• Limitations
• Configuration
• Prerequisite software for running a Job
• Stages
  • Config
  • Prepare
  • Run
  • Cleanup
• Terminating and killing executables
• Error handling
  • Build Failure
  • System Failure
• Driver examples

The Custom executor

GitLab Runner provides the Custom executor for environments that it doesn’t support natively, for example, Podman or Libvirt.

This gives you the control to create your own executor by configuring GitLab Runner to use some executable to provision, run, and clean up your environment.

The scripts you configure for the custom executor are called Drivers. For example, you could create a Podman driver, an LXD driver or a Libvirt driver.

Limitations

Below are some current limitations when using the Custom executor:

• No support for services. See #4358 for more details.
• No Interactive Web Terminal support.

Configuration

There are a few configuration keys that you can choose from. Some of them are optional.

Below is an example of configuration for the Custom executor using all available configuration keys:

[[runners]]
  name = "custom"
  url = "https://gitlab.com"
  token = "TOKEN"
  executor = "custom"
  builds_dir = "/builds"
  cache_dir = "/cache"
  [runners.custom]
    config_exec = "/path/to/config.sh"
    config_args = [ "SomeArg" ]
    config_exec_timeout = 200

    prepare_exec = "/path/to/script.sh"
    prepare_args = [ "SomeArg" ]
    prepare_exec_timeout = 200

    run_exec = "/path/to/binary"
    run_args = [ "SomeArg" ]

    cleanup_exec = "/path/to/executable"
    cleanup_args = [ "SomeArg" ]
    cleanup_exec_timeout = 200

    graceful_kill_timeout = 200
    force_kill_timeout = 200

For field definitions and which ones are required, see [runners.custom] section configuration.

In addition both builds_dir and cache_dir inside of the [[runners]] are required fields.

Prerequisite software for running a Job

The user must set up the environment, including the following that must be present in the PATH:

• Git: Used to clone the repositories.
• Git LFS: Pulls any LFS objects that might be in the repository.
• GitLab Runner: Used to download/update artifacts and cache.

Stages

The Custom executor provides the stages for you to configure some details of the job, prepare and cleanup the environment and run the job script within it. Each stage is responsible for specific things and has different things to keep in mind.

Each stage executed by the Custom executor is executed at the time a builtin GitLab Runner executor would execute them.

For each step that will be executed, specific environment variables are exposed to the executable, which can be used to get information about the specific Job that is running. All stages will have the following environment variables available to them:

• Standard CI/CD environment variables, including predefined variables.
• All environment variables provided by the Custom Runner host system.

Both CI/CD environment variables and predefined variables are prefixed with CUSTOM_ENV_ to prevent conflicts with system environment variables. For example, CI_BUILDS_DIR will be available as CUSTOM_ENV_CI_BUILDS_DIR.

The stages run in the following sequence:

1. config_exec
2. prepare_exec
3. run_exec
4. cleanup_exec

Config

The Config stage is executed by config_exec.

Sometimes you might want to set some settings during execution time. For example settings a build directory depending on the project ID. config_exec reads from STDOUT and expects a valid JSON string with specific keys.

For example:

#!/usr/bin/env bash

cat << EOS
{
  "builds_dir": "/builds/${CUSTOM_ENV_CI_CONCURRENT_PROJECT_ID}/${CUSTOM_ENV_CI_PROJECT_PATH_SLUG}",
  "cache_dir": "/cache/${CUSTOM_ENV_CI_CONCURRENT_PROJECT_ID}/${CUSTOM_ENV_CI_PROJECT_PATH_SLUG}",
  "builds_dir_is_shared": true,
  "hostname": "custom-hostname",
  "driver": {
    "name": "test driver",
    "version": "v0.0.1"
  }
}
EOS

Any additional keys inside of the JSON string will be ignored. If it’s not a valid JSON string the stage will fail and be retried two more times.

The STDERR of the executable will print to the job log.

The user can set config_exec_timeout if they want to set a deadline for how long GitLab Runner should wait to return the JSON string before terminating the process.

If any of the config_exec_args are defined, these will be added in order to the executable defined in config_exec. For example we have the config.toml content below:

...
[runners.custom]
  ...
  config_exec = "/path/to/config"
  config_args = [ "Arg1", "Arg2" ]
  ...

GitLab Runner would execute it as /path/to/config Arg1 Arg2.

Prepare

The Prepare stage is executed by prepare_exec.

At this point, GitLab Runner knows everything about the job (where and how it’s going to run). The only thing left is for the environment to be set up so the job can run. GitLab Runner will execute the executable that is specified in prepare_exec.

This is responsible for setting up the environment (for example, creating the virtual machine or container, or anything else). After this is done, we expect that the environment is ready to run the job.

This stage is executed only once, in a job execution.

The user can set prepare_exec_timeout if they want to set a deadline for how long GitLab Runner should wait to prepare the environment before terminating the process.

The STDOUT and STDERR returned from this executable will print to the job log.

If any of the prepare_exec_args are defined, these will be added in order to the executable defined in prepare_exec. For example we have the config.toml content below:

...
[runners.custom]
  ...
  prepare_exec = "/path/to/bin"
  prepare_args = [ "Arg1", "Arg2" ]
  ...

GitLab Runner would execute it as /path/to/bin Arg1 Arg2.

Run

The Run stage is executed by run_exec.

The STDOUT and STDERR returned from this executable will print to the job log.

Unlike the other stages, the run_exec stage is executed multiple times, since it’s split into sub stages listed below in sequential order:

1. prepare_script
2. get_sources
3. restore_cache
4. download_artifacts
5. step_*
6. build_script
7. step_*
8. after_script
9. archive_cache
10. upload_artifacts_on_success OR upload_artifacts_on_failure

For each stage mentioned above, the run_exec executable will be executed with:

• The usual environment variables.
• Two arguments:
  • The path to the script that GitLab Runner creates for the Custom executor to run.
  • Name of the stage.

For example:

/path/to/run_exec.sh /path/to/tmp/script1 prepare_executor
/path/to/run_exec.sh /path/to/tmp/script1 prepare_script
/path/to/run_exec.sh /path/to/tmp/script1 get_sources

If you have run_args defined, they are the first set of arguments passed to the run_exec executable, then GitLab Runner adds others. For example, suppose we have the following config.toml:

...
[runners.custom]
  ...
  run_exec = "/path/to/run_exec.sh"
  run_args = [ "Arg1", "Arg2" ]
  ...

GitLab Runner will execute the executable with the following arguments:

/path/to/run_exec.sh Arg1 Arg2 /path/to/tmp/script1 prepare_executor
/path/to/run_exec.sh Arg1 Arg2 /path/to/tmp/script1 prepare_script
/path/to/run_exec.sh Arg1 Arg2 /path/to/tmp/script1 get_sources

This executable should be responsible for executing the scripts that are specified in the first argument. They contain all the scripts any GitLab Runner executor would run normally to clone, download artifacts, run user scripts and all the other steps described below. The scripts can be of the following shells:

• Bash
• PowerShell Desktop
• PowerShell Core
• Batch (deprecated)

We generate the script using the shell configured by shell inside of [[runners]]. If none is provided the defaults for the OS platform are used.

The table below is a detailed explanation of what each script does and what the main goal of that script is.

Cleanup

The Cleanup stage is executed by cleanup_exec.

This final stage is executed even if one of the previous stages failed. The main goal for this stage is to clean up any of the environments that might have been set up. For example, turning off VMs or deleting containers.

The result of cleanup_exec does not affect job statuses. For example, a job will be marked as successful even if the following occurs:

• Both prepare_exec and run_exec are successful.
• cleanup_exec fails.

The user can set cleanup_exec_timeout if they want to set some kind of deadline of how long GitLab Runner should wait to clean up the environment before terminating the process.

The STDOUT of this executable will be printed to GitLab Runner logs at a DEBUG level. The STDERR will be printed to the logs at a WARN level.

If any of the cleanup_exec_args are defined, these will be added in order to the executable defined in cleanup_exec. For example we have the config.toml content below:

...
[runners.custom]
  ...
  cleanup_exec = "/path/to/bin"
  cleanup_args = [ "Arg1", "Arg2" ]
  ...

GitLab Runner would execute it as /path/to/bin Arg1 Arg2.

Terminating and killing executables

GitLab Runner will try to gracefully terminate an executable under any of the following conditions:

• config_exec_timeout, prepare_exec_timeout or cleanup_exec_timeout are met.
• The job times out.
• The job is cancelled.

When a timeout is reached, a SIGTERM is sent to the executable, and the countdown for exec_terminate_timeout starts. The executable should listen to this signal to make sure it cleans up any resources. If exec_terminate_timeout passes and the process is still running, a SIGKILL is sent to kill the process and exec_force_kill_timeout will start. If the process is still running after exec_force_kill_timeout has finished, GitLab Runner will abandon the process and will not try to stop/kill anymore. If both these timeouts are reached during config_exec, prepare_exec or run_exec the build is marked as failed.

As of GitLab 13.1 any child process that is spawned by the driver will also receive the graceful termination process explained above on UNIX based systems. This is achieved by having the main process set as a process group which all the child processes belong too.

Error handling

There are two types of errors that GitLab Runner can handle differently. These errors are only handled when the executable inside of config_exec, prepare_exec, run_exec, and cleanup_exec exits with these codes. If the user exits with a non-zero exit code, it should be propagated as one of the error codes below.

If the user script exits with one of these code it has to be propagated to the executable exit code.

Build Failure

GitLab Runner provides BUILD_FAILURE_EXIT_CODE environment variable which should be used by the executable as an exit code to inform GitLab Runner that there is a failure on the users job. If the executable exits with the code from BUILD_FAILURE_EXIT_CODE, the build is marked as a failure appropriately in GitLab CI.

If the script that the user defines inside of .gitlab-ci.yml file exits with a non-zero code, run_exec should exit with BUILD_FAILURE_EXIT_CODE value.

System Failure

You can send a system failure to GitLab Runner by exiting the process with the error code specified in the SYSTEM_FAILURE_EXIT_CODE. If this error code is returned, on certain stages GitLab Runner will retry the stage, if none of the retries are successful the job will be marked as failed.

Below is a table of what stages are retried, and by how many times.

Driver examples

A set of example drivers using the Custom executor can be found in the examples page.