New issue
Advanced search Search tips

Issue 894300 link

Starred by 1 user

Issue metadata

Status: Untriaged
Owner: ----
Cc:
Components:
EstimatedDays: ----
NextAction: 2018-10-15
OS: ----
Pri: 1
Type: Bug

Blocking:
issue 893828



Sign in to add a comment

Erroneous memory usage alerts on android-go-perf

Project Member Reported by m...@chromium.org, Oct 11

Issue description

Graphs show a distinct increase in resident memory size: https://chromeperf.appspot.com/group_report?bug_id=893828

However, Pinpoint's numbers agree with the pre-regression values: https://pinpoint-dot-chromeperf.appspot.com/job/112dc132e40000

And, Pinpoint sees no regression.

Therefore, perhaps there is something wrong with the bot that generates measurements for the dashboard?
 
The NextAction date has arrived: 2018-10-15
Cc: perezju@chromium.org
Components: -Speed>Bisection Speed>Benchmarks
This is a fairly common occurrence. I think we mostly just WontFix in these scenarios.

Nothing really stands out to me as the cause. I see that:
* The bisect ran 40 repeats on many different devices, so probably not random chance.
* The regression range of the bisect was pretty wide, so it's not a misplaced alert.
* The device on the perf waterfall did not change.

Side note: I think the Pinpoint results look more like the post-regression values than the pre-regression values, but they doesn't seem to completely match up with either.
Yeah, not sure there is much we can do about this. Dave, do we have any stats on how common is this kind of issue?
FWIW, it seems to happen every time I am perf sheriff. It's frustrating because the graphs often step up like there's a clear regression point, and then the bisects always fail to find anything. It would help if the benchmarks all had "ref" datasets so we could just ignore the alerts. (Although, having to set the ignore status on hundreds of alerts can be tedious too.)

As this is common, we really should root-cause this issue. Otherwise, disable alarming on these.

Sign in to add a comment