This changes the BackgroundMigration worker so it checks for the health
of the DB before performing a background migration. This in turn allows
us to reduce the minimum interval, without having to worry about blowing
things up if we schedule too many migrations.
In this setup, the BackgroundMigration worker will reschedule jobs as
long as the database is considered to be in an unhealthy state. Once the
database has recovered, the migration can be performed.
To determine if the database is in a healthy state, we look at the
replication lag of any replication slots defined on the primary. If the
lag is deemed to great (100 MB by default) for too many slots, the
migration is rescheduled for a later point in time.
The health checking code is hidden behind a feature flag, allowing us to
disable it if necessary.
This adds a minimum interval to BackgroundMigrationWorker, ensuring
background migrations of the same class only run once every 5 minutes.
This prevents a thundering herd problem where scheduled migrations all
run at once due to their delays having been expired (e.g. as the result
of a queue being paused for a long time).
If a job was recently executed it's rescheduled with a delay that equals
the remaining time of the job's lease. This means that if the lease
expires in two minutes we only need to wait two minutes, instead of
five.
Fixes https://gitlab.com/gitlab-org/gitlab-ce/issues/41624
