To fix a race between split and repair here c1de4859d8, a new sstable
generated during streaming can be split before being attached to the sstable
set. That's to prevent an unsplit sstable from reaching the set after the
tablet map is resized.
So we can think this split is an extension of the sstable writer. A failure
during split means the new sstable won't be added. Also, the duration of split
is also adding to the time erm is held. For example, repair writer will only
release its erm once the split sstable is added into the set.
This single-sstable split is going through run_custom_job(), which serializes
with other maintenance tasks. That was a terrible decision, since the split may
have to wait for ongoing maintenance task to finish, which means holding erm
for longer. Additionally, if split monitor decides to run split on the entire
compaction group, it can cause single-sstable split to be aborted since the
former wants to select all sstables, propagating a failure to the streaming
writer.
That results in new sstable being leaked and may cause problems on restart,
since the underlying tablet may have moved elsewhere or multiple splits may
have happened. We have some fragility today in cleaning up leaked sstables on
streaming failure, but this single-sstable split made it worse since the
failure can happen during normal operation, when there's e.g. no I/O error.
It makes sense to kill run_custom_job() usage, since the single-sstable split
is offline and an extension of sstable writing, therefore it makes no sense to
serialize with maintenance tasks. It must also inherit the sched group of the
process writing the new sstable. The inheritance happens today, but is fragile.
Fixes#20626.
Closesscylladb/scylladb#20737
* github.com:scylladb/scylladb:
tablet: Fix single-sstable split when attaching new unsplit sstables
replica: Fix tablet split execute after restart
(cherry picked from commit bca8258150)
Ref scylladb/scylladb#21415
a3a0ffbcd0