We migrated away from WorkManager over 2 years ago. We needed it at the
time because we wanted to migrate jobs that were scheduled on
WorkManager into the new system. However, at this point, the user's
client would have been expired for 2 years at the point of upgrade, and
there wouldn't be any jobs that need migrating.
This was a bug that was most notable during the attachment pre-upload
process: if an attachment failed to upload, the subsequently-enqueued
PushMediaSendJob would still send. This is because the attachment jobs
were enqueued first and failed *before* we enqueued the PushMediaSendJob
as a dependency.
This will use the JobTracker to determine if a dependency already failed
at the time of enqueueing a job like this. This isn't perfect, because
the JobTracker is memory-only and has a limited buffer (currently 1000),
but in practice this should be sufficient for our use cases. I imagine
it'd only fall apart if we somehow enqueued a dependent job *much*
later, or somehow enqueued it based on a job ID that we persisted on
disk through an app restart. We don't do any of these things, currently,
and probably never should.
Also took the opportunity to patch a case where we weren't failing
dependent jobs when canceling a job, since I was giving the failure
stuff a look-over.
Currently the maxInstancesForQueue limit checks the count of all jobs in a
given queue. If there are already too many jobs, the new job is discarded.
However this is not the expected behavior for the two jobs where it's used:
GroupCallPeekWorkerJob and AutomaticSessionResetJob
For both the expected behavior is that there aren't too many jobs of them
started, but that there will be at least one instance of them started.
Both of them use the same queue as the PushProcessMessageJob and the MarkerJob.
Those two jobs are often in the queue at the same time, effectively preventing
the GroupCallPeekWorkerJob and AutomaticSessionResetJob from being enqueued.
Some devices actually enforce a scheduling rate, and will crash if you
submit more than, say, 250 jobs in 1 minute. This can happen when
catching up with messages and scheduling a lot of
PushDecryptMessageJobs.
While it'd be tricky to limit jobs with constraints, this just does the
simple thing of not enqueueing unnecessary jobs for constraint-less
jobs.
