[tbb-dev] Thoughts about future Tor Browser plans

Matthew Finkel sysrqb at torproject.org
Thu Apr 8 13:45:28 UTC 2021


On Thu, Apr 08, 2021 at 05:53:16AM +0000, Georg Koppen wrote:
> Matthew Finkel:
> > Hi everyone,
> > 
> > (tldr; read the last paragraph)
> > 
> > In October [0] we were very excited about the prospect of all Tor
> > Browser platforms following the Firefox Rapid Release schedule. However,
> > in April (now), Android is still the only platform following the rapid
> > release train and Windows, macOS and Linux remain on the extended
> > support release (ESR).
> > 
> > As we move closer to the next ESR transition, this year it is beginning
> > at Firefox 91 in May [1], I am wondering whether we should reverse
> > course and slow down. At this point, we cannot safely transition all
> > platforms onto the rapid release train before October (when 78esr
> > reaches its EOL), so the only option is moving all desktop platforms
> > onto FF91esr and then evaluate migrating onto the rapid release train
> > after that.
> > 
> > Unfortunately, we are still a very small team, and the current situation
> > is already pushing our team to its limits. Yesterday I spoke with Georg
> > about another subject, and he briefly mentioned the idea of keeping the
> > desktop platforms on ESR instead of moving onto the rapid releases. This
> > alone wouldn't make much difference, and I previously discarded this
> > idea, because we're already fighting all of the code churn for Android
> > (although some of the code, like the automatic updater, is not used on
> > Android). The important piece of this plan is reducing Fenix's burden,
> > too. 
> > 
> > My current thought is that we move desktop platforms from 78esr to
> > 91esr, and we begin rebasing our Fenix branches (and dependencies) only
> > every 2-3 months. The only exception is geckoview where we continue
> 
> So, for the timeline for that plan: this change is going to happen with
> the ESR91 transition? Or now? Or after it?

I would begin this next month, when we start the 91esr transition. We
could begin slowing our rebasing of Fenix this month, but we already
started that effort, so we might as well finish it.

> 
> > rebasing our geckoview branches on the existing schedule. In addition,
> > we drop desktop patches from the geckoview branches, and Android patches
> > from the ESR branches. We could've (and should've) done the latter
> > simplification earlier, but the former change makes future ESR
> > transitions a little more complicated. I think we can live with that.
> 
> I have three concerns with that plan:
> 
> 1) Getting rid of mobile patches on desktop and vice versa sounds like a
> good idea on first glance. However, that would mean extra work for our
> test suite part as right now we require Linux builds (and thus desktop
> patches, too) to be sure non of our (GecKoView) patches requires
> (further) changes. One could argue we could drop that part as those
> tests are, to a large extent, desktop related tests anyway. I am not
> sure whether you wanted to make that point, though (and there *are* a
> bunch of tests that test functionality for mobile as well)

I agree this is a current problem we should solve. I'm curious how many
desktop-only patches are needed for successfully running the testsuite's
tests that cover shared functionality between linux and geckoview. I
suspect not many. In particular, if we can drop half of our current
geckoview patches, then that is a significant win.

> 
> 2) We have seen in the past while working on toolchain updates that the
> GeckoView, android-components, and Fenix parts are quite tightly
> coupled. You see that at the Gradle dependencies they require and that
> things broke already during build when the wrong GeckoView version got
> mixed with android-components/Fenix due to API incompatibilities. Just
> focusing on continuous GeckoView rebasing while ignoring the other parts
> will likely hit that problem again causing us to have extra work fixing
> that up. (Even without *that* extra work there will be additional work
> needed for de-coupling the those above mentioned tightly coupled
> projects on the build side, even though we might feel this could be
> acceptable)

I partially agree with this. GeckoView provides a deprecation period for
its (breaking) API changes. That period is usually 3 versions. As for
the Fenix and MozAC incompatibility, those issues came from newer
Fenix/MozAC branches and an older geckoview branch, if I remember
correctly. I believe the reverse situation should be safer.

> 
> 3) Somewhat related to 2) I am worried about instability we introduce by
> de-coupling GeckoView, android-components, and Fenix. It's often not
> seen (as it is hardly visible on desktop either) but we greatly benefit
> from all the testing Mozilla is running on those three components. But
> that testing is *only* valid if we follow the GeckoView +
> android-components + Fenix pipeline as close as we can. What happens
> with your plan is that me move straightly into untested territory and I
> doubt we can compensate for that given the resource issues you cite.

This may be the strongest argument for not following the new plan. I
suspect that running more tests (on vanilla Fenix and MozAC) should be
sufficient and will be less time consuming than our current situation,
but we can run an experiment before committing to this process change.

> 
> Georg
> 
> > - Matt
> > 
> > [0] https://lists.torproject.org/pipermail/tbb-dev/2020-October/001154.html
> > [1] https://wiki.mozilla.org/Release_Management/Calendar
> > _______________________________________________
> > tbb-dev mailing list
> > tbb-dev at lists.torproject.org
> > https://lists.torproject.org/cgi-bin/mailman/listinfo/tbb-dev
> > 
> 
> 




> _______________________________________________
> tbb-dev mailing list
> tbb-dev at lists.torproject.org
> https://lists.torproject.org/cgi-bin/mailman/listinfo/tbb-dev



More information about the tbb-dev mailing list