Re: [PATCH 4/4] ci: Add check-migration-quick to the clang job

[Fabiano Rosas <farosas@suse.de>]

Daniel P. Berrangé <berrange@redhat.com> writes:

> On Fri, Oct 18, 2024 at 10:51:03AM -0300, Fabiano Rosas wrote:
>> Daniel P. Berrangé <berrange@redhat.com> writes:
>> 
>> > On Fri, Oct 18, 2024 at 10:46:55AM +0100, Peter Maydell wrote:
>> >> On Fri, 18 Oct 2024 at 10:01, Daniel P. Berrangé <berrange@redhat.com> wrote:
>> >> >
>> >> > On Thu, Oct 17, 2024 at 01:29:35PM -0300, Fabiano Rosas wrote:
>> >> > > Daniel P. Berrangé <berrange@redhat.com> writes:
>> >> > >
>> >> > > > On Thu, Oct 17, 2024 at 11:32:11AM -0300, Fabiano Rosas wrote:
>> >> > > >> Recent changes to how we invoke the migration tests have
>> >> > > >> (intentionally) caused them to not be part of the check-qtest target
>> >> > > >> anymore. Add the check-migration-quick target so we don't lose
>> >> > > >> migration code testing in this job.
>> >> > > >
>> >> > > > But 'check-migration-quick' is only the subset of migration tests,
>> >> > > > 'check-migration' is all of the migration tests. So surely this is
>> >> > > > a massive regressions in covage in CI pipelines.
>> >> > >
>> >> > > I'm not sure it is. There are tests there already for all the major
>> >> > > parts of the code: precopy, postcopy, multifd, socket. Besides, we can
>> >> > > tweak migration-quick to cover spots where we think we're losing
>> >> > > coverage.
>> >> >
>> >> > Each of the tests in migration-test  were added for a good reason,
>> >> > generally to address testing gaps where we had functional regressions
>> >> > in the past. I don't think its a good idea to stop running such tests
>> >> > in CI as gating on new contributions. Any time we've had optional
>> >> > tests in QEMU, we've seen repeated regressions in the area in question.
>> >> >
>> >> > > Since our CI offers nothing in terms of reproducibility or
>> >> > > debuggability, I don't think it's productive to have an increasing
>> >> > > amount of tests running in CI if that means we'll be dealing with
>> >> > > timeouts and intermittent crashes constantly.
>> >> >
>> >> > Test reliability is a different thing. If a particular test is
>> >> > flaky, it needs to either be fixed or disabled. Splitting into
>> >> > a fast & slow grouping doesn't address reliability, just hides
>> >> > the problem from view.
>> >> 
>> >> A lot of the current reliability issue is timeouts -- sometimes
>> >> our CI runners just run really slow (I have seen an example where
>> >> between a normal and a slow run on the same commit both the
>> >> compile and test times were 10x different...) So any test
>> >> that is not a fast-to-complete is much much more likely to
>> >> hit its timeout if the runner is running slowly. When I am
>> >> doing CI testing for merges "migration test timed out again"
>> >> is really really common.
>> >
>> > If its frequently timing out, then we've got the timeouts
>> > wrong, or we have some genuine bugs in there to be fixed.
>> >
>> >> > > No disagreement here. But then I'm going to need advice on what to do
>> >> > > when other maintainers ask us to stop writing migration tests because
>> >> > > they take too long. I cannot send contributors away nor merge code
>> >> > > without tests.
>> >> >
>> >> > In general, I think it is unreasonable for other maintainers to
>> >> > tell us to stop adding test coverage for migration, and would
>> >> > push back against such a request.
>> >> 
>> >> We do not have infinite CI resources, unfortunately. Migration
>> >> is competing with everything else for time on CI. You have to
>> >> find a balance between "what do we run every time" and "what
>> >> do we only run when specifically testing a migration pullreq".
>> >> Similarly, there's a lot of iotests but we don't run all of them
>> >> for every block backend for every CI job via "make check".
>> >
>> > The combos we don't run for iotests are a good source of
>> > regressions too :-(
>> >
>> >> Long test times for tests run under "make check" are also bad
>> >> for individual developers -- if I'm running "make check" to
>> >> test a target/arm change I've made I don't really want that
>> >> to then spend 15 minutes testing the migration code that
>> >> I haven't touched and that is vanishingly unlikely to be
>> >> affected by my patches.
>> >
>> > Migration-test *used* to take 15 minutes to run, but that was a
>> > very long time ago. A run of it today is around 1m20.
>> >
>> > That said, if you are building multiple system emulators, we
>> > run the same test multiple times, and with the number of
>> > targets we have, that will be painful.
>> >
>> > That could be a good reason to split the migration-test into
>> > two distinct programs. One program that runs for every target,
>> > and one that is only run once, for some arbitrary "primary"
>> > target ?
>> 
>> What do you mean by distinct programs? It's not the migration-test that
>> decides on which targets it runs, it's meson.build. We register a test()
>> for each target, same as with any other qtest. Maybe I misunderstood
>> you...
>
> If we split, we could have meson.build register "migration-smoketest"
> for every target while registering "migration-bigtest" for just 1 target.

Isn't that a bunch of shuffling code around just to have two different
invocations of migration-test?

There's the possibility of using the gtester path like
/x86_64/migration/smoke and passing '-r' when invoking via meson. That
requires way fewer changes in the C code. It moves the complexity into
meson.build, which might be worse.

Alex's idea of full set for KVM arch + a TCG arch would probably be
trickier in meson.

>
>> >  Or could we make use of glib's g_test_thorough
>> > for this - a primary target runs with "SPEED=through" and
>> > all other targets with normal settings. That would give us
>> > a way to optimize any of the qtests to reduce redundant
>> > testing where appropriate.
>> 
>> This still requires a new make target I think. Otherwise we'd run *all*
>> thorough tests for a QEMU target and not only migration-test in thorough
>> mode.
>
> Yes, that's true, having separate programs is probably an easier
> option than playing games with "SPEED" settings.
>
>
> With regards,
> Daniel