Details
-
Bug
-
Resolution: Incomplete
-
Critical
-
4.5.0
-
Triaged
-
No
Description
Recent Array Indexing results have been published here:
http://showfast.sc.couchbase.com/#/timeline
Please review Q2, stale=ok results.
If you see title has compaction_mode = full, this means circular compaction is disabled.
For your reference, this is the html report:
http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=hades_450-1613-enterprise_a66_access
A few things I noticed:
We noticed that avg query latency in Array Q2(stale=ok) is close to 100ms.
For comparison, for Q2(stale=ok), avg query latency is 8~10ms.
A few things we noticed in the showfast. (Only focused on Q2, stale=ok results now)
Indexer CPU utilization is relatively low. For array indexing Q2(stale=ok), it is 2 cores. For Q2(stale=ok), it is close to 12 cores.
Avg scan latency is 20ms in the query log. But in show fast, avg query latency is 100ms.
John's analysis here;
===
Look at indexer CPU graph, it does not look like there is a lot GC going on. So it is not clear how the next 2i build will help (with GC optimization).
>> Indexer CPU utilization is relatively low. For array indexing Q2(stale=ok), it is 2 cores. For Q2(stale=ok), it is close to 12 cores.
Since throughput is more than 10x slower than Q2, so I would expect indexer CPU utilization is lower as well.
"Keshav, cbq cpu could be maxed out at 24 cores (note that you have hyper threading turned off in this cluster). "
====
Question for the query team:
Based on the query log, the avg index scan latency is only 20ms, why avg query latency is 100ms?
This is the index/query definition:
https://github.com/couchbase/perfrunner/blob/master/tests/n1ql/n1ql_thr_array_indexing_Q2_20M_gsi_ok.test.
This is one sample run:
http://perf.jenkins.couchbase.com/job/hades/2915/
Please let me know if you need anything else.