Kubernetes integration - development guidelines

This document provides various guidelines when developing for the GitLab Kubernetes integration.



Some Kubernetes operations, such as creating restricted project namespaces are performed on the GitLab Rails application. These operations are performed using a client library, and carry an element of risk. The operations are run as the same user running the GitLab Rails application. For more information, read the security section below.

Some Kubernetes operations, such as installing cluster applications are performed on one-off pods on the Kubernetes cluster itself. These installation pods are named install-<application_name> and are created within the gitlab-managed-apps namespace.

In terms of code organization, we generally add objects that represent Kubernetes resources in lib/gitlab/kubernetes.

Client library

We use the kubeclient gem to perform Kubernetes API calls. As the kubeclient gem does not support different API Groups (such as apis/rbac.authorization.k8s.io) from a single client, we have created a wrapper class, Gitlab::Kubernetes::KubeClient that enable you to achieve this.

Selected Kubernetes API groups are supported. Do add support for new API groups or methods to Gitlab::Kubernetes::KubeClient if you need to use them. New API groups or API group versions can be added to SUPPORTED_API_GROUPS - internally, this creates an internal client for that group. New methods can be added as a delegation to the relevant internal client.

Performance considerations

All calls to the Kubernetes API must be in a background process. Don’t perform Kubernetes API calls within a web request. This blocks webserver, and can lead to a denial-of-service (DoS) attack in GitLab as the Kubernetes cluster response times are outside of our control.

The easiest way to ensure your calls happen a background process is to delegate any such work to happen in a Sidekiq worker.

You may want to make calls to Kubernetes and return the response, but a background worker isn’t a good fit. Consider using reactive caching. For example:

  def calculate_reactive_cache!
    { pods: cluster.platform_kubernetes.kubeclient.get_pods }

  def pods
    with_reactive_cache do |data|


We have some WebMock stubs in KubernetesHelpers which can help with mocking out calls to Kubernetes API in your tests.

Amazon EKS integration

This section outlines the process for allowing a GitLab instance to create EKS clusters.

The following prerequisites are required:

A Customer AWS account. The EKS cluster is created in this account. The following resources must be present:

  • A provisioning role that has permissions to create the cluster and associated resources. It must list the GitLab AWS account as a trusted entity.
  • A VPC, management role, security group, and subnets for use by the cluster.

A GitLab AWS account. This is the account which performs the provisioning actions. The following resources must be present:

  • A service account with permissions to assume the provisioning role in the Customer account above.
  • Credentials for this service account configured in GitLab via the kubernetes section of gitlab.yml.

The process for creating a cluster is as follows:

  1. Using the :provision_role_external_id, GitLab assumes the role provided by :provision_role_arn and stores a set of temporary credentials on the provider record. By default these credentials are valid for one hour.
  2. A CloudFormation stack is created, based on the AWS CloudFormation EKS template. This triggers creation of all resources required for an EKS cluster.
  3. GitLab polls the status of the stack until all resources are ready, which takes somewhere between 10 and 15 minutes in most cases.
  4. When the stack is ready, GitLab stores the cluster details and generates another set of temporary credentials, this time to allow connecting to the cluster via kubeclient. These credentials are valid for one minute.
  5. GitLab configures the worker nodes so that they are able to authenticate to the cluster, and creates a service account for itself for future operations.
  6. Credentials that are no longer required are removed. This deletes the following attributes:

    • access_key_id
    • secret_access_key
    • session_token


Server Side Request Forgery (SSRF) attacks

As URLs for Kubernetes clusters are user controlled it is easily susceptible to Server Side Request Forgery (SSRF) attacks. You should understand the mitigation strategies if you are adding more API calls to a cluster.

Mitigation strategies include:

  1. Not allowing redirects to attacker controller resources: Kubeclient::KubeClient can be configured to prevent any redirects by passing in http_max_redirects: 0 as an option.
  2. Not exposing error messages: by doing so, we prevent attackers from triggering errors to expose results from attacker controlled requests. For example, we do not expose (or store) raw error messages:

    rescue Kubernetes::HttpError => e
      # bad
      # app.make_errored!("Kubernetes error: #{e.message}")
      # good
      app.make_errored!("Kubernetes error: #{e.error_code}")

Debugging Kubernetes integrations

Logs related to the Kubernetes integration can be found in kubernetes.log. On a local GDK install, these logs are present in log/kubernetes.log.

You can also follow the installation logs to debug issues related to installation. Once the installation/upgrade is underway, wait for the pod to be created. Then run the following to obtain the pods logs as they are written:

kubectl logs <pod_name> --follow -n gitlab-managed-apps