[Feature flag] Roll out ci_job_artifacts_backlog_work

Summary

This issue is to rollout the specific code path to update and remove ci_job_artifact records where locked is "unknown" (enum value 2) on production, that is currently behind the ci_job_artifacts_backlog_work feature flag.

This code was originally introduced behind the ci_destroy_unlocked_job_artifacts flag, except the existing code there is useful and this new code proved to be problematic. So we want a new flag to enable the old code that prevents our expired ci_job_artifacts backlog from growing while not enabling this more dangerous code to work on the existing backlog.

Owners

• Team: grouppipeline execution but more specifically the devopsverify Engineering Allocation
• Most appropriate slack channel to reach out to: #verify-reliability-engineering-allocation
• Best individual to reach out to: @drew and/or @mattkasa
• PM: @jheimbuck_gl

Stakeholders

Are there any other stages or teams involved that need to be kept in the loop?

• #g_testing and @shampton (JobArtifact feature owners)
• @sgoldstein & @avielle (managing this particular chunk of the Engineering Allocation)

Expectations

What are we expecting to happen?

We are expecting to see mildly higher index tuple reads on the ci_job_artifacts table. Preferably under 100k here, and we won't let the reads exceed 1m.

While this is happening, we'll also expect to see the number of records in the ci_job_artifacts table decrease.

We should see:

• SELECT COUNT(*) FROM ci_job_artifacts WHERE locked = 0 Decrease
• SELECT COUNT(*) FROM ci_job_artifacts WHERE locked = 1 Increase
• SELECT COUNT(*) FROM ci_job_artifacts WHERE locked = 2 Decrease
• SELECT COUNT(*) FROM ci_job_artifacts Decrease

Rather than run these actual counts, we can look at some of the table states that Postgres makes available. We'll be able to see the overall table size decrease while the conditional counts move proportionately to the others.

When is the feature viable?

We'll turn the feature flag on to 100% and observe the execution of the ExpireBuildArtifactsWorker. A lower percentage rollout doesn't make sense here, because only one ExpireBuildArtifactsWorker ever runs at a time. A 10% rollout would just mean that we only have a 10% chance of the new code running in any given 7-minute window.

We're starting with a low record limit for execution (10,000 records per 7 minutes), with another flag that will enable us to bump the limit up to 50,000. When the limit was 100,000 we expired 50,000-60,000 records per execution so we expect that proportion to remain similar. We've also added logging to count the number of records that are updated to locked, and so not expired, so we'll have a more complete picture of how many records we're querying in a single worker execution.

Increasing the LOOP_LIMIT in the service limit is our intended method of accelerating the pace of the worker, rather than the percentage rollout of the feature.

If we need to slow down the pace further, we can use a percentage rollout of this flag to lower the average number of records expired in any given execution, thereby lowering the number of records we expire in a given day. We believe this table is only vacuumed about once per day, so this might be a useful strategy.

What might happen if this goes wrong?

If this goes wrong, query performance on indexes of the ci_job_artifacts table will degrade. Tuple reads will increase sharply, query planning time will increase, and overall throughput will drop. Apdex on gitlab.com will suffer. It will look exactly like gitlab-com/gl-infra/production#5952 (closed).

If this happens, we'll need to both shut the flag off and immediately manually VACUUM the ci_job_artifacts table. An SRE who can do this should be present during the rollout to monitor and possibly perform this mitigation.

What can we check for monitoring production both during and after rollouts?

• Kibana logs reporting the number of updates and removed ci_job_artifact records through each execution path.
• Index tuple reads on ci_job_artifacts in Thanos

Rollout Steps

Rollout on non-production environments

• Ensure that the feature MRs have been deployed to non-production environments.
  • /chatops run auto_deploy status <merge-commit-of-your-feature>
• Enable the feature globally on non-production environments.
  • /chatops run feature set ci_job_artifacts_backlog_work true --dev
  • /chatops run feature set ci_job_artifacts_backlog_work true --staging
• Verify that the feature works as expected. Posting the QA result in this issue is preferable.

Specific rollout on production

• Ensure that the feature MRs have been deployed to both production and canary.
  • /chatops run auto_deploy status <merge-commit-of-your-feature>
• If you're using project-actor, you must enable the feature on these entries:
  • /chatops run feature set --project=gitlab-org/gitlab ci_job_artifacts_backlog_work true
  • /chatops run feature set --project=gitlab-org/gitlab-foss ci_job_artifacts_backlog_work true
  • /chatops run feature set --project=gitlab-com/www-gitlab-com ci_job_artifacts_backlog_work true
• If you're using group-actor, you must enable the feature on these entries:
  • /chatops run feature set --group=gitlab-org ci_job_artifacts_backlog_work true
  • /chatops run feature set --group=gitlab-com ci_job_artifacts_backlog_work true
• If you're using user-actor, you must enable the feature on these entries:
  • /chatops run feature set --user=<your-username> <feature-flag-name> true
• Verify that the feature works on the specific entries. Posting the QA result in this issue is preferable.

Preparation before global rollout

• Check if the feature flag change needs to be accompanied with a change management issue. Cross link the issue here if it does.
• Ensure that you or a representative in development can be available for at least 2 hours after feature flag updates in production. If a different developer will be covering, or an exception is needed, please inform the oncall SRE by using the @sre-oncall Slack alias.
• Ensure that documentation has been updated (More info).
• Announce on the feature issue an estimated time this will be enabled on GitLab.com.
• Notify #support_gitlab-com and your team channel. This is necessary. We caused an outage and blocked deploys last time.

Global rollout on production

For visibility, all /chatops commands that target production should be executed in the #production slack channel and cross-posted (with the command results) to the responsible team's slack channel (#g_TEAM_NAME).

• Incrementally roll out the feature.
  • If the feature flag in code has an actor, perform actor-based rollout.
    • /chatops run feature set ci_job_artifacts_backlog_work <rollout-percentage> --actors
  • If the feature flag in code does NOT have an actor, perform time-based rollout (random rollout).
    • /chatops run feature set ci_job_artifacts_backlog_work <rollout-percentage>
  • Enable the feature globally on production environment.
    • /chatops run feature set ci_job_artifacts_backlog_work true
• Announce on the feature issue that the feature has been globally enabled.
• Wait for at least one day for the verification term.

(Optional) Release the feature with the feature flag

If you're still unsure whether the feature is deemed stable but want to release it in the current milestone, you can change the default state of the feature flag to be enabled. To do so, follow these steps:

• Create a merge request with the following changes. Ask for review and merge it.
  • Set the default_enabled attribute in the feature flag definition to true.
  • Create a changelog entry.
• Ensure that the default-enabling MR has been deployed to both production and canary. If the merge request was deployed before the code cutoff, the feature can be officially announced in a release blog post.
  • /chatops run auto_deploy status <merge-commit-of-default-enabling-mr>
• Close the feature issue to indicate the feature will be released in the current milestone.
• Set the next milestone to this rollout issue for scheduling the flag removal.
• (Optional) You can create a separate issue for scheduling the steps below to Release the feature.
  • Set the title to "[Feature flag] Cleanup ci_job_artifacts_backlog_work".
  • Execute the /copy_metadata <this-rollout-issue-link> quick action to copy the labels from this rollout issue.
  • Link this rollout issue as a related issue.
  • Close this rollout issue.

WARNING: This approach has the downside that it makes it difficult for us to clean up the flag. For example, on-premise users could disable the feature on their GitLab instance. But when you remove the flag at some point, they suddenly see the feature as enabled and they can't roll it back to the previous behavior. To avoid this potential breaking change, use this approach only for urgent matters.

Release the feature

After the feature has been deemed stable, the clean up should be done as soon as possible to permanently enable the feature and reduce complexity in the codebase.

• Create a merge request to remove <feature-flag-name> feature flag. Ask for review and merge it.
  • Remove all references to the feature flag from the codebase.
  • Remove the YAML definitions for the feature from the repository.
  • Create a changelog entry.
• Ensure that the cleanup MR has been deployed to both production and canary. If the merge request was deployed before the code cutoff, the feature can be officially announced in a release blog post.
  • /chatops run auto_deploy status <merge-commit-of-cleanup-mr>
• Close the feature issue to indicate the feature will be released in the current milestone.
• Clean up the feature flag from all environments by running these chatops command in #production channel:
  • /chatops run feature delete ci_job_artifacts_backlog_work --dev
  • /chatops run feature delete ci_job_artifacts_backlog_work --staging
  • /chatops run feature delete ci_job_artifacts_backlog_work
• Close this rollout issue.

Rollback Steps

• This feature can be disabled by running the following Chatops command:

/chatops run feature set ci_job_artifacts_backlog_work false

• An SRE should immediately VACUUM the ci_job_artifacts table if this feature gets shut off.