

It is required to kubernetes as the hostname, because this is part of the certificate common name of the Kubernetes apiserver.

After annotating the service, it will be exposed in the tailnet and can be shown by running tailscale status:

# tailscale status
...
100.83.240.121  tailscale-operator   tagged-devices linux   -
100.96.191.87   kubernetes           tagged-devices linux   idle, tx 19548 rx 71656
...

Modify the kubeconfig

In order to access the cluster via the VPN, you must modify the kubeconfig to point to the Kubernetes service exposed in the tailnet, by changing the server entry to https://kubernetes.

---
apiVersion: v1
clusters:
  - cluster:
      certificate-authority-data: <base64 encoded secret>
      server: https://kubernetes
    name: my-cluster
...

Enable ACLs to Block All IPs

Now you are ready to use your cluster from every device which is part of your tailnet. Therefore you can now block all access to the Kubernetes apiserver with the ACL extension.

Caveats

Multiple Kubernetes Clusters

You can actually not join multiple Kubernetes Clusters to join your tailnet because the kubernetes service in every cluster would overlap.

Headscale

It is possible to host a tailscale coordination by your own if you do not want to rely on the service tailscale.com offers. The headscale project is a open source implementation of this.

This works for basic tailscale VPN setups, but not for the tailscale operator described here, because headscale does not implement all required API endpoints for the tailscale operator. The details can be found in this Github Issue.

2.5 - Monitor and Troubleshoot

2.5.1 - Analyzing Node Removal and Failures

Overview

Sometimes operators want to find out why a certain node got removed. This guide helps to identify possible causes. There are a few potential reasons why nodes can be removed:

• broken node: a node becomes unhealthy and machine-controller-manager terminates it in an attempt to replace the unhealthy node with a new one
• scale-down: cluster-autoscaler sees that a node is under-utilized and therefore scales down a worker pool
• node rolling: configuration changes to a worker pool (or cluster) require all nodes of one or all worker pools to be rolled and thus all nodes to be replaced. Some possible changes are:
  • the K8s/OS version
  • changing machine types

Helpful information can be obtained by using the logging stack. See Logging Stack for how to utilize the logging information in Gardener.

Find Out Whether the Node Was unhealthy

Check the Node Events

A good first indication on what happened to a node can be obtained from the node’s events. Events are scraped and ingested into the logging system, so they can be found in the explore tab of Grafana (make sure to select loki as datasource) with a query like {job="event-logging"} | unpack | object="Node/<node-name>" or find any event mentioning the node in question via a broader query like {job="event-logging"}|="<node-name>".

A potential result might reveal:

{"_entry":"Node ip-10-55-138-185.eu-central-1.compute.internal status is now: NodeNotReady","count":1,"firstTimestamp":"2023-04-05T12:02:08Z","lastTimestamp":"2023-04-05T12:02:08Z","namespace":"default","object":"Node/ip-10-55-138-185.eu-central-1.compute.internal","origin":"shoot","reason":"NodeNotReady","source":"node-controller","type":"Normal"}

Check machine-controller-manager Logs

If a node was getting unhealthy, the last conditions can be found in the logs of the machine-controller-manager by using a query like {pod_name=~"machine-controller-manager.*"}|="<node-name>".

Caveat: every node resource is backed by a corresponding machine resource managed by machine-controller-manager. Usually two corresponding node and machine resources have the same name with the exception of AWS. Here you first need to find with the above query the corresponding machine name, typically via a log like this

2023-04-05 12:02:08 {"log":"Conditions of Machine \"shoot--demo--cluster-pool-z1-6dffc-jh4z4\" with providerID \"aws:///eu-central-1/i-0a6ad1ca4c2e615dc\" and backing node \"ip-10-55-138-185.eu-central-1.compute.internal\" are changing","pid":"1","severity":"INFO","source":"machine_util.go:629"}

This reveals that node ip-10-55-138-185.eu-central-1.compute.internal is backed by machine shoot--demo--cluster-pool-z1-6dffc-jh4z4. On infrastructures other than AWS you can omit this step.

With the machine name at hand, now search for log entries with {pod_name=~"machine-controller-manager.*"}|="<machine-name>". In case the node had failing conditions, you’d find logs like this:

2023-04-05 12:02:08 {"log":"Machine shoot--demo--cluster-pool-z1-6dffc-jh4z4 is unhealthy - changing MachineState to Unknown. Node conditions: [{Type:ClusterNetworkProblem Status:False LastHeartbeatTime:2023-04-05 11:58:39 +0000 UTC LastTransitionTime:2023-03-23 11:59:29 +0000 UTC Reason:NoNetworkProblems Message:no cluster network problems} ... {Type:Ready Status:Unknown LastHeartbeatTime:2023-04-05 11:55:27 +0000 UTC LastTransitionTime:2023-04-05 12:02:07 +0000 UTC Reason:NodeStatusUnknown Message:Kubelet stopped posting node status.}]","pid":"1","severity":"WARN","source":"machine_util.go:637"}

In the example above, the reason for an unhealthy node was that kubelet failed to renew its heartbeat. Typical reasons would be either a broken VM (that couldn’t execute kubelet anymore) or a broken network. Note that some VM terminations performed by the infrastructure provider are actually expected (e.g., scheduled events on AWS).

In both cases, the infrastructure provider might be able to provide more information on particular VM or network failures.

Whatever the failure condition might have been, if a node gets unhealthy, it will be terminated by machine-controller-manager after the machineHealthTimeout has elapsed (this parameter can be configured in your shoot spec).

Check the Node Logs

For each node the kernel and kubelet logs, as well as a few others, are scraped and can be queried with this query {nodename="<node-name>"} This might reveal OS specific issues or, in the absence of any logs (e.g., after the node went unhealthy), might indicate a network disruption or sudden VM termination. Note that some VM terminations performed by the infrastructure provider are actually expected (e.g., scheduled events on AWS).

Infrastructure providers might be able to provide more information on particular VM failures in such cases.

Check the Network Problem Detector Dashboard

If your Gardener installation utilizes gardener-extension-shoot-networking-problemdetector, you can check the dashboard named “Network Problem Detector” in Grafana for hints on network issues on the node of interest.

Scale-Down

In general, scale-downs are managed by the cluster-autoscaler, its logs can be found with the query {container_name="cluster-autoscaler"}. Attempts to remove a node can be found with the query {container_name="cluster-autoscaler"}|="Scale-down: removing empty node"

If a scale-down has caused disruptions in your workload, consider protecting your workload by adding PodDisruptionBudgets (see the autoscaler FAQ for more options).

Node Rolling

Node rolling can be caused by, e.g.:

• change of the K8s minor version of the cluster or a worker pool
• change of the OS version of the cluster or a worker pool
• change of the disk size/type or machine size/type of a worker pool
• change of node labels

Changes like the above are done by altering the shoot specification and thus are recorded in the external auditlog system that is configured for the garden cluster.

2.5.2 - Get a Shell to a Gardener Shoot Worker Node

Overview

To troubleshoot certain problems in a Kubernetes cluster, operators need access to the host of the Kubernetes node. This can be required if a node misbehaves or fails to join the cluster in the first place.

With access to the host, it is for instance possible to check the kubelet logs and interact with common tools such as systemctl and journalctl.

The first section of this guide explores options to get a shell to the node of a Gardener Kubernetes cluster. The options described in the second section do not rely on Kubernetes capabilities to get shell access to a node and thus can also be used if an instance failed to join the cluster.

This guide only covers how to get access to the host, but does not cover troubleshooting methods.

• Overview
• Get a Shell to an Operational Cluster Node
  • Gardener Dashboard
    • Result
  • Gardener Ops Toolbelt
  • Custom Root Pod
• SSH Access to a Node That Failed to Join the Cluster
  • Identifying the Problematic Instance
  • gardenctl ssh
  • SSH with a Manually Created Bastion on AWS
    • Create the Bastion Security Group
    • Create the Bastion Instance
  • Connecting to the Target Instance
• Cleanup

Get a Shell to an Operational Cluster Node

The following describes four different approaches to get a shell to an operational Shoot worker node. As a prerequisite to troubleshooting a Kubernetes node, the node must have joined the cluster successfully and be able to run a pod. All of the described approaches involve scheduling a pod with root permissions and mounting the root filesystem.

Gardener Dashboard

Prerequisite: the terminal feature is configured for the Gardener dashboard.

1. Navigate to the cluster overview page and find the Terminal in the Access tile.

Select the target Cluster (Garden, Seed / Control Plane, Shoot cluster) depending on the requirements and access rights (only certain users have access to the Seed Control Plane).

1. To open the terminal configuration, interact with the top right-hand corner of the screen.

1. Set the Terminal Runtime to “Privileged”. Also, specify the target node from the drop-down menu.

Result

The Dashboard then schedules a pod and opens a shell session to the node.

To get access to the common binaries installed on the host, prefix the command with chroot /hostroot. Note that the path depends on where the root path is mounted in the container. In the default image used by the Dashboard, it is under /hostroot.

Gardener Ops Toolbelt

Prerequisite: kubectl is available.

The Gardener ops-toolbelt can be used as a convenient way to deploy a root pod to a node. The pod uses an image that is bundled with a bunch of useful troubleshooting tools. This is also the same image that is used by default when using the Gardener Dashboard terminal feature as described in the previous section.

The easiest way to use the Gardener ops-toolbelt is to execute the ops-pod script in the hacks folder. To get root shell access to a node, execute the aforementioned script by supplying the target node name as an argument:

<path-to-ops-toolbelt-repo>/hacks/ops-pod <target-node>

Custom Root Pod

Alternatively, a pod can be assigned to a target node and a shell can be opened via standard Kubernetes means. To enable root access to the node, the pod specification requires proper securityContext and volume properties.

For instance, you can use the following pod manifest, after changing with the name of the node you want this pod attached to:

apiVersion: v1
kind: Pod
metadata:
  name: privileged-pod
  namespace: default
spec:
  nodeSelector:
    kubernetes.io/hostname: <target-node-name>
  containers:
  - name: busybox
    image: busybox
    stdin: true
    securityContext:
      privileged: true
    volumeMounts:
    - name: host-root-volume
      mountPath: /host
      readOnly: true
  volumes:
  - name: host-root-volume
    hostPath:
      path: /
  hostNetwork: true
  hostPID: true
  restartPolicy: Never

SSH Access to a Node That Failed to Join the Cluster

This section explores two options that can be used to get SSH access to a node that failed to join the cluster. As it is not possible to schedule a pod on the node, the Kubernetes-based methods explored so far cannot be used in this scenario.

Additionally, Gardener typically provisions worker instances in a private subnet of the VPC, hence - there is no public IP address that could be used for direct SSH access.

For this scenario, cloud providers typically have extensive documentation (e.g., AWS & GCP and in some cases tooling support). However, these approaches are mostly cloud provider specific, require interaction via their CLI and API or sometimes the installation of a cloud provider specific agent on the node.

Alternatively, gardenctl can be used providing a cloud provider agnostic and out-of-the-box support to get ssh access to an instance in a private subnet. Currently gardenctl supports AWS, GCP, Openstack, Azure and Alibaba Cloud.

Identifying the Problematic Instance

First, the problematic instance has to be identified. In Gardener, worker pools can be created in different cloud provider regions, zones, and accounts.

The instance would typically show up as successfully started / running in the cloud provider dashboard or API and it is not immediately obvious which one has a problem. Instead, we can use the Gardener API / CRDs to obtain the faulty instance identifier in a cloud-agnostic way.

Gardener uses the Machine Controller Manager to create the Shoot worker nodes. For each worker node, the Machine Controller Manager creates a Machine CRD in the Shoot namespace in the respective Seed cluster. Usually the problematic instance can be identified, as the respective Machine CRD has status pending.

The instance / node name can be obtained from the Machine .status field:

kubectl get machine <machine-name> -o json | jq -r .status.node

This is all the information needed to go ahead and use gardenctl ssh to get a shell to the node. In addition, the used cloud provider, the specific identifier of the instance, and the instance region can be identified from the Machine CRD.

Get the identifier of the instance via:

kubectl get machine <machine-name> -o json | jq -r .spec.providerID // e.g aws:///eu-north-1/i-069733c435bdb4640

The identifier shows that the instance belongs to the cloud provider aws with the ec2 instance-id i-069733c435bdb4640 in region eu-north-1.

To get more information about the instance, check out the MachineClass (e.g., AWSMachineClass) that is associated with each Machine CRD in the Shoot namespace of the Seed cluster.

The AWSMachineClass contains the machine image (ami), machine-type, iam information, network-interfaces, subnets, security groups and attached volumes.

Of course, the information can also be used to get the instance with the cloud provider CLI / API.

gardenctl ssh

Using the node name of the problematic instance, we can use the gardenctl ssh command to get SSH access to the cloud provider instance via an automatically set up bastion host. gardenctl takes care of spinning up the bastion instance, setting up the SSH keys, ports and security groups and opens a root shell on the target instance. After the SSH session has ended, gardenctl deletes the created cloud provider resources.

Use the following commands:

1. First, target a Garden cluster containing all the Shoot definitions.

gardenctl target garden <target-garden>

1. Target an available Shoot by name. This sets up the context, configures the kubeconfig file of the Shoot cluster and downloads the cloud provider credentials. Subsequent commands will execute in this context.

gardenctl target shoot <target-shoot>

1. This uses the cloud provider credentials to spin up the bastion and to open a shell on the target instance.

gardenctl ssh <target-node>

SSH with a Manually Created Bastion on AWS

In case you are not using gardenctl or want to control the bastion instance yourself, you can also manually set it up. The steps described here are generally the same as those used by gardenctl internally. Despite some cloud provider specifics, they can be generalized to the following list:

• Open port 22 on the target instance.
• Create an instance / VM in a public subnet (the bastion instance needs to have a public IP address).
• Set-up security groups and roles, and open port 22 for the bastion instance.

The following diagram shows an overview of how the SSH access to the target instance works:

This guide demonstrates the setup of a bastion on AWS.

Prerequisites:

• The AWS CLI is set up.

• Obtain target instance-id (see Identifying the Problematic Instance).

• Obtain the VPC ID the Shoot resources are created in. This can be found in the Infrastructure CRD in the Shoot namespace in the Seed.

• Make sure that port 22 on the target instance is open (default for Gardener deployed instances).

  • Extract security group via:

  aws ec2 describe-instances --instance-ids <instance-id>
  

  • Check for rule that allows inbound connections on port 22:

  aws ec2 describe-security-groups --group-ids=<security-group-id>
  

  • If not available, create the rule with the following comamnd:

  aws ec2 authorize-security-group-ingress --group-id <security-group-id>  --protocol tcp --port 22 --cidr 0.0.0.0/0
  

Create the Bastion Security Group

1. The common name of the security group is <shoot-name>-bsg. Create the security group:

aws ec2 create-security-group --group-name <bastion-security-group-name>  --description ssh-access --vpc-id <VPC-ID>

1. Optionally, create identifying tags for the security group:

aws ec2 create-tags --resources <bastion-security-group-id> --tags Key=component,Value=<tag>

1. Create a permission in the bastion security group that allows ssh access on port 22:

aws ec2 authorize-security-group-ingress --group-id <bastion-security-group-id>  --protocol tcp --port 22 --cidr 0.0.0.0/0

1. Create an IAM role for the bastion instance with the name <shoot-name>-bastions:

aws iam create-role --role-name <shoot-name>-bastions

The content should be:

{
"Version": "2012-10-17",
"Statement": [
    {
        "Effect": "Allow",
        "Action": [
            "ec2:DescribeRegions"
        ],
        "Resource": [
            "*"
        ]
    }
]
}

1. Create the instance profile and name it <shoot-name>-bastions:

aws iam create-instance-profile --instance-profile-name <name>

1. Add the created role to the instance profile:

aws iam add-role-to-instance-profile --instance-profile-name <instance-profile-name> --role-name <role-name>

Create the Bastion Instance

Next, in order to be able to ssh into the bastion instance, the instance has to be set up with a user with a public ssh key. Create a user gardener that has the same Gardener-generated public ssh key as the target instance.

1. First, we need to get the public part of the Shoot ssh-key. The ssh-key is stored in a secret in the the project namespace in the Garden cluster. The name is: <shoot-name>-ssh-publickey. Get the key via:

kubectl get secret aws-gvisor.ssh-keypair -o json | jq -r .data.\"id_rsa.pub\"

1. A script handed over as user-data to the bastion ec2 instance, can be used to create the gardener user and add the ssh-key. For your convenience, you can use the following script to generate the user-data.

#!/bin/bash -eu
saveUserDataFile () {
  ssh_key=$1

cat > gardener-bastion-userdata.sh <<EOF
#!/bin/bash -eu
id gardener || useradd gardener -mU
mkdir -p /home/gardener/.ssh
echo "$ssh_key" > /home/gardener/.ssh/authorized_keys
chown gardener:gardener /home/gardener/.ssh/authorized_keys
echo "gardener ALL=(ALL) NOPASSWD:ALL" >/etc/sudoers.d/99-gardener-user
EOF
}


if [ -p /dev/stdin ]; then
    read -r input
    cat | saveUserDataFile "$input"
else
    pbpaste | saveUserDataFile "$input"
fi

1. Use the script by handing-over the public ssh-key of the Shoot cluster:

kubectl get secret aws-gvisor.ssh-keypair -o json | jq -r .data.\"id_rsa.pub\" | ./generate-userdata.sh

This generates a file called gardener-bastion-userdata.sh in the same directory containing the user-data.

1. The following information is needed to create the bastion instance:

bastion-IAM-instance-profile-name - Use the created instance profile with the name <shoot-name>-bastions

image-id - It is possible to use the same image-id as the one used for the target instance (or any other image). Has cloud provider specific format (AWS: ami).

ssh-public-key-name

- This is the ssh key pair already created in the Shoot's cloud provider account by Gardener during the `Infrastructure` CRD reconciliation.
- The name is usually: `<shoot-name>-ssh-publickey`

subnet-id - Choose a subnet that is attached to an Internet Gateway and NAT Gateway (bastion instance must have a public IP). - The Gardener created public subnet with the name <shoot-name>-public-utility-<xy> can be used. Please check the created subnets with the cloud provider.

bastion-security-group-id - Use the id of the created bastion security group.

file-path-to-userdata - Use the filepath to the user-data file generated in the previous step.

• bastion-instance-name
  • Optionaly, you can tag the instance.
  • Usually <shoot-name>-bastions

1. Create the bastion instance via:

ec2 run-instances --iam-instance-profile Name=<bastion-IAM-instance-profile-name> --image-id <image-id>  --count 1 --instance-type t3.nano --key-name <ssh-public-key-name>  --security-group-ids <bastion-security-group-id> --subnet-id <subnet-id> --associate-public-ip-address --user-data <file-path-to-userdata> --tag-specifications ResourceType=instance,Tags=[{Key=Name,Value=<bastion-instance-name>},{Key=component,Value=<mytag>}] ResourceType=volume,Tags=[{Key=component,Value=<mytag>}]"

Capture the instance-id from the response and wait until the ec2 instance is running and has a public IP address.

Connecting to the Target Instance

1. Save the private key of the ssh-key-pair in a temporary local file for later use:

umask 077

kubectl get secret <shoot-name>.ssh-keypair -o json | jq -r .data.\"id_rsa\" | base64 -d > id_rsa.key

1. Use the private ssh key to ssh into the bastion instance:

ssh -i <path-to-private-key> gardener@<public-bastion-instance-ip> 

1. If that works, connect from your local terminal to the target instance via the bastion:

ssh  -i <path-to-private-key> -o ProxyCommand="ssh -W %h:%p -i <private-key> -o IdentitiesOnly=yes -o StrictHostKeyChecking=no gardener@<public-ip-bastion>" gardener@<private-ip-target-instance> -o IdentitiesOnly=yes -o StrictHostKeyChecking=no

Cleanup

Do not forget to cleanup the created resources. Otherwise Gardener will eventually fail to delete the Shoot.

2.5.3 - How to Debug a Pod

Introduction

Kubernetes offers powerful options to get more details about startup or runtime failures of pods as e.g. described in Application Introspection and Debugging or Debug Pods and Replication Controllers.

In order to identify pods with potential issues, you could, e.g., run kubectl get pods --all-namespaces | grep -iv Running to filter out the pods which are not in the state Running. One of frequent error state is CrashLoopBackOff, which tells that a pod crashes right after the start. Kubernetes then tries to restart the pod again, but often the pod startup fails again.

Here is a short list of possible reasons which might lead to a pod crash:

1. Error during image pull caused by e.g. wrong/missing secrets or wrong/missing image
2. The app runs in an error state caused e.g. by missing environmental variables (ConfigMaps) or secrets
3. Liveness probe failed
4. Too high resource consumption (memory and/or CPU) or too strict quota settings
5. Persistent volumes can’t be created/mounted
6. The container image is not updated

Basically, the commands kubectl logs ... and kubectl describe ... with different parameters are used to get more detailed information. By calling e.g. kubectl logs --help you can get more detailed information about the command and its parameters.

In the next sections you’ll find some basic approaches to get some ideas what went wrong.

Remarks:

• Even if the pods seem to be running, as the status Running indicates, a high counter of the Restarts shows potential problems
• You can get a good overview of the troubleshooting process with the interactive tutorial Troubleshooting with Kubectl available which explains basic debugging activities
• The examples below are deployed into the namespace default. In case you want to change it, use the optional parameter --namespace <your-namespace> to select the target namespace. The examples require a Kubernetes release ≥ 1.8.

Prerequisites

Your deployment was successful (no logical/syntactical errors in the manifest files), but the pod(s) aren’t running.

Error Caused by Wrong Image Name

Start by running kubectl describe pod <your-pod> <your-namespace> to get detailed information about the pod startup.

In the Events section, you should get an error message like Failed to pull image ... and Reason: Failed. The pod is in state ImagePullBackOff.

The example below is based on a demo in the Kubernetes documentation. In all examples, the default namespace is used.

First, perform a cleanup with:

kubectl delete pod termination-demo

Next, create a resource based on the yaml content below:

apiVersion: v1
kind: Pod 
metadata:
  name: termination-demo
spec:
  containers:
  - name: termination-demo-container
    image: debiann
    command: ["/bin/sh"]
    args: ["-c", "sleep 10 && echo Sleep expired > /dev/termination-log"]

kubectl describe pod termination-demo lists in the Event section the content

Events:
  FirstSeen	LastSeen	Count	From							SubObjectPath					Type		Reason			Message
  ---------	--------	-----	----							-------------					--------	------			-------
  2m		2m		1	default-scheduler											Normal		Scheduled		Successfully assigned termination-demo to ip-10-250-17-112.eu-west-1.compute.internal
  2m		2m		1	kubelet, ip-10-250-17-112.eu-west-1.compute.internal							Normal		SuccessfulMountVolume	MountVolume.SetUp succeeded for volume "default-token-sgccm" 
  2m		1m		4	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Normal		Pulling			pulling image "debiann"
  2m		1m		4	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Warning		Failed			Failed to pull image "debiann": rpc error: code = Unknown desc = Error: image library/debiann:latest not found
  2m		54s		10	kubelet, ip-10-250-17-112.eu-west-1.compute.internal							Warning		FailedSync		Error syncing pod
  2m		54s		6	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Normal		BackOff			Back-off pulling image "debiann"

The error message with Reason: Failed tells you that there is an error during pulling the image. A closer look at the image name indicates a misspelling.

The App Runs in an Error State Caused, e.g., by Missing Environmental Variables (ConfigMaps) or Secrets

This example illustrates the behavior in the case when the app expects environment variables but the corresponding Kubernetes artifacts are missing.

First, perform a cleanup with:

kubectl delete deployment termination-demo
kubectl delete configmaps app-env

Next, deploy the following manifest:

apiVersion: apps/v1beta2 
kind: Deployment
metadata:
  name: termination-demo
  labels:
     app: termination-demo
spec:
  replicas: 1
  selector:
    matchLabels:
      app: termination-demo
  template:
    metadata:
      labels:
        app: termination-demo
    spec:
      containers:
      - name: termination-demo-container
        image: debian
        command: ["/bin/sh"]
        args: ["-c", "sed \"s/foo/bar/\" < $MYFILE"]

Now, the command kubectl get pods lists the pod termination-demo-xxx in the state Error or CrashLoopBackOff. The command kubectl describe pod termination-demo-xxx tells you that there is no error during startup but gives no clue about what caused the crash.

Events:
  FirstSeen	LastSeen	Count	From							SubObjectPath					Type		Reason		Message
  ---------	--------	-----	----							-------------					--------	------		-------
  19m		19m		1	default-scheduler											Normal		Scheduled	Successfully assigned termination-demo-5fb484867d-xz2x9 to ip-10-250-17-112.eu-west-1.compute.internal
  19m		19m		1	kubelet, ip-10-250-17-112.eu-west-1.compute.internal							Normal		SuccessfulMountVolume	MountVolume.SetUp succeeded for volume "default-token-sgccm" 
  19m		19m		4	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Normal		Pulling		pulling image "debian"
  19m		19m		4	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Normal		Pulled		Successfully pulled image "debian"
  19m		19m		4	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Normal		Created		Created container
  19m		19m		4	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Normal		Started		Started container
  19m		14m		24	kubelet, ip-10-250-17-112.eu-west-1.compute.internal	spec.containers{termination-demo-container}	Warning		BackOff		Back-off restarting failed container
  19m		4m		69	kubelet, ip-10-250-17-112.eu-west-1.compute.internal							Warning		FailedSync	Error syncing pod

The command kubectl get logs termination-demo-xxx gives access to the output, the application writes on stderr and stdout. In this case, you should get an output similar to:

/bin/sh: 1: cannot open : No such file

So you need to have a closer look at the application. In this case, the environmental variable MYFILE is missing. To fix this issue, you could e.g. add a ConfigMap to your deployment as is shown in the manifest listed below:

apiVersion: v1
kind: ConfigMap
metadata:
  name: app-env
data:
  MYFILE: "/etc/profile"
---
apiVersion: apps/v1beta2 
kind: Deployment
metadata:
  name: termination-demo
  labels:
     app: termination-demo
spec:
  replicas: 1
  selector:
    matchLabels:
      app: termination-demo
  template:
    metadata:
      labels:
        app: termination-demo
    spec:
      containers:
      - name: termination-demo-container
        image: debian
        command: ["/bin/sh"]
        args: ["-c", "sed \"s/foo/bar/\" < $MYFILE"]
        envFrom:
        - configMapRef:
            name: app-env 

Note that once you fix the error and re-run the scenario, you might still see the pod in a CrashLoopBackOff status. It is because the container finishes the command sed ... and runs to completion. In order to keep the container in a Running status, a long running task is required, e.g.:

apiVersion: v1
kind: ConfigMap
metadata:
  name: app-env
data:
  MYFILE: "/etc/profile"
  SLEEP: "5"
---
apiVersion: apps/v1beta2
kind: Deployment
metadata:
  name: termination-demo
  labels:
     app: termination-demo
spec:
  replicas: 1
  selector:
    matchLabels:
      app: termination-demo
  template:
    metadata:
      labels:
        app: termination-demo
    spec:
      containers:
      - name: termination-demo-container
        image: debian
        command: ["/bin/sh"]
        # args: ["-c", "sed \"s/foo/bar/\" < $MYFILE"]
        args: ["-c", "while true; do sleep $SLEEP; echo sleeping; done;"]
        envFrom:
        - configMapRef:
            name: app-env

Too High Resource Consumption (Memory and/or CPU) or Too Strict Quota Settings

You can optionally specify the amount of memory and/or CPU your container gets during runtime. In case these settings are missing, the default requests settings are taken: CPU: 0m (in Milli CPU) and RAM: 0Gi, which indicate no other limits other than the ones of the node(s) itself. For more details, e.g. about how to configure limits, see Configure Default Memory Requests and Limits for a Namespace.

In case your application needs more resources, Kubernetes distinguishes between requests and limit settings: requests specify the guaranteed amount of resource, whereas limit tells Kubernetes the maximum amount of resource the container might need. Mathematically, both settings could be described by the relation 0 <= requests <= limit. For both settings you need to consider the total amount of resources your nodes provide. For a detailed description of the concept, see Resource Quality of Service in Kubernetes.

Use kubectl describe nodes to get a first overview of the resource consumption in your cluster. Of special interest are the figures indicating the amount of CPU and Memory Requests at the bottom of the output.

The next example demonstrates what happens in case the CPU request is too high in order to be managed by your cluster.

First, perform a cleanup with:

kubectl delete deployment termination-demo
kubectl delete configmaps app-env

Next, adapt the cpu below in the yaml below to be slightly higher than the remaining CPU resources in your cluster and deploy this manifest. In this example, 600m (milli CPUs) are requested in a Kubernetes system with a single 2 core worker node which results in an error message.

apiVersion: apps/v1beta2 
kind: Deployment
metadata:
  name: termination-demo
  labels:
     app: termination-demo
spec:
  replicas: 1
  selector:
    matchLabels:
      app: termination-demo
  template:
    metadata:
      labels:
        app: termination-demo
    spec:
      containers:
      - name: termination-demo-container
        image: debian
        command: ["/bin/sh"]
        args: ["-c", "sleep 10 && echo Sleep expired > /dev/termination-log"]
        resources:
          requests:
            cpu: "600m" 

The command kubectl get pods lists the pod termination-demo-xxx in the state Pending. More details on why this happens could be found by using the command kubectl describe pod termination-demo-xxx:

$ kubectl describe po termination-demo-fdb7bb7d9-mzvfw
Name:           termination-demo-fdb7bb7d9-mzvfw
Namespace:      default
...
Containers:
  termination-demo-container:
    Image:      debian
    Port:       <none>
    Host Port:  <none>
    Command:
      /bin/sh
    Args:
      -c
      sleep 10 && echo Sleep expired > /dev/termination-log
    Requests:
      cpu:        6
    Environment:  <none>
    Mounts:
      /var/run/secrets/kubernetes.io/serviceaccount from default-token-t549m (ro)
Conditions:
  Type           Status
  PodScheduled   False
Events:
  Type     Reason            Age               From               Message
  ----     ------            ----              ----               -------
  Warning  FailedScheduling  9s (x7 over 40s)  default-scheduler  0/2 nodes are available: 2 Insufficient cpu.

You can find more details in:

• Managing Compute Resources for Containters
• Resource Quality of Service in Kubernetes

Remarks:

• This example works similarly when specifying a too high request for memory
• In case you configured an autoscaler range when creating your Kubernetes cluster, another worker node will be spinned up automatically if you didn’t reach the maximum number of worker nodes
• In case your app is running out of memory (the memory settings are too small), you will typically find an OOMKilled (Out Of Memory) message in the Events section of the kubectl describe pod ... output

The Container Image Is Not Updated

You applied a fix in your app, created a new container image and pushed it into your container repository. After redeploying your Kubernetes manifests, you expected to get the updated app, but the same bug is still in the new deployment present.

This behavior is related to how Kubernetes decides whether to pull a new docker image or to use the cached one.

In case you didn’t change the image tag, the default image policy IfNotPresent tells Kubernetes to use the cached image (see Images).

As a best practice, you should not use the tag latest and change the image tag in case you changed anything in your image (see Configuration Best Practices).

For more information, see Container Image Not Updating.

Related Links

• Application Introspection and Debugging
• Debug Pods and Replication Controllers
• Logging Architecture
• Configure Default Memory Requests and Limits for a Namespace
• Managing Compute Resources for Containters
• Resource Quality of Service in Kubernetes
• Interactive Tutorial Troubleshooting with Kubectl
• Images
• Kubernetes Best Practices

2.5.4 - tail -f /var/log/my-application.log

Problem

One thing that always bothered me was that I couldn’t get logs of several pods at once with kubectl. A simple tail -f <path-to-logfile> isn’t possible at all. Certainly, you can use kubectl logs -f <pod-id>, but it doesn’t help if you want to monitor more than one pod at a time.

This is something you really need a lot, at least if you run several instances of a pod behind a deployment. This is even more so if you don’t have a Kibana or a similar setup.

Solution

Luckily, there are smart developers out there who always come up with solutions. The finding of the week is a small bash script that allows you to aggregate log files of several pods at the same time in a simple way. The script is called kubetail and is available at GitHub.