Race condition when pushing manifest lists with the same digest

Reported

We have a Maven project using jib to build and push a Docker image with two tags (current version and "latest") to GitLab Container Registry.

This fails frequently with

Failed to execute goal com.google.cloud.tools:jib-maven-plugin:3.3.2:build (docker-push) on project management-console: 
500 Internal Server Error
PUT https://registry.gitlab.com/v2/chromaway/core-tools/management-console/pmc/manifests/3.15.0
{"errors":[{"code":"UNKNOWN","message":"unknown error","detail":{}}]}

It usually works if rerunning the job.

The project is here: https://gitlab.com/chromaway/core-tools/management-console

Seems to be similar to fix(datastore): race condition pushing two or m... (#640 - closed)

Context

The container registry on GitLab.com is backed by a metadata database. The manifests table keeps track of these manifests.

Only one manifest per digest and repository can exist as defined by the constraint top_level_namespace_id_repository_id_digest_key.

Problem

A race condition happens when pushing a manifest list with the same digest but different tags at the same time, as seen in the report. This is how jib builds the lists and pushes them to the registry. The problem is similar to #640 (closed).

Solution

In fix(datastore): race condition pushing two or m... (#640 - closed), we fixed this by exposing and using a CreateOrFind method to the manifests store, and then updated the handler to use this new method.

For this problem, we can do the same but the handler for manifests lists is a bit different so we did not catch that in the first solution.