fix(server): deprioritize METADATA_EXTRACTION to let LINK_LIVE_PHOTOS complete first

[mPyKen]

Least intrusive change to partially fix #12306.

When stopping and clearing the METADATA_EXTRACTION queue midway, METADATA_EXTRACTION queue must be forcefully filled and run again to LINK_LIVE_PHOTOS as LINK_LIVE_PHOTOS jobs are processed after all the METADATA_EXTRACTION jobs. Hence, LINK_LIVE_PHOTOS jobs should be prioritized over METADATA_EXTRACTION.

This also fixes UX such that the queue size actually decreases on the job status page evenly over time.

[mertalev]

Hi! Thank you for the PR, but we can't merge this as it is. Metadata extraction is required for most other jobs, so deprioritizing it is not a good idea.

The situation you're describing can be solved in other ways instead, such as by tracking live photo linking separately from metadata extraction in the job status table.

[jrasm91]

Closing per merts comment.

[mPyKen]

@mertalev My understanding is that priority affects only jobs in the same queue, not across queues. So in this case, deprioritizing metadata extraction only means prioritizing JobName.QUEUE_METADATA_EXTRACTION and JobName.LINK_LIVE_PHOTOS over JobName.METADATA_EXTRACTION as these 3 are the only ones appearing in QueueName.METADATA_EXTRACTION at the moment. I don't see why this would be a problem.

If you say you want to fix the original issue differently, sure. This was supposed to be a quick fix so that others do not run into the same issue as I did.

[mertalev]

The issue is that the jobs in the other queues can't run until metadata extraction is finished, so lowering its priority has a cascade effect on the job system.

[mPyKen]

For JobName.QUEUE_METADATA_EXTRACTION Q, JobName.LINK_LIVE_PHOTOS L and JobName.METADATA_EXTRACTION M, we currently complete jobs of METADATA_EXTRACTION queue in this order:
QMMMMMLLLLL
whereas this PR changes this order to
QMLMLMLMLML
The first jobs in other queues will still start after the first M job. As LINK_LIVE_PHOTOS is just a couple of database queries, I thought it wouldn't affect the job system that much. Is that not the case?

[mertalev]

I've seen cases where metadata extraction was the bottleneck and the other queues were waiting on it. I just think a different approach would accomplish the same result without that tradeoff.