Github user revans2 commented on the issue:
https://github.com/apache/storm/pull/2203
I did a quick analysis of the latency vs CPU usage for the branch from
@HeartSaVioR and this branch based off of the test results posted by
@HeartSaVioR (just to get more of an apples to apples comparison, but my test
results look similar).
It looks like metrics_v2_replace_meters_to_counters had an increased
latency over 1.2.0-SNAPSHOT by about 2.1% to 4.0% (depending on if I do a mean
vs a median of the reported mean latency values) Similarly for CPU usage it
was between 1.7% to 3.8% more CPU usage.
When using the normal metrics v2 it was a 7.2% to 8.0% increase in latency
and 7.0% to 10.1% increase in CPU. So this is a huge step in the right
direction.
I really would love to see an overhead of < 1%, but I could probably live
with these numbers.
Would it be possible to try an experiment where we also do random
sub-sampling with these metrics? I know it is a giant confusion for people to
not have an exact number for metrics, but I would be curious to see what
difference if any it would make so we know what the issue is.
---
