Re: [RFC PATCH 4/4] ci: Add the new migration device tests

[Peter Xu <peterx@redhat.com>]

On Mon, May 27, 2024 at 08:59:00PM -0300, Fabiano Rosas wrote:
> Peter Xu <peterx@redhat.com> writes:
> 
> > On Thu, May 23, 2024 at 05:19:22PM -0300, Fabiano Rosas wrote:
> >> We have two new migration tests that check cross version
> >> compatibility. One uses the vmstate-static-checker.py script to
> >> compare the vmstate structures from two different QEMU versions. The
> >> other runs a simple migration with a few devices present in the VM, to
> >> catch obvious breakages.
> >> 
> >> Add both tests to the migration-compat-common job.
> >> 
> >> Signed-off-by: Fabiano Rosas <farosas@suse.de>
> >> ---
> >>  .gitlab-ci.d/buildtest.yml | 43 +++++++++++++++++++++++++++++++-------
> >>  1 file changed, 36 insertions(+), 7 deletions(-)
> >> 
> >> diff --git a/.gitlab-ci.d/buildtest.yml b/.gitlab-ci.d/buildtest.yml
> >> index 91c57efded..bc7ac35983 100644
> >> --- a/.gitlab-ci.d/buildtest.yml
> >> +++ b/.gitlab-ci.d/buildtest.yml
> >> @@ -202,18 +202,47 @@ build-previous-qemu:
> >>    needs:
> >>      - job: build-previous-qemu
> >>      - job: build-system-opensuse
> >> -  # The old QEMU could have bugs unrelated to migration that are
> >> -  # already fixed in the current development branch, so this test
> >> -  # might fail.
> >> +  # This test is allowed to fail because:
> >> +  #
> >> +  # - The old QEMU could have bugs unrelated to migration that are
> >> +  #   already fixed in the current development branch.
> >
> > Did you ever hit a real failure with this?  I'm wondering whether we can
> > remove this allow_failure thing.
> >
> 
> I haven't. But when it fails we'll go through an entire release cycle
> with this thing showing red for every person that runs the CI. Remember,
> this is a CI failure to which there's no fix aside from waiting for the
> release to happen. Even if we're quick to react and disable the job, I
> feel it might create some confusion already.

My imagination was if needed we'll get complains and we add that until
then for that broken release only, and remove in the next release again.

> 
> >> +  #
> >> +  # - The vmstate-static-checker script trips on renames and other
> >> +  #   backward-compatible changes to the vmstate structs.
> >
> > I think I keep my preference per last time we talked on this. :)
> 
> Sorry, I'm not trying to force this in any way, I just wrote these to
> use in the pull-request and thought I'd put it out there. At the very
> least we can have your concerns documented. =)

Yep that's fine.  I think we should keep such discussion on the list,
especially we have different opinions, while none of us got convinced yet
so far. :)

> 
> > I still think it's too early to involve a test that can report false
> > negative.
> 
> (1)
> Well, we haven't seen any false negatives, we've seen fields being
> renamed. If that happens, then we'll ask the person to update the
> script. Is that not acceptable to you? Or are you thinking about other
> sorts of issues?

Then question is how to update the script. So far it's treated as failure
on rename, even if it's benign. Right now we have this:

            print("Section \"" + sec + "\",", end=' ')
            print("Description \"" + desc + "\":", end=' ')
            print("expected field \"" + s_item["field"] + "\",", end=' ')
            print("got \"" + d_item["field"] + "\"; skipping rest")
            bump_taint()
            break

Do you want to introduce a list of renamed vmsd fields in this script and
maintain that?  IMHO it's an overkill and unnecessary burden to other
developers.

> 
> > I'd still keep running this before soft-freeze like I used to
> > do, throw issues to others and urge them to fix before release.
> 
> Having hidden procedures that maintainers run before a release is bad
> IMHO, it just delays the catching of bugs and frustrates
> contributors. Imagine working on a series, everything goes well with
> reviews, CI passes, patch gets queued and merged and a month later you
> get a ping about something you should have done to avoid breaking
> migration. Right during freeze.

I understand your point, however I don't yet see a way CI could cover
everything.  CI won't cover performance test, and I still ran multifd tests
at least smoke it too before soft-freeze.

If there's something done wrong, we notify the sooner the better.  Now it
looks to me the best trade-off is like that - we notify at soft-freeze once
per release considering that's pretty rare too, e.g. 9.0 has no such outlier.

Again I'm happy if we have a solution to not report false negatives; that's
the only concern I have.

> 
> > Per my
> > previous experience that doesn't consume me a lot of time, and it's not
> > common to see issues either.
> >
> > So I want people to really pay attention when someone sees a migration CI
> > test failed, rather than we help people form the habit in "oh migration CI
> > failed again?  I think that's fine, it allows failing anyway".
> 
> That's a good point. I don't think it applies here though. See my point
> in (1).

Yes, if you have an answer to (1) I'm ok too.  Maybe I misunderstood your
plan there.

Thanks,

> 
> > So far I still don't see as much benefit to adding this if we need to pay
> > for the other false negative issue.  I'll fully support it if e.g. we can
> > fix the tool to avoid reporting false negatives, but that may take effort
> > that I didn't check.
> >
> 

-- 
Peter Xu