CI configuration performance

Interruptible pipelines

By default, all jobs are interruptible, except the dont-interrupt-me job which runs automatically on main, and is manual otherwise.

If you want a running pipeline to finish even if you push new commits to a merge request, be sure to start the dont-interrupt-me job before pushing.

Git fetch caching

Because GitLab.com uses the pack-objects cache, concurrent Git fetches of the same pipeline ref are deduplicated on the Gitaly server (always) and served from cache (when available).

This works well for the following reasons:

  • The pack-objects cache is enabled on all Gitaly servers on GitLab.com.
  • The CI/CD Git strategy setting for gitlab-org/gitlab is Git clone, causing all jobs to fetch the same data, which maximizes the cache hit ratio.
  • We use shallow clone to avoid downloading the full Git history for every job.

Caching strategy

  1. All jobs must only pull caches by default.
  2. All jobs must be able to pass with an empty cache. In other words, caches are only there to speed up jobs.
  3. We currently have several different cache definitions defined in .gitlab/ci/global.gitlab-ci.yml, with fixed keys:
    • .setup-test-env-cache
    • .ruby-cache
    • .static-analysis-cache
    • .rubocop-cache
    • .coverage-cache
    • .ruby-node-cache
    • .qa-cache
    • .yarn-cache
    • .assets-compile-cache (the key includes ${NODE_ENV} so it’s actually two different caches).
  4. These cache definitions are composed of multiple atomic caches.
  5. Only the following jobs, running in 2-hourly maintenance scheduled pipelines, are pushing (that is, updating) to the caches:
  6. These jobs can also be forced to run in merge requests with the pipeline:update-cache label (this can be useful to warm the caches in a MR that updates the cache keys).

Artifacts strategy

We limit the artifacts that are saved and retrieved by jobs to the minimum to reduce the upload/download time and costs, as well as the artifacts storage.

Components caching

Some external components (GitLab Workhorse and frontend assets) of GitLab need to be built from source as a preliminary step for running tests.

cache-workhorse

In this MR, and then this MR, we introduced a new cache-workhorse job that:

  • runs automatically for all GitLab.com gitlab-org/gitlab scheduled pipelines
  • runs automatically for any master commit that touches the workhorse/ folder
  • is manual for GitLab.com’s gitlab-org’s MRs that touches caching-related files

This job tries to download a generic package that contains GitLab Workhorse binaries needed in the GitLab test suite (under tmp/tests/gitlab-workhorse).

  • If the package URL returns a 404:
    1. It runs scripts/setup-test-env, so that the GitLab Workhorse binaries are built.
    2. It then creates an archive which contains the binaries and upload it as a generic package.
  • Otherwise, if the package already exists, it exits the job successfully.

We also changed the setup-test-env job to:

  1. First download the GitLab Workhorse generic package build and uploaded by cache-workhorse.
  2. If the package is retrieved successfully, its content is placed in the right folder (for example, tmp/tests/gitlab-workhorse), preventing the building of the binaries when scripts/setup-test-env is run later on.
  3. If the package URL returns a 404, the behavior doesn’t change compared to the current one: the GitLab Workhorse binaries are built as part of scripts/setup-test-env.
note
The version of the package is the workhorse tree SHA (for example, git rev-parse HEAD:workhorse).

cache-assets

In this MR, we introduced three new cache-assets:test, cache-assets:test as-if-foss, and cache-assets:production jobs that:

  • never run unless $CACHE_ASSETS_AS_PACKAGE == "true"
  • runs automatically for all GitLab.com gitlab-org/gitlab scheduled pipelines
  • runs automatically for any master commit that touches the assets-related folders
  • is manual for GitLab.com’s gitlab-org’s MRs that touches caching-related files

This job tries to download a generic package that contains GitLab compiled assets needed in the GitLab test suite (under app/assets/javascripts/locale/**/app.js, and public/assets).

  • If the package URL returns a 404:
    1. It runs bin/rake gitlab:assets:compile, so that the GitLab assets are compiled.
    2. It then creates an archive which contains the assets and uploads it as a generic package. The package version is set to the assets folders’ hash sum.
  • Otherwise, if the package already exists, it exits the job successfully.

compile-*-assets

We also changed the compile-test-assets, compile-test-assets as-if-foss, and compile-production-assets jobs to:

  1. First download the “native” cache assets, which contain:
  2. We then we compute the SHA256 hexdigest of all the source files the assets depend on, for the current checked out branch. We store the hexdigest in the GITLAB_ASSETS_HASH variable.
  3. If $CACHE_ASSETS_AS_PACKAGE == "true", we download the generic package built and uploaded by cache-assets:*.
    • If the cache is up-to-date for the checked out branch, we download the native cache and the cache package. We could optimize that by not downloading the genetic package but the native cache is actually very often outdated because it’s rebuilt only every 2 hours.

  4. We run the assets_compile_script function, which itself runs the assets:compile Rake task.

    This task is responsible for deciding if assets need to be compiled or not. It compares the HEAD SHA256 hexdigest from $GITLAB_ASSETS_HASH with the master hexdigest from cached-assets-hash.txt.

  5. If the hashes are the same, we don’t compile anything. If they’re different, we compile the assets.