1189504 - are telemetry/fhrV4 missing subsessions associated with crash-on-startup?

Status: RESOLVED FIXED

(Toolkit :: Telemetry, defect, P1)

Description

[Sam Penrose]

+++ This bug was initially created as a clone of Bug #1189502 +++

Are the pings investigated in bug 1171268 associated with startup crashes? Can we extract uptime from crash pings?

Comment 1

[Sam Penrose]

80% are not associated with any crash at all:

  http://nbviewer.ipython.org/gist/vitillo/3047c0d896b08f75c403

90% of clients without missing pings do not have a crash.

Comment 2

[Benjamin Smedberg]

No reason for this to be confidential. I'm having trouble matching up your statement that most missing subessions aren't related to a crash with rvitillo's assertion that missing subsessions are highly correlated with crash pings.

Comment 3

[Sam Penrose]

(In reply to Benjamin Smedberg  [:bsmedberg] from comment #2)
> No reason for this to be confidential. I'm having trouble matching up your
> statement that most missing subessions aren't related to a crash with
> rvitillo's assertion that missing subsessions are highly correlated with
> crash pings.

Roberto: clients with missing pings are twice (20% vs 10%) as likely to have crashed.

Me: 80% of clients with missing pings did not have a crash.

Coincidentally, the crash rate Roberto found (5%) times 20% gives 1%, which is the threshold you identified for this-issue-is-not-a-blocker.

Finally, both crashes and missing pings seem prima facie likely to be associated with increased use. It may be that Roberto has accounted for this factor; I don't see it in a quick scan of his notebook. Perhaps I will check that.

Comment 4

[Sam Penrose]

I meant to say: "20% of the has-missing-subsession-ping rate (5%) is 1%, which ...

(In reply to Sam Penrose from comment #3)
> Coincidentally, the crash rate Roberto found (5%) times 20% gives 1%, which
> is the threshold you identified for this-issue-is-not-a-blocker.

Comment 5

[Roberto Agostino Vitillo (:rvitillo)]

(In reply to Sam Penrose from comment #3)
> (In reply to Benjamin Smedberg  [:bsmedberg] from comment #2)
> > No reason for this to be confidential. I'm having trouble matching up your
> > statement that most missing subessions aren't related to a crash with
> > rvitillo's assertion that missing subsessions are highly correlated with
> > crash pings.
> 
> Roberto: clients with missing pings are twice (20% vs 10%) as likely to have
> crashed.
> 
> Me: 80% of clients with missing pings did not have a crash.

I was concerned that crashes might have been underreported as I have seen pings with STARTUP_CRASH_DETECTED set to true but without a corresponding crash ping. But then again this might be due to the way I associated a crash ping to a broken session chain using timestamps, since we don't have a good way yet to do that.

> Coincidentally, the crash rate Roberto found (5%) times 20% gives 1%, which
> is the threshold you identified for this-issue-is-not-a-blocker.
> 
> Finally, both crashes and missing pings seem prima facie likely to be
> associated with increased use. It may be that Roberto has accounted for this
> factor; I don't see it in a quick scan of his notebook. Perhaps I will check
> that.

That's a good point, I just checked the median number of fragments for clients with missing pings and without and indeed clients with missing pings have 2x more fragments.

Comment 6

[Georg Fritzsche [:gfritzsche]]

(In reply to Roberto Agostino Vitillo (:rvitillo) from comment #5)
> (In reply to Sam Penrose from comment #3)
> > (In reply to Benjamin Smedberg  [:bsmedberg] from comment #2)
> > > No reason for this to be confidential. I'm having trouble matching up your
> > > statement that most missing subessions aren't related to a crash with
> > > rvitillo's assertion that missing subsessions are highly correlated with
> > > crash pings.
> > 
> > Roberto: clients with missing pings are twice (20% vs 10%) as likely to have
> > crashed.
> > 
> > Me: 80% of clients with missing pings did not have a crash.
> 
> I was concerned that crashes might have been underreported as I have seen
> pings with STARTUP_CRASH_DETECTED set to true but without a corresponding
> crash ping. But then again this might be due to the way I associated a crash
> ping to a broken session chain using timestamps, since we don't have a good
> way yet to do that.

Bug 1187270 takes care of that.