If we compile locally before testing in automation, we're no longer testing
the same bits that we're going to be shipping.  That's a major change in
direction, and one that may have significant consequences.

An additional historical reason for keeping development tools off test
machines: we once shipped a new runtime dependency that we didn't detect
because the test machines all had development environments.  Users on that
platform were unable to run Firefox without installing a 3rd party library.

I very much agree with the goal of simplifying and streamlining how we run
tests.  I think we do have to keep the above in mind when we go about that
simplification.


On Thu, Sep 22, 2016 at 3:54 PM, Gregory Szorc <g...@mozilla.com> wrote:

> On Thu, Sep 22, 2016 at 2:49 PM, Nicholas Alexander <
> nalexan...@mozilla.com>
> wrote:
>
> >
> >
> > On Thu, Sep 22, 2016 at 1:20 PM, Gregory Szorc <g...@mozilla.com> wrote:
> >
> >> For years it has been a common pattern for local builds and Firefox
> >> automation to copy test files to the object directory (or some other
> >> staging area) and run them from there. This creates several
> inefficiencies:
> >>
> >> * The build system must symlink or copy thousands of test related files
> >> during the build
> >> * Running tests requires checking if test files in the objdir are up to
> >> date (can add significant latency to the edit-test loop)
> >> * Builds in automation must create and upload archives containing
> >> thousands of test files
> >>   * This makes it harder to skip build jobs if all that has changed is a
> >> test file
> >> * Test jobs in automation must download and extract archives containing
> >> test files
> >> * Test jobs in automation typically don't have access to a full source
> >> checkout and can't easily piggyback on existing infrastructure (such as
> >> mach commands, vendored Python packages, etc). This leads to wheel
> >> reinvention and inconsistency between local dev environments and
> automation.
> >>
> >> The historical reasons for doing things this way were valid. But with
> >> moz.build, other improvements to the build system, automation scheduling
> >> and configuration living in-tree, and better scaling of Mercurial in
> >> automation, these historical reasons are largely no longer valid.
> >>
> >> Going forward, I'm requesting we change our policy regarding tests and
> >> automation to be source checkout first. This means:
> >>
> >> * Tests should be designed to run from a source checkout with minimal
> >> "build" actions required (no file copying/symlinking, no preprocessing
> into
> >> a new file, etc)
> >> * Automation should run everything from a source checkout (as opposed to
> >> downloading archives containing files derived from source checkouts)
> >>
> >> Does anyone have any concerns or objections to this?
> >>
> >
> > This only makes sense for a subset of our harnesses.  Tests that need to
> > be compiled (gtest?, Android unit and Robocop tests) can never achieve
> > this.  So as long as we're clear that we want to remove preprocessing and
> > special mozharness-foo to work around packaging tests but we're not
> pushing
> > *everything* to source-only, I'm good.
> >
>
> The goal is to reduce the amount of work needed to run tests. You are
> correct that compiled tests inherently need a build step to run them and we
> can't simply run them from a source checkout.
>
> That being said, it isn't out of the question to defer the compilation of
> these tests to test jobs: the build job delays execution of nearly all
> tests and making the build job faster by moving work out of that job could
> make end-to-end times faster, even if we're doing extra work in individual
> test jobs (that have to compile tests). Historically, we didn't have a
> compiler available on test machines. But with TaskCluster+Docker and
> tooltool, it is much easier to put needed files on testers. So compiling
> tests on test machines is something we should consider.
> _______________________________________________
> tools mailing list
> to...@lists.mozilla.org
> https://lists.mozilla.org/listinfo/tools
>
_______________________________________________
dev-builds mailing list
dev-builds@lists.mozilla.org
https://lists.mozilla.org/listinfo/dev-builds

Reply via email to