In theory the case could happen that the initial linking of the pool
fails and so do all the retries that Sidekiq performs. This could lead
to data loss.
To prevent that case, linking is done before Gits GC too. This makes
sure that case doesn't happen.
When a project is forked, the new repository used to be a deep copy of everything
stored on disk by leveraging `git clone`. This works well, and makes isolation
between repository easy. However, the clone is at the start 100% the same as the
origin repository. And in the case of the objects in the object directory, this
is almost always going to be a lot of duplication.
Object Pools are a way to create a third repository that essentially only exists
for its 'objects' subdirectory. This third repository's object directory will be
set as alternate location for objects. This means that in the case an object is
missing in the local repository, git will look in another location. This other
location is the object pool repository.
When Git performs garbage collection, it's smart enough to check the
alternate location. When objects are duplicated, it will allow git to
throw one copy away. This copy is on the local repository, where to pool
remains as is.
These pools have an origin location, which for now will always be a
repository that itself is not a fork. When the root of a fork network is
forked by a user, the fork still clones the full repository. Async, the
pool repository will be created.
Either one of these processes can be done earlier than the other. To
handle this race condition, the Join ObjectPool operation is
idempotent. Given its idempotent, we can schedule it twice, with the
same effect.
To accommodate the holding of state two migrations have been added.
1. Added a state column to the pool_repositories column. This column is
managed by the state machine, allowing for hooks on transitions.
2. pool_repositories now has a source_project_id. This column in
convenient to have for multiple reasons: it has a unique index allowing
the database to handle race conditions when creating a new record. Also,
it's nice to know who the host is. As that's a short link to the fork
networks root.
Object pools are only available for public project, which use hashed
storage and when forking from the root of the fork network. (That is,
the project being forked from itself isn't a fork)
In this commit message I use both ObjectPool and Pool repositories,
which are alike, but different from each other. ObjectPool refers to
whatever is on the disk stored and managed by Gitaly. PoolRepository is
the record in the database.
Dumping too many jobs in the same queue (e.g. the "default" queue) is a
dangerous setup. Jobs that take a long time to process can effectively
block any other work from being performed given there are enough of
these jobs.
Furthermore it becomes harder to monitor the jobs as a single queue
could contain jobs for different workers. In such a setup the only
reliable way of getting counts per job is to iterate over all jobs in a
queue, which is a rather time consuming process.
By using separate queues for various workers we have better control over
throughput, we can add weight to queues, and we can monitor queues
better. Some workers still use the same queue whenever their work is
related. For example, the various CI pipeline workers use the same
"pipeline" queue.
This commit includes a Rails migration that moves Sidekiq jobs from the
old queues to the new ones. This migration also takes care of doing the
inverse if ever needed. This does require downtime as otherwise new jobs
could be scheduled in the old queues after this migration completes.
This commit also includes an RSpec test that blacklists the use of the
"default" queue and ensures cron workers use the "cronjob" queue.
Fixesgitlab-org/gitlab-ce#23370
Due to a stale NFS cache, it's possible that a branch lookup fails
while `git gc` is running and causes missing branches in merge requests.
Possible workaround for #15392
