Details
-
Bug
-
Resolution: Fixed
-
Major
-
7.0.0, 7.0.1, 7.0.2, 7.0.3, 7.1.0
-
Triaged
-
1
-
Unknown
-
KV 2022-Jan
Description
While investigating MB-50389, when concurrent compaction tasks were set to 2 in ep-engine, it was observed that 2 tasks were not always running concurrently, even though there were outstanding tasks - from the phosphor trace:
Initially there are two concurrent CompactTasks running (as expected); however when they finished (around the same time) only one CompactTask is observed to be scheduled after. This continues; with just a single compactor running.
Examining the wakeup phase at the end of Compaction (by adding some additional trace spans), we can observe that both of the Compact tasks which were running at 28,650ms attempted to wake up the same Compact task (vb:2):
Title updateCompactionTasks.wake
|
Category ep-engine/task
|
User Friendly Category other
|
Start 28,666.847 ms
|
Wall Duration 0.005 ms
|
Args
|
vbid 2
|
The next (single) CompactTask to run at 28,671ms in indeed vb:2. Instead of each finishing CompactTask waking up a new, different task; they both wake up the same, so we end up only starting one new task - and the concurrency drops.
Attachments
Issue Links
- relates to
-
MB-50389 Increase in tail latencies for SyncWrite / Observe operations
- Closed