Details
-
Bug
-
Resolution: Unresolved
-
Major
-
None
-
3.0.4
-
None
-
1
Description
While running latency tests, I noticed a weird behaviour in one of the Latency-SSL tests: SET latency (ms), 4 nodes, 1 bucket x 20M x 1KB, 0/100 R/W, 100 threads, 2000 ops/sec, ReplicateTo=1 PersistTo=0, s=1 c=1, TLS".
As can be seen from the image, it has significantly higher latency than the equivalent test without collections. This behaviour is consistent for all versions of the couchbase server. However, this doesn't happen for other durability collection tests, where the results are in line with the durability tests without collections.
This is the latency chart from the cbmonitor report of the run with collections: http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=ares_710-1695_access_e1a8
And this is the latency chart from the cbmonitor report of the run without collections:
http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=ares_710-1695_access_0602
Looking at both charts, it seems that for the run with collection, 30-35% of the ops have low latency(under 1ms), in line with what is expected from the run without collection. However, at around 40% the latency quickly grows to more than 20ms.
Additionally, to make sure that tls is not the cause of the issues, I added a similar durability latency test with collections, but without tls: https://review.couchbase.org/c/perfrunner/+/168838. I ran it: http://perf.jenkins.couchbase.com/job/ares-dev/368/console and the latency is once again huge compared to the test without collections.
However, it seems that throughput is not affected by this bug.