Remove replica records when deleting a repository

Prafect currently identifies stale replicas of deleted repositories
by the replicas having entries in 'storage_repositories' when the
there's no corresponding entry for the repository in the 'repositories'
table. Prior to having database records for every repository with the
'per_repository' election strategy, this was necessary to differentiate
replicas of deleted repositories from repositories without database
records. This then allowed Praefect to retry deleting stale replicas by
rescheduling the deletion jobs in the reconciler. Nowadays, Praefect
expects to have database records for every repository that should be
present on the cluster. Given that, there's no problem differentiating
repositories without database records from stale replicas of deleted
repositories anymore. If a repository doesn't have a database entry, it's
eligible for deletion.

There can also be stale repositories on the disks of the Gitalys due to
failed operations. For example, repository creation could succeed far
enough to create the repository but fail to acknowledge the creation to
Praefect. Such repository would not be recorded as existing, and should
be deleted from the disks. In these cases, Praefect wouldn't have database
records for the stale replicas.

These both cases can be handled by a clean up crawler that is being
implemented in Praefect. The crawler will walk the disks of the Gitaly
nodes and identify repositories which do not have database records. This
allows us to simplify our database model. When a repository is deleted,
all of its replica records can be deleted in the same transaction.

This commit implements this by cascading the delete from 'repositories'
table to the 'storage_repositories' table. 'DeleteRepository' is updated
to delete just the records in 'repositories' which then cascades to the
'storage_repositories' table. The method is updated to return the path
where the replicas are stored and the storages which are known
to hold a replica. This allows a later commit to remove replicas
from the storages immediately so the cluster remains clean even between
the clean up crawlers runs.

The schema still allows orphaned records in 'storage_repositories'. In a
later release, we can remove the remaining orphaned records and
enforce all 'storage_repositories' records have a repository_id set.

1 files changed, 2 insertions, 3 deletions

diff --git a/cmd/praefect/subcmd_remove_repository_test.go b/cmd/praefect/subcmd_remove_repository_test.go
index 1953c90a0..519ab4764 100644
--- a/cmd/praefect/subcmd_remove_repository_test.go
+++ b/cmd/praefect/subcmd_remove_repository_test.go
@@ -144,9 +144,8 @@ func TestRemoveRepository_Exec(t *testing.T) {
 	t.Run("no info about repository on praefect", func(t *testing.T) {
 		repo := createRepo(t, ctx, repoClient, praefectStorage, t.Name())
 		repoStore := datastore.NewPostgresRepositoryStore(db.DB, nil)
-		require.NoError(t, repoStore.DeleteRepository(
-			ctx, repo.StorageName, repo.RelativePath, []string{g1Cfg.Storages[0].Name, g2Cfg.Storages[0].Name},
-		))
+		_, _, err := repoStore.DeleteRepository(ctx, repo.StorageName, repo.RelativePath)
+		require.NoError(t, err)
 
 		logger := testhelper.NewTestLogger(t)
 		loggerHook := test.NewLocal(logger)