Configuring runners in GitLab

In GitLab CI/CD, runners run the code defined in .gitlab-ci.yml. A runner is a lightweight, highly-scalable agent that picks up a CI job through the coordinator API of GitLab CI/CD, runs the job, and sends the result back to the GitLab instance.

Runners are created by an administrator and are visible in the GitLab UI. Runners can be specific to certain projects or available to all projects.

This documentation is focused on using runners in GitLab. If you need to install and configure GitLab Runner, see the GitLab Runner documentation.

Types of runners

In the GitLab UI there are three types of runners, based on who you want to have access:

Shared runners are available to all groups and projects in a GitLab instance.
Group runners are available to all projects and subgroups in a group.
Specific runners are associated with specific projects. Typically, specific runners are used for one project at a time.

Shared runners

Shared runners are available to every project in a GitLab instance.

Use shared runners when you have multiple jobs with similar requirements. Rather than having multiple runners idling for many projects, you can have a few runners that handle multiple projects.

If you are using a self-managed instance of GitLab:

Your administrator can install and register shared runners by going to your project's Settings > CI/CD, expanding the Runners section, and clicking Show runner installation instructions. These instructions are also available in the documentation.
The administrator can also configure a maximum number of shared runner pipeline minutes for each group.

If you are using GitLab.com:

You can select from a list of shared runners that GitLab maintains.
The shared runners consume the pipelines minutes included with your account.

How shared runners pick jobs

Shared runners process jobs by using a fair usage queue. This queue prevents projects from creating hundreds of jobs and using all available shared runner resources.

The fair usage queue algorithm assigns jobs based on the projects that have the fewest number of jobs already running on shared runners.

Example 1

If these jobs are in the queue:

Job 1 for Project 1
Job 2 for Project 1
Job 3 for Project 1
Job 4 for Project 2
Job 5 for Project 2
Job 6 for Project 3

The fair usage algorithm assigns jobs in this order:

Job 1 is chosen first, because it has the lowest job number from projects with no running jobs (that is, all projects).
Job 4 is next, because 4 is now the lowest job number from projects with no running jobs (Project 1 has a job running).
Job 6 is next, because 6 is now the lowest job number from projects with no running jobs (Projects 1 and 2 have jobs running).
Job 2 is next, because, of projects with the lowest number of jobs running (each has 1), it is the lowest job number.
Job 5 is next, because Project 1 now has 2 jobs running and Job 5 is the lowest remaining job number between Projects 2 and 3.
Finally is Job 3... because it's the only job left.

Example 2

If these jobs are in the queue:

Job 1 for Project 1
Job 2 for Project 1
Job 3 for Project 1
Job 4 for Project 2
Job 5 for Project 2
Job 6 for Project 3

The fair usage algorithm assigns jobs in this order:

Job 1 is chosen first, because it has the lowest job number from projects with no running jobs (that is, all projects).
We finish Job 1.
Job 2 is next, because, having finished Job 1, all projects have 0 jobs running again, and 2 is the lowest available job number.
Job 4 is next, because with Project 1 running a Job, 4 is the lowest number from projects running no jobs (Projects 2 and 3).
We finish Job 4.
Job 5 is next, because having finished Job 4, Project 2 has no jobs running again.
Job 6 is next, because Project 3 is the only project left with no running jobs.
Lastly we choose Job 3... because, again, it's the only job left.

Enable shared runners

On GitLab.com, shared runners are enabled in all projects by default.

On self-managed instances of GitLab, an administrator must install and register them.

You can also enable shared runners for individual projects.

To enable shared runners:

Go to the project's Settings > CI/CD and expand the Runners section.
Select Enable shared runners for this project.

Disable shared runners

You can disable shared runners for individual projects or for groups. You must have Owner permissions for the project or group.

To disable shared runners for a project:

Go to the project's Settings > CI/CD and expand the Runners section.
In the Shared runners area, select Enable shared runners for this project so the toggle is grayed-out.

Shared runners are automatically disabled for a project:

If the shared runners setting for the parent group is disabled, and
If overriding this setting is not permitted at the project level.

To disable shared runners for a group:

Go to the group's Settings > CI/CD and expand the Runners section.
In the Shared runners area, turn off the Enable shared runners for this group toggle.
Optionally, to allow shared runners to be enabled for individual projects or subgroups, click Allow projects and subgroups to override the group setting.

NOTE: To re-enable the shared runners for a group, turn on the Enable shared runners for this group toggle. Then, an owner or maintainer must explicitly change this setting for each project subgroup or project.

Group runners

Use Group runners when you want all projects in a group to have access to a set of runners.

Group runners process jobs by using a first in, first out (FIFO) queue.

Create a group runner

You can create a group runner for your self-managed GitLab instance or for GitLab.com. You must have Owner permissions for the group.

To create a group runner:

Install GitLab Runner.
Go to the group you want to make the runner work for.
Go to Settings > CI/CD and expand the Runners section.
Note the URL and token.
Register the runner.

View and manage group runners

Introduced in GitLab 13.2.

You can view and manage all runners for a group, its subgroups, and projects. You can do this for your self-managed GitLab instance or for GitLab.com. You must have Owner permissions for the group.

Go to the group where you want to view the runners.
Go to Settings > CI/CD and expand the Runners section.

The following fields are displayed.

Attribute	Description
Type	One or more of the following states: shared, group, specific, locked, or paused
Runner token	Token used to identify the runner, and that the runner uses to communicate with the GitLab instance
Description	Description given to the runner when it was created
Version	GitLab Runner version
IP address	IP address of the host on which the runner is registered
Projects	The count of projects to which the runner is assigned
Jobs	Total of jobs run by the runner
Tags	Tags associated with the runner
Last contact	Timestamp indicating when the GitLab instance last contacted the runner

From this page, you can edit, pause, and remove runners from the group, its subgroups, and projects.

Pause or remove a group runner

You can pause or remove a group runner for your self-managed GitLab instance or for GitLab.com. You must have Owner permissions for the group.

Go to the group you want to remove or pause the runner for.
Go to Settings > CI/CD and expand the Runners section.
Click Pause or Remove runner.
- If you pause a group runner that is used by multiple projects, the runner pauses for all projects.
- From the group view, you cannot remove a runner that is assigned to more than one project. You must remove it from each project first.
On the confirmation dialog, click OK.

Specific runners

Use Specific runners when you want to use runners for specific projects. For example, when you have:

Jobs with specific requirements, like a deploy job that requires credentials.
Projects with a lot of CI activity that can benefit from being separate from other runners.

You can set up a specific runner to be used by multiple projects. Specific runners must be enabled for each project explicitly.

Specific runners process jobs by using a first in, first out (FIFO) queue.

NOTE: Specific runners do not get shared with forked projects automatically. A fork does copy the CI/CD settings of the cloned repository.

Create a specific runner

You can create a specific runner for your self-managed GitLab instance or for GitLab.com. You must have Owner permissions for the project.

To create a specific runner:

Install runner.
Go to the project's Settings > CI/CD and expand the Runners section.
Note the URL and token.
Register the runner.

Enable a specific runner for a specific project

A specific runner is available in the project it was created for. An administrator can enable a specific runner to apply to additional projects.

You must have Owner permissions for the project.
The specific runner must not be locked.

To enable or disable a specific runner for a project:

Go to the project's Settings > CI/CD and expand the Runners section.
Click Enable for this project or Disable for this project.

Prevent a specific runner from being enabled for other projects

You can configure a specific runner so it is "locked" and cannot be enabled for other projects. This setting can be enabled when you first register a runner, but can also be changed later.

To lock or unlock a runner:

Go to the project's Settings > CI/CD and expand the Runners section.
Find the runner you want to lock or unlock. Make sure it's enabled.
Click the pencil button.
Check the Lock to current projects option.
Click Save changes.

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).

When not configured, runners do not override the project timeout.

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:

In a project, go to Settings > CI/CD > Runners.
Select your specific runner to edit the settings.
Enter a value under Maximum job timeout.
Select Save changes.

How this feature works:

Example 1 - Runner timeout bigger than project timeout

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

Example 2 - Runner timeout not configured

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

Example 3 - Runner timeout smaller than project timeout

You set the maximum job timeout for a runner to 30 minutes
You set the CI/CD Timeout for a project to 2 hours
You start a job
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:

Go to the project's Settings > CI/CD and expand the Runners section.
Find the runner you want to protect or unprotect. Make sure it's enabled.
Click the pencil button.
Check the Protected option.
Click 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 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 token:

Go to the project's Settings > CI/CD.
Expand the General pipelines settings section.
Find the Runner token form field and click the Reveal value button.
Delete the value and save the form.
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.

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 admin access to the GitLab instance. To determine this:

Visit Admin Area > Overview > Runners.
Look for the runner in the table and you should see a column for IP Address.

Determine the IP address of a specific runner

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

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

Use tags to limit the number of jobs using the runner

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 tags are not the same as Git tags. GitLab CI 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.

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

To make a runner pick untagged jobs:

Go to the project's Settings > CI/CD and expand the Runners section.
Find the runner you want to pick untagged jobs and make sure it's enabled.
Click the pencil button.
Check the Run untagged jobs option.
Click the Save changes button 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:

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

Example 2:

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

Example 3:

The runner is configured to run only tagged jobs and has the docker tag.
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:

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

Example 2:

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

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_DEPTH (shallow cloning)
GIT_CLONE_PATH (custom build directories)

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

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.

It does not work for the kubernetes executor, but a feature proposal exists. The kubernetes executor always clones into an temporary directory.

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.

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.

The GIT_FETCH_EXTRA_FLAGS variable is used 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.

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:

Variable	Description
`ARTIFACT_DOWNLOAD_ATTEMPTS`	Number of attempts to download artifacts running a job
`EXECUTOR_JOB_SECTION_ATTEMPTS`	In GitLab 12.10 and later, the number of attempts to run a section in a job after a `No Such Container` error (Docker executor only).
`GET_SOURCES_ATTEMPTS`	Number of attempts to fetch sources running a job
`RESTORE_CACHE_ATTEMPTS`	Number of attempts to restore the cache running a job

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 also be enabled to provide the rate of transfer for uploads and downloads.

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"

Variable	Description
`TRANSFER_METER_FREQUENCY`	Specify how often to print the meter's transfer rate. It can be set to a duration (for example, `1s` or `1m30s`). A duration of `0` disables the meter (default). When a value is set, the pipeline shows a progress meter for artifact and cache uploads and downloads.
`ARTIFACT_COMPRESSION_LEVEL`	To adjust compression ratio, set to `fastest`, `fast`, `default`, `slow`, or `slowest`. This setting works with the Fastzip archiver only, so the GitLab Runner feature flag `FF_USE_FASTZIP` must also be enabled.
`CACHE_COMPRESSION_LEVEL`	To adjust compression ratio, set to `fastest`, `fast`, `default`, `slow`, or `slowest`. This setting works with the Fastzip archiver only, so the GitLab Runner feature flag `FF_USE_FASTZIP` must also be enabled.