New issue
Advanced search Search tips

Issue 765336 link

Starred by 1 user

Issue metadata

Status: Assigned
Owner:
Cc:
Components:
EstimatedDays: ----
NextAction: ----
OS: ----
Pri: 2
Type: Bug



Sign in to add a comment

Few RTT observations may dominate the estimate resulting in incorrect estimate

Project Member Reported by tbansal@chromium.org, Sep 14 2017

Issue description

Under certain circumstances, it is possible that a few observations may have more recent timestamps while most of the old observations have very old timestamps. This may cause some of the observations to have very high weight, and dominate the estimate computation. In this case, if the few estimations are incorrect (e.g., have high RTT value due to hanging GETs), then the computed estimate may have incorrect value.

Data to gather to determine how frequently :
1. Record the weight of the observation with max. weight in UMA.
2. Record the time delta since the most recent observation in UMA.
3. Record the # of observations in the buffer in UMA.

Possible solutions:
If any observation has more than 0.X weight (say X = 0.33), then may be set weight of all observations to be equal?

 

Comment 1 by bengr@chromium.org, Sep 14 2017

Status: Available (was: Untriaged)
Please expand upon what (1) means, explain when all three would be recorded, and how these would help understand this issue.

Also, I don't think that solution is quite right.
Yeah, I am not sure about solution either. We should collect some UMA before trying different solutions.

We can collect UMA every time ECT is recomputed. The goal is to collect data on how frequently a very small subset of observations dominate the estimate computation. If there are N observations, each with different weight, then one way is to record the weight of the observation that has the highest weight.

Comment 3 by bengr@chromium.org, Dec 1 2017

What's the plan?
Cc: tbansal@chromium.org
tbansal, I've been getting hit with previews on my WiFi again. I think UNKNOWN should be used until we have a confident estimate.
ryansturm: Please try to get a netlog or just a dump of chrome://histograms. You can file the in-app feedback report, and that will automatically capture the data from chrome://histograms.
Talked with ryansturm@ offline sometime back. The NQ was correctly computed (as seen in chrome://interventions-internals log), but we suspected that there was some other bug somewhere in the previews pipeline.
Tarun, any updates on this issue?

Comment 8 by bengr@chromium.org, Jun 22 2018

Labels: -Pri-3 M-70 Pri-2
Owner: tbansal@chromium.org
Status: Assigned (was: Available)
Refreshed during triage.
Refreshed during triage.

Refreshed during triage.

Comment 12 by tbansal@chromium.org, Yesterday (25 hours ago)

Refreshed during triage.

Sign in to add a comment