amd64 workers shouldn't abort after i386 test bed failures

[Brian Murray]

I was looking at the results of i386 tests today and noticed a pattern like the following:

Mar 15 05:43:57 juju-7f2275-prod-proposed-migration-environment-2 /home/ubuntu/autopkgtest-cloud/worker/worker[3291087]: ERROR: Three tmpfails in a row, aborting worker. Log follows:
...
                          320s Creating nova instance adt-jammy-i386-libxml-parser-perl-20240315-051351-juju-7f2275-prod-proposed-migration-environment-2 from image adt/ubuntu-jammy-amd64-se>
                          320s autopkgtest [05:43:56]: ERROR: testbed failure: Can't resolve build dependencies on testbed

Not being able to resolve build dependencies for i386 is a common occurrence and my concern here is how much time are we losing by aborting the worker. Is it down for 5, 10, 15 minutes until it gets restarted again? Any amount of downtime for this is wasteful.

Additionally, retrying two more times is also wasteful because they aren't likely to get sorted.

Finally, the ideal solution would probably involve not running tests where the build-deps are gonna fail at all but that seems a lot harder.

I think we should look at the FAIL_STRINGS stuff here and try to adjust it for i386.

[Paride Legovini (paride)]

Just a note to avoid confusion: the log speaks about build-deps because, in the case of cross-arch testing, the build-dependency resolver is used to install the test dependencies, instead of the "normal" dependency resolver.

[Tim Andersson (andersson123)]

I'm going to look into this bug today.

[Tim Andersson (andersson123)]

MP created and in review.

[Tim Andersson (andersson123)]

Decided not to fix for now. Leaving the work as is until we come back to this.