Scheduling the merge of a GitLab merge request
We want to be able to merge GitLab merge requests at preset times. This enables us to roll out changes during maintenance windows in an automated fashion.
While there is an open request for such a feature in GitLab, the implementation isn’t on the horizon yet, and no suitable functionality exists at present.
We develop a new service, to be run on our infrastructure, which periodically checks all open MRs that have a specific label set.
In addition to the label, a merge window needs to be specified. This could be either via an in-repo config (similar to Renovate automerge schedules), or via a command that can be commented into the merge request, or both.
If the label is set, and the merge window is currently active, and the MR has no failing pipelines or file conflicts, the MR gets merged.
This automation should run frequently enough so it’s unlikely to miss a merge window. For example, every 10 minutes.
In the future, the automation could also automatically rebase all MRs with the label, if necessary. This could happen outside of the configured window, to ensure the MR is up-to-date once the window rolls around.
We develop a Job that checks a cluster’s tenant repository for MRs with a specific label, and merges them via the GitLab API. The Job is run as an UpgradeJob hook during cluster maintenance. Any MR that needs to be merged during maintenance simply needs to have the corresponding label.
This would marry the merge automation to the cluster upgrade process, which would automatically ensure it runs during cluster maintenance while adding little complexity. However, in this case, there is no merge automation for arbitrary other cases (for example, for scheduling a rollout in global-defaults). Only tenant repos could benefit from this automation. Furthermore, it’s also not straightforward to teach this automation to keep the MRs up-to-date.
Each tenant repository would ideally have its own access token for this automation, which then somehow needs to be made available to the automation. Ideally, this would be automated.
We should create our own automation to schedule the merging of arbitrary MRs based on in-repo configs and labels.
Creating a service that’s separate from the cluster upgrade process doesn’t incur that much extra complexity, but gives us a lot more flexibility. Marrying the MR scheduling to the cluster upgrade process is in some ways a convenient solution, but seems like a coupling of systems that shouldn’t be connected in this way.
Furthermore, any system that can merge MRs for us will require a GitLab access token. Managing such a token is much easier with one centralized service, as opposed to maintaining a separate token for each tenant.
We already rely on in-repo configurations to schedule Renovate MRs. It makes sense to schedule other MRs in the same way, so the use of the new system will be familiar. If we later require the ability to quickly create very specific schedules for individual MRs, we can still add a feature to override the configured schedule with a comment to the MR.