Open andrwng opened 8 months ago
Removing from the rc5 milestone after some discussion with Michal. I was originally nervous this had contributed to higher latencies in a test cluster, but the underestimate used today means that compactions are run at even lower priority.
This issue hasn't seen activity in 3 months. If you want to keep it open, post a comment or remove the stale
label – otherwise this will be closed in two weeks.
Still needs to be worked on.
The current compaction controller backlog https://github.com/redpanda-data/redpanda/blob/90598fa8d23847af2f37b7f6ddfdb03f002ba8cb/src/v/storage/disk_log_impl.cc#L2218-L2355 estimates the amount of work compactions will do in order to adjust the priority of the scheduling group and io_priority_class.
This doesn't yet account for windowed compaction, which does more work than self+merge compaction. The result is that it's likely that compaction is likely being scheduled at too high a priority (as it's underestimating the amount of work done per compaction), which may in turn affect e2e latencies, etc.
JIRA Link: CORE-1680