The Docker executor

GitLab Runner can use Docker to run jobs on user provided images. This is possible with the use of Docker executor.

The Docker executor when used with GitLab CI, connects to Docker Engine and runs each build in a separate and isolated container using the predefined image that is set up in .gitlab-ci.yml and in accordance in config.toml.

That way you can have a simple and reproducible build environment that can also run on your workstation. The added benefit is that you can test all the commands that we will explore later from your shell, rather than having to test them on a dedicated CI server.

The following configurations are supported:

Runner is installed on: Executor is: Container is running:
Windows docker-windows Windows
Windows docker Linux
Linux docker Linux

These configurations are not supported:

Runner is installed on: Executor is: Container is running:
Linux docker-windows Linux
Linux docker Windows
Linux docker-windows Windows
Windows docker Windows
Windows docker-windows Linux
GitLab Runner uses Docker Engine API v1.25 to talk to the Docker Engine. This means the minimum supported version of Docker on a Linux server is 1.13.0, on Windows Server it needs to be more recent to identify the Windows Server version.

Using Windows containers

Introduced in GitLab Runner 11.11.

To use Windows containers with the Docker executor, note the following information about limitations, supported Windows versions, and configuring a Windows Docker executor.

Nanoserver support

Introduced in GitLab Runner 13.6.

With the support for Powershell Core introduced in the Windows helper image, it is now possible to leverage the nanoserver variants for the helper image.

Limitations of Docker executor on Windows

The following are some limitations of using Windows containers with Docker executor:

  • Docker-in-Docker is not supported, since it’s not supported by Docker itself.
  • Interactive web terminals are not supported.
  • Host device mounting not supported.
  • When mounting a volume directory it has to exist, or Docker will fail to start the container, see #3754 for additional detail.
  • docker-windows executor can be run only using GitLab Runner running on Windows.
  • Linux containers on Windows are not supported, since they are still experimental. Read the relevant issue for more details.
  • Because of a limitation in Docker, if the destination path drive letter is not c:, paths are not supported for:

    This means values such as f:\\cache_dir are not supported, but f: is supported. However, if the destination path is on the c: drive, paths are also supported (for example c:\\cache_dir).

Supported Windows versions

GitLab Runner only supports the following versions of Windows which follows our support lifecycle for Windows:

  • Windows Server 20H2.
  • Windows Server 2004.
  • Windows Server 1809.

For future Windows Server versions, we have a future version support policy.

You can only run containers based on the same OS version that the Docker daemon is running on. For example, the following Windows Server Core images can be used:


Supported Docker versions

A Windows Server running GitLab Runner must be running a recent version of Docker because GitLab Runner uses Docker to detect what version of Windows Server is running.

A known version of Docker that doesn’t work with GitLab Runner is Docker 17.06 since Docker does not identify the version of Windows Server resulting in the following error:

unsupported Windows Version: Windows Server Datacenter

Read more about troubleshooting this.

Configuring a Windows Docker executor

When a runner is registered with c:\\cache as a source directory when passing the --docker-volumes or DOCKER_VOLUMES environment variable, there is a known issue.

Below is an example of the configuration for a simple Docker executor running Windows.

  name = "windows-docker-2019"
  url = ""
  token = "xxxxxxx"
  executor = "docker-windows"
    image = ""
    volumes = ["c:\\cache"]

For other configuration options for the Docker executor, see the advanced configuration section.


In GitLab Runner 12.9 and later, you can use services by enabling a network for each job.


The Docker executor divides the job into multiple steps:

  1. Prepare: Create and start the services.
  2. Pre-job: Clone, restore cache and download artifacts from previous stages. This is run on a special Docker image.
  3. Job: User build. This is run on the user-provided Docker image.
  4. Post-job: Create cache, upload artifacts to GitLab. This is run on a special Docker Image.

The special Docker image is based on Alpine Linux and contains all the tools required to run the prepare, pre-job, and post-job steps, like the Git and the GitLab Runner binaries for supporting caching and artifacts. You can find the definition of this special image in the official GitLab Runner repository.

The image keyword

The image keyword is the name of the Docker image that is present in the local Docker Engine (list all images with docker images) or any image that can be found at Docker Hub. For more information about images and Docker Hub please read the Docker Fundamentals documentation.

In short, with image we refer to the Docker image, which will be used to create a container on which your build will run.

If you don’t specify the namespace, Docker implies library which includes all official images. That’s why you’ll see many times the library part omitted in .gitlab-ci.yml and config.toml. For example you can define an image like image: ruby:2.7, which is a shortcut for image: library/ruby:2.7.

Then, for each Docker image there are tags, denoting the version of the image. These are defined with a colon (:) after the image name. For example, for Ruby you can see the supported tags at If you don’t specify a tag (like image: ruby), latest is implied.

The image you choose to run your build in via image directive must have a working shell in its operating system PATH. Supported shells are sh, bash, and pwsh (since 13.9) for Linux, and PowerShell for Windows. GitLab Runner cannot execute a command using the underlying OS system calls (such as exec).

The services keyword

The services keyword defines just another Docker image that is run during your build and is linked to the Docker image that the image keyword defines. This allows you to access the service image during build time.

The service image can run any application, but the most common use case is to run a database container, e.g., mysql. It’s easier and faster to use an existing image and run it as an additional container than install mysql every time the project is built.

You can see some widely used services examples in the relevant documentation of CI services examples.

If needed, you can assign an alias to each service.


Networking is required to connect services to a CI/CD job. Networking can also be used to run jobs in user-defined networks. You can use either legacy container links, or create a network for each job. We recommend creating a network for each job.

The default network mode uses Legacy container links with the default Docker bridge mode to link the job container with the services.

This mode can be used to configure how the networking stack is set up for the containers by using network_mode configuration parameter with one of the following values:

  • One of the standard Docker networking modes:
    • bridge: use the bridge network (default)
    • host: use the host’s network stack inside the container
    • none: no networking (not recommended)
  • Any other network_mode value is taken as the name of an already existing Docker network, which the build container should connect to.

For name resolution to work, Docker manipulates the /etc/hosts file in the container to include the service container hostname and alias. However, the service container is not able to resolve the container name. To resolve the container name, create a network for each job.

Linked containers share their environment variables.

Create a network for each job

Introduced in GitLab Runner 12.9.

This networking mode creates and uses a new user-defined Docker bridge network for each job. User-defined bridge networks are covered in detail in the Docker documentation.

Unlike legacy container links used in other network modes, Docker environment variables are not shared across the containers.

Docker networks might conflict with other networks on the host, including other Docker networks, if the CIDR ranges are already in use. The default Docker address pool can be configured by using default-address-pool in dockerd.

To enable this mode you must enable the FF_NETWORK_PER_BUILD feature flag.

When a job starts, a bridge network is created (similar to docker network create <network>). Upon creation, the service containers and the build job container are connected to this network.

Both the container running the job and the containers running the service can resolve each other’s hostnames and aliases. This functionality is provided by Docker.

The job container is resolvable by using the build alias as well, because the hostname is assigned by GitLab.

The network is removed at the end of the job.

Define image and services from .gitlab-ci.yml

You can simply define an image that will be used for all jobs and a list of services that you want to use during build time.

image: ruby:2.7

  - postgres:9.3

  - bundle install

  - bundle exec rake spec

It is also possible to define different images and services per job:

  - bundle install

  image: ruby:2.6
  - postgres:9.3
  - bundle exec rake spec

  image: ruby:2.7
  - postgres:9.4
  - bundle exec rake spec

Define image and services in config.toml

Look for the [runners.docker] section:

  image = "ruby:2.7"

  name = "mysql:latest"
  alias = "db"

  name = "redis:latest"
  alias = "cache"

The example above uses the array of tables syntax.

The image and services defined this way will be added to all builds run by that runner, so even if you don’t define an image inside .gitlab-ci.yml, the one defined in config.toml will be used.

Define an image from a private Docker registry

Starting with GitLab Runner 0.6.0, you are able to define images located to private registries that could also require authentication.

All you have to do is be explicit on the image definition in .gitlab-ci.yml.

image: my.registry.tld:5000/namepace/image:tag

In the example above, GitLab Runner will look at my.registry.tld:5000 for the image namespace/image:tag.

If the repository is private you need to authenticate your GitLab Runner in the registry. Read more on using a private Docker registry.

Accessing the services

Let’s say that you need a Wordpress instance to test some API integration with your application.

You can then use for example the tutum/wordpress as a service image in your .gitlab-ci.yml:

- tutum/wordpress:latest

When the build is run, tutum/wordpress will be started first and you will have access to it from your build container under the hostname tutum__wordpress and tutum-wordpress.

The GitLab Runner creates two alias hostnames for the service that you can use alternatively. The aliases are taken from the image name following these rules:

  1. Everything after : is stripped.
  2. For the first alias, the slash (/) is replaced with double underscores (__).
  3. For the second alias, the slash (/) is replaced with a single dash (-).

Using a private service image will strip any port given and apply the rules as described above. A service will result in hostname registry.gitlab-wp.com__tutum__wordpress and

Configuring services

Many services accept environment variables which allow you to easily change database names or set account names depending on the environment.

GitLab Runner 0.5.0 and up passes all YAML-defined variables to the created service containers.

For all possible configuration variables check the documentation of each image provided in their corresponding Docker Hub page.

All variables are passed to all services containers. It’s not designed to distinguish which variable should go where. Secure variables are only passed to the build container.

Mounting a directory in RAM

You can mount a path in RAM using tmpfs. This can speed up the time required to test if there is a lot of I/O related work, such as with databases. If you use the tmpfs and services_tmpfs options in the runner configuration, you can specify multiple paths, each with its own options. See the Docker reference for details. This is an example config.toml to mount the data directory for the official Mysql container in RAM.

  # For the main container
      "/var/lib/mysql" = "rw,noexec"

  # For services
      "/var/lib/mysql" = "rw,noexec"

Build directory in service

Since version 1.5 GitLab Runner mounts a /builds directory to all shared services.

See an issue:

PostgreSQL service example

See the specific documentation for using PostgreSQL as a service.

MySQL service example

See the specific documentation for using MySQL as a service.

The services health check

After the service is started, GitLab Runner waits some time for the service to be responsive. Currently, the Docker executor tries to open a TCP connection to the first exposed service in the service container.

You can see how it is implemented by checking this Go command.

The builds and cache storage

The Docker executor by default stores all builds in /builds/<namespace>/<project-name> and all caches in /cache (inside the container). You can overwrite the /builds and /cache directories by defining the builds_dir and cache_dir options under the [[runners]] section in config.toml. This will modify where the data are stored inside the container.

If you modify the /cache storage path, you also need to make sure to mark this directory as persistent by defining it in volumes = ["/my/cache/"] under the [runners.docker] section in config.toml.

Clearing Docker cache

Introduced in GitLab Runner 13.9, all created runner resources cleaned up.

GitLab Runner provides the clear-docker-cache script to remove old containers and volumes that can unnecessarily consume disk space.

Run clear-docker-cache regularly (using cron once per week, for example), ensuring a balance is struck between:

  • Maintaining some recent containers in the cache for performance.
  • Reclaiming disk space.

clear-docker-cache can remove old or unused containers and volumes that are created by the GitLab Runner. For a list of options, run the script with help option:

clear-docker-cache help

The default option is prune-volumes which the script will remove all unused containers (both dangling and unreferenced) and volumes.

Clearing old build images

The clear-docker-cache script will not remove the Docker images as they are not tagged by the GitLab Runner. You can however confirm the space that can be reclaimed by running the script with the space option as illustrated below:

clear-docker-cache space

Show docker disk usage

Images          14        9         1.306GB   545.8MB (41%)
Containers      19        18        115kB     0B (0%)
Local Volumes   0         0         0B        0B
Build Cache     0         0         0B        0B

Once you have confirmed the reclaimable space, run the docker system prune command that will remove all unused containers, networks, images (both dangling and unreferenced), and optionally, volumes that are not tagged by the GitLab Runner.

The persistent storage

The Docker executor can provide a persistent storage when running the containers. All directories defined under volumes = will be persistent between builds.

The volumes directive supports two types of storage:

  1. <path> - the dynamic storage. The <path> is persistent between subsequent runs of the same concurrent job for that project. The data is attached to a custom cache volume: runner-<short-token>-project-<id>-concurrent-<concurrency-id>-cache-<md5-of-path>.
  2. <host-path>:<path>[:<mode>] - the host-bound storage. The <path> is bound to <host-path> on the host system. The optional <mode> can specify that this storage is read-only or read-write (default).

The persistent storage for builds

If you make the /builds directory a host-bound storage, your builds will be stored in: /builds/<short-token>/<concurrent-id>/<namespace>/<project-name>, where:

  • <short-token> is a shortened version of the Runner’s token (first 8 letters)
  • <concurrent-id> is a unique number, identifying the local job ID on the particular runner in context of the project

The privileged mode

The Docker executor supports a number of options that allows fine-tuning of the build container. One of these options is the privileged mode.

Use Docker-in-Docker with privileged mode

The configured privileged flag is passed to the build container and all services, thus allowing to easily use the Docker-in-Docker approach.

First, configure your runner (config.toml) to run in privileged mode:

  executor = "docker"
    privileged = true

Then, make your build script (.gitlab-ci.yml) to use Docker-in-Docker container:

image: docker:git
- docker:dind

  - docker build -t my-image .
  - docker push my-image


The Docker executor doesn’t overwrite the ENTRYPOINT of a Docker image.

That means that if your image defines the ENTRYPOINT and doesn’t allow running scripts with CMD, the image will not work with the Docker executor.

With the use of ENTRYPOINT it is possible to create special Docker image that would run the build script in a custom environment, or in secure mode.

You may think of creating a Docker image that uses an ENTRYPOINT that doesn’t execute the build script, but does execute a predefined set of commands, for example to build the Docker image from your directory. In that case, you can run the build container in privileged mode, and make the build environment of the runner secure.

Consider the following example:

  1. Create a new Dockerfile:

    FROM docker:dind
    ADD / /
    ENTRYPOINT ["/bin/sh", "/"]
  2. Create a bash script ( that will be used as the ENTRYPOINT:

    dind docker daemon
        --host=unix:///var/run/docker.sock \
        --host=tcp:// \
        --storage-driver=vf &
    docker build -t "$BUILD_IMAGE" .
    docker push "$BUILD_IMAGE"
  3. Push the image to the Docker registry.

  4. Run Docker executor in privileged mode. In config.toml define:

      executor = "docker"
        privileged = true
  5. In your project use the following .gitlab-ci.yml:

      BUILD_IMAGE: my.image
      image: my/docker-build:image
      - Dummy Script

This is just one of the examples. With this approach the possibilities are limitless.

How pull policies work

When using the docker or docker+machine executors, you can set the pull_policy parameter in the runner config.toml file as described in the configuration docs’ Docker section.

This parameter defines how the runner works when pulling Docker images (for both image and services keywords). You can set it to a single value, or a list of pull policies, which will be attempted in order until an image is pulled successfully.

If you don’t set any value for the pull_policy parameter, then the runner will use the always pull policy as the default value.

Now let’s see how these policies work.

Using the never pull policy

The never pull policy disables images pulling completely. If you set the pull_policy parameter of a runner to never, then users will be able to use only the images that have been manually pulled on the Docker host the runner runs on.

If an image cannot be found locally, then the runner will fail the build with an error similar to:

Pulling docker image local_image:latest ...
ERROR: Build failed: Error: image local_image:latest not found

When to use the never pull policy

The never pull policy should be used if you want or need to have a full control on which images are used by the runner’s users. It is a good choice for private runners that are dedicated to a project where only specific images can be used (not publicly available on any registries).

When not to use the never pull policy

The never pull policy will not work properly with most of auto-scaled Docker executor use cases. Because of how auto-scaling works, the never pull policy may be usable only when using a pre-defined cloud instance images for chosen cloud provider. The image needs to contain installed Docker Engine and local copy of used images.

Using the if-not-present pull policy

When the if-not-present pull policy is used, the runner will first check if the image is present locally. If it is, then the local version of image will be used. Otherwise, the runner will try to pull the image.

When to use the if-not-present pull policy

The if-not-present pull policy is a good choice if you want to use images pulled from remote registries, but you want to reduce time spent on analyzing image layers difference when using heavy and rarely updated images. In that case, you will need once in a while to manually remove the image from the local Docker Engine store to force the update of the image.

It is also the good choice if you need to use images that are built and available only locally, but on the other hand, also need to allow to pull images from remote registries.

When not to use the if-not-present pull policy

The if-not-present pull policy should not be used if your builds use images that are updated frequently and need to be used in most recent versions. In such a situation, the network load reduction created by this policy may be less worthy than the necessity of the very frequent deletion of local copies of images.

This pull policy should also not be used if your runner can be used by different users which should not have access to private images used by each other. Especially do not use this pull policy for shared runners.

To understand why the if-not-present pull policy creates security issues when used with private images, read the security considerations documentation.

Using the always pull policy

The always pull policy will ensure that the image is always pulled. When always is used, the runner will try to pull the image even if a local copy is available. The caching semantics of the underlying image provider make this policy efficient. The pull attempt is fast because all image layers are cached.

If the image is not found, then the build will fail with an error similar to:

Pulling docker image registry.tld/my/image:latest ...
ERROR: Build failed: Error: image registry.tld/my/image:latest not found

When using the always pull policy in GitLab Runner versions older than v1.8, it could fall back to the local copy of an image and print a warning:

Pulling docker image registry.tld/my/image:latest ...
WARNING: Cannot pull the latest version of image registry.tld/my/image:latest : Error: image registry.tld/my/image:latest not found
WARNING: Locally found image will be used instead.

This was changed in GitLab Runner v1.8.

When to use the always pull policy

The always pull policy should be used if your runner is publicly available and configured as a shared runner in your GitLab instance. It is the only pull policy that can be considered as secure when the runner will be used with private images.

This is also a good choice if you want to force users to always use the newest images.

Also, this will be the best solution for an auto-scaled configuration of the runner.

When not to use the always pull policy

The always pull policy will definitely not work if you need to use locally stored images. In this case, the runner will skip the local copy of the image and try to pull it from the remote registry. If the image was built locally and doesn’t exist in any public registry (and especially in the default Docker registry), the build will fail with:

Pulling docker image local_image:latest ...
ERROR: Build failed: Error: image local_image:latest not found

Using multiple pull policies

Introduced in GitLab Runner 13.8.

The pull_policy parameter allows you to specify a list of pull policies. The policies in the list will be attempted in order from left to right until a pull attempt is successful, or the list is exhausted.

This functionality can be useful when the Docker registry is not available and you need to increase job resiliency. If you use the always policy and the registry is not available, the job fails even if the desired image is cached locally.

To overcome that behavior, you can add additional fallback pull policies that execute in case of failure. By adding a second pull policy value of if-not-present, the runner finds any locally-cached Docker image layers:

  pull_policy = ["always", "if-not-present"]

Any failure to fetch the Docker image causes the runner to attempt the following pull policy. Examples include an HTTP 403 Forbidden or an HTTP 500 Internal Server Error response from the repository.

Note that the security implications mentioned in the When not to use this pull policy? sub-section of the Using the if-not-present pull policy section still apply, so you should be aware of the security implications and read the security considerations documentation.

Using Docker executor with image alpine:latest ...
Pulling docker image alpine:latest ...
WARNING: Failed to pull image with policy "always": Error response from daemon: received unexpected HTTP status: 502 Bad Gateway (docker.go:143:0s)
Attempt #2: Trying "if-not-present" pull policy
Using locally found image version due to "if-not-present" pull policy

Docker vs Docker-SSH (and Docker+Machine vs Docker-SSH+Machine)

Starting with GitLab Runner 10.0, both Docker-SSH and Docker-SSH+machine executors are deprecated and will be removed in one of the upcoming releases.

We provided a support for a special type of Docker executor, namely Docker-SSH (and the autoscaled version: Docker-SSH+Machine). Docker-SSH uses the same logic as the Docker executor, but instead of executing the script directly, it uses an SSH client to connect to the build container.

Docker-SSH then connects to the SSH server that is running inside the container using its internal IP.

This executor is no longer maintained and will be removed in the near future.