Configuring runners (FREE)

If you have installed your own runners, you can configure and secure them in GitLab.

If you need to configure runners on the machine where you installed GitLab Runner, see the GitLab Runner documentation.

Manually clear the runner cache

Read clearing the cache.

Set maximum job timeout for a runner

For each runner, you can specify a maximum job timeout. This timeout, if smaller than the project defined timeout, takes precedence.

This feature can be used to prevent your shared runner from being overwhelmed by a project that has jobs with a long timeout (for example, one week).

On GitLab.com, you cannot override the job timeout for shared runners and must use the project defined timeout.

To set the maximum job timeout:

1. In a project, go to Settings > CI/CD > Runners.
2. Select your specific runner to edit the settings.
3. Enter a value under Maximum job timeout. Must be 10 minutes or more. If not defined, the project's job timeout setting is used.
4. Select Save changes.

How this feature works:

Example 1 - Runner timeout bigger than project timeout

1. You set the maximum job timeout for a runner to 24 hours
2. You set the CI/CD Timeout for a project to 2 hours
3. You start a job
4. The job, if running longer, times out after 2 hours

Example 2 - Runner timeout not configured

1. You remove the maximum job timeout configuration from a runner
2. You set the CI/CD Timeout for a project to 2 hours
3. You start a job
4. The job, if running longer, times out after 2 hours

Example 3 - Runner timeout smaller than project timeout

1. You set the maximum job timeout for a runner to 30 minutes
2. You set the CI/CD Timeout for a project to 2 hours
3. You start a job
4. The job, if running longer, times out after 30 minutes

Be careful with sensitive information

With some runner executors, if you can run a job on the runner, you can get full access to the file system, and thus any code it runs as well as the token of the runner. With shared runners, this means that anyone that runs jobs on the runner, can access anyone else's code that runs on the runner.

In addition, because you can get access to the runner token, it is possible to create a clone of a runner and submit false jobs, for example.

The above is easily avoided by restricting the usage of shared runners on large public GitLab instances, controlling access to your GitLab instance, and using more secure runner executors.

Prevent runners from revealing sensitive information

Introduced in GitLab 10.0.

You can protect runners so they don't reveal sensitive information. When a runner is protected, the runner picks jobs created on protected branches or protected tags only, and ignores other jobs.

To protect or unprotect a runner:

1. Go to the project's Settings > CI/CD and expand the Runners section.
2. Find the runner you want to protect or unprotect. Make sure it's enabled.
3. Select the pencil button.
4. Check the Protected option.
5. Select Save changes.

Forks

Whenever a project is forked, it copies the settings of the jobs that relate to it. This means that if you have shared runners set up for a project and someone forks that project, the shared runners serve jobs of this project.

Attack vectors in runners

Mentioned briefly earlier, but the following things of runners can be exploited. We're always looking for contributions that can mitigate these Security Considerations.

Reset the runner registration token for a project

If you think that a registration token for a project was revealed, you should reset it. A registration token can be used to register another runner for the project. That new runner may then be used to obtain the values of secret variables or to clone project code.

To reset the registration token:

1. Go to the project's Settings > CI/CD.
2. Expand the General pipelines settings section.
3. Find the Runner token form field and select Reveal value.
4. Delete the value and save the form.
5. After the page is refreshed, expand the Runners settings section and check the registration token - it should be changed.

From now on the old token is no longer valid and does not register any new runners to the project. If you are using any tools to provision and register new runners, the tokens used in those tools should be updated to reflect the value of the new token.

Reset the runner authentication token

If you think that an authentication token for a runner was revealed, you should reset it. An attacker could use the token to clone a runner.

To reset the authentication token, unregister the runner and then register it again.

To verify that the previous authentication token has been revoked, use the Runners API.

Determine the IP address of a runner

Introduced in GitLab 10.6.

It may be useful to know the IP address of a runner so you can troubleshoot issues with that runner. GitLab stores and displays the IP address by viewing the source of the HTTP requests it makes to GitLab when polling for jobs. The IP address is always kept up to date so if the runner IP changes it automatically updates in GitLab.

The IP address for shared runners and specific runners can be found in different places.

Determine the IP address of a shared runner

To view the IP address of a shared runner you must have administrator access to the GitLab instance. To determine this:

1. On the top bar, select Menu > Admin.
2. On the left sidebar, select Overview > Runners.
3. Find the runner in the table and view the IP Address column.

Determine the IP address of a specific runner

To can find the IP address of a runner for a specific project, you must have the Owner role for the project.

1. Go to the project's Settings > CI/CD and expand the Runners section.
2. On the details page you should see a row for IP Address.

Use tags to control which jobs a runner can run

You must set up a runner to be able to run all the different types of jobs that it may encounter on the projects it's shared over. This would be problematic for large amounts of projects, if it weren't for tags.

GitLab CI/CD tags are not the same as Git tags. GitLab CI/CD tags are associated with runners. Git tags are associated with commits.

By tagging a runner for the types of jobs it can handle, you can make sure shared runners will only run the jobs they are equipped to run.

For instance, at GitLab we have runners tagged with rails if they contain the appropriate dependencies to run Rails test suites.

Set a runner to run untagged jobs

When you register a runner, its default behavior is to only pick tagged jobs. To change this, you must have the Owner role for the project.

To make a runner pick untagged jobs:

1. Go to the project's Settings > CI/CD and expand the Runners section.
2. Find the runner you want to pick untagged jobs and make sure it's enabled.
3. Select the pencil button.
4. Check the Run untagged jobs option.
5. Select Save changes for the changes to take effect.

NOTE: The runner tags list can not be empty when it's not allowed to pick untagged jobs.

Below are some example scenarios of different variations.

Runner runs only tagged jobs

The following examples illustrate the potential impact of the runner being set to run only tagged jobs.

Example 1:

1. The runner is configured to run only tagged jobs and has the docker tag.
2. A job that has a hello tag is executed and stuck.

Example 2:

1. The runner is configured to run only tagged jobs and has the docker tag.
2. A job that has a docker tag is executed and run.

Example 3:

1. The runner is configured to run only tagged jobs and has the docker tag.
2. A job that has no tags defined is executed and stuck.

Runner is allowed to run untagged jobs

The following examples illustrate the potential impact of the runner being set to run tagged and untagged jobs.

Example 1:

1. The runner is configured to run untagged jobs and has the docker tag.
2. A job that has no tags defined is executed and run.
3. A second job that has a docker tag defined is executed and run.

Example 2:

1. The runner is configured to run untagged jobs and has no tags defined.
2. A job that has no tags defined is executed and run.
3. A second job that has a docker tag defined is stuck.

Use tags to run jobs on different platforms

You can use tags to run different jobs on different platforms. For example, if you have an OS X runner with tag osx and a Windows runner with tag windows, you can run a job on each platform:

windows job:
  stage:
    - build
  tags:
    - windows
  script:
    - echo Hello, %USERNAME%!

osx job:
  stage:
    - build
  tags:
    - osx
  script:
    - echo "Hello, $USER!"

Use CI/CD variables in tags

Introduced in GitLab 14.1.

You can use CI/CD variables with tags for dynamic runner selection:

variables:
  KUBERNETES_RUNNER: kubernetes

  job:
    tags:
      - docker
      - $KUBERNETES_RUNNER
    script:
      - echo "Hello runner selector feature"

Runner statuses

A runner can have one of the following statuses.

Configure runner behavior with variables

You can use CI/CD variables to configure runner Git behavior globally or for individual jobs:

• GIT_STRATEGY
• GIT_SUBMODULE_STRATEGY
• GIT_CHECKOUT
• GIT_CLEAN_FLAGS
• GIT_FETCH_EXTRA_FLAGS
• GIT_SUBMODULE_UPDATE_FLAGS
• GIT_DEPTH (shallow cloning)
• GIT_CLONE_PATH (custom build directories)
• TRANSFER_METER_FREQUENCY (artifact/cache meter update frequency)
• ARTIFACT_COMPRESSION_LEVEL (artifact archiver compression level)
• CACHE_COMPRESSION_LEVEL (cache archiver compression level)
• CACHE_REQUEST_TIMEOUT (cache request timeout)

You can also use variables to configure how many times a runner attempts certain stages of job execution.

When using the Kubernetes executor, you can use variables to override Kubernetes CPU and memory allocations for requests and limits.

Git strategy

• Introduced in GitLab 8.9 as an experimental feature.
• GIT_STRATEGY=none requires GitLab Runner v1.7+.

You can set the GIT_STRATEGY used to fetch the repository content, either globally or per-job in the variables section:

variables:
  GIT_STRATEGY: clone

There are three possible values: clone, fetch, and none. If left unspecified, jobs use the project's pipeline setting.

clone is the slowest option. It clones the repository from scratch for every job, ensuring that the local working copy is always pristine. If an existing worktree is found, it is removed before cloning.

fetch is faster as it re-uses the local working copy (falling back to clone if it does not exist). git clean is used to undo any changes made by the last job, and git fetch is used to retrieve commits made after the last job ran.

However, fetch does require access to the previous worktree. This works well when using the shell or docker executor because these try to preserve worktrees and try to re-use them by default.

This has limitations when using the Docker Machine executor.

A Git strategy of none also re-uses the local working copy, but skips all Git operations normally done by GitLab. GitLab Runner pre-clone scripts are also skipped, if present. This strategy could mean you need to add fetch and checkout commands to your .gitlab-ci.yml script.

It can be used for jobs that operate exclusively on artifacts, like a deployment job. Git repository data may be present, but it's likely out of date. You should only rely on files brought into the local working copy from cache or artifacts.

Git submodule strategy

Requires GitLab Runner v1.10+.

The GIT_SUBMODULE_STRATEGY variable is used to control if / how Git submodules are included when fetching the code before a build. You can set them globally or per-job in the variables section.

There are three possible values: none, normal, and recursive:

• none means that submodules are not included when fetching the project code. This is the default, which matches the pre-v1.10 behavior.

• normal means that only the top-level submodules are included. It's equivalent to:

  git submodule sync
  git submodule update --init

• recursive means that all submodules (including submodules of submodules) are included. This feature needs Git v1.8.1 and later. When using a GitLab Runner with an executor not based on Docker, make sure the Git version meets that requirement. It's equivalent to:

  git submodule sync --recursive
  git submodule update --init --recursive

For this feature to work correctly, the submodules must be configured (in .gitmodules) with either:

• the HTTP(S) URL of a publicly-accessible repository, or
• a relative path to another repository on the same GitLab server. See the Git submodules documentation.

You can provide additional flags to control advanced behavior using GIT_SUBMODULE_UPDATE_FLAGS.

Git checkout

Introduced in GitLab Runner 9.3.

The GIT_CHECKOUT variable can be used when the GIT_STRATEGY is set to either clone or fetch to specify whether a git checkout should be run. If not specified, it defaults to true. You can set them globally or per-job in the variables section.

If set to false, the runner:

• when doing fetch - updates the repository and leaves the working copy on the current revision,
• when doing clone - clones the repository and leaves the working copy on the default branch.

If GIT_CHECKOUT is set to true, both clone and fetch work the same way. The runner checks out the working copy of a revision related to the CI pipeline:

variables:
  GIT_STRATEGY: clone
  GIT_CHECKOUT: "false"
script:
  - git checkout -B master origin/master
  - git merge $CI_COMMIT_SHA

Git clean flags

Introduced in GitLab Runner 11.10

The GIT_CLEAN_FLAGS variable is used to control the default behavior of git clean after checking out the sources. You can set it globally or per-job in the variables section.

GIT_CLEAN_FLAGS accepts all possible options of the git clean command.

git clean is disabled if GIT_CHECKOUT: "false" is specified.

If GIT_CLEAN_FLAGS is:

• Not specified, git clean flags default to -ffdx.
• Given the value none, git clean is not executed.

For example:

variables:
  GIT_CLEAN_FLAGS: -ffdx -e cache/
script:
  - ls -al cache/

Git fetch extra flags

Introduced in GitLab Runner 13.1.

Use the GIT_FETCH_EXTRA_FLAGS variable to control the behavior of git fetch. You can set it globally or per-job in the variables section.

GIT_FETCH_EXTRA_FLAGS accepts all options of the git fetch command. However, GIT_FETCH_EXTRA_FLAGS flags are appended after the default flags that can't be modified.

The default flags are:

• GIT_DEPTH.
• The list of refspecs.
• A remote called origin.

If GIT_FETCH_EXTRA_FLAGS is:

• Not specified, git fetch flags default to --prune --quiet along with the default flags.
• Given the value none, git fetch is executed only with the default flags.

For example, the default flags are --prune --quiet, so you can make git fetch more verbose by overriding this with just --prune:

variables:
  GIT_FETCH_EXTRA_FLAGS: --prune
script:
  - ls -al cache/

The configuration above results in git fetch being called this way:

git fetch origin $REFSPECS --depth 50  --prune

Where $REFSPECS is a value provided to the runner internally by GitLab.

Git submodule update flags

Introduced in GitLab Runner 14.8.

Use the GIT_SUBMODULE_UPDATE_FLAGS variable to control the behavior of git submodule update when GIT_SUBMODULE_STRATEGY is set to either normal or recursive. You can set it globally or per-job in the variables section.

GIT_SUBMODULE_UPDATE_FLAGS accepts all options of the git submodule update subcommand. However, note that GIT_SUBMODULE_UPDATE_FLAGS flags are appended after a few default flags:

• --init, if GIT_SUBMODULE_STRATEGY was set to normal or recursive.
• --recursive, if GIT_SUBMODULE_STRATEGY was set to recursive.
• GIT_DEPTH. See the default value below.

Git honors the last occurrence of a flag in the list of arguments, so manually providing them in GIT_SUBMODULE_UPDATE_FLAGS will also override these default flags.

You can use this variable to fetch the latest remote HEAD instead of the commit tracked, in the repository, or to speed up the checkout by fetching submodules in multiple parallel jobs:

variables:
  GIT_SUBMODULE_STRATEGY: recursive
  GIT_SUBMODULE_UPDATE_FLAGS: --remote --jobs 4
script:
  - ls -al .git/modules/

The configuration above results in git submodule update being called this way:

git submodule update --init --depth 50 --recursive --remote --jobs 4

WARNING: You should be aware of the implications for the security, stability, and reproducibility of your builds when using the --remote flag. In most cases, it is better to explicitly track submodule commits as designed, and update them using an auto-remediation/dependency bot.

Shallow cloning

Introduced in GitLab 8.9 as an experimental feature.

You can specify the depth of fetching and cloning using GIT_DEPTH. GIT_DEPTH does a shallow clone of the repository and can significantly speed up cloning. It can be helpful for repositories with a large number of commits or old, large binaries. The value is passed to git fetch and git clone.

In GitLab 12.0 and later, newly-created projects automatically have a default git depth value of 50.

If you use a depth of 1 and have a queue of jobs or retry jobs, jobs may fail.

Git fetching and cloning is based on a ref, such as a branch name, so runners can't clone a specific commit SHA. If multiple jobs are in the queue, or you're retrying an old job, the commit to be tested must be within the Git history that is cloned. Setting too small a value for GIT_DEPTH can make it impossible to run these old commits and unresolved reference is displayed in job logs. You should then reconsider changing GIT_DEPTH to a higher value.

Jobs that rely on git describe may not work correctly when GIT_DEPTH is set since only part of the Git history is present.

To fetch or clone only the last 3 commits:

variables:
  GIT_DEPTH: "3"

You can set it globally or per-job in the variables section.

Custom build directories

Introduced in GitLab Runner 11.10.

By default, GitLab Runner clones the repository in a unique subpath of the $CI_BUILDS_DIR directory. However, your project might require the code in a specific directory (Go projects, for example). In that case, you can specify the GIT_CLONE_PATH variable to tell the runner the directory to clone the repository in:

variables:
  GIT_CLONE_PATH: $CI_BUILDS_DIR/project-name

test:
  script:
    - pwd

The GIT_CLONE_PATH has to always be within $CI_BUILDS_DIR. The directory set in $CI_BUILDS_DIR is dependent on executor and configuration of runners.builds_dir setting.

This can only be used when custom_build_dir is enabled in the runner's configuration. This is the default configuration for the docker and kubernetes executors.

Handling concurrency

An executor that uses a concurrency greater than 1 might lead to failures. Multiple jobs might be working on the same directory if the builds_dir is shared between jobs.

The runner does not try to prevent this situation. It's up to the administrator and developers to comply with the requirements of runner configuration.

To avoid this scenario, you can use a unique path within $CI_BUILDS_DIR, because runner exposes two additional variables that provide a unique ID of concurrency:

• $CI_CONCURRENT_ID: Unique ID for all jobs running within the given executor.
• $CI_CONCURRENT_PROJECT_ID: Unique ID for all jobs running within the given executor and project.

The most stable configuration that should work well in any scenario and on any executor is to use $CI_CONCURRENT_ID in the GIT_CLONE_PATH. For example:

variables:
  GIT_CLONE_PATH: $CI_BUILDS_DIR/$CI_CONCURRENT_ID/project-name

test:
  script:
    - pwd

The $CI_CONCURRENT_PROJECT_ID should be used in conjunction with $CI_PROJECT_PATH as the $CI_PROJECT_PATH provides a path of a repository. That is, group/subgroup/project. For example:

variables:
  GIT_CLONE_PATH: $CI_BUILDS_DIR/$CI_CONCURRENT_ID/$CI_PROJECT_PATH

test:
  script:
    - pwd

Nested paths

The value of GIT_CLONE_PATH is expanded once and nesting variables within is not supported.

For example, you define both the variables below in your .gitlab-ci.yml file:

variables:
  GOPATH: $CI_BUILDS_DIR/go
  GIT_CLONE_PATH: $GOPATH/src/namespace/project

The value of GIT_CLONE_PATH is expanded once into $CI_BUILDS_DIR/go/src/namespace/project, and results in failure because $CI_BUILDS_DIR is not expanded.

Job stages attempts

Introduced in GitLab, it requires GitLab Runner v1.9+.

You can set the number of attempts that the running job tries to execute the following stages:

The default is one single attempt.

Example:

variables:
  GET_SOURCES_ATTEMPTS: 3

You can set them globally or per-job in the variables section.

System calls not available on GitLab.com shared runners

GitLab.com shared runners run on CoreOS. This means that you cannot use some system calls, like getlogin, from the C standard library.

Artifact and cache settings

Introduced in GitLab Runner 13.9.

Artifact and cache settings control the compression ratio of artifacts and caches. Use these settings to specify the size of the archive produced by a job.

• On a slow network, uploads might be faster for smaller archives.
• On a fast network where bandwidth and storage are not a concern, uploads might be faster using the fastest compression ratio, despite the archive produced being larger.

For GitLab Pages to serve HTTP Range requests, artifacts should use the ARTIFACT_COMPRESSION_LEVEL: fastest setting, as only uncompressed zip archives support this feature.

A meter can be enabled to provide the rate of transfer for uploads and downloads.

You can set a maximum time for cache upload and download with the CACHE_REQUEST_TIMEOUT setting. This setting can be useful when slow cache uploads substantially increase the duration of your job.

variables:
  # output upload and download progress every 2 seconds
  TRANSFER_METER_FREQUENCY: "2s"

  # Use fast compression for artifacts, resulting in larger archives
  ARTIFACT_COMPRESSION_LEVEL: "fast"

  # Use no compression for caches
  CACHE_COMPRESSION_LEVEL: "fastest"

  # Set maximum duration of cache upload and download
  CACHE_REQUEST_TIMEOUT: 5

Artifact attestation

Introduced in GitLab Runner 15.1.

GitLab Runner can generate and produce attestation metadata for all build artifacts. To enable this feature, you must set the RUNNER_GENERATE_ARTIFACTS_METADATA environment variable to true. This variable can either be set globally or it can be set for individual jobs. The metadata is in rendered in a plain text .json file that's stored with the artifact. The file name is as follows: {JOB_ID}-artifacts-metadata.json.

Attestation format

The attestation metadata is generated in the in-toto attestation format for spec version v0.1. The following fields are populated by default:

An example of an attestation that the GitLab Runner might generate is as follows:

{
    "_type": "https://gitlab.com/gitlab-org/gitlab-runner/-/blob/v15.1.0/PROVENANCE.md",
    "subject": [
        {
            "name": "script.sh",
            "digest": {
                "sha256": "f5ae5ced234922eebe6461d32228ba8ab9c3d0c0f3983a3bef707e6e1a1ab52a"
            }
        }
    ],
    "predicateType": "https://slsa.dev/provenance/v0.2",
    "predicate": {
        "buildType": "https://gitlab.com/gitlab-org/gitlab-runner/-/blob/v15.1.0/PROVENANCE.md",
        "builder": {
            "id": "https://gitlab.com/ggeorgiev_gitlab/playground/-/runners/14811533"
        },
        "invocation": {
            "configSource": {
                "uri": "https://gitlab.com/ggeorgiev_gitlab/playground",
                "digest": {
                    "sha256": "f0582e2c9a16b5cc2cde90e8be8f1b50fd67c631"
                },
                "entryPoint": "whoami shell"
            },
            "environment": {
                "name": "local",
                "executor": "shell",
                "architecture": "amd64"
            },
            "parameters": {
                "CI_PIPELINE_ID": "",
                "CI_PIPELINE_URL": "",
                // All other CI variable names are listed here. Values are always represented as empty strings to avoid leaking secrets.
            }
        },
        "metadata": {
            "buildStartedOn": "2022-06-17T00:47:27+03:00",
            "buildFinishedOn": "2022-06-17T00:47:28+03:00",
            "completeness": {
                "parameters": true,
                "environment": true,
                "materials": false
            },
            "reproducible": false
        },
        "materials": []
    }
}

Staging directory

Introduced in GitLab Runner 15.0.

If you do not want to archive cache and artifacts in the system's default temporary directory, you can specify a different directory.

You might need to change the directory if your system's default temporary path has constraints. If you use a fast disk for the directory location, it can also improve performance.

To change the directory, set ARCHIVER_STAGING_DIR as a variable in your CI job, or use a runner variable when you register the runner (gitlab register --env ARCHIVER_STAGING_DIR=<dir>).

The directory you specify is used as the location for downloading artifacts prior to extraction. If the fastzip archiver is used, this location is also used as scratch space when archiving.

Configure fastzip to improve performance

Introduced in GitLab Runner 15.0.

To tune fastzip, ensure the FF_USE_FASTZIP flag is enabled. Then use any of the following environment variables.

Files in a zip archive are appended sequentially. This makes concurrent compression challenging. fastzip works around this limitation by compressing files concurrently to disk first, and then copying the result back to zip archive sequentially.

To avoid writing to disk and reading the contents back for smaller files, a small buffer per concurrency is used. This setting can be controlled with FASTZIP_ARCHIVER_BUFFER_SIZE. The default size for this buffer is 2 MiB, therefore, a concurrency of 16 will allocate 32 MiB. Data that exceeds the buffer size will be written to and read back from disk. Therefore, using no buffer, FASTZIP_ARCHIVER_BUFFER_SIZE: 0, and only scratch space is a valid option.

FASTZIP_ARCHIVER_CONCURRENCY controls how many files are compressed concurrency. As mentioned above, this setting therefore can increase how much memory is being used, but also how much temporary data is written to the scratch space. The default is the number of CPUs available, but given the memory ramifications, this may not always be the best setting.

FASTZIP_EXTRACTOR_CONCURRENCY controls how many files are decompressed at once. Files from a zip archive can natively be read from concurrency, so no additional memory is allocated in additional to what the decompressor requires. This defaults to the number of CPUs available.

Clean up stale runners (ULTIMATE)

Introduced in GitLab 15.1.

You can clean up group runners that have been inactive for more than three months.

Group runners are those that were created at the group level.

1. On the top bar, select Menu > Groups and find your group.
2. On the left sidebar, select Settings > CI/CD.
3. Expand Runners.
4. Turn on the Enable stale runner cleanup toggle.

View stale runner cleanup logs

You can check the Sidekiq logs to see the cleanup result. In Kibana you can use the following query:

{
  "query": {
    "match_phrase": {
      "json.class.keyword": "Ci::Runners::StaleGroupRunnersPruneCronWorker"
    }
  }
}

Filter entries where stale runners were removed:

{
  "query": {
    "range": {
      "json.extra.ci_runners_stale_group_runners_prune_cron_worker.total_pruned": {
        "gte": 1,
        "lt": null
      }
    }
  }
}