Closed Bug 1463858 Opened 6 years ago Closed 5 years ago

Regression in SEARCH_SERVICE_INIT_MS on or around 2018-04-22

Categories

(Firefox :: Search, defect, P3)

defect

Tracking

()

RESOLVED WONTFIX
Tracking Status
firefox62 --- affected

People

(Reporter: chutten, Unassigned)

References

Details

There was a sudden change in the Telemetry probe SEARCH_SERVICE_INIT_MS which seems to have occurred in Nightly 20180423[2][3].

There was a shift to the right in the distribution of timings without a commensurate change in the number of samples or ping. This is consistent with a performance regression in the code under measurement.

Is this actually a regression?

Is this intentional? Is this expected?

Is this probe still measuring something useful? Its "high" value seems rather lower than most of the samples recorded by users in the field.

[2]: https://mzl.la/2s7L2fj
[3]: https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=8d4cf28964f6956cd22d8710b316456e2c7c848d&tochange=dfb15917c057f17e5143f7d7c6e1972ba53efc49
Most likely candidate looks to be bug 1449317.
Blocks: 1449317
As far as I can tell, the regression on the median is around 400ms.

The patch in bug 1449317 doesn't touch the actual initialisation process, though it may start the initialisation slightly earlier.

It does add an additional observer, but if I'm reading the code correctly, then I think they won't be counted in the timing.

So my only thought is that this is a result of the initialisation being kicked off earlier, other things still running, and causing it to be overall slower.

Florian might have some ideas when he gets back.
Priority: -- → P3

I suspect it might be a bit late to do anything on this now. Though just in case... Florian, any ideas?

Flags: needinfo?(florian)

Yeah, it's a bit late. I agree with comment 2; starting this earlier at a time when the process is busier could have made the whole init overall slower. Measuring the length of something async that can be mixed with other random things happening in the middle is not a very stable measurement.
It's frustrating to not understand exactly what happened, but I doubt we can make this bug actionable at this point. If we think search service init takes too long, the best thing to do would be to profile it on slow hardware and see what happens.

Status: NEW → RESOLVED
Closed: 5 years ago
Flags: needinfo?(florian)
Resolution: --- → WONTFIX
You need to log in before you can comment on or make changes to this bug.