Australis update

Matt Woodrow matt.woodrow at gmail.com
Wed Nov 13 22:58:15 UTC 2013


Alright, I'm convinced. :)

I have a few leads that I'm working on today which I'd like to finish
and get try runs for.

I'll prepare backout patches with the intent of landing them tomorrow if
that doesn't come through.

Sound reasonable?

- Matt

On 14/11/13 11:47 AM, Justin Dolske wrote:
> On 11/13/13 2:03 PM, Matt Woodrow wrote:
>> I agree that this shouldn't block the australis landing.
>>
>> I don't think that we need to rush into a backout just yet though.
>> Reverting the changing is incredibly simple, and it's very unlikely that
>> we'll regress anything on the existing path (since all other platforms
>> are still using it).
>>
>> I'd much prefer to leave it in for now, so that we can get more user
>> coverage of the changes and potentially find out more information on the
>> effects the change has. We already missed one regression on
>> mozilla-central, it's fairly plausible that there are other issues not
>> covered by tests on any branch.
>
> It should be backed out (or pref'd off) as soon as possible, for a few
> reasons:
>
> * Tree policy is pretty clear about the need to back out regressions.
> This is a bit of a special circumstance, but had the order been
> reversed (with Australis landing a few days before your patch) there
> would be no question about it.
>
> * Our bug/perf tracking tools are... not great, and are geared around
> detecting regressions when they land. Not for making sure that future
> improvements restore previous performance. (Part of the reason for the
> backout policy.) This only gets more complicated the longer things
> stay in the tree.
>
> * When Australis was discovered to regress TART back in August --
> coincidentally, also around 40% -- there was extreme pushback from all
> quarters stressing the importance of not regressing performance. Even
> though I said that would mean we'd have to redo a bunch of work and
> blow out the schedule by months (which is exactly what happened). So
> the tolerance for regression has already been set, and should apply
> here as well.
>
> Justin
>



More information about the firefox-dev mailing list