Job artifact troubleshooting for administrators

When administering job artifacts, you might encounter the following issues.

Job artifacts using too much disk space

Job artifacts can fill up your disk space quicker than expected. Some possible reasons are:

In these and other cases, identify the projects most responsible for disk space usage, figure out what types of artifacts are using the most space, and in some cases, manually delete job artifacts to reclaim disk space.

Artifacts housekeeping

Artifacts housekeeping is the process that identifies which artifacts are expired and can be deleted.

Housekeeping disabled in GitLab 15.0 to 15.2

Artifact housekeeping was significantly improved in GitLab 15.0, introduced behind feature flags disabled by default. The flags were enabled by default in GitLab 15.3.

If artifacts housekeeping does not seem to be working in GitLab 15.0 to GitLab 15.2, you should check if the feature flags are enabled.

To check if the feature flags are enabled:

  1. Start a Rails console.

  2. Check if the feature flags are enabled. 

    Feature.enabled?(:ci_detect_wrongly_expired_artifacts)
Feature.enabled?(:ci_update_unlocked_job_artifacts)
Feature.enabled?(:ci_job_artifacts_backlog_work)

  3. If any of the feature flags are disabled, enable them: 

    Feature.enable(:ci_detect_wrongly_expired_artifacts)
Feature.enable(:ci_update_unlocked_job_artifacts)
Feature.enable(:ci_destroy_unlocked_job_artifacts)

These changes include switching artifacts from unlocked to locked if they should be retained.

Artifacts with unknown status

Artifacts created before housekeeping was updated have a status of unknown. After they expire, these artifacts are not processed by the new housekeeping.

You can check the database to confirm if your instance has artifacts with the unknown status:

  1. Start a database console:

    Linux package (Omnibus)
    sudo gitlab-psql
    Helm chart (Kubernetes)
    # Find the toolbox pod
kubectl --namespace <namespace> get pods -lapp=toolbox
# Connect to the PostgreSQL console
kubectl exec -it <toolbox-pod-name> -- /srv/gitlab/bin/rails dbconsole --include-password --database main
    Docker
    sudo docker exec -it <container_name> /bin/bash
gitlab-psql
    Self-compiled (source)
    sudo -u git -H psql -d gitlabhq_production

  2. Run the following query: 

    select expire_at, file_type, locked, count(*) from ci_job_artifacts
where expire_at is not null and
file_type != 3
group by expire_at, file_type, locked having count(*) > 1;

If records are returned, then there are artifacts which the housekeeping job is unable to process. For example: 

           expire_at           | file_type | locked | count
-------------------------------+-----------+--------+--------
 2021-06-21 22:00:00+00        |         1 |      2 |  73614
 2021-06-21 22:00:00+00        |         2 |      2 |  73614
 2021-06-21 22:00:00+00        |         4 |      2 |   3522
 2021-06-21 22:00:00+00        |         9 |      2 |     32
 2021-06-21 22:00:00+00        |        12 |      2 |    163

Artifacts with locked status 2 are unknown. Check issue #346261 for more details.

Clean up unknown artifacts

The Sidekiq worker that processes all unknown artifacts is enabled by default in GitLab 15.3 and later. It analyzes the artifacts returned by the above database query and determines which should be locked or unlocked. Artifacts are then deleted by that worker if needed.

The worker can be enabled on self-managed instances:

  1. Start a Rails console.

  2. Check if the feature is enabled. 

    Feature.enabled?(:ci_job_artifacts_backlog_work)

  3. Enable the feature, if needed: 

    Feature.enable(:ci_job_artifacts_backlog_work)

The worker processes 10,000 unknown artifacts every seven minutes, or roughly two million in 24 hours.

There is a related ci_job_artifacts_backlog_large_loop_limit feature flag which causes the worker to process unknown artifacts in batches that are five times larger. This flag is not recommended for use on self-managed instances.

List projects and builds with artifacts with a specific expiration (or no expiration)

Using a Rails console, you can find projects that have job artifacts with either:

  • No expiration date.
  • An expiration date more than 7 days in the future.

Similar to deleting artifacts, use the following example time frames and alter them as needed:

  • 7.days.from_now
  • 10.days.from_now
  • 2.weeks.from_now
  • 3.months.from_now
  • 1.year.from_now

Each of the following scripts also limits the search to 50 results with .limit(50), but this number can also be changed as needed: 

# Find builds & projects with artifacts that never expire
builds_with_artifacts_that_never_expire = Ci::Build.with_downloadable_artifacts.where(artifacts_expire_at: nil).limit(50)
builds_with_artifacts_that_never_expire.find_each do |build|
  puts "Build with id #{build.id} has artifacts that don't expire and belongs to project #{build.project.full_path}"
end

# Find builds & projects with artifacts that expire after 7 days from today
builds_with_artifacts_that_expire_in_a_week = Ci::Build.with_downloadable_artifacts.where('artifacts_expire_at > ?', 7.days.from_now).limit(50)
builds_with_artifacts_that_expire_in_a_week.find_each do |build|
  puts "Build with id #{build.id} has artifacts that expire at #{build.artifacts_expire_at} and belongs to project #{build.project.full_path}"
end

List projects by total size of job artifacts stored

List the top 20 projects, sorted by the total size of job artifacts stored, by running the following code in the Rails console: 

include ActionView::Helpers::NumberHelper
ProjectStatistics.order(build_artifacts_size: :desc).limit(20).each do |s|
  puts "#{number_to_human_size(s.build_artifacts_size)} \t #{s.project.full_path}"
end

You can change the number of projects listed by modifying .limit(20) to the number you want.

List largest artifacts in a single project

List the 50 largest job artifacts in a single project by running the following code in the Rails console: 

include ActionView::Helpers::NumberHelper
project = Project.find_by_full_path('path/to/project')
Ci::JobArtifact.where(project: project).order(size: :desc).limit(50).map { |a| puts "ID: #{a.id} - #{a.file_type}: #{number_to_human_size(a.size)}" }

You can change the number of job artifacts listed by modifying .limit(50) to the number you want.

List artifacts in a single project

List the artifacts for a single project, sorted by artifact size. The output includes the:

  • ID of the job that created the artifact
  • artifact size
  • artifact file type
  • artifact creation date
  • on-disk location of the artifact 
p = Project.find_by_id(<project_id>)
arts = Ci::JobArtifact.where(project: p)

list = arts.order(size: :desc).limit(50).each do |art|
    puts "Job ID: #{art.job_id} - Size: #{art.size}b - Type: #{art.file_type} - Created: #{art.created_at} - File loc: #{art.file}"
end

To change the number of job artifacts listed, change the number in limit(50).

Delete old builds and artifacts

caution
These commands remove data permanently from database and storage. Before running them, we highly recommend seeking guidance from a Support Engineer, or running them in a test environment with a backup of the instance ready to be restored, just in case.

Delete old artifacts for a project

This step also erases artifacts that users have chosen to keep: 

project = Project.find_by_full_path('path/to/project')
builds_with_artifacts =  project.builds.with_downloadable_artifacts
builds_with_artifacts.where("finished_at < ?", 1.year.ago).each_batch do |batch|
  batch.each do |build|
    Ci::JobArtifacts::DeleteService.new(build).execute
  end

  batch.update_all(artifacts_expire_at: Time.current)
end

In GitLab 15.3 and earlier, use the following instead: 

project = Project.find_by_full_path('path/to/project')
builds_with_artifacts =  project.builds.with_downloadable_artifacts
builds_with_artifacts.where("finished_at < ?", 1.year.ago).each_batch do |batch|
  batch.each do |build|
    build.artifacts_expire_at = Time.current
    build.erase_erasable_artifacts!
  end
end

Delete old artifacts instance wide

This step also erases artifacts that users have chosen to keep: 

builds_with_artifacts = Ci::Build.with_downloadable_artifacts
builds_with_artifacts.where("finished_at < ?", 1.year.ago).each_batch do |batch|
  batch.each do |build|
    Ci::JobArtifacts::DeleteService.new(build).execute
  end

  batch.update_all(artifacts_expire_at: Time.current)
end

In GitLab 15.3 and earlier, use the following instead: 

builds_with_artifacts =  Ci::Build.with_downloadable_artifacts
builds_with_artifacts.where("finished_at < ?", 1.year.ago).each_batch do |batch|
  batch.each do |build|
    build.artifacts_expire_at = Time.current
    build.erase_erasable_artifacts!
  end
end

Delete old job logs and artifacts for a project 

project = Project.find_by_full_path('path/to/project')
builds =  project.builds
admin_user = User.find_by(username: 'username')
builds.where("finished_at < ?", 1.year.ago).each_batch do |batch|
  batch.each do |build|
    print "Ci::Build ID #{build.id}... "

    if build.erasable?
      Ci::BuildEraseService.new(build, admin_user).execute
      puts "Erased"
    else
      puts "Skipped (Nothing to erase or not erasable)"
    end
  end
end

Delete old job logs and artifacts instance wide 

builds = Ci::Build.all
admin_user = User.find_by(username: 'username')
builds.where("finished_at < ?", 1.year.ago).each_batch do |batch|
  batch.each do |build|
    print "Ci::Build ID #{build.id}... "

    if build.erasable?
      Ci::BuildEraseService.new(build, admin_user).execute
      puts "Erased"
    else
      puts "Skipped (Nothing to erase or not erasable)"
    end
  end
end

In GitLab 15.3 and earlier, replace Ci::BuildEraseService.new(build, admin_user).execute with build.erase(erased_by: admin_user).

1.year.ago is a Rails ActiveSupport::Duration method. Start with a long duration to reduce the risk of accidentally deleting artifacts that are still in use. Rerun the deletion with shorter durations as needed, for example 3.months.ago, 2.weeks.ago, or 7.days.ago.

The method erase_erasable_artifacts! is synchronous, and upon execution the artifacts are immediately removed; they are not scheduled by a background queue.

Job artifact upload fails with error 500

If you are using object storage for artifacts and a job artifact fails to upload, review:

  • The job log for an error message similar to: 

    WARNING: Uploading artifacts as "archive" to coordinator... failed id=12345 responseStatus=500 Internal Server Error status=500 token=abcd1234

  • The workhorse log for an error message similar to: 

    {"error":"MissingRegion: could not find region configuration","level":"error","msg":"error uploading S3 session","time":"2021-03-16T22:10:55-04:00"}

In both cases, you might need to add region to the job artifact object storage configuration.

Job artifact upload fails with 500 Internal Server Error (Missing file)

Bucket names that include folder paths are not supported with consolidated object storage. For example, bucket/path. If a bucket name has a path in it, you might receive an error similar to: 

WARNING: Uploading artifacts as "archive" to coordinator... POST https://gitlab.example.com/api/v4/jobs/job_id/artifacts?artifact_format=zip&artifact_type=archive&expire_in=1+day: 500 Internal Server Error (Missing file)
FATAL: invalid argument

If a job artifact fails to upload with the above error when using consolidated object storage, make sure you are using separate buckets for each data type.

Job artifacts fail to upload with FATAL: invalid argument when using Windows mount

If you are using a Windows mount with CIFS for job artifacts, you may see an invalid argument error when the runner attempts to upload artifacts: 

WARNING: Uploading artifacts as "dotenv" to coordinator... POST https://<your-gitlab-instance>/api/v4/jobs/<JOB_ID>/artifacts: 500 Internal Server Error  id=1296 responseStatus=500 Internal Server Error status=500 token=*****
FATAL: invalid argument

To work around this issue, you can try:

  • Switching to an ext4 mount instead of CIFS.
  • Upgrading to at least Linux kernel 5.15 which contains a number of important bug fixes relating to CIFS file leases.
  • For older kernels, using the nolease mount option to disable file leasing.

For more information, see the investigation details.

Usage quota shows incorrect artifact storage usage

Sometimes the artifacts storage usage displays an incorrect value for the total storage space used by artifacts. To recalculate the artifact usage statistics for all projects in the instance, you can run this background script: 

gitlab-rake gitlab:refresh_project_statistics_build_artifacts_size[https://example.com/path/file.csv]

The https://example.com/path/file.csv file must list the project IDs for all projects for which you want to recalculate artifact storage usage. Use this format for the file: 

PROJECT_ID
1
2

The artifact usage value can fluctuate to 0 while the script is running. After recalculation, usage should display as expected again.

Artifact download flow diagrams

The following flow diagrams illustrate how job artifacts work. These diagrams assume object storage is configured for job artifacts.

Proxy download disabled

With proxy_download set to false, GitLab redirects the runner to download artifacts from object storage with a pre-signed URL. It is usually faster for runners to fetch from the source directly so this configuration is generally recommended. It should also reduce bandwidth usage because the data does not have to be fetched by GitLab and sent to the runner. However, it does require giving runners direct access to object storage.

The request flow looks like:

sequenceDiagram autonumber participant C as Runner participant O as Object Storage participant W as Workhorse participant R as Rails participant P as PostgreSQL C->>+W: GET /api/v4/jobs/:id/artifacts?direct_download=true Note over C,W: gitlab-ci-token@<CI_JOB_TOKEN> W-->+R: GET /api/v4/jobs/:id/artifacts?direct_download=true Note over W,R: gitlab-ci-token@<CI_JOB_TOKEN> R->>P: Look up job for CI_JOB_TOKEN R->>P: Find user who triggered job R->>R: Does user have :read_build access? alt Yes R->>W: Send 302 redirect to object storage presigned URL R->>C: 302 redirect C->>O: GET <presigned URL> else No R->>W: 401 Unauthorized W->>C: 401 Unauthorized end

In this diagram:

  1. First, the runner attempts to fetch a job artifact by using the GET /api/v4/jobs/:id/artifacts endpoint. The runner attaches the direct_download=true query parameter on the first attempt to indicate that it is capable of downloading from object storage directly. Direct downloads can be disabled in the runner configuration via the FF_USE_DIRECT_DOWNLOAD feature flag. This flag is set to true by default.

  2. The runner sends the GET request using HTTP Basic Authentication with the gitlab-ci-token username and an auto-generated CI/CD job token as the password. This token is generated by GitLab and given to the runner at the start of a job.

  3. The GET request gets passed to the GitLab API, which looks up the token in the database and finds the user who triggered the job.

  4. In steps 5-8:

    • If the user has access to the build, then GitLab generates a presigned URL and sends a 302 Redirect with the Location set to that URL. The runner follows the 302 Redirect and downloads the artifacts.

    • If the job cannot be found or the user does not have access to the job, then the API returns 401 Unauthorized.

    The runner does not retry if it receives the following HTTP status codes:

    • 200 OK
    • 401 Unauthorized
    • 403 Forbidden
    • 404 Not Found

    However, if the runner receives any other status code, such as a 500 error, it re-attempts to download the artifacts two more times, sleeping 1 second between each attempt. The subsequent attempts omit direct_download=true.

Proxy download enabled

If proxy_download is true, GitLab always fetches the artifacts from object storage and send the data to the runner, even if the runner sends the direct_download=true query parameter. Proxy downloads might be desirable if runners have restricted network access.

The following diagram is similar to the disabled proxy download example, except at steps 6-9, GitLab does not send a 302 Redirect to the runner. Instead, GitLab instructs Workhorse to fetch the data and stream it back to the runner. From the runner perspective, the original GET request to /api/v4/jobs/:id/artifacts returns the binary data directly.

sequenceDiagram autonumber participant C as Runner participant O as Object Storage participant W as Workhorse participant R as Rails participant P as PostgreSQL C->>+W: GET /api/v4/jobs/:id/artifacts?direct_download=true Note over C,W: gitlab-ci-token@<CI_JOB_TOKEN> W-->+R: GET /api/v4/jobs/:id/artifacts?direct_download=true Note over W,R: gitlab-ci-token@<CI_JOB_TOKEN> R->>P: Look up job for CI_JOB_TOKEN R->>P: Find user who triggered job R->>R: Does user have :read_build access? alt Yes R->>W: SendURL with object storage presigned URL W->>O: GET <presigned URL> O->>W: <artifacts data> W->>C: <artifacts data> else No R->>W: 401 Unauthorized W->>C: 401 Unauthorized end