Details
-
Bug
-
Resolution: Fixed
-
Major
-
7.6.0, 7.2.0, 7.2.1
-
None
-
Untriaged
-
0
-
Yes
Description
As I was working on a fix for MB-56634, I realized that this could cause issues for XDCR's high/low priority replication switches. This is because the XDCR's high/low priority design depends on real time and accurate accounting of the process's CPU usages.
Thus, when I looked at showfast, I saw some odd regression w.r.t. the high/low performance numbers when compared to pre 7.2.0.
As a follow up, I built a toy-build on 7.6 that contains the fix for MB-56634 (which means CPU account should be now accurate) and ran the same performance test http://perf.jenkins.couchbase.com/job/titan_5node1/1920/console
The result from the toy run comes out with:
high: 128222
low: 15947
Which brings the result of the showfast test back in line with the original pre-regression numbers.
The fix is already done in 7.6 as the commit https://review.couchbase.org/c/goxdcr/+/194226 is checked into master (7.6 and above).
However, it seem that 7.2.1 is still affected. This MB is to document the issue and document the fixes that has gone in.
I will be testing a toy build against 7.2.1 to validate