New issue
Advanced search Search tips
Note: Color blocks (like or ) mean that a user may not be available. Tooltip shows the reason.

Issue 654771 link

Starred by 1 user

Issue metadata

Status: Started
Owner:
Last visit > 30 days ago
Cc:
Components:
EstimatedDays: ----
NextAction: ----
OS: Windows
Pri: 2
Type: Bug



Sign in to add a comment

Crashes reported with the sampled-out experiment.

Project Member Reported by jwd@chromium.org, Oct 11 2016

Issue description

Crashes reported with the sampled-out experiment. The study "MetricsAndCrashSampling" is intended to control sampling of metrics and crash reporting. If a client is in the experiment "OutOfReportingSample", it shouldn't be reporting crash or metrics. While running a test on canary, we noticed a suspiciously high number of crashes tagged with the "OutOfReportingSample" experiment. No such clients were detected in metrics data.

 

Comment 1 by jwd@chromium.org, Oct 11 2016

There were few enough crashes that it seemed like sampling was having some effect, but the numbers were still suspiciously high.

Comment 2 by jwd@chromium.org, Oct 11 2016

Cc: asvitk...@chromium.org
I've found one repro for this. If chrome is running in the OutOfReportingSample state, then it's launched again with a different user data dir, that would put it in the in-sample state, the original chrome will report a crash with OutOfReportiSample. This is because crashpad uses the default user data dir to store the crash database, and will thus be shared between instances of chrome.

Based on the crash data, this doesn't seem very likely though. There is not many crashes reporting with the --user-data-dir flag, in either the Out or In experiments.

One bug that mentions the crashpad issue:  crbug.com/565446 
And a CL comment that also mentions it: https://codereview.chromium.org/1481703002/#msg31

Comment 3 by jwd@chromium.org, Oct 17 2016

A more likely cause has been found. The installer shares the crash database with chrome, and the installer crash client doesn't know about sampling. When the installer runs to install an update, it will turn on crash reporting.

I've repro'd locally by leaving a canary with that was sampled out running over the weekend. Installer was run at some point for new versions of carry. Checked on monday, and that client can report crashes, and reports the sampled out experiment.

I'm going to update the installer's crash client to check sampling.
You started fixing this bug over two years ago. Are you still working on it? 

Sign in to add a comment