This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.


1 - Changing the API

Changing the API

This document describes the steps that need to be performed when changing the API. It provides guidance for API changes to both (Gardener system in general or component configurations).

Generally, as Gardener is a Kubernetes-native extension, it follows the same API conventions and guidelines like Kubernetes itself. The Kubernetes API Conventions as well as Changing the API topics already provide a good overview and general explanation of the basic concepts behind it. We are following the same approaches.

Gardener API

The Gardener API is defined in the pkg/apis/{core,extensions,settings} directories and is the main point of interaction with the system. It must be ensured that the API is always backwards-compatible.

Changing the API

Checklist when changing the API:

  1. Modify the field(s) in the respective Golang files of all external versions and the internal version.
    1. Make sure new fields are being added as “optional” fields, i.e., they are of pointer types, they have the // +optional comment, and they have the omitempty JSON tag.
    2. Make sure that the existing field numbers in the protobuf tags are not changed.
  2. If necessary, implement/adapt the conversion logic defined in the versioned APIs (e.g., pkg/apis/core/v1beta1/conversions*.go).
  3. If necessary, implement/adapt defaulting logic defined in the versioned APIs (e.g., pkg/apis/core/v1beta1/defaults*.go).
  4. Run the code generation: make generate
  5. If necessary, implement/adapt validation logic defined in the internal API (e.g., pkg/apis/core/validation/validation*.go).
  6. If necessary, adapt the exemplary YAML manifests of the Gardener resources defined in example/*.yaml.
  7. In most cases, it makes sense to add/adapt the documentation for administrators/operators and/or end-users in the docs folder to provide information on purpose and usage of the added/changed fields.
  8. When opening the pull request, always add a release note so that end-users are becoming aware of the changes.

Removing a Field

If fields shall be removed permanently from the API, then a proper deprecation period must be adhered to so that end-users have enough time to adapt their clients.

Once the deprecation period is over, the field should be dropped from the API in a two-step process, i.e., in two release cycles. In the first step, all the usages in the code base should be dropped. In the second step, the field should be dropped from API. We need to follow this two-step process cause there can be the case where gardener-apiserver is upgraded to a new version in which the field has been removed but other controllers are still on the old version of Gardener. This can lead to nil pointer exceptions or other unexpected behaviour.

The steps for removing a field from the code base is:

  1. The field in the external version(s) has to be commented out with appropriate doc string that the protobuf number of the corresponding field is reserved. Example:

    -	SeedTemplate *gardencorev1beta1.SeedTemplate `json:"seedTemplate,omitempty" protobuf:"bytes,2,opt,name=seedTemplate"`
    +	// SeedTemplate is tombstoned to show why 2 is reserved protobuf tag.
    +	// SeedTemplate *gardencorev1beta1.SeedTemplate `json:"seedTemplate,omitempty" protobuf:"bytes,2,opt,name=seedTemplate"`

    The reasoning behind this is to prevent the same protobuf number being used by a new field. Introducing a new field with the same protobuf number would be a breaking change for clients still using the old protobuf definitions that have the old field for the given protobuf number. The field in the internal version can be removed.

  2. A unit test has to be added to make sure that a new field does not reuse the already reserved protobuf tag.

Example of field removal can be found in the Remove seedTemplate field from ManagedSeed API PR.

Component Configuration APIs

Most Gardener components have a component configuration that follows similar principles to the Gardener API. Those component configurations are defined in pkg/{controllermanager,gardenlet,scheduler},pkg/apis/config. Hence, the above checklist also applies for changes to those APIs. However, since these APIs are only used internally and only during the deployment of Gardener, the guidelines with respect to changes and backwards-compatibility are slightly relaxed. If necessary, it is allowed to remove fields without a proper deprecation period if the release note uses the breaking operator keywords.

In addition to the above checklist:

  1. If necessary, then adapt the Helm chart of Gardener defined in charts/gardener. Adapt the values.yaml file as well as the manifest templates.

2 - Component Checklist

Checklist For Adding New Components

Adding new components that run in the garden, seed, or shoot cluster is theoretically quite simple - we just need a Deployment (or other similar workload resource), the respective container image, and maybe a bit of configuration. In practice, however, there are a couple of things to keep in mind in order to make the deployment production-ready. This document provides a checklist for them that you can walk through.


  1. Avoid usage of Helm charts (example)

    Nowadays, we use Golang components instead of Helm charts for deploying components to a cluster. Please find a typical structure of such components in the provided metrics_server.go file (configuration values are typically managed in a Values structure). There are a few exceptions (e.g., Istio) still using charts, however the default should be using a Golang-based implementation. For the exceptional cases, use Golang’s embed package to embed the Helm chart directory (example 1, example 2).

  2. Choose the proper deployment way (example 1 (direct application w/ client), example 2 (using ManagedResource), example 3 (mixed scenario))

    For historic reasons, resources related to shoot control plane components are applied directly with the client. All other resources (seed or shoot system components) are deployed via gardener-resource-manager’s Resource controller (ManagedResources) since it performs health checks out-of-the-box and has a lot of other features (see its documentation for more information). Components that can run as both seed system component or shoot control plane component (e.g., VPA or kube-state-metrics) can make use of these utility functions.

  3. Do not hard-code container image references (example 1, example 2, example 3)

    We define all image references centrally in the charts/images.yaml file. Hence, the image references must not be hard-coded in the pod template spec but read from this so-called image vector instead.

  4. Use unique ConfigMaps/Secrets (example 1, example 2)

    Unique ConfigMaps/Secrets are immutable for modification and have a unique name. This has a couple of benefits, e.g. the kubelet doesn’t watch these resources, and it is always clear which resource contains which data since it cannot be changed. As a consequence, unique/immutable ConfigMaps/Secret are superior to checksum annotations on the pod templates. Stale/unused ConfigMaps/Secrets are garbage-collected by gardener-resource-manager’s GarbageCollector. There are utility functions (see examples above) for using unique ConfigMaps/Secrets in Golang components. It is essential to inject the annotations into the workload resource to make the garbage-collection work.
    Note that some ConfigMaps/Secrets should not be unique (e.g., those containing monitoring or logging configuration). The reason is that the old revision stays in the cluster even if unused until the garbage-collector acts. During this time, they would be wrongly aggregated to the full configuration.

  5. Manage certificates/secrets via secrets manager (example)

    You should use the secrets manager for the management of any kind of credentials. This makes sure that credentials rotation works out-of-the-box without you requiring to think about it. Generally, do not use client certificates (see the Security section).

  6. Consider hibernation when calculating replica count (example)

    Shoot clusters can be hibernated meaning that all control plane components in the shoot namespace in the seed cluster are scaled down to zero and all worker nodes are terminated. If your component runs in the seed cluster then you have to consider this case and provide the proper replica count. There is a utility function available (see example).

  7. Ensure task dependencies are as precise as possible in shoot flows (example 1, example 2)

    Only define the minimum of needed dependency tasks in the shoot reconciliation/deletion flows.

  8. Handle shoot system components

    Shoot system components deployed by gardener-resource-manager are labelled with gardener. This makes Gardener adding required label selectors and tolerations so that non-DaemonSet managed Pods will exclusively run on selected nodes (for more information, see System Components Webhook). DaemonSets on the other hand, should generally tolerate any NoSchedule or NoExecute taints so that they can run on any Node, regardless of user added taints.


  1. Use a dedicated ServiceAccount and disable auto-mount (example)

    Components that need to talk to the API server of their runtime cluster must always use a dedicated ServiceAccount (do not use default), with automountServiceAccountToken set to false. This makes gardener-resource-manager’s TokenInvalidator invalidate the static token secret and its ProjectedTokenMount webhook inject a projected token automatically.

  2. Use shoot access tokens instead of a client certificates (example)

    For components that need to talk to a target cluster different from their runtime cluster (e.g., running in seed cluster but talking to shoot) the gardener-resource-manager’s TokenRequestor should be used to manage a so-called “shoot access token”.

  3. Define RBAC roles with minimal privileges (example)

    The component’s ServiceAccount (if it exists) should have as little privileges as possible. Consequently, please define proper RBAC roles for it. This might include a combination of ClusterRoles and Roles. Please do not provide elevated privileges due to laziness (e.g., because there is already a ClusterRole that can be extended vs. creating a Role only when access to a single namespace is needed).

  4. Use NetworkPolicys to restrict network traffic (example)

    You should restrict both ingress and egress traffic to/from your component as much as possible to ensure that it only gets access to/from other components if really needed. Gardener provides a few default policies for typical usage scenarios. For more information, see Seed Network Policies and Shoot Network Policies.

  5. Do not run components in privileged mode (example 1, example 2)

    Avoid running components with privileged=true. Instead, define the needed Linux capabilities.

  6. Choose the proper Seccomp profile (example 1, example 2)

    The Seccomp profile will be defaulted by gardener-resource-manager’s SeccompProfile webhook which works well for the majority of components. However, in some special cases you might need to overwrite it.

  7. Define PodSecurityPolicys (example)

    PodSecurityPolicys are deprecated, however Gardener still supports shoot clusters with older Kubernetes versions (ref). To make sure that such clusters can run with .spec.kubernetes.allowPrivilegedContainers=false, you have to define proper PodSecurityPolicys. For more information, see Pod Security.

High Availability / Stability

  1. Specify the component type label for high availability (example)

    To support high-availability deployments, gardener-resource-managers HighAvailabilityConfig webhook injects the proper specification like replica or topology spread constraints. You only need to specify the type label. For more information, see High Availability Of Deployed Components.

  2. Define a PodDisruptionBudget (example)

    Closely related to high availability but also to stability in general: The definition of a PodDisruptionBudget with maxUnavailable=1 should be provided by default.

  3. Choose the right PriorityClass (example)

    Each cluster runs many components with different priorities. Gardener provides a set of default PriorityClasses. For more information, see Priority Classes.

  4. Consider defining liveness and readiness probes (example)

    To ensure smooth rolling update behaviour, consider the definition of liveness and/or readiness probes.


  1. Provide resource requirements (example)

    All components should have resource requirements. Generally, they should always request CPU and memory, while only memory shall be limited (no CPU limits!).

  2. Define a VerticalPodAutoscaler (example)

    We typically perform vertical auto-scaling via the VPA managed by the Kubernetes community. Each component should have a respective VerticalPodAutoscaler with “min allowed” resources, “auto update mode”, and “requests only”-mode. VPA is always enabled in garden or seed clusters, while it is optional for shoot clusters.

  3. Define a HorizontalPodAutoscaler if needed (example)

    If your component is capable of scaling horizontally, you should consider defining a HorizontalPodAutoscaler.

Observability / Operations Productivity

  1. Provide monitoring scrape config and alerting rules (example 1, example 2)

    Components should provide scrape configuration and alerting rules for Prometheus/Alertmanager if appropriate. This should be done inside a dedicated monitoring.go file. Extensions should follow the guidelines described in Extensions Monitoring Integration.

  2. Provide logging parsers and filters (example 1, example 2)

    Components should provide parsers and filters for fluent-bit, if appropriate. This should be done inside a dedicated logging.go file. Extensions should follow the guidelines described in Fluent-bit log parsers and filters.

  3. Set the revisionHistoryLimit to 2 for Deployments (example)

    In order to allow easy inspection of two ReplicaSets to quickly find the changes that lead to a rolling update, the revision history limit should be set to 2.

  4. Define health checks (example 1, example 2)

    gardenlet’s care controllers regularly check the health status of system or control plane components. You need to enhance the lists of components to check if your component related to the seed system or shoot control plane (shoot system components are automatically checked via their respective ManagedResource conditions), see examples above.

  5. Configure automatic restarts in shoot maintenance time window (example 1, example 2)

    Gardener offers to restart components during the maintenance time window. For more information, see Restart Control Plane Controllers and Restart Some Core Addons. You can consider adding the needed label to your control plane component to get this automatic restart (probably not needed for most components).

3 - Dependencies

Dependency Management

We are using go modules for depedency management. In order to add a new package dependency to the project, you can perform go get <PACKAGE>@<VERSION> or edit the go.mod file and append the package along with the version you want to use.

Updating Dependencies

The Makefile contains a rule called revendor which performs go mod tidy and go mod vendor:

  • go mod tidy makes sure go.mod matches the source code in the module. It adds any missing modules necessary to build the current module’s packages and dependencies, and it removes unused modules that don’t provide any relevant packages.
  • go mod vendor resets the main module’s vendor directory to include all packages needed to build and test all the main module’s packages. It does not include test code for vendored packages.
make revendor

The dependencies are installed into the vendor folder, which should be added to the VCS.

⚠️ Make sure that you test the code after you have updated the dependencies!

Exported Packages

This repository contains several packages that could be considered “exported packages”, in a sense that they are supposed to be reused in other Go projects. For example:

  • Gardener’s API packages: pkg/apis
  • Library for building Gardener extensions: extensions
  • Gardener’s Test Framework: test/framework

There are a few more folders in this repository (non-Go sources) that are reused across projects in the Gardener organization:

  • GitHub templates: .github
  • Concourse / cc-utils related helpers: hack/.ci
  • Development, build and testing helpers: hack

These packages feature a dummy doc.go file to allow other Go projects to pull them in as go mod dependencies.

These packages are explicitly not supposed to be used in other projects (consider them as “non-exported”):

  • API validation packages: pkg/apis/*/*/validation
  • Operation package (main Gardener business logic regarding Seed and Shoot clusters): pkg/operation
  • Third party code: third_party

Currently, we don’t have a mechanism yet for selectively syncing out these exported packages into dedicated repositories like kube’s staging mechanism (publishing-bot).

Import Restrictions

We want to make sure that other projects can depend on this repository’s “exported” packages without pulling in the entire repository (including “non-exported” packages) or a high number of other unwanted dependencies. Hence, we have to be careful when adding new imports or references between our packages.

ℹ️ General rule of thumb: the mentioned “exported” packages should be as self-contained as possible and depend on as few other packages in the repository and other projects as possible.

In order to support that rule and automatically check compliance with that goal, we leverage import-boss. The tool checks all imports of the given packages (including transitive imports) against rules defined in .import-restrictions files in each directory. An import is allowed if it matches at least one allowed prefix and does not match any forbidden prefixes.

Note: '' (the empty string) is a prefix of everything. For more details, see the import-boss topic.

import-boss is executed on every pull request and blocks the PR if it doesn’t comply with the defined import restrictions. You can also run it locally using make check.

Import restrictions should be changed in the following situations:

  • We spot a new pattern of imports across our packages that was not restricted before but makes it more difficult for other projects to depend on our “exported” packages. In that case, the imports should be further restricted to disallow such problematic imports, and the code/package structure should be reworked to comply with the newly given restrictions.
  • We want to share code between packages, but existing import restrictions prevent us from doing so. In that case, please consider what additional dependencies it will pull in, when loosening existing restrictions. Also consider possible alternatives, like code restructurings or extracting shared code into dedicated packages for minimal impact on dependent projects.

4 - Getting Started Locally

Developing Gardener Locally

This document will walk you through running Gardener on your local machine for development purposes. If you encounter difficulties, please open an issue so that we can make this process easier.

Gardener runs in any Kubernetes cluster. In this guide, we will start a KinD cluster which is used as both garden and seed cluster (please refer to the architecture overview) for simplicity.

The Gardener components, however, will be run as regular processes on your machine (hence, no container images are being built).

Architecture Diagram


When developing Gardener on your local machine you might face several limitations:

  • Your machine doesn’t have enough compute resources (see prerequisites) for hosting a second seed cluster or multiple shoot clusters.
  • Developing Gardener’s IPv6 features requires a Linux machine and native IPv6 connectivity to the internet, but you’re on macOS or don’t have IPv6 connectivity in your office environment or via your home ISP.

In these cases, you might want to check out one of the following options that run the setup described in this guide elsewhere for circumventing these limitations:


  • Make sure that you have followed the Local Setup guide up until the Get the sources step.

  • Make sure your Docker daemon is up-to-date, up and running and has enough resources (at least 4 CPUs and 4Gi memory; see here how to configure the resources for Docker for Mac).

    Please note that 4 CPU / 4Gi memory might not be enough for more than one Shoot cluster, i.e., you might need to increase these values if you want to run additional Shoots. If you plan on following the optional steps to create a second seed cluster, the required resources will be more - at least 10 CPUs and 16Gi memory.

    Additionally, please configure at least 120Gi of disk size for the Docker daemon.

    Tip: With docker system df and docker system prune -a you can cleanup unused data.

  • Make sure the kind docker network is using the CIDR

    • If the network does not exist, it can be created with docker network create kind --subnet
    • If the network already exists, the CIDR can be checked with docker network inspect kind | jq '.[].IPAM.Config[].Subnet'. If it is not, delete the network with docker network rm kind and create it with the command above.
  • Make sure that you increase the maximum number of open files on your host:

    • On Mac, run sudo launchctl limit maxfiles 65536 200000

    • On Linux, extend the /etc/security/limits.conf file with

      * hard nofile 97816
      * soft nofile 97816

      and reload the terminal.

Setting Up the KinD Cluster (Garden and Seed)

make kind-up KIND_ENV=local

If you want to setup an IPv6 KinD cluster, use make kind-up IPFAMILY=ipv6 instead.

This command sets up a new KinD cluster named gardener-local and stores the kubeconfig in the ./example/gardener-local/kind/local/kubeconfig file.

It might be helpful to copy this file to $HOME/.kube/config since you will need to target this KinD cluster multiple times. Alternatively, make sure to set your KUBECONFIG environment variable to ./example/gardener-local/kind/local/kubeconfig for all future steps via export KUBECONFIG=example/gardener-local/kind/local/kubeconfig.

All following steps assume that you are using this kubeconfig.

Additionally, this command also deploys a local container registry to the cluster as well as a few registry mirrors, that are set up as a pull-through cache for all upstream registries Gardener uses by default. This is done to speed up image pulls across local clusters. The local registry can be accessed as localhost:5001 for pushing and pulling. The storage directories of the registries are mounted to the host machine under dev/local-registry. With this, mirrored images don’t have to be pulled again after recreating the cluster.

The command also deploys a default calico installation as the cluster’s CNI implementation with NetworkPolicy support (the default kindnet CNI doesn’t provide NetworkPolicy support). Furthermore, it deploys the metrics-server in order to support HPA and VPA on the seed cluster.

Outgoing IPv6 Single-Stack Networking (optional)

If you want to test IPv6-related features, we need to configure NAT for outgoing traffic from the kind network to the internet. After make kind-up IPFAMILY=ipv6, check the network created by kind:

$ docker network inspect kind | jq '.[].IPAM.Config[].Subnet'

Determine which device is used for outgoing internet traffic by looking at the default route:

$ ip route show default
default via dev enp3s0 proto dhcp src metric 100

Configure NAT for traffic from the kind cluster to the internet using the IPv6 range and the network device from the previous two steps:

ip6tables -t nat -A POSTROUTING -o enp3s0 -s fc00:f853:ccd:e793::/64 -j MASQUERADE

Setting Up Gardener

In a terminal pane, run:

make dev-setup                                                                # preparing the environment (without webhooks for now)
kubectl wait --for=condition=ready pod -l run=etcd -n garden --timeout 2m     # wait for etcd to be ready
make start-apiserver                                                          # starting gardener-apiserver

In a new terminal pane, run:

kubectl wait --for=condition=available apiservice # wait for gardener-apiserver to be ready
make start-admission-controller                                               # starting gardener-admission-controller

In a new terminal pane, run:

make dev-setup DEV_SETUP_WITH_WEBHOOKS=true                                   # preparing the environment with webhooks
make start-controller-manager                                                 # starting gardener-controller-manager

(Optional): In a new terminal pane, run:

make start-scheduler                                                          # starting gardener-scheduler

In a new terminal pane, run:

make register-local-env                                                       # registering the local environment (CloudProfile, Seed, etc.)
make start-gardenlet SEED_NAME=local                                          # starting gardenlet

In a new terminal pane, run:

make start-extension-provider-local                                           # starting gardener-extension-provider-local

ℹ️ The provider-local is started with elevated privileges since it needs to manipulate your /etc/hosts file to enable you accessing the created shoot clusters from your local machine, see this for more details.

Creating a Shoot Cluster

You can wait for the Seed to become ready by running:

kubectl wait --for=condition=gardenletready --for=condition=extensionsready --for=condition=bootstrapped seed local --timeout=5m

Alternatively, you can run kubectl get seed local and wait for the STATUS to indicate readiness:

local   Ready    local      local    4m42s   vX.Y.Z-dev    v1.21.1

In order to create a first shoot cluster, just run:

kubectl apply -f example/provider-local/shoot.yaml

You can wait for the Shoot to be ready by running:

kubectl wait --for=condition=apiserveravailable --for=condition=controlplanehealthy --for=condition=observabilitycomponentshealthy --for=condition=everynodeready --for=condition=systemcomponentshealthy shoot local -n garden-local --timeout=10m

Alternatively, you can run kubectl -n garden-local get shoot local and wait for the LAST OPERATION to reach 100%:

local   local          local      local    1.21.0        Awake         Create Processing (43%)   healthy   94s

(Optional): You could also execute a simple e2e test (creating and deleting a shoot) by running:

make test-e2e-local-simple KUBECONFIG="$PWD/example/gardener-local/kind/local/kubeconfig"

When the shoot got successfully created you can access it as follows:

kubectl -n garden-local get secret local.kubeconfig -o jsonpath={.data.kubeconfig} | base64 -d > /tmp/kubeconfig-shoot-local.yaml
kubectl --kubeconfig=/tmp/kubeconfig-shoot-local.yaml get nodes

(Optional): Setting Up a Second Seed Cluster

There are cases where you would want to create a second seed cluster in your local setup. For example, if you want to test the control plane migration feature. The following steps describe how to do that.

Add a new IP address on your loopback device which will be necessary for the new KinD cluster that you will create. On Mac, the default loopback device is lo0.

sudo ip addr add dev lo0                                     # adding ip to the loopback interface

Next, setup the second KinD cluster:

make kind2-up KIND_ENV=local

This command sets up a new KinD cluster named gardener-local2 and stores its kubeconfig in the ./example/gardener-local/kind/local2/kubeconfig file. You will need this file when starting the provider-local extension controller for the second seed cluster.

make register-kind2-env                                           # registering the local2 seed
make start-gardenlet SEED_NAME=local2                             # starting gardenlet for the local2 seed

In a new terminal pane, run:

export KUBECONFIG=./example/gardener-local/kind/local2/kubeconfig       # setting KUBECONFIG to point to second kind cluster
make start-extension-provider-local \
  WEBHOOK_CERT_DIR=/tmp/gardener-extension-provider-local2 \
  HEALTH_BIND_ADDRESS=:8083                                       # starting gardener-extension-provider-local

If you want to perform a control plane migration you can follow the steps outlined in the Control Plane Migration topic to migrate the shoot cluster to the second seed you just created.

Deleting the Shoot Cluster

./hack/usage/delete shoot local garden-local

(Optional): Tear Down the Second Seed Cluster

make tear-down-kind2-env
make kind2-down

Tear Down the Gardener Environment

make tear-down-local-env
make kind-down

Remote Local Setup

Just like Prow is executing the KinD based integration tests in a K8s pod, it is possible to interactively run this KinD based Gardener development environment aka “local setup” in a “remote” K8s pod.

k apply -f docs/development/content/remote-local-setup.yaml
k exec -it deployment/remote-local-setup -- sh

tmux -u a


Please refer to the TMUX documentation for working effectively inside the remote-local-setup pod.

To access Grafana, Prometheus, or other components in a browser, two port forwards are needed:

The port forward from the laptop to the pod:

k port-forward deployment/remote-local-setup 3000

The port forward in the remote-local-setup pod to the respective component:

k port-forward -n shoot--local--local deployment/grafana-operators 3000

5 - High Availability

High Availability of Deployed Components

gardenlets and extension controllers are deploying components via Deployments, StatefulSets, etc., as part of the shoot control plane, or the seed or shoot system components.

Some of the above component deployments must be further tuned to improve fault tolerance / resilience of the service. This document outlines what needs to be done to achieve this goal.

Please be forwarded to the Convenient Application Of These Rules section, if you want to take a shortcut to the list of actions that require developers’ attention.

Seed Clusters

The worker nodes of seed clusters can be deployed to one or multiple availability zones. The Seed specification allows you to provide the information which zones are available:

    region: europe-1
    - europe-1a
    - europe-1b
    - europe-1c

Independent of the number of zones, seed system components like the gardenlet or the extension controllers themselves, or others like etcd-druid, dependency-watchdog, etc., should always be running with multiple replicas.

Concretely, all seed system components should respect the following conventions:

  • Replica Counts

    Component Type< 3 Zones>= 3 ZonesComment
    Observability (Monitoring, Logging)11Downtimes accepted due to cost reasons
    (Webhook) Servers22/

    Apart from the above, there might be special cases where these rules do not apply, for example:

    • istio-ingressgateway is scaled horizontally, hence the above numbers are the minimum values.
    • nginx-ingress-controller in the seed cluster is used to advertise all shoot observability endpoints, so due to performance reasons it runs with 2 replicas at all times. In the future, this component might disappear in favor of the istio-ingressgateway anyways.
  • Topology Spread Constraints

    When the component has >= 2 replicas …

    • … then it should also have a topologySpreadConstraint, ensuring the replicas are spread over the nodes:

        - maxSkew: 1
          whenUnsatisfiable: ScheduleAnyway
          matchLabels: ...

      Hence, the node spread is done on best-effort basis only.

    • … and the seed cluster has >= 2 zones, then the component should also have a second topologySpreadConstraint, ensuring the replicas are spread over the zones:

        - maxSkew: 1
          whenUnsatisfiable: DoNotSchedule
          matchLabels: ...

According to these conventions, even seed clusters with only one availability zone try to be highly available “as good as possible” by spreading the replicas across multiple nodes. Hence, while such seed clusters obviously cannot handle zone outages, they can at least handle node failures.

Shoot Clusters

The Shoot specification allows configuring “high availability” as well as the failure tolerance type for the control plane components, see Highly Available Shoot Control Plane for details.

Regarding the seed cluster selection, the only constraint is that shoot clusters with failure tolerance type zone are only allowed to run on seed clusters with at least three zones. All other shoot clusters (non-HA or those with failure tolerance type node) can run on seed clusters with any number of zones.

Control Plane Components

All control plane components should respect the following conventions:

  • Replica Counts

    Component Typew/o HAw/ HA (node)w/ HA (zone)Comment
    Observability (Monitoring, Logging)111Downtimes accepted due to cost reasons
    (Webhook) Servers222/

    Apart from the above, there might be special cases where these rules do not apply, for example:

    • etcd is a server, though the most critical component of a cluster requiring a quorum to survive failures. Hence, it should have 3 replicas even when the failure tolerance is node only.
    • kube-apiserver is scaled horizontally, hence the above numbers are the minimum values (even when the shoot cluster is not HA, there might be multiple replicas).
  • Topology Spread Constraints

    When the component has >= 2 replicas …

    • … then it should also have a topologySpreadConstraint ensuring the replicas are spread over the nodes:

        - maxSkew: 1
          whenUnsatisfiable: ScheduleAnyway
          matchLabels: ...

      Hence, the node spread is done on best-effort basis only.

      However, if the shoot cluster has defined a failure tolerance type, the whenUnsafisfiable field should be set to DoNotSchedule.

    • … and the failure tolerance type of the shoot cluster is zone, then the component should also have a second topologySpreadConstraint ensuring the replicas are spread over the zones:

        - maxSkew: 1
          whenUnsatisfiable: DoNotSchedule
          matchLabels: ...
  • Node Affinity

    The gardenlet annotates the shoot namespace in the seed cluster with the annotation.

    • If the shoot cluster is non-HA or has failure tolerance type node, then the value will be always exactly one zone (e.g.,
    • If the shoot cluster has failure tolerance type zone, then the value will always contain exactly three zones (e.g.,,europe-1b,europe-1c).

    For backwards-compatibility, this annotation might contain multiple zones for shoot clusters created before gardener/gardener@v1.60 and not having failure tolerance type zone. This is because their volumes might already exist in multiple zones, hence pinning them to only one zone would not work.

    Hence, in case this annotation is present, the components should have the following node affinity:

            - matchExpressions:
              - key:
                operator: In
                - europe-1a
              # - ...

    This is to ensure all pods are running in the same (set of) availability zone(s) such that cross-zone network traffic is avoided as much as possible (such traffic is typically charged by the underlying infrastructure provider).

System Components

The availability of system components is independent of the control plane since they run on the shoot worker nodes while the control plane components run on the seed worker nodes (for more information, see the Kubernetes architecture overview). Hence, it only depends on the number of availability zones configured in the shoot worker pools via .spec.provider.workers[].zones. Concretely, the highest number of zones of a worker pool with systemComponents.allow=true is considered.

All system components should respect the following conventions:

  • Replica Counts

    Component Type1 or 2 Zones>= 3 Zones
    (Webhook) Servers22

    Apart from the above, there might be special cases where these rules do not apply, for example:

    • coredns is scaled horizontally (today), hence the above numbers are the minimum values (possibly, scaling these components vertically may be more appropriate, but that’s unrelated to the HA subject matter).
    • Optional addons like nginx-ingress or kubernetes-dashboard are only provided on best-effort basis for evaluation purposes, hence they run with 1 replica at all times.
  • Topology Spread Constraints

    When the component has >= 2 replicas …

    • … then it should also have a topologySpreadConstraint ensuring the replicas are spread over the nodes:

        - maxSkew: 1
          whenUnsatisfiable: ScheduleAnyway
          matchLabels: ...

      Hence, the node spread is done on best-effort basis only.

    • … and the cluster has >= 2 zones, then the component should also have a second topologySpreadConstraint ensuring the replicas are spread over the zones:

        - maxSkew: 1
          whenUnsatisfiable: DoNotSchedule
          matchLabels: ...

Convenient Application of These Rules

According to above scenarios and conventions, the replicas, topologySpreadConstraints or affinity settings of the deployed components might need to be adapted.

In order to apply those conveniently and easily for developers, Gardener installs a mutating webhook into both seed and shoot clusters which reacts on Deployments and StatefulSets deployed to namespaces with the label set.

The following actions have to be taken by developers:

  1. Check if components are prepared to run concurrently with multiple replicas, e.g. controllers usually use leader election to achieve this.

  2. All components should be generally equipped with PodDisruptionBudgets with .spec.maxUnavailable=1:

  maxUnavailable: 1
    matchLabels: ...
  1. Add the label to deployments or statefulsets, as well as optionally involved horizontalpodautoscalers or HVPAs where the following two values are possible:
  • controller
  • server

Type server is also preferred if a component is a controller and (webhook) server at the same time.

You can read more about the webhook’s internals in High Availability Config.

gardenlet Internals

Make sure you have read the above document about the webhook internals before continuing reading this section.

Seed Controller

The gardenlet performs the following changes on all namespaces running seed system components:

  • adds the label
  • adds the annotation<zones>, where <zones> is the list provided in .spec.provider.zones[] in the Seed specification.

Note that neither the, nor the annotations are set, hence the node affinity would never be touched by the webhook.

The only exception to this rule are the istio ingress gateway namespaces. This includes the default istio ingress gateway when SNI is enabled, as well as analogous namespaces for exposure classes and zone-specific istio ingress gateways. Those namespaces will additionally be annotated with set to true, resulting in the node affinities and the topology spread constraints being set. The replicas are not touched, as the istio ingress gateways are scaled by a horizontal autoscaler instance.

Shoot Controller

Control Plane

The gardenlet performs the following changes on the namespace running the shoot control plane components:

  • adds the label This makes the webhook mutate the replica count and the topology spread constraints.
  • adds the annotation with value equal to .spec.controlPlane.highAvailability.failureTolerance.type (or "", if .spec.controlPlane.highAvailability=nil). This makes the webhook mutate the node affinity according to the specified zone(s).
  • adds the annotation<zones>, where <zones> is a …
    • … random zone chosen from the .spec.provider.zones[] list in the Seed specification (always only one zone (even if there are multiple available in the seed cluster)) in case the Shoot has no HA setting (i.e., spec.controlPlane.highAvailability=nil) or when the Shoot has HA setting with failure tolerance type node.
    • … list of three randomly chosen zones from the .spec.provider.zones[] list in the Seed specification in case the Shoot has HA setting with failure tolerance type zone.

System Components

The gardenlet performs the following changes on all namespaces running shoot system components:

  • adds the label This makes the webhook mutate the replica count and the topology spread constraints.
  • adds the annotation<zones> where <zones> is the merged list of zones provided in .zones[] with systemComponents.allow=true for all worker pools in .spec.provider.workers[] in the Shoot specification.

Note that neither the, nor the annotations are set, hence the node affinity would never be touched by the webhook.

6 - Kubernetes Clients

Kubernetes Clients in Gardener

This document aims at providing a general developer guideline on different aspects of using Kubernetes clients in a large-scale distributed system and project like Gardener. The points included here are not meant to be consulted as absolute rules, but rather as general rules of thumb that allow developers to get a better feeling about certain gotchas and caveats. It should be updated with lessons learned from maintaining the project and running Gardener in production.


Please familiarize yourself with the following basic Kubernetes API concepts first, if you’re new to Kubernetes. A good understanding of these basics will help you better comprehend the following document.

Client Types: Client-Go, Generated, Controller-Runtime

For historical reasons, you will find different kinds of Kubernetes clients in Gardener:

Client-Go Clients

client-go is the default/official client for talking to the Kubernetes API in Golang. It features the so called “client sets” for all built-in Kubernetes API groups and versions (e.g. v1 (aka core/v1), apps/v1, etc.). client-go clients are generated from the built-in API types using client-gen and are composed of interfaces for every known API GroupVersionKind. A typical client-go usage looks like this:

var (
  ctx        context.Context
  c          kubernetes.Interface // ""
  deployment *appsv1.Deployment   // ""

updatedDeployment, err := c.AppsV1().Deployments("default").Update(ctx, deployment, metav1.UpdateOptions{})

Important characteristics of client-go clients:

  • clients are specific to a given API GroupVersionKind, i.e., clients are hard-coded to corresponding API-paths (don’t need to use the discovery API to map GVK to a REST endpoint path).
  • client’s don’t modify the passed in-memory object (e.g. deployment in the above example). Instead, they return a new in-memory object.
    This means that controllers have to continue working with the new in-memory object or overwrite the shared object to not lose any state updates.

Generated Client Sets for Gardener APIs

Gardener’s APIs extend the Kubernetes API by registering an extension API server (in the garden cluster) and CustomResourceDefinitions (on Seed clusters), meaning that the Kubernetes API will expose additional REST endpoints to manage Gardener resources in addition to the built-in API resources. In order to talk to these extended APIs in our controllers and components, client-gen is used to generate client-go-style clients to pkg/client/{core,extensions,seedmanagement,...}.

Usage of these clients is equivalent to client-go clients, and the same characteristics apply. For example:

var (
  ctx   context.Context
  c     gardencoreclientset.Interface // ""
  shoot *gardencorev1beta1.Shoot      // ""

updatedShoot, err := c.CoreV1beta1().Shoots("garden-my-project").Update(ctx, shoot, metav1.UpdateOptions{})

Controller-Runtime Clients

controller-runtime is a Kubernetes community project (kubebuilder subproject) for building controllers and operators for custom resources. Therefore, it features a generic client that follows a different approach and does not rely on generated client sets. Instead, the client can be used for managing any Kubernetes resources (built-in or custom) homogeneously. For example:

var (
  ctx        context.Context
  c          client.Client            // ""
  deployment *appsv1.Deployment       // ""
  shoot      *gardencorev1beta1.Shoot // ""

err := c.Update(ctx, deployment)
// or
err = c.Update(ctx, shoot)

A brief introduction to controller-runtime and its basic constructs can be found at the official Go documentation.

Important characteristics of controller-runtime clients:

  • The client functions take a generic client.Object or client.ObjectList value. These interfaces are implemented by all Golang types, that represent Kubernetes API objects or lists respectively which can be interacted with via usual API requests. [1]
  • The client first consults a runtime.Scheme (configured during client creation) for recognizing the object’s GroupVersionKind (this happens on the client-side only).
    A runtime.Scheme is basically a registry for Golang API types, defaulting and conversion functions. Schemes are usually provided per GroupVersion (see this example for apps/v1) and can be combined to one single scheme for further usage (example). In controller-runtime clients, schemes are used only for mapping a typed API object to its GroupVersionKind.
  • It then consults a meta.RESTMapper (also configured during client creation) for mapping the GroupVersionKind to a RESTMapping, which contains the GroupVersionResource and Scope (namespaced or cluster-scoped). From these values, the client can unambiguously determine the REST endpoint path of the corresponding API resource. For instance: appsv1.DeploymentList is available at /apis/apps/v1/deployments or /apis/apps/v1/namespaces/<namespace>/deployments respectively.
    • There are different RESTMapper implementations, but generally they are talking to the API server’s discovery API for retrieving RESTMappings for all API resources known to the API server (either built-in, registered via API extension or CustomResourceDefinitions).
    • The default implementation of a controller-runtime (which Gardener uses as well) is the dynamic RESTMapper. It caches discovery results (i.e. RESTMappings) in-memory and only re-discovers resources from the API server when a client tries to use an unknown GroupVersionKind, i.e., when it encounters a No{Kind,Resource}MatchError.
  • The client writes back results from the API server into the passed in-memory object.
    • This means that controllers don’t have to worry about copying back the results and should just continue to work on the given in-memory object.
    • This is a nice and flexible pattern, and helper functions should try to follow it wherever applicable. Meaning, if possible accept an object param, pass it down to clients and keep working on the same in-memory object instead of creating a new one in your helper function.
    • The benefit is that you don’t lose updates to the API object and always have the last-known state in memory. Therefore, you don’t have to read it again, e.g., for getting the current resourceVersion when working with optimistic locking, and thus minimize the chances for running into conflicts.
    • However, controllers must not use the same in-memory object concurrently in multiple goroutines. For example, decoding results from the API server in multiple goroutines into the same maps (e.g., labels, annotations) will cause panics because of “concurrent map writes”. Also, reading from an in-memory API object in one goroutine while decoding into it in another goroutine will yield non-atomic reads, meaning data might be corrupt and represent a non-valid/non-existing API object.
    • Therefore, if you need to use the same in-memory object in multiple goroutines concurrently (e.g., shared state), remember to leverage proper synchronization techniques like channels, mutexes, atomic.Value and/or copy the object prior to use. The average controller however, will not need to share in-memory API objects between goroutines, and it’s typically an indicator that the controller’s design should be improved.
  • The client decoder erases the object’s TypeMeta (apiVersion and kind fields) after retrieval from the API server, see kubernetes/kubernetes#80609, kubernetes-sigs/controller-runtime#1517. Unstructured and metadata-only requests objects are an exception to this because the contained TypeMeta is the only way to identify the object’s type. Because of this behavior, obj.GetObjectKind().GroupVersionKind() is likely to return an empty GroupVersionKind. I.e., you must not rely on TypeMeta being set or GetObjectKind() to return something usable.
    If you need to identify an object’s GroupVersionKind, use a scheme and its ObjectKinds function instead (or the helper function apiutil.GVKForObject). This is not specific to controller-runtime clients and applies to client-go clients as well.

[1] Other lower level, config or internal API types (e.g., such as AdmissionReview) don’t implement client.Object. However, you also can’t interact with such objects via the Kubernetes API and thus also not via a client, so this can be disregarded at this point.

Metadata-Only Clients

Additionally, controller-runtime clients can be used to easily retrieve metadata-only objects or lists. This is useful for efficiently checking if at least one object of a given kind exists, or retrieving metadata of an object, if one is not interested in the rest (e.g., spec/status).
The Accept header sent to the API server then contains application/json;as=PartialObjectMetadataList;;v=v1, which makes the API server only return metadata of the retrieved object(s). This saves network traffic and CPU/memory load on the API server and client side. If the client fully lists all objects of a given kind including their spec/status, the resulting list can be quite large and easily exceed the controllers available memory. That’s why it’s important to carefully check if a full list is actually needed, or if metadata-only list can be used instead.

For example:

var (
  ctx       context.Context
  c         client.Client                         // ""
  shootList = &metav1.PartialObjectMetadataList{} // ""

if err := c.List(ctx, shootList, client.InNamespace("garden-my-project"), client.Limit(1)); err != nil {
  return err

if len(shootList.Items) > 0 {
  // project has at least one shoot
} else {
  // project doesn't have any shoots

Gardener’s Client Collection, ClientMaps

The Gardener codebase has a collection of clients (kubernetes.Interface), which can return all the above mentioned client types. Additionally, it contains helpers for rendering and applying helm charts (ChartRender, ChartApplier) and retrieving the API server’s version (Version).
Client sets are managed by so called ClientMaps, which are a form of registry for all client set for a given type of cluster, i.e., Garden, Seed and Shoot. ClientMaps manage the whole lifecycle of clients: they take care of creating them if they don’t exist already, running their caches, refreshing their cached server version and invalidating them when they are no longer needed.

var (
  ctx   context.Context
  cm    clientmap.ClientMap // ""
  shoot *gardencorev1beta1.Shoot

cs, err := cm.GetClient(ctx, keys.ForShoot(shoot)) // kubernetes.Interface
if err != nil {
  return err

c := cs.Client() // client.Client

The client collection mainly exist for historical reasons (there used to be a lot of code using the client-go style clients). However, Gardener is in the process of moving more towards controller-runtime and only using their clients, as they provide many benefits and are much easier to use. Also, gardener/gardener#4251 aims at refactoring our controller and admission components to native controller-runtime components.

⚠️ Please always prefer controller-runtime clients over other clients when writing new code or refactoring existing code.

Cache Types: Informers, Listers, Controller-Runtime Caches

Similar to the different types of client(set)s, there are also different kinds of Kubernetes client caches. However, all of them are based on the same concept: Informers. An Informer is a watch-based cache implementation, meaning it opens watch connections to the API server and continuously updates cached objects based on the received watch events (ADDED, MODIFIED, DELETED). Informers offer to add indices to the cache for efficient object lookup (e.g., by name or labels) and to add EventHandlers for the watch events. The latter is used by controllers to fill queues with objects that should be reconciled on watch events.

Informers are used in and created via several higher-level constructs:

SharedInformerFactories, Listers

The generated clients (built-in as well as extended) feature a SharedInformerFactory for every API group, which can be used to create and retrieve Informers for all GroupVersionKinds. Similarly, it can be used to retrieve Listers that allow getting and listing objects from the Informer’s cache. However, both of these constructs are only used for historical reasons, and we are in the process of migrating away from them in favor of cached controller-runtime clients (see gardener/gardener#2414, gardener/gardener#2822). Thus, they are described only briefly here.

Important characteristics of Listers:

  • Objects read from Informers and Listers can always be slightly out-out-date (i.e., stale) because the client has to first observe changes to API objects via watch events (which can intermittently lag behind by a second or even more).
  • Thus, don’t make any decisions based on data read from Listers if the consequences of deciding wrongfully based on stale state might be catastrophic (e.g. leaking infrastructure resources). In such cases, read directly from the API server via a client instead.
  • Objects retrieved from Informers or Listers are pointers to the cached objects, so they must not be modified without copying them first, otherwise the objects in the cache are also modified.

Controller-Runtime Caches

controller-runtime features a cache implementation that can be used equivalently as their clients. In fact, it implements a subset of the client.Client interface containing the Get and List functions. Under the hood, a cache.Cache dynamically creates Informers (i.e., opens watches) for every object GroupVersionKind that is being retrieved from it.

Note that the underlying Informers of a controller-runtime cache (cache.Cache) and the ones of a SharedInformerFactory (client-go) are not related in any way. Both create Informers and watch objects on the API server individually. This means that if you read the same object from different cache implementations, you may receive different versions of the object because the watch connections of the individual Informers are not synced.

⚠️ Because of this, controllers/reconcilers should get the object from the same cache in the reconcile loop, where the EventHandler was also added to set up the controller. For example, if a SharedInformerFactory is used for setting up the controller then read the object in the reconciler from the Lister instead of from a cached controller-runtime client.

By default, the client.Client created by a controller-runtime Manager is a DelegatingClient. It delegates Get and List calls to a Cache, and all other calls to a client that talks directly to the API server. Exceptions are requests with *unstructured.Unstructured objects and object kinds that were configured to be excluded from the cache in the DelegatingClient.

ℹ️ kubernetes.Interface.Client() returns a DelegatingClient that uses the cache returned from kubernetes.Interface.Cache() under the hood. This means that all Client() usages need to be ready for cached clients and should be able to cater with stale cache reads.

Important characteristics of cached controller-runtime clients:

  • Like for Listers, objects read from a controller-runtime cache can always be slightly out of date. Hence, don’t base any important decisions on data read from the cache (see above).
  • In contrast to Listers, controller-runtime caches fill the passed in-memory object with the state of the object in the cache (i.e., they perform something like a “deep copy into”). This means that objects read from a controller-runtime cache can safely be modified without unintended side effects.
  • Reading from a controller-runtime cache or a cached controller-runtime client implicitly starts a watch for the given object kind under the hood. This has important consequences:
    • Reading a given object kind from the cache for the first time can take up to a few seconds depending on size and amount of objects as well as API server latency. This is because the cache has to do a full list operation and wait for an initial watch sync before returning results.
    • ⚠️ Controllers need appropriate RBAC permissions for the object kinds they retrieve via cached clients (i.e., list and watch).
    • ⚠️ By default, watches started by a controller-runtime cache are cluster-scoped, meaning it watches and caches objects across all namespaces. Thus, be careful which objects to read from the cache as it might significantly increase the controller’s memory footprint.
  • There is no interaction with the cache on writing calls (Create, Update, Patch and Delete), see below.

Uncached objects, filtered caches, APIReaders:

In order to allow more granular control over which object kinds should be cached and which calls should bypass the cache, controller-runtime offers a few mechanisms to further tweak the client/cache behavior:

  • When creating a DelegatingClient, certain object kinds can be configured to always be read directly from the API instead of from the cache. Note that this does not prevent starting a new Informer when retrieving them directly from the cache.
  • Watches can be restricted to a given (set of) namespace(s) by using cache.MultiNamespacedCacheBuilder or setting cache.Options.Namespace.
  • Watches can be filtered (e.g., by label) per object kind by configuring cache.Options.SelectorsByObject on creation of the cache.
  • Retrieving metadata-only objects or lists from a cache results in a metadata-only watch/cache for that object kind.
  • The APIReader can be used to always talk directly to the API server for a given Get or List call (use with care and only as a last resort!).

To Cache or Not to Cache

Although watch-based caches are an important factor for the immense scalability of Kubernetes, it definitely comes at a price (mainly in terms of memory consumption). Thus, developers need to be careful when introducing new API calls and caching new object kinds. Here are some general guidelines on choosing whether to read from a cache or not:

  • Always try to use the cache wherever possible and make your controller able to tolerate stale reads.
    • Leverage optimistic locking: use deterministic naming for objects you create (this is what the Deployment controller does [2]).
    • Leverage optimistic locking / concurrency control of the API server: send updates/patches with the last-known resourceVersion from the cache (see below). This will make the request fail, if there were concurrent updates to the object (conflict error), which indicates that we have operated on stale data and might have made wrong decisions. In this case, let the controller handle the error with exponential backoff. This will make the controller eventually consistent.
    • Track the actions you took, e.g., when creating objects with generateName (this is what the ReplicaSet controller does [3]). The actions can be tracked in memory and repeated if the expected watch events don’t occur after a given amount of time.
    • Always try to write controllers with the assumption that data will only be eventually correct and can be slightly out of date (even if read directly from the API server!).
    • If there is already some other code that needs a cache (e.g., a controller watch), reuse it instead of doing extra direct reads.
    • Don’t read an object again if you just sent a write request. Write requests (Create, Update, Patch and Delete) don’t interact with the cache. Hence, use the current state that the API server returned (filled into the passed in-memory object), which is basically a “free direct read” instead of reading the object again from a cache, because this will probably set back the object to an older resourceVersion.
  • If you are concerned about the impact of the resulting cache, try to minimize that by using filtered or metadata-only watches.
  • If watching and caching an object type is not feasible, for example because there will be a lot of updates, and you are only interested in the object every ~5m, or because it will blow up the controllers memory footprint, fallback to a direct read. This can either be done by disabling caching the object type generally or doing a single request via an APIReader. In any case, please bear in mind that every direct API call results in a quorum read from etcd, which can be costly in a heavily-utilized cluster and impose significant scalability limits. Thus, always try to minimize the impact of direct calls by filtering results by namespace or labels, limiting the number of results and/or using metadata-only calls.

[2] The Deployment controller uses the pattern <deployment-name>-<podtemplate-hash> for naming ReplicaSets. This means, the name of a ReplicaSet it tries to create/update/delete at any given time is deterministically calculated based on the Deployment object. By this, it is insusceptible to stale reads from its ReplicaSets cache.

[3] In simple terms, the ReplicaSet controller tracks its CREATE pod actions as follows: when creating new Pods, it increases a counter of expected ADDED watch events for the corresponding ReplicaSet. As soon as such events arrive, it decreases the counter accordingly. It only creates new Pods for a given ReplicaSet once all expected events occurred (counter is back to zero) or a timeout has occurred. This way, it prevents creating more Pods than desired because of stale cache reads and makes the controller eventually consistent.

Conflicts, Concurrency Control, and Optimistic Locking

Every Kubernetes API object contains the metadata.resourceVersion field, which identifies an object’s version in the backing data store, i.e., etcd. Every write to an object in etcd results in a newer resourceVersion. This field is mainly used for concurrency control on the API server in an optimistic locking fashion, but also for efficient resumption of interrupted watch connections.

Optimistic locking in the Kubernetes API sense means that when a client wants to update an API object, then it includes the object’s resourceVersion in the request to indicate the object’s version the modifications are based on. If the resourceVersion in etcd has not changed in the meantime, the update request is accepted by the API server and the updated object is written to etcd. If the resourceVersion sent by the client does not match the one of the object stored in etcd, there were concurrent modifications to the object. Consequently, the request is rejected with a conflict error (status code 409, API reason Conflict), for example:

  "kind": "Status",
  "apiVersion": "v1",
  "metadata": {},
  "status": "Failure",
  "message": "Operation cannot be fulfilled on configmaps \"foo\": the object has been modified; please apply your changes to the latest version and try again",
  "reason": "Conflict",
  "details": {
    "name": "foo",
    "kind": "configmaps"
  "code": 409

This concurrency control is an important mechanism in Kubernetes as there are typically multiple clients acting on API objects at the same time (humans, different controllers, etc.). If a client receives a conflict error, it should read the object’s latest version from the API server, make the modifications based on the newest changes, and retry the update. The reasoning behind this is that a client might choose to make different decisions based on the concurrent changes made by other actors compared to the outdated version that it operated on.

Important points about concurrency control and conflicts:

  • The resourceVersion field carries a string value and clients must not assume numeric values (the type and structure of versions depend on the backing data store). This means clients may compare resourceVersion values to detect whether objects were changed. But they must not compare resourceVersions to figure out which one is newer/older, i.e., no greater/less-than comparisons are allowed.
  • By default, update calls (e.g. via client-go and controller-runtime clients) use optimistic locking as the passed in-memory usually object contains the latest resourceVersion known to the controller, which is then also sent to the API server.
  • API servers can also choose to accept update calls without optimistic locking (i.e., without a resourceVersion in the object’s metadata) for any given resource. However, sending update requests without optimistic locking is strongly discouraged, as doing so overwrites the entire object, discarding any concurrent changes made to it.
  • On the other side, patch requests can always be executed either with or without optimistic locking, by (not) including the resourceVersion in the patched object’s metadata. Sending patch requests without optimistic locking might be safe and even desirable as a patch typically updates only a specific section of the object. However, there are also situations where patching without optimistic locking is not safe (see below).

Don’t Retry on Conflict

Similar to how a human would typically handle a conflict error, there are helper functions implementing RetryOnConflict-semantics, i.e., try an update call, then re-read the object if a conflict occurs, apply the modification again and retry the update. However, controllers should generally not use RetryOnConflict-semantics. Instead, controllers should abort their current reconciliation run and let the queue handle the conflict error with exponential backoff. The reasoning behind this is that a conflict error indicates that the controller has operated on stale data and might have made wrong decisions earlier on in the reconciliation. When using a helper function that implements RetryOnConflict-semantics, the controller doesn’t check which fields were changed and doesn’t revise its previous decisions accordingly. Instead, retrying on conflict basically just ignores any conflict error and blindly applies the modification.

To properly solve the conflict situation, controllers should immediately return with the error from the update call. This will cause retries with exponential backoff so that the cache has a chance to observe the latest changes to the object. In a later run, the controller will then make correct decisions based on the newest version of the object, not run into conflict errors, and will then be able to successfully reconcile the object. This way, the controller becomes eventually consistent.

The other way to solve the situation is to modify objects without optimistic locking in order to avoid running into a conflict in the first place (only if this is safe). This can be a preferable solution for controllers with long-running reconciliations (which is actually an anti-pattern but quite unavoidable in some of Gardener’s controllers). Aborting the entire reconciliation run is rather undesirable in such cases, as it will add a lot of unnecessary waiting time for end users and overhead in terms of compute and network usage.

However, in any case, retrying on conflict is probably not the right option to solve the situation (there are some correct use cases for it, though, they are very rare). Hence, don’t retry on conflict.

To Lock or Not to Lock

As explained before, conflicts are actually important and prevent clients from doing wrongful concurrent updates. This means that conflicts are not something we generally want to avoid or ignore. However, in many cases controllers are exclusive owners of the fields they want to update and thus it might be safe to run without optimistic locking.

For example, the gardenlet is the exclusive owner of the spec section of the Extension resources it creates on behalf of a Shoot (e.g., the Infrastructure resource for creating VPC, etc.). Meaning, it knows the exact desired state and no other actor is supposed to update the Infrastructure’s spec fields. When the gardenlet now updates the Infrastructures spec section as part of the Shoot reconciliation, it can simply issue a PATCH request that only updates the spec and runs without optimistic locking. If another controller concurrently updated the object in the meantime (e.g., the status section), the resourceVersion got changed, which would cause a conflict error if running with optimistic locking. However, concurrent status updates would not change the gardenlet’s mind on the desired spec of the Infrastructure resource as it is determined only by looking at the Shoot’s specification. If the spec section was changed concurrently, it’s still fine to overwrite it because the gardenlet should reconcile the spec back to its desired state.

Generally speaking, if a controller is the exclusive owner of a given set of fields and they are independent of concurrent changes to other fields in that object, it can patch these fields without optimistic locking. This might ignore concurrent changes to other fields or blindly overwrite changes to the same fields, but this is fine if the mentioned conditions apply. Obviously, this applies only to patch requests that modify only a specific set of fields but not to update requests that replace the entire object.

In such cases, it’s even desirable to run without optimistic locking as it will be more performant and save retries. If certain requests are made with high frequency and have a good chance of causing conflicts, retries because of optimistic locking can cause a lot of additional network traffic in a large-scale Gardener installation.

Updates, Patches, Server-Side Apply

There are different ways of modifying Kubernetes API objects. The following snippet demonstrates how to do a given modification with the most frequently used options using a controller-runtime client:

var (
  ctx   context.Context
  c     client.Client
  shoot *gardencorev1beta1.Shoot

// update
shoot.Spec.Kubernetes.Version = "1.22"
err := c.Update(ctx, shoot)

// json merge patch
patch := client.MergeFrom(shoot.DeepCopy())
shoot.Spec.Kubernetes.Version = "1.22"
err = c.Patch(ctx, shoot, patch)

// strategic merge patch
patch = client.StrategicMergeFrom(shoot.DeepCopy())
shoot.Spec.Kubernetes.Version = "1.22"
err = c.Patch(ctx, shoot, patch)

Important characteristics of the shown request types:

  • Update requests always send the entire object to the API server and update all fields accordingly. By default, optimistic locking is used (resourceVersion is included).
  • Both patch types run without optimistic locking by default. However, it can be enabled explicitly if needed:
    // json merge patch + optimistic locking
    patch := client.MergeFromWithOptions(shoot.DeepCopy(), client.MergeFromWithOptimisticLock{})
    // ...
    // strategic merge patch + optimistic locking
    patch = client.StrategicMergeFrom(shoot.DeepCopy(), client.MergeFromWithOptimisticLock{})
    // ...
  • Patch requests only contain the changes made to the in-memory object between the copy passed to client.*MergeFrom and the object passed to Client.Patch(). The diff is calculated on the client-side based on the in-memory objects only. This means that if in the meantime some fields were changed on the API server to a different value than the one on the client-side, the fields will not be changed back as long as they are not changed on the client-side as well (there will be no diff in memory).
  • Thus, if you want to ensure a given state using patch requests, always read the object first before patching it, as there will be no diff otherwise, meaning the patch will be empty. For more information, see gardener/gardener#4057 and the comments in gardener/gardener#4027.
  • Also, always send updates and patch requests even if your controller hasn’t made any changes to the current state on the API server. I.e., don’t make any optimization for preventing empty patches or no-op updates. There might be mutating webhooks in the system that will modify the object and that rely on update/patch requests being sent (even if they are no-op). Gardener’s extension concept makes heavy use of mutating webhooks, so it’s important to keep this in mind.
  • JSON merge patches always replace lists as a whole and don’t merge them. Keep this in mind when operating on lists with merge patch requests. If the controller is the exclusive owner of the entire list, it’s safe to run without optimistic locking. Though, if you want to prevent overwriting concurrent changes to the list or its items made by other actors (e.g., additions/removals to the metadata.finalizers list), enable optimistic locking.
  • Strategic merge patches are able to make more granular modifications to lists and their elements without replacing the entire list. It uses Golang struct tags of the API types to determine which and how lists should be merged. See Update API Objects in Place Using kubectl patch or the strategic merge patch documentation for more in-depth explanations and comparison with JSON merge patches. With this, controllers might be able to issue patch requests for individual list items without optimistic locking, even if they are not exclusive owners of the entire list. Remember to check the patchStrategy and patchMergeKey struct tags of the fields you want to modify before blindly adding patch requests without optimistic locking.
  • Strategic merge patches are only supported by built-in Kubernetes resources and custom resources served by Extension API servers. Strategic merge patches are not supported by custom resources defined by CustomResourceDefinitions (see this comparison). In that case, fallback to JSON merge patches.
  • Server-side Apply is yet another mechanism to modify API objects, which is supported by all API resources (in newer Kubernetes versions). However, it has a few problems and more caveats preventing us from using it in Gardener at the time of writing. See gardener/gardener#4122 for more details.

Generally speaking, patches are often the better option compared to update requests because they can save network traffic, encoding/decoding effort, and avoid conflicts under the presented conditions. If choosing a patch type, consider which type is supported by the resource you’re modifying and what will happen in case of a conflict. Consider whether your modification is safe to run without optimistic locking. However, there is no simple rule of thumb on which patch type to choose.

On Helper Functions

Here is a note on some helper functions, that should be avoided and why:

controllerutil.CreateOrUpdate does a basic get, mutate and create or update call chain, which is often used in controllers. We should avoid using this helper function in Gardener, because it is likely to cause conflicts for cached clients and doesn’t send no-op requests if nothing was changed, which can cause problems because of the heavy use of webhooks in Gardener extensions (see above). That’s why usage of this function was completely replaced in gardener/gardener#4227 and similar PRs.

controllerutil.CreateOrPatch is similar to CreateOrUpdate but does a patch request instead of an update request. It has the same drawback as CreateOrUpdate regarding no-op updates. Also, controllers can’t use optimistic locking or strategic merge patches when using CreateOrPatch. Another reason for avoiding use of this function is that it also implicitly patches the status section if it was changed, which is confusing for others reading the code. To accomplish this, the func does some back and forth conversion, comparison and checks, which are unnecessary in most of our cases and simply wasted CPU cycles and complexity we want to avoid.

There were some Try{Update,UpdateStatus,Patch,PatchStatus} helper functions in Gardener that were already removed by gardener/gardener#4378 but are still used in some extension code at the time of writing. The reason for eliminating these functions is that they implement RetryOnConflict-semantics. Meaning, they first get the object, mutate it, then try to update and retry if a conflict error occurs. As explained above, retrying on conflict is a controller anti-pattern and should be avoided in almost every situation. The other problem with these functions is that they read the object first from the API server (always do a direct call), although in most cases we already have a recent version of the object at hand. So, using this function generally does unnecessary API calls and therefore causes unwanted compute and network load.

For the reasons explained above, there are similar helper functions that accomplish similar things but address the mentioned drawbacks: controllerutils.{GetAndCreateOrMergePatch,GetAndCreateOrStrategicMergePatch}. These can be safely used as replacements for the aforementioned helper funcs. If they are not fitting for your use case, for example because you need to use optimistic locking, just do the appropriate calls in the controller directly.

These resources are only partially related to the topics covered in this doc, but might still be interesting for developer seeking a deeper understanding of Kubernetes API machinery, architecture and foundational concepts.

7 - Local Setup


Conceptually, all Gardener components are designed to run as a Pod inside a Kubernetes cluster. The Gardener API server extends the Kubernetes API via the user-aggregated API server concepts. However, if you want to develop it, you may want to work locally with the Gardener without building a Docker image and deploying it to a cluster each and every time. That means that the Gardener runs outside a Kubernetes cluster which requires providing a Kubeconfig in your local filesystem and point the Gardener to it when starting it (see below).

Further details can be found in

  1. Principles of Kubernetes, and its components
  2. Kubernetes Development Guide
  3. Architecture of Gardener

This guide is split into three main parts:

Preparing the Setup

[macOS only] Installing homebrew

The copy-paste instructions in this guide are designed for macOS and use the package manager Homebrew.

On macOS run

/bin/bash -c "$(curl -fsSL"

Installing git

We use git as VCS which you need to install. On macOS run

brew install git

For other OS, please check the Git installation documentation.

Installing Go

Install the latest version of Go. On macOS run

brew install go

For other OS, please check Go installation documentation.

Installing kubectl

Install kubectl. Please make sure that the version of kubectl is at least v1.20.x. On macOS run

brew install kubernetes-cli

For other OS, please check the kubectl installation documentation.

Installing Docker

You need to have docker installed and running. On macOS run

brew install --cask docker

For other OS please check the docker installation documentation.

Installing iproute2

iproute2 provides a collection of utilities for network administration and configuration. On macOS run

brew install iproute2mac

Installing jq

brew install jq

Installing GNU Parallel

GNU Parallel is a shell tool for executing jobs in parallel, used by the code generation scripts (make generate). On macOS run

brew install parallel

[macOS only] Install GNU Core Utilities

When running on macOS, install the GNU core utilities and friends:

brew install coreutils gnu-sed gnu-tar grep

This will create symbolic links for the GNU utilities with g prefix in /usr/local/bin, e.g., gsed or gbase64. To allow using them without the g prefix please put /usr/local/opt/coreutils/libexec/gnubin etc., at the beginning of your PATH environment variable, e.g., export PATH=/usr/local/opt/coreutils/libexec/gnubin:$PATH (brew will print out instructions for each installed formula).

export PATH=/usr/local/opt/coreutils/libexec/gnubin:$PATH
export PATH=/usr/local/opt/gnu-sed/libexec/gnubin:$PATH
export PATH=/usr/local/opt/gnu-tar/libexec/gnubin:$PATH
export PATH=/usr/local/opt/grep/libexec/gnubin:$PATH

[Windows Only] WSL2

Apart from Linux distributions and macOS, the local gardener setup can also run on the Windows Subsystem for Linux 2.

While WSL1, plain docker for Windows and various Linux distributions and local Kubernetes environments may be supported, this setup was verified with:

The Gardener repository and all the above-mentioned tools (git, golang, kubectl, …) should be installed in your WSL2 distro, according to the distribution-specific Linux installation instructions.

Start Gardener Locally

Get the Sources

Clone the repository from GitHub into your $GOPATH.

mkdir -p $(go env GOPATH)/src/
cd $(go env GOPATH)/src/
git clone
cd gardener

Note: Gardener is using Go modules and cloning the repository into $GOPATH is not a hard requirement. However it is still recommended to clone into $GOPATH because does not work yet outside of $GOPATH - kubernetes/kubernetes#86753.

Start the Gardener

ℹ️ In the following guide, you have to define the configuration (CloudProfiles, SecretBindings, Seeds, etc.) manually for the infrastructure environment you want to develop against. Additionally, you have to register the respective Gardener extensions manually. If you are rather looking for a quick start guide to develop entirely locally on your machine (no real cloud provider or infrastructure involved), then you should rather follow this guide.

Start a Local Kubernetes Cluster

For the development of Gardener you need a Kubernetes API server on which you can register Gardener’s own Extension API Server as APIService. This cluster doesn’t need any worker nodes to run pods, though, therefore, you can use the “nodeless Garden cluster setup” residing in hack/local-garden. This will start all minimally required components of a Kubernetes cluster (etcd, kube-apiserver, kube-controller-manager) and an etcd Instance for the gardener-apiserver as Docker containers. This is the easiest way to get your Gardener development setup up and running.

Using the nodeless cluster setup

Use the provided Makefile rules to start your local Garden:

make local-garden-up
Starting gardener-dev kube-etcd cluster..!
Starting gardener-dev kube-apiserver..!
Starting gardener-dev kube-controller-manager..!
Starting gardener-dev gardener-etcd cluster..!
namespace/garden created created created

ℹ️ [Optional] If you want to develop the SeedAuthorization feature then you have to run make ACTIVATE_SEEDAUTHORIZER=true local-garden-up. However, please note that this forces you to start the gardener-admission-controller via make start-admission-controller.

To tear down the local Garden cluster and remove the Docker containers, simply run:

make local-garden-down
Alternative: Using a local Kubernetes cluster

Instead of starting a Kubernetes API server and etcd as docker containers, you can also opt for running a local Kubernetes cluster, provided by e.g. minikube, kind or docker desktop.

Note: Gardener requires self-contained kubeconfig files because of a security issue. You can configure your minikube to create self-contained kubeconfig files via:

minikube config set embed-certs true

or when starting the local cluster

minikube start --embed-certs
Alternative: Using a remote Kubernetes cluster

For some testing scenarios, you may want to use a remote cluster instead of a local one as your Garden cluster. To do this, you can use the “remote Garden cluster setup” residing in hack/remote-garden. This will start an etcd instance for the gardener-apiserver as a Docker container, and open tunnels for accessing local gardener components from the remote cluster.

To avoid mistakes, the remote cluster must have a garden namespace labeled with You must create the garden namespace and label it manually before running make remote-garden-up as described below.

Use the provided Makefile rules to bootstrap your remote Garden:

export KUBECONFIG=<path to kubeconfig>
make remote-garden-up
# Start gardener etcd used to store gardener resources (e.g., seeds, shoots)
Starting gardener-dev-remote gardener-etcd cluster!
# Open tunnels for accessing local gardener components from the remote cluster

To close the tunnels and remove the locally-running Docker containers, run:

make remote-garden-down

ℹ️ [Optional] If you want to use the remote Garden cluster setup with the SeedAuthorization feature, you have to adapt the kube-apiserver process of your remote Garden cluster. To do this, perform the following steps after running make remote-garden-up:

  • Create an authorization webhook configuration file using the IP of the garden/quic-server pod running in your remote Garden cluster and port 10444 that tunnels to your locally running gardener-admission-controller process.

    apiVersion: v1
    kind: Config
    current-context: seedauthorizer
    - name: gardener-admission-controller
        insecure-skip-tls-verify: true
        server: https://<quic-server-pod-ip>:10444/webhooks/auth/seed
    - name: kube-apiserver
      user: {}
    - name: seedauthorizer
        cluster: gardener-admission-controller
        user: kube-apiserver
  • Change or add the following command line parameters to your kube-apiserver process:

    • --authorization-mode=<...>,Webhook
    • --authorization-webhook-config-file=<path to config file>
    • --authorization-webhook-cache-authorized-ttl=0
    • --authorization-webhook-cache-unauthorized-ttl=0
  • Delete the cluster role and rolebinding from your remote Garden cluster.

If your remote Garden cluster is a Gardener shoot, and you can access the seed on which this shoot is scheduled, you can automate the above steps by running the enable-seed-authorizer script and passing the kubeconfig of the seed cluster and the shoot namespace as parameters:

hack/local-development/remote-garden/enable-seed-authorizer <seed kubeconfig> <namespace>

Note: This script is not working anymore, as the ReversedVPN feature can’t be disabled. The annotation on Shoots is no longer respected.

To prevent Gardener from reconciling the shoot and overwriting your changes, add the annotation 'true' to the remote Garden shoot. Note that this annotation takes effect only if it is enabled via the constollers.shoot.respectSyncPeriodOverwrite: true option in the gardenlet configuration.

To disable the seed authorizer again, run the same script with -d as a third parameter:

hack/local-development/remote-garden/enable-seed-authorizer <seed kubeconfig> <namespace> -d

If the seed authorizer is enabled, you also have to start the gardener-admission-controller via make start-admission-controller.

⚠️ In the remote garden setup all Gardener components run with administrative permissions, i.e., there is no fine-grained access control via RBAC (as opposed to productive installations of Gardener).

Prepare the Gardener

Now, that you have started your local cluster, we can go ahead and register the Gardener API Server. Just point your KUBECONFIG environment variable to the cluster you created in the previous step and run:

make dev-setup
namespace/garden created
namespace/garden-dev created
deployment.apps/etcd created
service/etcd created
service/gardener-apiserver created
service/gardener-admission-controller created
endpoints/gardener-apiserver created
endpoints/gardener-admission-controller created created created created created

ℹ️ [Optional] If you want to enable logging, in the gardenlet configuration add:

  enabled: true

The Gardener exposes the API servers of Shoot clusters via Kubernetes services of type LoadBalancer. In order to establish stable endpoints (robust against changes of the load balancer address), it creates DNS records pointing to these load balancer addresses. They are used internally and by all cluster components to communicate. You need to have control over a domain (or subdomain) for which these records will be created. Please provide an internal domain secret (see this for an example) which contains credentials with the proper privileges. Further information can be found in Gardener Configuration and Usage.

kubectl apply -f example/10-secret-internal-domain-unmanaged.yaml
secret/internal-domain-unmanaged created

Run the Gardener

Next, run the Gardener API Server, the Gardener Controller Manager (optionally), the Gardener Scheduler (optionally), and the gardenlet in different terminal windows/panes using rules in the Makefile.

make start-apiserver
I0306 15:23:51.044421   74536 plugins.go:84] Registered admission plugin "ResourceReferenceManager"
I0306 15:23:51.044523   74536 plugins.go:84] Registered admission plugin "DeletionConfirmation"
I0306 15:23:51.626836   74536 secure_serving.go:116] Serving securely on [::]:8443

(Optional) Now you are ready to launch the Gardener Controller Manager.

make start-controller-manager
time="2019-03-06T15:24:17+02:00" level=info msg="Starting Gardener controller manager..."
time="2019-03-06T15:24:17+02:00" level=info msg="Feature Gates: "
time="2019-03-06T15:24:17+02:00" level=info msg="Starting HTTP server on"
time="2019-03-06T15:24:17+02:00" level=info msg="Acquired leadership, starting controllers."
time="2019-03-06T15:24:18+02:00" level=info msg="Starting HTTPS server on"
time="2019-03-06T15:24:18+02:00" level=info msg="Found internal domain secret internal-domain-unmanaged for domain"
time="2019-03-06T15:24:18+02:00" level=info msg="Successfully bootstrapped the Garden cluster."
time="2019-03-06T15:24:18+02:00" level=info msg="Gardener controller manager (version 1.0.0-dev) initialized."
time="2019-03-06T15:24:18+02:00" level=info msg="ControllerRegistration controller initialized."
time="2019-03-06T15:24:18+02:00" level=info msg="SecretBinding controller initialized."
time="2019-03-06T15:24:18+02:00" level=info msg="Project controller initialized."
time="2019-03-06T15:24:18+02:00" level=info msg="Quota controller initialized."
time="2019-03-06T15:24:18+02:00" level=info msg="CloudProfile controller initialized."

(Optional) Now you are ready to launch the Gardener Scheduler.

make start-scheduler
time="2019-05-02T16:31:50+02:00" level=info msg="Starting Gardener scheduler ..."
time="2019-05-02T16:31:50+02:00" level=info msg="Starting HTTP server on"
time="2019-05-02T16:31:50+02:00" level=info msg="Acquired leadership, starting scheduler."
time="2019-05-02T16:31:50+02:00" level=info msg="Gardener scheduler initialized (with Strategy: SameRegion)"
time="2019-05-02T16:31:50+02:00" level=info msg="Scheduler controller initialized."

The Gardener should now be ready to operate on Shoot resources. You can use

kubectl get shoots
No resources found.

to operate against your local running Gardener API Server.

Note: It may take several seconds until the Gardener API server has been started and is available. No resources found is the expected result of our initial development setup.

Create a Shoot

The steps below describe the general process of creating a Shoot. Have in mind that the steps do not provide full example manifests. The reader needs to check the provider documentation and adapt the manifests accordingly.

1. Copy the Example Manifests

The next steps require modifications of the example manifests. These modifications are part of local setup and should not be git push-ed. To do not interfere with git, let’s copy the example manifests to dev/ which is ignored by git.

cp example/*.yaml dev/

2. Create a Project

Every Shoot is associated with a Project. Check the corresponding example manifests dev/00-namespace-garden-dev.yaml and dev/05-project-dev.yaml. Adapt them and create them.

kubectl apply -f dev/00-namespace-garden-dev.yaml
kubectl apply -f dev/05-project-dev.yaml

Make sure that the Project is successfully reconciled:

$ kubectl get project dev
NAME   NAMESPACE    STATUS   OWNER                  CREATOR            AGE
dev    garden-dev   Ready   kubernetes-admin   6s

3. Create a CloudProfile

The CloudProfile resource is provider specific and describes the underlying cloud provider (available machine types, regions, machine images, etc.). Check the corresponding example manifest dev/30-cloudprofile.yaml. Check also the documentation and example manifests of the provider extension. Adapt dev/30-cloudprofile.yaml and apply it.

kubectl apply -f dev/30-cloudprofile.yaml

4. Install Necessary Gardener Extensions

The Known Extension Implementations section contains a list of available extension implementations. You need to create a ControllerRegistration and ControllerDeployment for:

  • at least one infrastructure provider
  • a DNS provider (if the DNS for the Seed is not disabled)
  • at least one operating system extension
  • at least one network plugin extension

As a convention, the example ControllerRegistration manifest (containing also the necessary ControllerDeployment) for an extension is located under example/controller-registration.yaml in the corresponding repository (for example for AWS the ControllerRegistration can be found here). An example creation for provider-aws (make sure to replace <version> with the newest released version tag):

kubectl apply -f<version>/example/controller-registration.yaml

Instead of updating extensions manually you can use Gardener Extensions Manager to install and update extension controllers. This is especially useful if you want to keep and maintain your development setup for a longer time. Also, please refer to Registering Extension Controllers for further information about how extensions are registered in case you want to use other versions than the latest releases.

5. Register a Seed

Shoot controlplanes run in seed clusters, so we need to create our first Seed now.

Check the corresponding example manifest dev/40-secret-seed.yaml and dev/50-seed.yaml. Update dev/40-secret-seed.yaml with base64 encoded kubeconfig of the cluster that will be used as Seed (the scope of the permissions should be identical to the kubeconfig that the gardenlet creates during bootstrapping - for now, cluster-admin privileges are recommended).

kubectl apply -f dev/40-secret-seed.yaml

Adapt dev/50-seed.yaml - adjust .spec.secretRef to refer the newly created Secret, adjust .spec.provider with the Seed cluster provider and revise the other fields.

kubectl apply -f dev/50-seed.yaml

6. Start the gardenlet

Once the Seed is created, start the gardenlet to reconcile it. The make start-gardenlet command will automatically configure the local gardenlet process to use the Seed and its kubeconfig. If you have multiple Seeds, you have to specify which to use by setting the SEED_NAME environment variable like in make start-gardenlet SEED_NAME=my-first-seed.

make start-gardenlet
time="2019-11-06T15:24:17+02:00" level=info msg="Starting Gardenlet..."
time="2019-11-06T15:24:17+02:00" level=info msg="Feature Gates: HVPA=true, Logging=true"
time="2019-11-06T15:24:17+02:00" level=info msg="Acquired leadership, starting controllers."
time="2019-11-06T15:24:18+02:00" level=info msg="Found internal domain secret internal-domain-unmanaged for domain"
time="2019-11-06T15:24:18+02:00" level=info msg="Gardenlet (version 1.0.0-dev) initialized."
time="2019-11-06T15:24:18+02:00" level=info msg="ControllerInstallation controller initialized."
time="2019-11-06T15:24:18+02:00" level=info msg="Shoot controller initialized."
time="2019-11-06T15:24:18+02:00" level=info msg="Seed controller initialized."

The gardenlet will now reconcile the Seed. Check the progess from time to time until it’s Ready:

kubectl get seed
seed-aws   Ready     aws         eu-west-1   4m     v1.61.0-dev   v1.24.8

7. Create a Shoot

A Shoot requires a SecretBinding. The SecretBinding refers to a Secret that contains the cloud provider credentials. The Secret data keys are provider specific and you need to check the documentation of the provider to find out which data keys are expected (for example for AWS the related documentation can be found at Provider Secret Data). Adapt dev/70-secret-provider.yaml and dev/80-secretbinding.yaml and apply them.

kubectl apply -f dev/70-secret-provider.yaml
kubectl apply -f dev/80-secretbinding.yaml

After the SecretBinding creation, you are ready to proceed with the Shoot creation. You need to check the documentation of the provider to find out the expected configuration (for example for AWS the related documentation and example Shoot manifest can be found at Using the AWS provider extension with Gardener as end-user). Adapt dev/90-shoot.yaml and apply it.

To make sure that a specific Seed cluster will be chosen or to skip the scheduling (the sheduling requires Gardener Scheduler to be running), specify the .spec.seedName field (see here).

kubectl apply -f dev/90-shoot.yaml

Watch the progress of the operation and make sure that the Shoot will be successfully created.

watch kubectl get shoot --all-namespaces

8 - Log Parsers

How to Create Log Parser for Container into fluent-bit

If our log message is parsed correctly, it has to be showed in Grafana like this:

  {"log":"OpenAPI AggregationController: Processing item","pid":"1","severity":"INFO","source":"controller.go:107"}

Otherwise it will looks like this:

  \"level\":\"info\",\"ts\":\"2020-06-01T11:23:26.679Z\",\"logger\":\"\",\"msg\":\"Finished ManagedResource health checks\",\"object\":\"garden/provider-aws-dsm9r\"

Create a Custom Parser

  • First of all, we need to know how the log for the specific container looks like (for example, lets take a log from the alertmanager : level=info ts=2019-01-28T12:33:49.362015626Z caller=main.go:175 build_context="(go=go1.11.2, user=root@4ecc17c53d26, date=20181109-15:40:48))

  • We can see that this log contains 4 subfields(severity=info, timestamp=2019-01-28T12:33:49.362015626Z, source=main.go:175 and the actual message). So we have to write a regex which matches this log in 4 groups(We can use like helping tool). So, for this purpose our regex looks like this:

  • It’s time to apply our new regex into fluent-bit configuration. Go to fluent-bit-configmap.yaml and create new filter using the following template:
        Name                parser
        Match               kubernetes.<< pod-name >>*<< container-name >>*
        Key_Name            log
        Parser              << parser-name >>
        Reserve_Data        True
        Name                parser
        Match               kubernetes.alertmanager*alertmanager*
        Key_Name            log
        Parser              alermanagerParser
        Reserve_Data        True
  • Now lets check if there already exists parser with such a regex and time format that we need. If it doesn’t, create one:
        Name        << parser-name >>
        Format      regex
        Regex       << regex >>
        Time_Key    time
        Time_Format << time-format >>
        Name        alermanagerParser
        Format      regex
        Regex       ^level=(?<severity>\w+)\s+ts=(?<time>\d{4}-\d{2}-\d{2}[Tt].*[zZ])\s+caller=(?<source>[^\s]*+)\s+(?<log>.*)
        Time_Key    time
        Time_Format %Y-%m-%dT%H:%M:%S.%L
Follow your development setup to validate that the parsers are working correctly.

9 - Logging

Logging in Gardener Components

This document aims at providing a general developer guideline on different aspects of logging practices and conventions used in the Gardener codebase. It contains mostly Gardener-specific points, and references other existing and commonly accepted logging guidelines for general advice. Developers and reviewers should consult this guide when writing, refactoring, and reviewing Gardener code. If parts are unclear or new learnings arise, this guide should be adapted accordingly.

Logging Libraries / Implementations

Historically, Gardener components have been using logrus. There is a global logrus logger (logger.Logger) that is initialized by components on startup and used across the codebase. In most places, it is used as a printf-style logger and only in some instances we make use of logrus’ structured logging functionality.

In the process of migrating our components to native controller-runtime components (see gardener/gardener#4251), we also want to make use of controller-runtime’s built-in mechanisms for streamlined logging. controller-runtime uses logr, a simple structured logging interface, for library-internal logging and logging in controllers.

logr itself is only an interface and doesn’t provide an implementation out of the box. Instead, it needs to be backed by a logging implementation like zapr. Code that uses the logr interface is thereby not tied to a specific logging implementation and makes the implementation easily exchangeable. controller-runtime already provides a set of helpers for constructing zapr loggers, i.e., logr loggers backed by zap, which is a popular logging library in the go community. Hence, we are migrating our component logging from logrus to logr (backed by zap) as part of gardener/gardener#4251.

⚠️ logger.Logger (logrus logger) is deprecated in Gardener and shall not be used in new code – use logr loggers when writing new code! (also see Migration from logrus to logr)

ℹ️ Don’t use zap loggers directly, always use the logr interface in order to avoid tight coupling to a specific logging implementation.

gardener-apiserver differs from the other components as it is based on the apiserver library and therefore uses klog – just like kube-apiserver. As gardener-apiserver writes (almost) no logs in our coding (outside the apiserver library), there is currently no plan for switching the logging implementation. Hence, the following sections focus on logging in the controller and admission components only.

logcheck Tool

To ensure a smooth migration to logr and make logging in Gardener components more consistent, the logcheck tool was added. It enforces (parts of) this guideline and detects programmer-level errors early on in order to prevent bugs. Please check out the tool’s documentation for a detailed description.

Structured Logging

Similar to efforts in the Kubernetes project, we want to migrate our component logs to structured logging. As motivated above, we will use the logr interface instead of klog though.

You can read more about the motivation behind structured logging in logr’s background and FAQ (also see this blog post by Dave Cheney). Also, make sure to check out controller-runtime’s logging guideline with specifics for projects using the library. The following sections will focus on the most important takeaways from those guidelines and give general instructions on how to apply them to Gardener and its controller-runtime components.

Note: Some parts in this guideline differ slightly from controller-runtime’s document.

TL;DR of Structured Logging

❌ Stop using printf-style logging:

var logger *logrus.Logger
logger.Infof("Scaling deployment %s/%s to %d replicas", deployment.Namespace, deployment.Name, replicaCount)

✅ Instead, write static log messages and enrich them with additional structured information in form of key-value pairs:

var logger logr.Logger
logger.Info("Scaling deployment", "deployment", client.ObjectKeyFromObject(deployment), "replicas", replicaCount)

Log Configuration

Gardener components can be configured to either log in json (default) or text format: json format is supposed to be used in production, while text format might be nicer for development.

# json
{"level":"info","ts":"2021-12-16T08:32:21.059+0100","msg":"Hello botanist","garden":"eden"}

# text
2021-12-16T08:32:21.059+0100    INFO    Hello botanist  {"garden": "eden"}

Components can be set to one of the following log levels (with increasing verbosity): error, info (default), debug.

Log Levels

logr uses V-levels (numbered log levels), higher V-level means higher verbosity. V-levels are relative (in contrast to klog’s absolute V-levels), i.e., V(1) creates a logger, that is one level more verbose than its parent logger.

In Gardener components, the mentioned log levels in the component config (error, info, debug) map to the zap levels with the same names (see here). Hence, our loggers follow the same mapping from numerical logr levels to named zap levels like described in zapr, i.e.:

  • component config specifies debug ➡️ both V(0) and V(1) are enabled
  • component config specifies info ➡️ V(0) is enabled, V(1) will not be shown
  • component config specifies error ➡️ neither V(0) nor V(1) will be shown
  • Error() logs will always be shown

This mapping applies to the components’ root loggers (the ones that are not “derived” from any other logger; constructed on component startup). If you derive a new logger with e.g. V(1), the mapping will shift by one. For example, V(0) will then log at zap’s debug level.

There is no warning level (see Dave Cheney’s post). If there is an error condition (e.g., unexpected error received from a called function), the error should either be handled or logged at error if it is neither handled nor returned. If you have an error value at hand that doesn’t represent an actual error condition, but you still want to log it as an informational message, log it at info level with key err.

We might consider to make use of a broader range of log levels in the future when introducing more logs and common command line flags for our components (comparable to --v of Kubernetes components). For now, we stick to the mentioned two log levels like controller-runtime: info (V(0)) and debug (V(1)).

Logging in Controllers

Named Loggers

Controllers should use named loggers that include their name, e.g.:

controllerLogger := rootLogger.WithName("controller").WithName("shoot")
controllerLogger.Info("Deploying kube-apiserver")

results in

2021-12-16T09:27:56.550+0100    INFO    controller.shoot    Deploying kube-apiserver

Logger names are hierarchical. You can make use of it, where controllers are composed of multiple “subcontrollers”, e.g., controller.shoot.hibernation or controller.shoot.maintenance.

Using the global logger logf.Log directly is discouraged and should be rather exceptional because it makes correlating logs with code harder. Preferably, all parts of the code should use some named logger.

Reconciler Loggers

In your Reconcile function, retrieve a logger from the given context.Context. It inherits from the controller’s logger (i.e., is already named) and is preconfigured with name and namespace values for the reconciliation request:

func (r *reconciler) Reconcile(ctx context.Context, request reconcile.Request) (reconcile.Result, error) {
  log := logf.FromContext(ctx)
  log.Info("Reconciling Shoot")
  // ...
  return reconcile.Result{}, nil

results in

2021-12-16T09:35:59.099+0100    INFO    controller.shoot    Reconciling Shoot        {"name": "sunflower", "namespace": "garden-greenhouse"}

The logger is injected by controller-runtime’s Controller implementation. The logger returned by logf.FromContext is never nil. If the context doesn’t carry a logger, it falls back to the global logger (logf.Log), which might discard logs if not configured, but is also never nil.

⚠️ Make sure that you don’t overwrite the name or namespace value keys for such loggers, otherwise you will lose information about the reconciled object.

The controller implementation (controller-runtime) itself takes care of logging the error returned by reconcilers. Hence, don’t log an error that you are returning. Generally, functions should not return an error, if they already logged it, because that means the error is already handled and not an error anymore. See Dave Cheney’s post for more on this.


  • Log messages should be static. Don’t put variable content in there, i.e., no fmt.Sprintf or string concatenation (+). Use key-value pairs instead.
  • Log messages should be capitalized. Note: This contrasts with error messages, that should not be capitalized. However, both should not end with a punctuation mark.

Keys and Values

  • Use WithValues instead of repeatedly adding key-value pairs for multiple log statements. WithValues creates a new logger from the parent, that carries the given key-value pairs. E.g., use it when acting on one object in multiple steps and logging something for each step:

    log := parentLog.WithValues("infrastructure", client.ObjectKeyFromObject(infrastrucutre))
    // ...
    log.Info("Creating Infrastructure")
    // ...
    log.Info("Waiting for Infrastructure to be reconciled")
    // ...

Note: WithValues bypasses controller-runtime’s special zap encoder that nicely encodes ObjectKey/NamespacedName and runtime.Object values, see kubernetes-sigs/controller-runtime#1290. Thus, the end result might look different depending on the value and its Stringer implementation.

  • Use lowerCamelCase for keys. Don’t put spaces in keys, as it will make log processing with simple tools like jq harder.

  • Keys should be constant, human-readable, consistent across the codebase and naturally match parts of the log message, see logr guideline.

  • When logging object keys (name and namespace), use the object’s type as the log key and a client.ObjectKey/types.NamespacedName value as value, e.g.:

    var deployment *appsv1.Deployment
    log.Info("Creating Deployment", "deployment", client.ObjectKeyFromObject(deployment))

    which results in

    {"level":"info","ts":"2021-12-16T08:32:21.059+0100","msg":"Creating Deployment","deployment":{"name": "bar", "namespace": "foo"}}

    Earlier, we often used kutil.ObjectName() for logging object keys, which encodes them into a flat string like foo/bar. However, this flat string cannot be processed so easily by logging stacks (or jq) like a structured log. Hence, the use of kutil.ObjectName() for logging object keys is discouraged. Existing usages should be refactored to use client.ObjectKeyFromObject() instead.

  • There are cases where you don’t have the full object key or the object itself at hand, e.g., if an object references another object (in the same namespace) by name (think secretRef or similar). In such a cases, either construct the full object key including the implied namespace or log the object name under a key ending in Name, e.g.:

    var (
      // object to reconcile
      shoot *gardencorev1beta1.Shoot
      // retrieved via logf.FromContext, preconfigured by controller with namespace and name of reconciliation request
      log logr.Logger
    // option a: full object key, manually constructed
    log.Info("Shoot uses SecretBinding", "secretBinding", client.ObjectKey{Namespace: shoot.Namespace, Name: shoot.Spec.SecretBindingName})
    // option b: only name under respective *Name log key
    log.Info("Shoot uses SecretBinding", "secretBindingName", shoot.Spec.SecretBindingName)

    Both options result in well-structured logs, that are easy to interpret and process:

    {"level":"info","ts":"2022-01-18T18:00:56.672+0100","msg":"Shoot uses SecretBinding","name":"my-shoot","namespace":"garden-project","secretBinding":{"namespace":"garden-project","name":"aws"}}
    {"level":"info","ts":"2022-01-18T18:00:56.673+0100","msg":"Shoot uses SecretBinding","name":"my-shoot","namespace":"garden-project","secretBindingName":"aws"}
  • When handling generic client.Object values (e.g. in helper funcs), use object as key.

  • When adding timestamps to key-value pairs, use time.Time values. By this, they will be encoded in the same format as the log entry’s timestamp.
    Don’t use metav1.Time values, as they will be encoded in a different format by their Stringer implementation. Pass <someTimestamp>.Time to loggers in case you have a metav1.Time value at hand.

  • Same applies to durations. Use time.Duration values instead of *metav1.Duration. Durations can be handled specially by zap just like timestamps.

  • Event recorders not only create Event objects but also log them. However, both Gardener’s manually instantiated event recorders and the ones that controller-runtime provides log to debug level and use generic formats, that are not very easy to interpret or process (no structured logs). Hence, don’t use event recorders as replacements for well-structured logs. If a controller records an event for a completed action or important information, it should probably log it as well, e.g.:

    log.Info("Creating ManagedSeed", "replica", r.GetObjectKey())
    a.recorder.Eventf(managedSeedSet, corev1.EventTypeNormal, EventCreatingManagedSeed, "Creating ManagedSeed %s", r.GetFullName())

Logging in Test Code

  • If the tested production code requires a logger, you can pass logr.Discard() or logf.NullLogger{} in your test, which simply discards all logs.

  • logf.Log is safe to use in tests and will not cause a nil pointer deref, even if it’s not initialized via logf.SetLogger. It is initially set to a NullLogger by default, which means all logs are discarded, unless logf.SetLogger is called in the first 30 seconds of execution.

  • Pass zap.WriteTo(GinkgoWriter) in tests where you want to see the logs on test failure but not on success, for example:

    logf.SetLogger(logger.MustNewZapLogger(logger.DebugLevel, logger.FormatJSON, zap.WriteTo(GinkgoWriter)))
    log := logf.Log.WithName("test")

10 - Monitoring Stack

Extending the Monitoring Stack

This document provides instructions to extend the Shoot cluster monitoring stack by integrating new scrape targets, alerts and dashboards.

Please ensure that you have understood the basic principles of Prometheus and its ecosystem before you continue.

‼️ The purpose of the monitoring stack is to observe the behaviour of the control plane and the system components deployed by Gardener onto the worker nodes. Monitoring of custom workloads running in the cluster is out of scope.


Monitoring Architecture

Each Shoot cluster comes with its own monitoring stack. The following components are deployed into the seed and shoot:

In each Seed cluster there is a Prometheus in the garden namespace responsible for collecting metrics from the Seed kubelets and cAdvisors. These metrics are provided to each Shoot Prometheus via federation.

The alerts for all Shoot clusters hosted on a Seed are routed to a central Alertmanger running in the garden namespace of the Seed. The purpose of this central alertmanager is to forward all important alerts to the operators of the Gardener setup.

The Alertmanager in the Shoot namespace on the Seed is only responsible for forwarding alerts from its Shoot cluster to a cluster owner/cluster alert receiver via email. The Alertmanager is optional and the conditions for a deployment are already described in Alerting.

Adding New Monitoring Targets

After exploring the metrics which your component provides or adding new metrics, you should be aware which metrics are required to write the needed alerts and dashboards.

Prometheus prefers a pull based metrics collection approach and therefore the targets to observe need to be defined upfront. The targets are defined in charts/seed-monitoring/charts/prometheus/templates/config.yaml. New scrape jobs can be added in the section scrape_configs. Detailed information how to configure scrape jobs and how to use the kubernetes service discovery are available in the Prometheus documentation.

The job_name of a scrape job should be the name of the component e.g. kube-apiserver or vpn. The collection interval should be the default of 30s. You do not need to specify this in the configuration.

Please do not ingest all metrics which are provided by a component. Rather, collect only those metrics which are needed to define the alerts and dashboards (i.e. whitelist). This can be achieved by adding the following metric_relabel_configs statement to your scrape jobs (replace exampleComponent with component name).

    - job_name: example-component
{{ include "prometheus.keep-metrics.metric-relabel-config" .Values.allowedMetrics.exampleComponent | indent 6 }}

The whitelist for the metrics of your job can be maintained in charts/seed-monitoring/charts/prometheus/values.yaml in section allowedMetrics.exampleComponent (replace exampleComponent with component name). Check the following example:

  * metrics_name_1
  * metrics_name_2

Adding Alerts

The alert definitons are located in charts/seed-monitoring/charts/prometheus/rules. There are two approaches for adding new alerts.

  1. Adding additional alerts for a component which already has a set of alerts. In this case you have to extend the existing rule file for the component.
  2. Adding alerts for a new component. In this case a new rule file with name scheme example-component.rules.yaml needs to be added.
  3. Add the new alert to, add any required inhibition flows and render the new graph. To render the graph, run:
dot -Tpng ./content/ -o ./content/alertInhibitionGraph.png
  1. Create a test for the new alert. See Alert Tests.

Example alert:

* name: example.rules
  * alert: ExampleAlert
    expr: absent(up{job="exampleJob"} == 1)
    for: 20m
      service: example
      severity: critical # How severe is the alert? (blocker|critical|info|warning)
      type: shoot # For which topology is the alert relevant? (seed|shoot)
      visibility: all # Who should receive the alerts? (all|operator|owner)
      description: A longer description of the example alert that should also explain the impact of the alert.
      summary: Short summary of an example alert.

If the deployment of component is optional then the alert definitions needs to be added to charts/seed-monitoring/charts/prometheus/optional-rules instead. Furthermore the alerts for component need to be activatable in charts/seed-monitoring/charts/prometheus/values.yaml via rules.optional.example-component.enabled. The default should be true.

Basic instruction how to define alert rules can be found in the Prometheus documentation.

Routing Tree

The Alertmanager is grouping incoming alerts based on labels into buckets. Each bucket has its own configuration like alert receivers, initial delaying duration or resending frequency, etc. You can find more information about Alertmanager routing in the Prometheus/Alertmanager documentation. The routing trees for the Alertmanagers deployed by Gardener are depicted below.

Central Seed Alertmanager

∟ main route (all alerts for all shoots on the seed will enter)
  ∟ group by project and shoot name
    ∟ group by visibility "all" and "operator"
      ∟ group by severity "blocker", "critical", and "info" → route to Garden operators
      ∟ group by severity "warning" (dropped)
    ∟ group by visibility "owner" (dropped)

Shoot Alertmanager

∟ main route (only alerts for one Shoot will enter)
  ∟ group by visibility "all" and "owner"
    ∟ group by severity "blocker", "critical", and "info" → route to cluster alert receiver
    ∟ group by severity "warning" (dropped, will change soon → route to cluster alert receiver)
  ∟ group by visibility "operator" (dropped)

Alert Inhibition

All alerts related to components running on the Shoot workers are inhibited in case of an issue with the vpn connection, because those components can’t be scraped anymore and Prometheus will fire alerts in consequence. The components running on the workers are probably healthy and the alerts are presumably false positives. The inhibition flow is shown in the figure below. If you add a new alert, make sure to add it to the diagram.


Alert Attributes

Each alert rule definition has to contain the following annotations:

  • summary: A short description of the issue.
  • description: A detailed explanation of the issue with hints to the possible root causes and the impact assessment of the issue.

In addtion, each alert must contain the following labels:

  • type
    • shoot: Components running on the Shoot worker nodes in the kube-system namespace.
    • seed: Components running on the Seed in the Shoot namespace as part of/next to the control plane.
  • service
    • Name of the component (in lowercase) e.g. kube-apiserver, alertmanager or vpn.
  • severity
    • blocker: All issues which make the cluster entirely unusable, e.g. KubeAPIServerDown or KubeSchedulerDown
    • critical: All issues which affect single functionalities/components but do not affect the cluster in its core functionality e.g. VPNDown or KubeletDown.
    • info: All issues that do not affect the cluster or its core functionality, but if this component is down we cannot determine if a blocker alert is firing. (i.e. A component with an info level severity is a dependency for a component with a blocker severity)
    • warning: No current existing issue, rather a hint for situations which could lead to real issue in the close future e.g. HighLatencyApiServerToWorkers or ApiServerResponseSlow.

Alert Tests

To test the Prometheus alerts:

make test-prometheus

If you want to add alert tests:

  1. Create a new file in rules-tests in the form <alert-group-name>.rules.test.yaml or if the alerts are for an existing component with existing tests, simply add the tests to the appropriate files.

  2. Make sure that newly added tests succeed. See above.

Adding Grafana Dashboards

The dashboard definition files are located in charts/seed-monitoring/charts/grafana/dashboards. Every dashboard needs its own file.

If you are adding a new component dashboard please also update the overview dashboard by adding a chart for its current up/down status and with a drill down option to the component dashboard.

Dashboard Structure

The dashboards should be structured in the following way. The assignment of the component dashboards to the categories should be handled via dashboard tags.

  • Kubernetes control plane components (Tag: control-plane)
    • All components which are part of the Kubernetes control plane e. g. Kube API Server, Kube Controller Manager, Kube Scheduler and Cloud Controller Manager
    • ETCD + Backup/Restore
    • Kubernetes Addon Manager
  • Node/Machine components (Tag: node/machine)
    • All metrics which are related to the behaviour/control of the Kubernetes nodes and kubelets
    • Machine-Controller-Manager + Cluster Autoscaler
  • Networking components (Tag: network)
    • CoreDNS, KubeProxy, Calico, VPN, Nginx Ingress
  • Addon components (Tag: addon)
    • Cert Broker
  • Monitoring components (Tag: monitoring)
  • Logging components (Tag: logging)

Mandatory Charts for Component Dashboards

For each new component, its corresponding dashboard should contain the following charts in the first row, before adding custom charts for the component in the subsequent rows.

  1. Pod up/down status up{job="example-component"}
  2. Pod/containers cpu utilization
  3. Pod/containers memorty consumption
  4. Pod/containers network i/o

That information is provided by the cAdvisor metrics. These metrics are already integrated. Please check the other dashboards for detailed information on how to query.

Chart Requirements

Each chart needs to contain:

  • a meaningful name
  • a detailed description (for non trivial charts)
  • appropriate x/y axis descriptions
  • appropriate scaling levels for the x/y axis
  • proper units for the x/y axis
Dashboard Parameters

The following parameters should be added to all dashboards to ensure a homogeneous experience across all dashboards.

Dashboards have to:

  • contain a title which refers to the component name(s)
  • contain a timezone statement which should be the browser time
  • contain tags which express where the component is running (seed or shoot) and to which category the component belong (see dashboard structure)
  • contain a version statement with a value of 1
  • be immutable

Example dashboard configuration:

  "title": "example-component",
  "timezone": "utc",
  "tags": [
  "version": 1,
  "editable": "false"

Furthermore, all dashboards should contain the following time options:

  "time": {
    "from": "now-1h",
    "to": "now"
  "timepicker": {
    "refresh_intervals": [
    "time_options": [

11 - New Cloud Provider

Adding Cloud Providers

This document provides an overview of how to integrate a new cloud provider into Gardener. Each component that requires integration has a detailed description of how to integrate it and the steps required.

Cloud Components

Gardener is composed of 2 or more Kubernetes clusters:

  • Shoot: These are the end-user clusters, the regular Kubernetes clusters you have seen. They provide places for your workloads to run.
  • Seed: This is the “management” cluster. It manages the control planes of shoots by running them as native Kubernetes workloads.

These two clusters can run in the same cloud provider, but they do not need to. For example, you could run your Seed in AWS, while having one shoot in Azure, two in Google, two in Alicloud, and three in Equinix Metal.

The Seed cluster deploys and manages the Shoot clusters. Importantly, for this discussion, the etcd data store backing each Shoot runs as workloads inside the Seed. Thus, to use the above example, the clusters in Azure, Google, Alicloud and Equinix Metal will have their worker nodes and master nodes running in those clouds, but the etcd clusters backing them will run as separate deployments in the Seed Kubernetes cluster on AWS.

This distinction becomes important when preparing the integration to a new cloud provider.

Gardener Cloud Integration

Gardener and its related components integrate with cloud providers at the following key lifecycle elements:

  • Create/destroy/get/list machines for the Shoot.
  • Create/destroy/get/list infrastructure components for the Shoot, e.g. VPCs, subnets, routes, etc.
  • Backup/restore etcd for the Seed via writing files to and reading them from object storage.

Thus, the integrations you need for your cloud provider depend on whether you want to deploy Shoot clusters to the provider, Seed or both.

  • Shoot Only: machine lifecycle management, infrastructure
  • Seed: etcd backup/restore

Gardener API

In addition to the requirements to integrate with the cloud provider, you also need to enable the core Gardener app to receive, validate, and process requests to use that cloud provider.

  • Expose the cloud provider to the consumers of the Gardener API, so it can be told to use that cloud provider as an option.
  • Validate that API as requests come in.
  • Write cloud provider specific implementation (called “provider extension”).

Cloud Provider API Requirements

In order for a cloud provider to integrate with Gardener, the provider must have an API to perform machine lifecycle events, specifically:

  • Create a machine
  • Destroy a machine
  • Get information about a machine and its state
  • List machines

In addition, if the Seed is to run on the given provider, it also must have an API to save files to block storage and retrieve them, for etcd backup/restore.

The current integration with cloud providers is to add their API calls to Gardener and the Machine Controller Manager. As both Gardener and the Machine Controller Manager are written in go, the cloud provider should have a go SDK. However, if it has an API that is wrappable in go, e.g. a REST API, then you can use that to integrate.

The Gardener team is working on bringing cloud provider integrations out-of-tree, making them pluggable, which should simplify the process and make it possible to use other SDKs.


To add a new cloud provider, you need some or all of the following. Each repository contains instructions on how to extend it to a new cloud provider.

Seed or ShootMachine Lifecyclemachine-controller-managerMCM new cloud provider
Seed onlyetcd backup/restoreetcd-backup-restoreIn process
AllExtension implementationgardenerExtension controller

12 - New Kubernetes Version

Adding Support For a New Kubernetes Version

This document describes the steps needed to perform in order to confidently add support for a new Kubernetes minor version.

⚠️ Typically, once a minor Kubernetes version vX.Y is supported by Gardener, then all patch versions vX.Y.Z are also automatically supported without any required action. This is because patch versions do not introduce any new feature or API changes, so there is nothing that needs to be adapted in gardener/gardener code.

The Kubernetes community release a new minor version roughly every 4 months. Please refer to the official documentation about their release cycles for any additional information.

Shortly before a new release, an “umbrella” issue should be opened which is used to collect the required adaptations and to track the work items. For example, #5102 can be used as a template for the issue description. As you can see, the task of supporting a new Kubernetes version also includes the provider extensions maintained in the gardener GitHub organization and is not restricted to gardener/gardener only.

Generally, the work items can be split into two groups: The first group contains Kubernetes release-independent tasks, the second group contains tasks specific to the changes in the given Kubernetes release.

ℹ️ Upgrading the* and Golang dependencies is typically tracked and worked on separately (see e.g. #4772 or #5282).

Deriving Release-Specific Tasks

Most new minor Kubernetes releases incorporate API changes, deprecations, or new features. The community announces them via their change logs. In order to derive the release-specific tasks, the respective change log for the new version vX.Y has to be read and understood (for example, the changelog for v1.24).

As already mentioned, typical changes to watch out for are:

  • API version promotions or deprecations
  • Feature gate promotions or deprecations
  • CLI flag changes for Kubernetes components
  • New default values in resources
  • New available fields in resources
  • New features potentially relevant for the Gardener system
  • Changes of labels or annotations Gardener relies on

Obviously, this requires a certain experience and understanding of the Gardener project so that all “relevant changes” can be identified. While reading the change log, add the tasks (along with the respective PR in kubernetes/kubernetes to the umbrella issue).

ℹ️ Some of the changes might be specific to certain cloud providers. Pay attention to those as well and add related tasks to the issue.

List Of Release-Independent Tasks

The following paragraphs describe recurring tasks that need to be performed for each new release.

Make Sure a New hyperkube Image Is Released

The gardener/hyperkube repository is used to release container images consisting of the kubectl and kubelet binaries.

There is a CI/CD job that runs periodically and releases a new hyperkube image when there is a new Kubernetes release. Before proceeding with the next steps, make sure that a new hyperkube image is released for the corresponding new Kubernetes minor version. Make sure that container image is present in GCR.

Adapting Gardener

  • Allow instantiation of a Kubernetes client for the new minor version and update the
    • See this example commit.
  • Maintain the Kubernetes feature gates used for validation of Shoot resources:
    • The feature gates are maintained in this file.
    • To maintain this list for new Kubernetes versions, run hack/ <old-version> <new-version> (e.g. hack/ v1.22 v1.23).
    • It will present 3 lists of feature gates: those added and those removed in <new-version> compared to <old-version> and feature gates that got locked to default in <new-version>.
    • Add all added feature gates to the map with <new-version> as AddedInVersion and no RemovedInVersion.
    • For any removed feature gates, add <new-version> as RemovedInVersion to the already existing feature gate in the map.
    • For feature gates locked to default, add <new-version> as LockedToDefaultInVersion to the already existing feature gate in the map.
    • See this example commit.
  • Maintain the Kubernetes kube-apiserver admission plugins used for validation of Shoot resources:
    • The admission plugins are maintained in this file.
    • To maintain this list for new Kubernetes versions, run hack/ <old-version> <new-version> (e.g. hack/ 1.24 1.25).
    • It will present 2 lists of admission plugins: those added and those removed in <new-version> compared to <old-version>.
    • Add all added admission plugins to the admissionPluginsVersionRanges map with <new-version> as AddedInVersion and no RemovedInVersion.
    • For any removed admission plugins, add <new-version> as RemovedInVersion to the already existing admission plugin in the map.
    • Flag any admission plugins that are required (plugins that must not be disabled in the Shoot spec) by setting the Required boolean variable to true for the admission plugin in the map.
    • Flag any admission plugins that are forbidden by setting the Forbidden boolean variable to true for the admission plugin in the map.
  • Maintain the ServiceAccount names for the controllers part of kube-controller-manager:
    • The names are maintained in this file.
    • To maintain this list for new Kubernetes versions, run hack/ <old-version> <new-version> (e.g. hack/ 1.22 1.23).
    • It will present 2 lists of controllers: those added and those removed in <new-version> compared to <old-version>.
    • Double check whether such ServiceAccount indeed appears in the kube-system namespace when creating a cluster with <new-version>. Note that it sometimes might be hidden behind a default-off feature gate. You can create a local cluster with the new version using the local provider.
    • If it appears, add all added controllers to the list based on the Kubernetes version (example).
    • For any removed controllers, add them only to the Kubernetes version if it is low enough.
  • Bump the used Kubernetes version for local Shoot and local e2e test.
    • See this example commit.

Filing the Pull Request

Work on all the tasks you have collected and validate them using the local provider. Execute the e2e tests and if everything looks good, then go ahead and file the PR (example PR). Generally, it is great if you add the PRs also to the umbrella issue so that they can be tracked more easily.

Adapting Provider Extensions

After the PR in gardener/gardener for the support of the new version has been merged, you can go ahead and work on the provider extensions.

Actually, you can already start even if the PR is not yet merged and use the branch of your fork.

  • Revendor the dependency in the extension and update the
  • Work on release-specific tasks related to this provider.

Maintaining the cloud-controller-manager Images

Some of the cloud providers are not yet using upstream cloud-controller-manager images. Instead, we build and maintain them ourselves:

Until we switch to upstream images, you need to revendor the Kubernetes dependencies and release a new image. The required steps are as follows:

  • Checkout the legacy-cloud-provider branch of the respective repository
  • Bump the versions in the Dockerfile (example commit).
  • Update the VERSION to vX.Y.Z-dev where Z is the latest available Kubernetes patch version for the vX.Y minor version.
  • Update the* dependencies in the go.mod file to vX.Y.Z and run go mod vendor and go mod tidy (example commit).
  • Checkout a new release-vX.Y branch and release it (example)

As you are already on it, it is great if you also bump the* dependencies for the last three minor releases as well. In this case, you need to checkout the release-vX.{Y-{1,2,3}} branches and only perform the last three steps (example branch, example commit).

Now you need to update the new releases in the charts/images.yaml of the respective provider extension so that they are used (see this example commit for reference).

Filing the Pull Request

Again, work on all the tasks you have collected. This time, you cannot use the local provider for validation but should create real clusters on the various infrastructures. Typically, the following validations should be performed:

  • Create new clusters with versions < vX.Y
  • Create new clusters with version = vX.Y
  • Upgrade old clusters from version vX.{Y-1} to version vX.Y
  • Delete clusters with versions < vX.Y
  • Delete clusters with version = vX.Y

If everything looks good, then go ahead and file the PR (example PR). Generally, it is again great if you add the PRs also to the umbrella issue so that they can be tracked more easily.

13 - Priority Classes

PriorityClasses in Gardener Clusters

Gardener makes use of PriorityClasses to improve the overall robustness of the system. In order to benefit from the full potential of PriorityClasses, the gardenlet manages a set of well-known PriorityClasses with fine-granular priority values.

All components of the system should use these well-known PriorityClasses instead of creating and using separate ones with arbitrary values, which would compromise the overall goal of using PriorityClasses in the first place. The gardenlet manages the well-known PriorityClasses listed in this document, so that third parties (e.g., Gardener extensions) can rely on them to be present when deploying components to Seed and Shoot clusters.

The listed well-known PriorityClasses follow this rough concept:

  • Values are close to the maximum that can be declared by the user. This is important to ensure that Shoot system components have higher priority than the workload deployed by end-users.
  • Values have a bit of headroom in between to ensure flexibility when the need for intermediate priority values arises.
  • Values of PriorityClasses created on Seed clusters are lower than the ones on Shoots to ensure that Shoot system components have higher priority than Seed components, if the Seed is backed by a Shoot (ManagedSeed), e.g. coredns should have higher priority than gardenlet.
  • Names simply include the last digits of the value to minimize confusion caused by many (similar) names like critical, importance-high, etc.

Garden Clusters

When using the gardener-operator for managing the garden runtime and virtual cluster, the following PriorityClasses are available:

PriorityClasses for Garden Control Plane Components

NamePriorityAssociated Components (Examples)
gardener-garden-system-critical999999550gardener-operator, gardener-resource-manager
gardener-garden-system-500999999500virtual-garden-etcd-events, virtual-garden-etcd-main
gardener-garden-system-300999999300vpa-admission-controller, etcd-druid
gardener-garden-system-200999999200vpa-recommender, vpa-updater, hvpa-controller

Seed Clusters

PriorityClasses for Seed System Components

NamePriorityAssociated Components (Examples)
gardener-system-critical999998950gardenlet, gardener-resource-manager, istio-ingressgateway, istiod
gardener-system-900999998900Extensions, reversed-vpn-auth-server
gardener-system-800999998800dependency-watchdog-endpoint, dependency-watchdog-probe, etcd-druid, (auditlog-)mutator, vpa-admission-controller
gardener-system-700999998700auditlog-seed-controller, hvpa-controller, vpa-recommender, vpa-updater
gardener-system-600999998600aggregate-alertmanager, alertmanager, fluent-bit, grafana, kube-state-metrics, nginx-ingress-controller, nginx-k8s-backend, prometheus, loki, seed-prometheus
gardener-reserve-excess-capacity-5reserve-excess-capacity (ref)

PriorityClasses for Shoot Control Plane Components

NamePriorityAssociated Components (Examples)
gardener-system-500999998500etcd-events, etcd-main, kube-apiserver
gardener-system-300999998300cloud-controller-manager, cluster-autoscaler, csi-driver-controller, kube-controller-manager, kube-scheduler, machine-controller-manager, terraformer, vpn-seed-server
gardener-system-200999998200csi-snapshot-controller, csi-snapshot-validation, cert-controller-manager, shoot-dns-service, vpa-admission-controller, vpa-recommender, vpa-updater
gardener-system-100999998100alertmanager, grafana-operators, grafana-users, kube-state-metrics, prometheus, loki, event-logger

Shoot Clusters

PriorityClasses for Shoot System Components

NamePriorityAssociated Components (Examples)
system-node-critical (created by Kubernetes)2000001000calico-node, kube-proxy, apiserver-proxy, csi-driver, egress-filter-applier
system-cluster-critical (created by Kubernetes)2000000000calico-typha, calico-kube-controllers, coredns, vpn-shoot
gardener-shoot-system-800999999800calico-typha-horizontal-autoscaler, calico-typha-vertical-autoscaler
gardener-shoot-system-700999999700blackbox-exporter, node-exporter
gardener-shoot-system-600999999600addons-nginx-ingress-controller, addons-nginx-ingress-k8s-backend, kubernetes-dashboard, kubernetes-metrics-scraper

14 - Process

Releases, Features, Hotfixes

This document describes how to contribute features or hotfixes, and how new Gardener releases are usually scheduled, validated, etc.


The @gardener-maintainers are trying to provide a new release roughly every other week (depending on their capacity and the stability/robustness of the master branch).

Hotfixes are usually maintained for the latest three minor releases, though, there are no fixed release dates.

Release Responsible Plan

VersionWeek NoBegin Validation PhaseDue DateRelease Responsible
v1.63Week 01-04January 2, 2023January 29, 2023@shafeeqes
v1.64Week 05-06January 30, 2023February 12, 2023@ary1992
v1.65Week 07-08February 13, 2023February 26, 2023@timuthy
v1.66Week 09-10February 27, 2023March 12, 2023@plkokanov
v1.67Week 11-12March 13, 2023March 26, 2023@rfranzke
v1.68Week 13-14March 27, 2023April 9, 2023@acumino
v1.69Week 15-16April 10, 2023April 23, 2023@oliver-goetz
v1.70Week 17-18April 24, 2023May 7, 2023@ialidzhikov
v1.71Week 19-20May 8, 2023May 21, 2023@shafeeqes
v1.72Week 21-22May 22, 2023June 4, 2023@ary1992
v1.73Week 23-24June 5, 2023June 18, 2023@timuthy
v1.74Week 25-26June 19, 2023July 2, 2023@oliver-goetz
v1.75Week 27-28July 3, 2023July 16, 2023@rfranzke
v1.76Week 29-30July 17, 2023July 30, 2023@plkokanov
v1.77Week 31-32July 31, 2023August 13, 2023@ialidzhikov
v1.78Week 33-34August 14, 2023August 27, 2023@acumino

Apart from the release of the next version, the release responsible is also taking care of potential hotfix releases of the last three minor versions. The release responsible is the main contact person for coordinating new feature PRs for the next minor versions or cherry-pick PRs for the last three minor versions.

Click to expand the archived release responsible associations!
VersionWeek NoBegin Validation PhaseDue DateRelease Responsible
v1.17Week 07-08February 15, 2021February 28, 2021@rfranzke
v1.18Week 09-10March 1, 2021March 14, 2021@danielfoehrKn
v1.19Week 11-12March 15, 2021March 28, 2021@timebertt
v1.20Week 13-14March 29, 2021April 11, 2021@vpnachev
v1.21Week 15-16April 12, 2021April 25, 2021@timuthy
v1.22Week 17-18April 26, 2021May 9, 2021@BeckerMax
v1.23Week 19-20May 10, 2021May 23, 2021@ialidzhikov
v1.24Week 21-22May 24, 2021June 5, 2021@stoyanr
v1.25Week 23-24June 7, 2021June 20, 2021@rfranzke
v1.26Week 25-26June 21, 2021July 4, 2021@danielfoehrKn
v1.27Week 27-28July 5, 2021July 18, 2021@timebertt
v1.28Week 29-30July 19, 2021August 1, 2021@ialidzhikov
v1.29Week 31-32August 2, 2021August 15, 2021@timuthy
v1.30Week 33-34August 16, 2021August 29, 2021@BeckerMax
v1.31Week 35-36August 30, 2021September 12, 2021@stoyanr
v1.32Week 37-38September 13, 2021September 26, 2021@vpnachev
v1.33Week 39-40September 27, 2021October 10, 2021@voelzmo
v1.34Week 41-42October 11, 2021October 24, 2021@plkokanov
v1.35Week 43-44October 25, 2021November 7, 2021@kris94
v1.36Week 45-46November 8, 2021November 21, 2021@timebertt
v1.37Week 47-48November 22, 2021December 5, 2021@danielfoehrKn
v1.38Week 49-50December 6, 2021December 19, 2021@rfranzke
v1.39Week 01-04January 3, 2022January 30, 2022@ialidzhikov, @timuthy
v1.40Week 05-06January 31, 2022February 13, 2022@BeckerMax
v1.41Week 07-08February 14, 2022February 27, 2022@plkokanov
v1.42Week 09-10February 28, 2022March 13, 2022@kris94
v1.43Week 11-12March 14, 2022March 27, 2022@rfranzke
v1.44Week 13-14March 28, 2022April 10, 2022@timebertt
v1.45Week 15-16April 11, 2022April 24, 2022@acumino
v1.46Week 17-18April 25, 2022May 8, 2022@ialidzhikov
v1.47Week 19-20May 9, 2022May 22, 2022@shafeeqes
v1.48Week 21-22May 23, 2022June 5, 2022@ary1992
v1.49Week 23-24June 6, 2022June 19, 2022@plkokanov
v1.50Week 25-26June 20, 2022July 3, 2022@rfranzke
v1.51Week 27-28July 4, 2022July 17, 2022@timebertt
v1.52Week 29-30July 18, 2022July 31, 2022@acumino
v1.53Week 31-32August 1, 2022August 14, 2022@kris94
v1.54Week 33-34August 15, 2022August 28, 2022@ialidzhikov
v1.55Week 35-36August 29, 2022September 11, 2022@oliver-goetz
v1.56Week 37-38September 12, 2022September 25, 2022@shafeeqes
v1.57Week 39-40September 26, 2022October 9, 2022@ary1992
v1.58Week 41-42October 10, 2022October 23, 2022@plkokanov
v1.59Week 43-44October 24, 2022November 6, 2022@rfranzke
v1.60Week 45-46November 7, 2022November 20, 2022@acumino
v1.61Week 47-48November 21, 2022December 4, 2022@ialidzhikov
v1.62Week 49-50December 5, 2022December 18, 2022@oliver-goetz

Release Validation

The release phase for a new minor version lasts two weeks. Typically, the first week is used for the validation of the release. This phase includes the following steps:

  1. master (or latest release-* branch) is deployed to a development landscape that already hosts some existing seed and shoot clusters.
  2. An extended test suite is triggered by the “release responsible” which:
    1. executes the Gardener integration tests for different Kubernetes versions, infrastructures, and Shoot settings.
    2. executes the Kubernetes conformance tests.
    3. executes further tests like Kubernetes/OS patch/minor version upgrades.
  3. Additionally, every four hours (or on demand) more tests (e.g., including the Kubernetes e2e test suite) are executed for different infrastructures.
  4. The “release responsible” is verifying new features or other notable changes (derived of the draft release notes) in this development system.

Usually, the new release is triggered in the beginning of the second week if all tests are green, all checks were successful, and if all of the planned verifications were performed by the release responsible.

Contributing New Features or Fixes

Please refer to the Gardener contributor guide. Besides a lot of a general information, it also provides a checklist for newly created pull requests that may help you to prepare your changes for an efficient review process. If you are contributing a fix or major improvement, please take care to open cherry-pick PRs to all affected and still supported versions once the change is approved and merged in the master branch.

⚠️ Please ensure that your modifications pass the verification checks (linting, formatting, static code checks, tests, etc.) by executing

make verify

before filing your pull request.

The guide applies for both changes to the master and to any release-* branch. All changes must be submitted via a pull request and be reviewed and approved by at least one code owner.

Cherry Picks

This section explains how to initiate cherry picks on release branches within the gardener/gardener repository.


Before you initiate a cherry pick, make sure that the following prerequisites are accomplished.

  • A pull request merged against the master branch.
  • The release branch exists (check in the branches section).
  • Have the gardener/gardener repository cloned as follows:
    • the origin remote should point to your fork (alternatively this can be overwritten by passing FORK_REMOTE=<fork-remote>).
    • the upstream remote should point to the Gardener GitHub org (alternatively this can be overwritten by passing UPSTREAM_REMOTE=<upstream-remote>).
  • Have hub installed, which is most easily installed via go get assuming you have a standard golang development environment.
  • A GitHub token which has permissions to create a PR in an upstream branch.

Initiate a Cherry Pick

  • Run the [cherry pick script][cherry-pick-script].

    This example applies a master branch PR #3632 to the remote branch upstream/release-v3.14:

    GITHUB_USER=<your-user> hack/ upstream/release-v3.14 3632
    • Be aware the cherry pick script assumes you have a git remote called upstream that points at the Gardener GitHub org.

    • You will need to run the cherry pick script separately for each patch release you want to cherry pick to. Cherry picks should be applied to all active release branches where the fix is applicable.

    • When asked for your GitHub password, provide the created GitHub token rather than your actual GitHub password. Refer

  • cherry-pick-script

15 - Secrets Management

Secrets Management for Seed and Shoot Cluster

The gardenlet needs to create quite some amount of credentials (certificates, private keys, passwords, etc.) for seed and shoot clusters in order to ensure secure deployments. Such credentials typically should be renewed automatically when their validity expires, rotated regularly, and they potentially need to be persisted such that they don’t get lost in case of a control plane migration or a lost seed cluster.

SecretsManager Introduction

These requirements can be covered by using the SecretsManager package maintained in pkg/utils/secrets/manager. It is built on top of the ConfigInterface and DataInterface interfaces part of pkg/utils/secrets and provides the following functions:

  • Generate(context.Context, secrets.ConfigInterface, ...GenerateOption) (*corev1.Secret, error)

    This method either retrieves the current secret for the given configuration or it (re)generates it in case the configuration changed, the signing CA changed (for certificate secrets), or when proactive rotation was triggered. If the configuration describes a certificate authority secret then this method automatically generates a bundle secret containing the current and potentially the old certificate. Available GenerateOptions:

    • SignedByCA(string, ...SignedByCAOption): This is only valid for certificate secrets and automatically retrieves the correct certificate authority in order to sign the provided server or client certificate.
      • There are two SignedByCAOptions:
        • UseCurrentCA. This option will sign server certificates with the new/current CA in case of a CA rotation. For more information, please refer to the “Certificate Signing” section below.
        • UseOldCA. This option will sign client certificates with the old CA in case of a CA rotation. For more information, please refer to the “Certificate Signing” section below.
    • Persist(): This marks the secret such that it gets persisted in the ShootState resource in the garden cluster. Consequently, it should only be used for secrets related to a shoot cluster.
    • Rotate(rotationStrategy): This specifies the strategy in case this secret is to be rotated or regenerated (either InPlace which immediately forgets about the old secret, or KeepOld which keeps the old secret in the system).
    • IgnoreOldSecrets(): This specifies that old secrets should not be considered and loaded (contrary to the default behavior). It should be used when old secrets are no longer important and can be “forgotten” (e.g. in “phase 2” (t2) of the CA certificate rotation). Such old secrets will be deleted on Cleanup().
    • IgnoreOldSecretsAfter(time.Duration): This specifies that old secrets should not be considered and loaded once a given duration after rotation has passed. It can be used to clean up old secrets after automatic rotation (e.g. the Seed cluster CA is automatically rotated when its validity will soon end and the old CA will be cleaned up 24 hours after triggering the rotation).
    • Validity(time.Duration): This specifies how long the secret should be valid. For certificate secret configurations, the manager will automatically deduce this information from the generated certificate.
  • Get(string, ...GetOption) (*corev1.Secret, bool)

    This method retrieves the current secret for the given name. In case the secret in question is a certificate authority secret then it retrieves the bundle secret by default. It is important that this method only knows about secrets for which there were prior Generate calls. Available GetOptions:

    • Bundle (default): This retrieves the bundle secret.
    • Current: This retrieves the current secret.
    • Old: This retrieves the old secret.
  • Cleanup(context.Context) error

    This method deletes secrets which are no longer required. No longer required secrets are those still existing in the system which weren’t detected by prior Generate calls. Consequently, only call Cleanup after you have executed Generate calls for all desired secrets.

Some exemplary usages would look as follows:

secret, err := k.secretsManager.Generate(
        Name:                        "my-server-secret",
        CommonName:                  "server-abc",
        DNSNames:                    []string{"first-name", "second-name"},
        CertType:                    secrets.ServerCert,
        SkipPublishingCACertificate: true,
if err != nil {
    return err

As explained above, the caller does not need to care about the renewal, rotation or the persistence of this secret - all of these concerns are handled by the secrets manager. Automatic renewal of secrets happens when their validity approaches 80% or less than 10d are left until expiration.

In case a CA certificate is needed by some component, then it can be retrieved as follows:

caSecret, found := k.secretsManager.Get("my-ca")
if !found {
    return fmt.Errorf("secret my-ca not found")

As explained above, this returns the bundle secret for the CA my-ca which might potentially contain both the current and the old CA (in case of rotation/regeneration).

Certificate Signing

By default, client certificates are always signed by the current CA while server certificate are signed by the old CA (if it exists). This is to ensure a smooth exchange of certificate during a CA rotation (typically has two phases, ref GEP-18):

  • Client certificates:
    • In phase 1, clients get new certificates as soon as possible to ensure that all clients have been adapted before phase 2.
    • In phase 2, the respective server drops accepting certificates signed by the old CA.
  • Server certificates:
    • In phase 1, servers still use their old/existing certificates to allow clients to update their CA bundle used for verification of the servers’ certificates.
    • In phase 2, the old CA is dropped, hence servers need to get a certificate signed by the new/current CA. At this point in time, clients have already adapted their CA bundles.

Always Sign Server Certificates with Current CA

In case you control all clients and update them at the same time as the server, it is possible to make the secrets manager generate even server certificates with the new/current CA. This can help to prevent certificate mismatches when the CA bundle is already exchanged while the server still serves with a certificate signed by a CA no longer part of the bundle.

Let’s consider the two following examples:

  1. gardenlet deploys a webhook server (gardener-resource-manager) and a corresponding MutatingWebhookConfiguration at the same time. In this case, the server certificate should be generated with the new/current CA to avoid above mentioned certificate mismatches during a CA rotation.
  2. gardenlet deploys a server (etcd) in one step, and a client (kube-apiserver) in a subsequent step. In this case, the default behaviour should apply (server certificate should be signed by old/existing CA).

Always Sign Client Certificate with Old CA

In the unusual case where the client is deployed before the server, it might be useful to always use the old CA for signing the client’s certificate. This can help to prevent certificate mismatches when the client already gets a new certificate while the server still only accepts certificates signed by the old CA.

Let’s consider the following example:

  1. gardenlet deploys the kube-apiserver before the kubelet. However, the kube-apiserver has a client certificate signed by the ca-kubelet in order to communicate with it (e.g., when retrieving logs or forwarding ports). In this case, the client certificate should be generated with the old CA to avoid above mentioned certificate mismatches during a CA rotation.

Reusing the SecretsManager in Other Components

While the SecretsManager is primarily used by gardenlet, it can be reused by other components (e.g. extensions) as well for managing secrets that are specific to the component or extension. For example, provider extensions might use their own SecretsManager instance for managing the serving certificate of cloud-controller-manager.

External components that want to reuse the SecretsManager should consider the following aspects:

  • On initialization of a SecretsManager, pass an identity specific to the component, controller and purpose. For example, gardenlet’s shoot controller uses gardenlet as the SecretsManager’s identity, the Worker controller in provider-foo should use provider-foo-worker, and the ControlPlane controller should use provider-foo-controlplane-exposure for ControlPlane objects of purpose exposure. The given identity is added as a value for the manager-identity label on managed Secrets. This label is used by the Cleanup function to select only those Secrets that are actually managed by the particular SecretManager instance. This is done to prevent removing still needed Secrets that are managed by other instances.
  • Generate dedicated CAs for signing certificates instead of depending on CAs managed by gardenlet.
  • Names of Secrets managed by external SecretsManager instances must not conflict with Secret names from other instances (e.g. gardenlet).
  • For CAs that should be rotated in lock-step with the Shoot CAs managed by gardenlet, components need to pass information about the last rotation initiation time and the current rotation phase to the SecretsManager upon initialization. The relevant information can be retrieved from the Cluster resource under .spec.shoot.status.credentials.rotation.certificateAuthorities.
  • Independent of the specific identity, secrets marked with the Persist option are automatically saved in the ShootState resource by the gardenlet and are also restored by the gardenlet on Control Plane Migration to the new Seed.

Migrating Existing Secrets To SecretsManager

If you already have existing secrets which were not created with SecretsManager, then you can (optionally) migrate them by labeling them with secrets-manager-use-data-for=<config-name>. For example, if your SecretsManager generates a CertificateConfigSecret with name foo like this

secret, err := k.secretsManager.Generate(
        Name:                        "foo",
        // ...

and you already have an existing secret in your system whose data should be kept instead of regenerated, then labeling it with secrets-manager-use-data-for=foo will instruct SecretsManager accordingly.

⚠️ Caveat: You have to make sure that the existing data keys match with what SecretsManager uses:

Secret TypeData Keys
Basic Authusername, password, auth
CA Certificateca.crt, ca.key
Non-CA Certificatetls.crt, tls.key
Control Plane Secretca.crt, username, password, token, kubeconfig
ETCD Encryption Keykey, secret
RSA Private Keyid_rsa,
Static Tokenstatic_tokens.csv
VPN TLS Authvpn.tlsauth

Implementation Details

The source of truth for the secrets manager is the list of Secrets in the Kubernetes cluster it acts upon (typically, the seed cluster). The persisted secrets in the ShootState are only used if and only if the shoot is in the Restore phase - in this case all secrets are just synced to the seed cluster so that they can be picked up by the secrets manager.

In order to prevent kubelets from unneeded watches (thus, causing some significant traffic against the kube-apiserver), the Secrets are marked as immutable. Consequently, they have a unique, deterministic name which is computed as follows:

  • For CA secrets, the name is just exactly the name specified in the configuration (e.g., ca). This is for backwards-compatibility and will be dropped in a future release once all components depending on the static name have been adapted.
  • For all other secrets, the name specified in the configuration is used as prefix followed by an 8-digit hash. This hash is computed out of the checksum of the secret configuration and the checksum of the certificate of the signing CA (only for certificate configurations).

In all cases, the name of the secrets is suffixed with a 5-digit hash computed out of the time when the rotation for this secret was last started.

16 - Seed Network Policies

Network Policies in the Seed Cluster

This document describes the Kubernetes network policies deployed by Gardener into the Seed cluster. For network policies deployed into the Shoot kube-system namespace, please see the usage section.

Network policies deployed by Gardener have names and annotations describing their purpose, so this document only highlights a subset of the policies in detail.

Network Policies in the Shoot Namespace in the Seed

The network policies in the Shoot namespace in the Seed can roughly be grouped into policies required for the control plane components and policies required for logging & monitoring.

The network policy deny-all plays a special role. This policy denies all ingress and egress traffic from each pod in the Shoot namespace. So per default, a pod running in the control plane cannot talk to any other pod in the whole Seed cluster. This means the pod needs to have labels matching to appropriate network policies allowing it to talk to exactly the components required to execute its desired functionality. This has also implications for Gardener extensions that need to deploy additional components into the Shoot's control plane.

Network Policies for Control Plane Components

This section highlights a selection of network policies that exist in the Shoot namespace in the Seed cluster. In general, the control plane components serve different purposes and thus need access to different pods and network ranges.

In contrast to other network policies, the policy allow-to-shoot-networks is tailored to the individual Shoot cluster, because it is based on the network configuration in the Shoot manifest. It allows pods with the label to access pods in the Shoot pod, service and node CIDR range. This is used by the Shoot API Server and the Prometheus pods to communicate over VPN/proxy with pods in the Shoot cluster. This network policy is only useful if reversed vpn is disabled, as otherwise the vpn-seed-server pod in the control plane is the only pod with layer 3 routing to the shoot network.

The policy allow-to-blocked-cidrs allows pods with the label to access IPs that are explicitly blocked for all control planes in a Seed cluster (configurable via spec.networks.blockCIDRS). This is used for instance to block the cloud provider’s metadata service.

Another network policy to be highlighted is allow-to-runtime-apiserver. Some components need access to the Seed API Server. This can be allowed by labeling the pod with This policy allows exactly the IPs of the kube-apiserver of the Seed. While all other policies have a static set of permissions (do not change during the lifecycle of the Shoot), the policy allow-to-runtime-apiserver is reconciled to reflect the endpoints in the default namespace. This is required because endpoint IPs are not necessarily stable (think of scaling the Seed API Server pods or hibernating the Seed cluster (acting as a managed seed) in a local development environment).

Furthermore, the following network policies exist in the Shoot namespace. These policies are the same for every Shoot control plane.

NAME                              POD-SELECTOR      
# Pods that need to access the Shoot API server. Used by all Kubernetes control plane components.

# allows access to kube-dns/core-dns pods for DNS queries                       

# allows access to private IP address ranges 

# allows access to all but private IP address ranges 

# allows Ingress to etcd pods from the Shoot's Kubernetes API Server
allow-etcd                        app=etcd-statefulset,

# used by the Shoot API server to allows ingress from pods labeled
# with'', from Prometheus, and allows Egress to etcd pods
allow-kube-apiserver              app=kubernetes,,role=apiserver

Network Policies for Logging & Monitoring

Gardener currently introduces a logging stack based on Loki. So this section is subject to change. For more information, see the Loki Gardener Community Meeting .

These are the logging and monitoring related network policies:

NAME                              POD-SELECTOR                                                             
allow-grafana                     component=grafana,
allow-prometheus                  app=prometheus,,role=monitoring

For instance, let’s take a look at the network policy from-prometheus. As part of the shoot reconciliation flow, Gardener deploys a shoot-specific Prometheus into the shoot namespace. Each pod that should be scraped for metrics must be labeled with to allow incoming network requests by the Prometheus pod. Most components of the Shoot cluster’s control plane expose metrics and are therefore labeled appropriately.

Implications for Gardener Extensions

Gardener extensions sometimes need to deploy additional components into the Shoot namespace in the Seed hosting the control plane. For example, the Gardener extension provider-aws deploys the MachineControllerManager into the Shoot namespace, that is ultimately responsible to create the VMs with the cloud provider AWS.

Every Shoot namespace in the Seed contains the network policy deny-all. This requires a pod deployed by a Gardener extension to have labels from network policies that exist in the Shoot namespace, that allow the required network ranges.

Additionally, extensions could also deploy their own network policies. This is used e.g by the Gardener extension provider-aws to serve Admission Webhooks for the Shoot API server that need to be reachable from within the Shoot namespace.

The pod can use an arbitrary combination of network policies.

Network Policies in the garden Namespace

The network policies in the garden namespace are, with a few exceptions (e.g Kubernetes control plane specific policies), the same as in the Shoot namespaces. For your reference, these are all the deployed network policies.

NAME                              POD-SELECTOR  
allow-fluentbit                   app=fluent-bit,,role=logging              

This section describes the network policies that are unique to the garden namespace.

The network policy allow-to-all-shoot-apiservers allows pods to access every Shoot API server in the Seed. This is, for instance, used by the dependency watchdog to regularly check the health of all the Shoot API servers.

Gardener deploys a central Prometheus instance in the garden namespace that fetches metrics and data from all seed cluster nodes and all seed cluster pods. The network policies allow-to-aggregate-prometheus and allow-from-aggregate-prometheus allow traffic from and to this Prometheus instance.

Worth mentioning is, that the network policy allow-to-shoot-networks does not exist in the garden namespace. This is to forbid Gardener system components to talk to workload deployed in the Shoot VPC.

17 - Testing

Testing Strategy and Developer Guideline

This document walks you through:

  • What kind of tests we have in Gardener
  • How to run each of them
  • What purpose each kind of test serves
  • How to best write tests that are correct, stable, fast and maintainable
  • How to debug tests that are not working as expected

The document is aimed towards developers that want to contribute code and need to write tests, as well as maintainers and reviewers that review test code. It serves as a common guide that we commit to follow in our project to ensure consistency in our tests, good coverage for high confidence, and good maintainability.

The guidelines are not meant to be absolute rules. Always apply common sense and adapt the guideline if it doesn’t make much sense for some cases. If in doubt, don’t hesitate to ask questions during a PR review (as an author, but also as a reviewer). Add new learnings as soon as we make them!

Generally speaking, tests are a strict requirement for contributing new code. If you touch code that is currently untested, you need to add tests for the new cases that you introduce as a minimum. Ideally though, you would add the missing test cases for the current code as well (boy scout rule – “always leave the campground cleaner than you found it”).

Writing Tests (Relevant for All Kinds)

  • We follow BDD (behavior-driven development) testing principles and use Ginkgo, along with Gomega.
    • Make sure to check out their extensive guides for more information and how to best leverage all of their features
  • Use By to structure test cases with multiple steps, so that steps are easy to follow in the logs: example test
  • Call defer GinkgoRecover() if making assertions in goroutines: doc, example test
  • Use DeferCleanup instead of cleaning up manually (or use custom coding from the test framework): example test, example test
    • DeferCleanup makes sure to run the cleanup code in the right point in time, e.g., a DeferCleanup added in BeforeEach is executed with AfterEach.
  • Test failures should point to an exact location, so that failures in CI aren’t too difficult to debug/fix.
    • Use ExpectWithOffset for making assertions in helper funcs like expectSomethingWasCreated: example test
    • Make sure to add additional descriptions to Gomega matchers if necessary (e.g. in a loop): example test
  • Introduce helper functions for assertions to make test more readable where applicable: example test
  • Introduce custom matchers to make tests more readable where applicable: example matcher
  • Don’t rely on accurate timing of time.Sleep and friends.
  • Use the same client schemes that are also used by production code to avoid subtle bugs/regressions: example PR, production schemes, usage in test
  • Make sure that your test is actually asserting the right thing and it doesn’t pass if the exact bug is introduced that you want to prevent.
    • Use specific error matchers instead of asserting any error has happened, make sure that the corresponding branch in the code is tested, e.g., prefer
    • If you’re unsure about your test’s behavior, attaching the debugger can sometimes be helpful to make sure your test is correct.
  • About overwriting global variables:
    • This is a common pattern (or hack?) in go for faking calls to external functions.
    • However, this can lead to races, when the global variable is used from a goroutine (e.g., the function is called).
    • Alternatively, set fields on structs (passed via parameter or set directly): this is not racy, as struct values are typically (and should be) only used for a single test case.
    • An alternative to dealing with function variables and fields:
      • Add an interface which your code depends on
      • Write a fake and a real implementation (similar to clock.Clock.Sleep)
      • The real implementation calls the actual function (clock.RealClock.Sleep calls time.Sleep)
      • The fake implementation does whatever you want it to do for your test (clock.FakeClock.Sleep waits until the test code advanced the time)
  • Use constants in test code with care.
    • Typically, you should not use constants from the same package as the tested code, instead use literals.
    • If the constant value is changed, tests using the constant will still pass, although the “specification” is not fulfilled anymore.
    • There are cases where it’s fine to use constants, but keep this caveat in mind when doing so.
  • Creating sample data for tests can be a high effort.
    • If valuable, add a package for generating common sample data, e.g. Shoot/Cluster objects.
  • Make use of the testdata directory for storing arbitrary sample data needed by tests (helm charts, YAML manifests, etc.), example PR

Unit Tests

Running Unit Tests

Run all unit tests:

make test

Run all unit tests with test coverage:

make test-cov
open test.coverage.html
make test-cov-clean

Run unit tests of specific packages:

# run with same settings like in CI (race dector, timeout, ...)
./hack/ ./pkg/resourcemanager/controller/... ./pkg/utils/secrets/...

# freestyle
go test ./pkg/resourcemanager/controller/... ./pkg/utils/secrets/...
ginkgo run ./pkg/resourcemanager/controller/... ./pkg/utils/secrets/...

Debugging Unit Tests

Use ginkgo to focus on (a set of) test specs via code or via CLI flags. Remember to unfocus specs before contributing code, otherwise your PR tests will fail.

$ ginkgo run --focus "should delete the unused resources" ./pkg/resourcemanager/controller/garbagecollector
Will run 1 of 3 specs

Ran 1 of 3 Specs in 0.003 seconds
SUCCESS! -- 1 Passed | 0 Failed | 0 Pending | 2 Skipped

Use ginkgo to run tests until they fail:

$ ginkgo run --until-it-fails ./pkg/resourcemanager/controller/garbagecollector
Ran 3 of 3 Specs in 0.004 seconds
SUCCESS! -- 3 Passed | 0 Failed | 0 Pending | 0 Skipped

All tests passed...
Will keep running them until they fail.
This was attempt #58
No, seriously... you can probably stop now.

Use the stress tool for deflaking tests that fail sporadically in CI, e.g., due resource contention (CPU throttling):

# get the stress tool
go install

# build a test binary
ginkgo build ./pkg/resourcemanager/controller/garbagecollector
# alternatively
go test -c ./pkg/resourcemanager/controller/garbagecollector

# run the test in parallel and report any failures
stress -p 16 ./pkg/resourcemanager/controller/garbagecollector/garbagecollector.test -ginkgo.focus "should delete the unused resources"
5s: 1077 runs so far, 0 failures
10s: 2160 runs so far, 0 failures

stress will output a path to a file containing the full failure message when a test run fails.

Purpose of Unit Tests

  • Unit tests prove the correctness of a single unit according to the specification of its interface.
    • Think: Is the unit that I introduced doing what it is supposed to do for all cases?
  • Unit tests protect against regressions caused by adding new functionality to or refactoring of a single unit.
    • Think: Is the unit that was introduced earlier (by someone else) and that I changed still doing what it was supposed to do for all cases?
  • Example units: functions (conversion, defaulting, validation, helpers), structs (helpers, basic building blocks like the Secrets Manager), predicates, event handlers.
  • For these purposes, unit tests need to cover all important cases of input for a single unit and cover edge cases / negative paths as well (e.g., errors).
    • Because of the possible high dimensionality of test input, unit tests need to be fast to execute: individual test cases should not take more than a few seconds, test suites not more than 2 minutes.
    • Fuzzing can be used as a technique in addition to usual test cases for covering edge cases.
  • Test coverage can be used as a tool during test development for covering all cases of a unit.
  • However, test coverage data can be a false safety net.
    • Full line coverage doesn’t mean you have covered all cases of valid input.
    • We don’t have strict requirements for test coverage, as it doesn’t necessarily yield the desired outcome.
  • Unit tests should not test too large components, e.g. entire controller Reconcile functions.
    • If a function/component does many steps, it’s probably better to split it up into multiple functions/components that can be unit tested individually
    • There might be special cases for very small Reconcile functions.
    • If there are a lot of edge cases, extract dedicated functions that cover them and use unit tests to test them.
    • Usual-sized controllers should rather be tested in integration tests.
    • Individual parts (e.g. helper functions) should still be tested in unit test for covering all cases, though.
  • Unit tests are especially easy to run with a debugger and can help in understanding concrete behavior of components.

Writing Unit Tests

  • For the sake of execution speed, fake expensive calls/operations, e.g. secret generation: example test
  • Generally, prefer fakes over mocks, e.g., use controller-runtime fake client over mock clients.
    • Mocks decrease maintainability because they expect the tested component to follow a certain way to reach the desired goal (e.g., call specific functions with particular arguments), example consequence
    • Generally, fakes should be used in “result-oriented” test code (e.g., that a certain object was labelled, but the test doesn’t care if it was via patch or update as both a valid ways to reach the desired goal).
    • Although rare, there are valid use cases for mocks, e.g. if the following aspects are important for correctness:
      • Asserting that an exact function is called
      • Asserting that functions are called in a specific order
      • Asserting that exact parameters/values/… are passed
      • Asserting that a certain function was not called
      • Many of these can also be verified with fakes, although mocks might be simpler
    • Only use mocks if the tested code directly calls the mock; never if the tested code only calls the mock indirectly (e.g., through a helper package/function).
    • Keep in mind the maintenance implications of using mocks:
      • Can you make a valid non-behavioral change in the code without breaking the test or dependent tests?
    • It’s valid to mix fakes and mocks in the same test or between test cases.
  • Generally, use the go test package, i.e., declare package <production_package>_test:
    • Helps in avoiding cyclic dependencies between production, test and helper packages
    • Also forces you to distinguish between the public (exported) API surface of your code and internal state that might not be of interest to tests
    • It might be valid to use the same package as the tested code if you want to test unexported functions.
    • Helpers can also be exported if no one is supposed to import the containing package (e.g. controller package).

Integration Tests (envtests)

Integration tests in Gardener use the package. It sets up a temporary control plane (etcd + kube-apiserver) and runs the test against it. The test suites start their individual envtest environment before running the tested controller/webhook and executing test cases. Before exiting, the test suites tear down the temporary test environment.

Package augments the controller-runtime’s envtest package by starting and registering gardener-apiserver. This is used to test controllers that act on resources in the Gardener APIs (aggregated APIs).

Historically, test machinery tests have also been called “integration tests”. However, test machinery does not perform integration testing but rather executes a form of end-to-end tests against a real landscape. Hence, we tried to sharpen the terminology that we use to distinguish between “real” integration tests and test machinery tests but you might still find “integration tests” referring to test machinery tests in old issues or outdated documents.

Running Integration Tests

The test-integration make rule prepares the environment automatically by downloading the respective binaries (if not yet present) and setting the necessary environment variables.

make test-integration

If you want to run a specific set of integration tests, you can also execute them using ./hack/ directly instead of using the test-integration rule. Prior to execution, the PATH environment variable needs to be set to also included the tools binary directory. For example:

export PATH="$PATH:$PWD/hack/tools/bin"

source ./hack/test-integration.env
./hack/ ./test/integration/resourcemanager/tokenrequestor

The script takes care of preparing the environment for you. If you want to execute the test suites directly via go test or ginkgo, you have to point the KUBEBUILDER_ASSETS environment variable to the path that contains the etcd and kube-apiserver binaries. Alternatively, you can install the binaries to /usr/local/kubebuilder/bin. Additionally, the environment variables from hack/test-integration.env should be sourced.

Debugging Integration Tests

You can configure envtest to use an existing cluster or control plane instead of starting a temporary control plane that is torn down immediately after executing the test. This can be helpful for debugging integration tests because you can easily inspect what is going on in your test environment with kubectl.

While you can use an existing cluster (e.g., kind), some test suites expect that no controllers and no nodes are running in the test environment (as it is the case in envtest test environments). Hence, using a full-blown cluster with controllers and nodes might sometimes be impractical, as you would need to stop cluster components for the tests to work.

You can use make start-envtest to start an envtest test environment that is managed separately from individual test suites. This allows you to keep the test environment running for as long as you want, and to debug integration tests by executing multiple test runs in parallel or inspecting test runs using kubectl. When you are finished, just hit CTRL-C for tearing down the test environment. The kubeconfig for the test environment is placed in dev/envtest-kubeconfig.yaml.

make start-envtest brings up an envtest environment using the default configuration. If your test suite requires a different control plane configuration (e.g., disabled admission plugins or enabled feature gates), feel free to locally modify the configuration in test/start-envtest while debugging.

Run an envtest suite (not using gardener-apiserver) against an existing test environment:

make start-envtest

# in another terminal session:
export KUBECONFIG=$PWD/dev/envtest-kubeconfig.yaml

# run test with verbose output
./hack/ -v ./test/integration/resourcemanager/health -ginkgo.v

# in another terminal session:
export KUBECONFIG=$PWD/dev/envtest-kubeconfig.yaml
# watch test objects
k get managedresource -A -w

Run a gardenerenvtest suite (using gardener-apiserver) against an existing test environment:

# modify test/start-envtest to disable admission plugins and enable feature gates like in test suite...

make start-envtest ENVTEST_TYPE=gardener

# in another terminal session:
export KUBECONFIG=$PWD/dev/envtest-kubeconfig.yaml

# run test with verbose output
./hack/ -v ./test/integration/controllermanager/bastion -ginkgo.v

# in another terminal session:
export KUBECONFIG=$PWD/dev/envtest-kubeconfig.yaml
# watch test objects
k get bastion -A -w

Similar to debugging unit tests, the stress tool can help hunting flakes in integration tests. Though, you might need to run less tests in parallel though (specified via -p) and have a bit more patience. Generally, reproducing flakes in integration tests is easier when stress-testing against an existing test environment instead of starting temporary individual control planes per test run.

Stress-test an envtest suite (not using gardener-apiserver):

# build a test binary
ginkgo build ./test/integration/resourcemanager/health

# prepare a test environment to run the test against
make start-envtest

# in another terminal session:
export KUBECONFIG=$PWD/dev/envtest-kubeconfig.yaml

# use same timeout settings like in CI
source ./hack/test-integration.env

# switch to test package directory like `go test`
cd ./test/integration/resourcemanager/health

# run the test in parallel and report any failures
stress -ignore "unable to grab random port" -p 16 ./health.test

Stress-test a gardenerenvtest suite (using gardener-apiserver):

# modify test/start-envtest to disable admission plugins and enable feature gates like in test suite...

# build a test binary
ginkgo build ./test/integration/controllermanager/bastion

# prepare a test environment including gardener-apiserver to run the test against
make start-envtest ENVTEST_TYPE=gardener

# in another terminal session:
export KUBECONFIG=$PWD/dev/envtest-kubeconfig.yaml

# use same timeout settings like in CI
source ./hack/test-integration.env

# switch to test package directory like `go test`
cd ./test/integration/controllermanager/bastion

# run the test in parallel and report any failures
stress -ignore "unable to grab random port" -p 16 ./bastion.test

Purpose of Integration Tests

  • Integration tests prove that multiple units are correctly integrated into a fully-functional component of the system.
  • Example components with multiple units:
    • A controller with its reconciler, watches, predicates, event handlers, queues, etc.
    • A webhook with its server, handler, decoder, and webhook configuration.
  • Integration tests set up a full component (including used libraries) and run it against a test environment close to the actual setup.
    • e.g., start controllers against a real Kubernetes control plane to catch bugs that can only happen when talking to a real API server.
    • Integration tests are generally more expensive to run (e.g., in terms of execution time).
  • Integration tests should not cover each and every detailed case.
    • Rather than that, cover a good portion of the “usual” cases that components will face during normal operation (positive and negative test cases).
    • Also, there is no need to cover all failure cases or all cases of predicates -> they should be covered in unit tests already.
    • Generally, not supposed to “generate test coverage” but to provide confidence that components work well.
  • As integration tests typically test only one component (or a cohesive set of components) isolated from others, they cannot catch bugs that occur when multiple controllers interact (could be discovered by e2e tests, though).
  • Rule of thumb: a new integration tests should be added for each new controller (an integration test doesn’t replace unit tests though).

Writing Integration Tests

  • Make sure to have a clean test environment on both test suite and test case level:
    • Set up dedicated test environments (envtest instances) per test suite.
    • Use dedicated namespaces per test suite:
      • Use GenerateName with a test-specific prefix: example test
      • Restrict the controller-runtime manager to the test namespace by setting manager.Options.Namespace: example test
      • Alternatively, use a test-specific prefix with a random suffix determined upfront: example test
        • This can be used to restrict webhooks to a dedicated test namespace: example test
      • This allows running a test in parallel against the same existing cluster for deflaking and stress testing: example PR
    • If the controller works on cluster-scoped resources:
      • Label the resources with a label specific to the test run, e.g. the test namespace’s name: example test
      • Restrict the manager’s cache for these objects with a corresponding label selector: example test
      • Alternatively, use a checksum of a random UUID using uuid.NewUUID() function: example test
      • This allows running a test in parallel against the same existing cluster for deflaking and stress testing, even if it works with cluster-scoped resources that are visible to all parallel test runs: example PR
    • Use dedicated test resources for each test case:
      • Use GenerateName: example test
      • Alternatively, use a checksum of a random UUID using uuid.NewUUID() function: example test
      • Logging the created object names is generally a good idea to support debugging failing or flaky tests: example test
      • Always delete all resources after the test case (e.g., via DeferCleanup) that were created for the test case
      • This avoids conflicts between test cases and cascading failures which distract from the actual root failures
    • Don’t tolerate already existing resources (~dirty test environment), code smell: ignoring already exist errors
  • Don’t use a cached client in test code (e.g., the one from a controller-runtime manager), always construct a dedicated test client (uncached): example test
  • Use asynchronous assertions: Eventually and Consistently.
    • Never Expect anything to happen synchronously (immediately).
    • Don’t use retry or wait until functions -> use Eventually, Consistently instead: example test
    • This allows to override the interval/timeout values from outside instead of hard-coding this in the test (see hack/ example PR
    • Beware of the default Eventually / Consistently timeouts / poll intervals: docs
    • Don’t set custom (high) timeouts and intervals in test code: example PR
      • iInstead, shorten sync period of controllers, overwrite intervals of the tested code, or use fake clocks: example test
    • Pass g Gomega to Eventually/Consistently and use g.Expect in it: docs, example test, example PR
    • Don’t forget to call {Eventually,Consistently}.Should(), otherwise the assertions always silently succeeds without errors: onsi/gomega#561
  • When using Gardener’s envtest (envtest.GardenerTestEnvironment):
    • Disable gardener-apiserver’s admission plugins that are not relevant to the integration test itself by passing --disable-admission-plugins: example test
    • This makes setup / teardown code simpler and ensures to only test code relevant to the tested component itself (but not the entire set of admission plugins)
    • e.g., you can disable the ShootValidator plugin to create Shoots that reference non-existing SecretBindings or disable the DeletionConfirmation plugin to delete Gardener resources without adding a deletion confirmation first.
  • Use a custom rate limiter for controllers in integration tests: example test
    • This can be used for limiting exponential backoff to shorten wait times.
    • Otherwise, if using the default rate limiter, exponential backoff might exceed the timeout of Eventually calls and cause flakes.

End-to-End (e2e) Tests (Using provider-local)

We run a suite of e2e tests on every pull request and periodically on the master branch. It uses a KinD cluster and skaffold to boostrap a full installation of Gardener based on the current revision, including provider-local. This allows us to run e2e tests in an isolated test environment and fully locally without any infrastructure interaction. The tests perform a set of operations on Shoot clusters, e.g. creating, deleting, hibernating and waking up.

These tests are executed in our prow instance at, see job definition and job history.

Running e2e Tests

You can also run these tests on your development machine, using the following commands:

make kind-up
export KUBECONFIG=$PWD/example/gardener-local/kind/local/kubeconfig
make gardener-up
make test-e2e-local  # alternatively: make test-e2e-local-simple

If you want to run a specific set of e2e test cases, you can also execute them using ./hack/ directly in combination with ginkgo label filters. For example:

./hack/ --label-filter "Shoot && credentials-rotation"

If you want to use an existing shoot instead of creating a new one for the test case and deleting it afterwards, you can specify the existing shoot via the following flags. This can be useful to speed up the development of e2e tests.

./hack/ --label-filter "Shoot && credentials-rotation" -- --project-namespace=garden-local --existing-shoot-name=local

For more information, see Developing Gardener Locally and Deploying Gardener Locally.

Debugging e2e Tests

When debugging e2e test failures in CI, logs of the cluster components can be very helpful. Our e2e test jobs export logs of all containers running in the kind cluster to prow’s artifacts storage. You can find them by clicking the Artifacts link in the top bar in prow’s job view and navigating to artifacts. This directory will contain all cluster component logs grouped by node.

Pull all artifacts using gsutil for searching and filtering the logs locally (use the path displayed in the artifacts view):

gsutil cp -r gs://gardener-prow/pr-logs/pull/gardener_gardener/6136/pull-gardener-e2e-kind/1542030416616099840/artifacts/gardener-local-control-plane /tmp

Purpose of e2e Tests

  • e2e tests provide a high level of confidence that our code runs as expected by users when deployed to production.
  • They are supposed to catch bugs resulting from interaction between multiple components.
  • Test cases should be as close as possible to real usage by end users:
    • You should test “from the perspective of the user” (or operator).
    • Example: I create a Shoot and expect to be able to connect to it via the provided kubeconfig.
    • Accordingly, don’t assert details of the system.
      • e.g., the user also wouldn’t expect that there is a kube-apiserver deployment in the seed, they rather expect that they can talk to it no matter how it is deployed
      • Only assert details of the system if the tested feature is not fully visible to the end-user and there is no other way of ensuring that the feature works reliably
      • e.g., the Shoot CA rotation is not fully visible to the user but is assertable by looking at the secrets in the Seed.
  • Pro: can be executed by developers and users without any real infrastructure (provider-local).
  • Con: they currently cannot be executed with real infrastructure (e.g., provider-aws), we will work on this as part of #6016.
  • Keep in mind that the tested scenario is still artificial in a sense of using default configuration, only a few objects, only a few config/settings combinations are covered.
    • We will never be able to cover the full “test matrix” and this should not be our goal.
    • Bugs will still be released and will still happen in production; we can’t avoid it.
    • Instead, we should add test cases for preventing bugs in features or settings that were frequently regressed: example PR
  • Usually e2e tests cover the “straight-forward cases”.
    • However, negative test cases can also be included, especially if they are important from the user’s perspective.

Writing e2e Tests

  • Always wrap API calls and similar things in Eventually blocks: example test
    • At this point, we are pretty much working with a distributed system and failures can happen anytime.
    • Wrapping calls in Eventually makes tests more stable and more realistic (usually, you wouldn’t call the system broken if a single API call fails because of a short connectivity issue).
  • Most of the points from writing integration tests are relevant for e2e tests as well (especially the points about asynchronous assertions).
  • In contrast to integration tests, in e2e tests, it might make sense to specify higher timeouts for Eventually calls, e.g., when waiting for a Shoot to be reconciled.
    • Generally, try to use the default settings for Eventually specified via the environment variables.
    • Only set higher timeouts if waiting for long-running reconciliations to be finished.

Gardener Upgrade Tests (Using provider-local)

Gardener upgrade tests setup a kind cluster and deploy Gardener version vX.X.X before upgrading it to a given version vY.Y.Y.

This allows verifying whether the current (unreleased) revision/branch (or a specific release) is compatible with the latest (or a specific other) release. The GARDENER_PREVIOUS_RELEASE and GARDENER_NEXT_RELEASE environment variables are used to specify the respective versions.

This helps understanding what happens or how the system reacts when Gardener upgrades from versions vX.X.X to vY.Y.Y for existing shoots in different states (creation/hibernation/wakeup/deletion). Gardener upgrade tests also help qualifying releases for all flavors (non-HA or HA with failure tolerance node/zone).

Just like E2E tests, upgrade tests also use a KinD cluster and skaffold for bootstrapping a full Gardener installation based on the current revision/branch, including provider-local. This allows running e2e tests in an isolated test environment, fully locally without any infrastructure interaction. The tests perform a set of operations on Shoot clusters, e.g. create, delete, hibernate and wake up.

Below is a sequence describing how the tests are performed.

  • Create a kind cluster.
  • Install Gardener version vX.X.X.
  • Run gardener pre-upgrade tests which are labeled with pre-upgrade.
  • Upgrade Gardener version from vX.X.X to vY.Y.Y.
  • Run gardener post-upgrade tests which are labeled with post-upgrade
  • Tear down seed and kind cluster.

How to Run Upgrade Tests Between Two Gardener Releases

Sometimes, we need to verify/qualify two Gardener releases when we upgrade from one version to another.
This can performed by fetching the two Gardener versions from the GitHub Gardener release page and setting appropriate env variables GARDENER_PREVIOUS_RELEASE, GARDENER_NEXT_RELEASE.

GARDENER_PREVIOUS_RELEASE – This env variable refers to a source revision/branch (or a specific release) which has to be installed first and then upgraded to version GARDENER_NEXT_RELEASE. By default, it fetches the latest release version from GitHub Gardener release page.

GARDENER_NEXT_RELEASE – This env variable refers to the target revision/branch (or a specific release) to be upgraded to after successful installation of GARDENER_PREVIOUS_RELEASE. By default, it considers the local HEAD revision, builds code, and installs Gardener from the current revision where the Gardener upgrade tests triggered.

  • make ci-e2e-kind-upgrade GARDENER_PREVIOUS_RELEASE=v1.60.0 GARDENER_NEXT_RELEASE=v1.61.0
  • make ci-e2e-kind-ha-single-zone-upgrade GARDENER_PREVIOUS_RELEASE=v1.60.0 GARDENER_NEXT_RELEASE=v1.61.0
  • make ci-e2e-kind-ha-multi-zone-upgrade GARDENER_PREVIOUS_RELEASE=v1.60.0 GARDENER_NEXT_RELEASE=v1.61.0

Purpose of Upgrade Tests

  • Tests will ensure that shoot clusters reconciled with the previous version of Gardener work as expected even with the next Gardener version.
  • This will reproduce or catch actual issues faced by end users.
  • One of the test cases ensures no downtime is faced by the end-users for shoots while upgrading Gardener if the shoot’s control-plane is configured as HA.

Writing Upgrade Tests

  • Tests are divided into two parts and labeled with pre-upgrade and post-upgrade labels.
  • An example test case which ensures a shoot which was hibernated in a previous Gardener release should wakeup as expected in next release:
    • Creating a shoot and hibernating a shoot is pre-upgrade test case which should be labeled pre-upgrade label.
    • Then wakeup a shoot and delete a shoot is post-upgrade test case which should be labeled post-upgrade label.

Test Machinery Tests

Please see Test Machinery Tests.

Purpose of Test Machinery Tests

  • Test machinery tests have to be executed against full-blown Gardener installations.
  • They can provide a very high level of confidence that an installation is functional in its current state, this includes: all Gardener components, Extensions, the used Cloud Infrastructure, all relevant settings/configuration.
  • This brings the following benefits:
    • They test more realistic scenarios than e2e tests (real configuration, real infrastructure, etc.).
    • Tests run “where the users are”.
  • However, this also brings significant drawbacks:
    • Tests are difficult to develop and maintain.
    • Tests require a full Gardener installation and cannot be executed in CI (on PR-level or against master).
    • Tests require real infrastructure (think cloud provider credentials, cost).
    • Using TestDefinitions under .test-defs requires a full test machinery installation.
    • Accordingly, tests are heavyweight and expensive to run.
    • Testing against real infrastructure can cause flakes sometimes (e.g., in outage situations).
    • Failures are hard to debug, because clusters are deleted after the test (for obvious cost reasons).
    • Bugs can only be caught, once it’s “too late”, i.e., when code is merged and deployed.
  • Today, test machinery tests cover a bigger “test matrix” (e.g., Shoot creation across infrastructures, kubernetes versions, machine image versions, etc.).
  • Test machinery also runs Kubernetes conformance tests.
  • However, because of the listed drawbacks, we should rather focus on augmenting our e2e tests, as we can run them locally and in CI in order to catch bugs before they get merged.
  • It’s still a good idea to add test machinery tests if a feature that is depending on some installation-specific configuration needs to be tested.

Writing Test Machinery Tests

  • Generally speaking, most points from writing integration tests and writing e2e tests apply here as well.
  • However, test machinery tests contain a lot of technical debt and existing code doesn’t follow these best practices.
  • As test machinery tests are out of our general focus, we don’t intend on reworking the tests soon or providing more guidance on how to write new ones.

Manual Tests

  • Manual tests can be useful when the cost of trying to automatically test certain functionality are too high.
  • Useful for PR verification, if a reviewer wants to verify that all cases are properly tested by automated tests.
  • Currently, it’s the simplest option for testing upgrade scenarios.
    • e.g. migration coding is probably best tested manually, as it’s a high effort to write an automated test for little benefit
  • Obviously, the need for manual tests should be kept at a bare minimum.
    • Instead, we should add e2e tests wherever sensible/valuable.
    • We want to implement some form of general upgrade tests as part of #6016.

18 - Testmachinery Tests

Test Machinery Tests

In order to automatically qualify Gardener releases, we execute a set of end-to-end tests using Test Machinery. This requires a full Gardener installation including infrastructure extensions, as well as a setup of Test Machinery itself. These tests operate on Shoot clusters across different Cloud Providers, using different supported Kubernetes versions and various configuration options (huge test matrix).

This manual gives an overview about test machinery tests in Gardener.


Gardener test machinery tests are split into two test suites that can be found under test/testmachinery/suites:

  • The Gardener Test Suite contains all tests that only require a running gardener instance.
  • The Shoot Test Suite contains all tests that require a predefined running shoot cluster.

The corresponding tests of a test suite are defined in the import statement of the suite definition (see shoot/run_suite_test.go) and their source code can be found under test/testmachinery.

The test directory is structured as follows:

├── e2e           # end-to-end tests (using provider-local)
│  └── shoot
├── framework     # helper code shared across integration, e2e and testmachinery tests
├── integration   # integration tests (envtests)
│  ├── controllermanager
│  ├── envtest
│  ├── resourcemanager
│  ├── scheduler
│  ├── shootmaintenance
│  └── ...
└── testmachinery # test machinery tests
   ├── gardener   # actual test cases imported by suites/gardener
   │  └── security
   ├── shoots     # actual test cases imported by suites/shoot
   │  ├── applications
   │  ├── care
   │  ├── logging
   │  ├── operatingsystem
   │  ├── operations
   │  └── vpntunnel
   ├── suites     # suites that run agains a running garden or shoot cluster
   │  ├── gardener
   │  └── shoot
   └── system     # suites that are used for building a full test flow
      ├── complete_reconcile
      ├── managed_seed_creation
      ├── managed_seed_deletion
      ├── shoot_cp_migration
      ├── shoot_creation
      ├── shoot_deletion
      ├── shoot_hibernation
      ├── shoot_hibernation_wakeup
      └── shoot_update

A suite can be executed by running the suite definition with ginkgo’s focus and skip flags to control the execution of specific labeled test. See the example below:

go test -timeout=0 -mod=vendor ./test/testmachinery/suites/shoot \
      --v -ginkgo.v -ginkgo.progress \
      --report-file=/tmp/report.json \                     # write elasticsearch formatted output to a file
      --disable-dump=false \                               # disables dumping of teh current state if a test fails
      -kubecfg=/path/to/gardener/kubeconfig \
      -shoot-name=<shoot-name> \                           # Name of the shoot to test
      -project-namespace=<gardener project namespace> \    # Name of the gardener project the test shoot resides
      -ginkgo.focus="\[RELEASE\]" \                        # Run all tests that are tagged as release
      -ginkgo.skip="\[SERIAL\]|\[DISRUPTIVE\]"             # Exclude all tests that are tagged SERIAL or DISRUPTIVE

Add a New Test

To add a new test the framework requires the following steps (step 1. and 2. can be skipped if the test is added to an existing package):

  1. Create a new test file e.g. test/testmachinery/shoot/security/my-sec-test.go
  2. Import the test into the appropriate test suite (gardener or shoot): import _ ""
  3. Define your test with the testframework. The framework will automatically add its initialization, cleanup and dump functions.
var _ = ginkgo.Describe("my suite", func(){
  f := framework.NewShootFramework(nil)

  f.Beta().CIt("my first test", func(ctx context.Context) {
    // testing ...

The newly created test can be tested by focusing the test with the default ginkgo focus f.Beta().FCIt("my first test", func(ctx context.Context) and running the shoot test suite with:

go test -timeout=0 -mod=vendor ./test/testmachinery/suites/shoot \
      --v -ginkgo.v -ginkgo.progress \
      --report-file=/tmp/report.json \                     # write elasticsearch formatted output to a file
      --disable-dump=false \                               # disables dumping of the current state if a test fails
      -kubecfg=/path/to/gardener/kubeconfig \
      -shoot-name=<shoot-name> \                           # Name of the shoot to test
      -project-namespace=<gardener project namespace> \
      -fenced=<true|false>                                 # Tested shoot is running in a fenced environment and cannot be reached by gardener

or for the gardener suite with:

go test -timeout=0 -mod=vendor ./test/testmachinery/suites/gardener \
      --v -ginkgo.v -ginkgo.progress \
      --report-file=/tmp/report.json \                     # write elasticsearch formatted output to a file
      --disable-dump=false \                               # disables dumping of the current state if a test fails
      -kubecfg=/path/to/gardener/kubeconfig \
      -project-namespace=<gardener project namespace>

⚠️ Make sure that you do not commit any focused specs as this feature is only intended for local development! Ginkgo will fail the test suite if there are any focused specs.

Alternatively, a test can be triggered by specifying a ginkgo focus regex with the name of the test e.g.

go test -timeout=0 -mod=vendor ./test/testmachinery/suites/gardener \
      --v -ginkgo.v -ginkgo.progress \
      --report-file=/tmp/report.json \                     # write elasticsearch formatted output to a file
      -kubecfg=/path/to/gardener/kubeconfig \
      -project-namespace=<gardener project namespace> \
      -ginkgo.focus="my first test"                        # regex to match test cases

Test Labels

Every test should be labeled by using the predefined labels available with every framework to have consistent labeling across all test machinery tests.

The labels are applied to every new It()/CIt() definition by:

f := framework.NewCommonFramework()
f.Default().Serial().It("my test") => "[DEFAULT] [SERIAL] my test"

f := framework.NewShootFramework()
f.Default().Serial().It("my test") => "[DEFAULT] [SERIAL] [SHOOT] my test"

f := framework.NewGardenerFramework()
f.Default().Serial().It("my test") => "[DEFAULT] [GARDENER] [SERIAL] my test"


  • Beta: Newly created tests with no experience on stableness should be first labeled as beta tests. They should be watched (and probably improved) until stable enough to be promoted to Default.
  • Default: Tests that were Beta before and proved to be stable are promoted to Default eventually. Default tests run more often, produce alerts and are considered during the release decision although they don’t necessarily block a release.
  • Release: Test are release relevant. A failing Release test blocks the release pipeline. Therefore, these tests need to be stable. Only tests proven to be stable will eventually be promoted to Release.

Behavior Labels:

  • Serial: The test should always be executed in serial with no other tests running, as it may impact other tests.
  • Destructive: The test is destructive. Which means that is runs with no other tests and may break Gardener or the shoot. Only create such tests if really necessary, as the execution will be expensive (neither Gardener nor the shoot can be reused in this case for other tests).


The framework directory contains all the necessary functions / utilities for running test machinery tests. For example, there are methods for creation/deletion of shoots, waiting for shoot deletion/creation, downloading/installing/deploying helm charts, logging, etc.

The framework itself consists of 3 different frameworks that expect different prerequisites and offer context specific functionality.

  • CommonFramework: The common framework is the base framework that handles logging and setup of commonly needed resources like helm. It also contains common functions for interacting with Kubernetes clusters like Waiting for resources to be ready or Exec into a running pod.
  • GardenerFramework contains all functions of the common framework and expects a running Gardener instance with the provided Gardener kubeconfig and a project namespace. It also contains functions to interact with gardener like Waiting for a shoot to be reconciled or Patch a shoot or Get a seed.
  • ShootFramework: contains all functions of the common and the gardener framework. It expects a running shoot cluster defined by the shoot’s name and namespace (project namespace). This framework contains functions to directly interact with the specific shoot.

The whole framework also includes commonly used checks, ginkgo wrapper, etc., as well as commonly used tests. Theses common application tests (like the guestbook test) can be used within multiple tests to have a default application (with ingress, deployment, stateful backend) to test external factors.


Every framework commandline flag can also be defined by a configuration file (the value of the configuration file is only used if a flag is not specified by commandline). The test suite searches for a configuration file (yaml is preferred) if the command line flag --config=/path/to/config/file is provided. A framework can be defined in the configuration file by just using the flag name as root key e.g.

verbose: debug
kubecfg: /kubeconfig/path
project-namespace: garden-it


The framework automatically writes the ginkgo default report to stdout and a specifically structured elastichsearch bulk report file to a specified location. The elastichsearch bulk report will write one json document per testcase and injects the metadata of the whole testsuite. An example document for one test case would look like the following document:

    "suite": {
        "name": "Shoot Test Suite",
        "phase": "Succeeded",
        "tests": 3,
        "failures": 1,
        "errors": 0,
        "time": 87.427
    "name": "Shoot application testing  [DEFAULT] [RELEASE] [SHOOT] should download shoot kubeconfig successfully",
    "shortName": "should download shoot kubeconfig successfully",
    "labels": [
    "phase": "Succeeded",
    "time": 0.724512057


The resources directory contains all the templates, helm config files (e.g., repositories.yaml, charts, and cache index which are downloaded upon the start of the test), shoot configs, etc.

├── charts
├── repository
│   └── repositories.yaml
└── templates
    ├── guestbook-app.yaml.tpl
    └── logger-app.yaml.tpl

There are two special directories that are dynamically filled with the correct test files:

  • charts - the charts will be downloaded and saved in this directory
  • repository - contains the repository.yaml file that the target helm repos will be read from and the cache where the stable-index.yaml file will be created

System Tests

This directory contains the system tests that have a special meaning for the testmachinery with their own Test Definition. Currently, these system tests consist of:

  • Shoot creation
  • Shoot deletion
  • Shoot Kubernetes update
  • Gardener Full reconcile check

Shoot Creation Test

Create Shoot test is meant to test shoot creation.

Example Run

go test -mod=vendor -timeout=0 ./test/testmachinery/system/shoot_creation \
  --v -ginkgo.v -ginkgo.progress \
  -kubecfg=$HOME/.kube/config \
  -shoot-name=$SHOOT_NAME \
  -cloud-profile=$CLOUDPROFILE \
  -seed=$SEED \
  -secret-binding=$SECRET_BINDING \
  -provider-type=$PROVIDER_TYPE \
  -region=$REGION \
  -k8s-version=$K8S_VERSION \
  -project-namespace=$PROJECT_NAMESPACE \
  -annotations=$SHOOT_ANNOTATIONS \
  -infrastructure-provider-config-filepath=$INFRASTRUCTURE_PROVIDER_CONFIG_FILEPATH \
  -controlplane-provider-config-filepath=$CONTROLPLANE_PROVIDER_CONFIG_FILEPATH \
  -workers-config-filepath=$$WORKERS_CONFIG_FILEPATH \
  -worker-zone=$ZONE \
  -networking-pods=$NETWORKING_PODS \
  -networking-services=$NETWORKING_SERVICES \
  -networking-nodes=$NETWORKING_NODES \

Shoot Deletion Test

Delete Shoot test is meant to test the deletion of a shoot.

Example Run

go test -mod=vendor -timeout=0 -ginkgo.v -ginkgo.progress \
  ./test/testmachinery/system/shoot_deletion \
  -kubecfg=$HOME/.kube/config \
  -shoot-name=$SHOOT_NAME \

Shoot Update Test

The Update Shoot test is meant to test the Kubernetes version update of a existing shoot. If no specific version is provided, the next patch version is automatically selected. If there is no available newer version, this test is a noop.

Example Run

go test -mod=vendor -timeout=0 ./test/testmachinery/system/shoot_update \
  --v -ginkgo.v -ginkgo.progress \
  -kubecfg=$HOME/.kube/config \
  -shoot-name=$SHOOT_NAME \
  -project-namespace=$PROJECT_NAMESPACE \

Gardener Full Reconcile Test

The Gardener Full Reconcile test is meant to test if all shoots of a Gardener instance are successfully reconciled.

Example Run

go test -mod=vendor -timeout=0 ./test/testmachinery/system/complete_reconcile \
  --v -ginkgo.v -ginkgo.progress \
  -kubecfg=$HOME/.kube/config \
  -project-namespace=$PROJECT_NAMESPACE \
  -gardenerVersion=$GARDENER_VERSION # needed to validate the last acted gardener version of a shoot