Details
-
Bug
-
Resolution: Unresolved
-
Major
-
5.5.0
-
Triaged
-
No
Description
The stat gathering for scan uses atomics. But at very high concurrency, that doesn't seem to be scaling.
After adding new stats for aggregates, the scan latency got worse in showfast:
http://showfast.sc.couchbase.com/#/timeline/Linux/secondary/lat/MOI
http://showfast.sc.couchbase.com/#/timeline/Linux/secondary/lat/Plasma
The difference in results for build 5.5.0-2036 vs 5.5.0-2126.
It would be better to split the stats into a pool of stats which the scans can update concurrently by taking out an object out of the pool and the final stats can be aggregated at collection time.
Attachments
Issue Links
- relates to
-
MB-28017 Latency of aggregation queries is higher than expected
- Closed
Gerrit Reviews
For Gerrit Dashboard: MB-28698 | ||||||
---|---|---|---|---|---|---|
# | Subject | Branch | Project | Status | CR | V |
92806,3 | MB-28698 Collect expr eval stats per scan | unstable | indexing | Status: MERGED | +2 | +1 |
92860,2 | MB-28698 fix stat gathering for non expr scans | unstable | indexing | Status: MERGED | +2 | +1 |