cancel other arch builds if one fails

[ramcq]

At the moment if one arch fails, we continue to build the other 3 even though we're then going to fail the job and throw them away. We could cancel the others and save some worker CPU cycles.

[barthalion]

The benefit of current behavior is that built modules are cached anyway, so even if the architecture that failed to build is being disabled by the next commit, it's just a cache hit for others.

[ramcq]

I think disabling a broken architecture isn't the "normal" fix - a build failure is more likely to cross all architectures and need some patch to the JSON which could invalidate a lot of the caches anyway.

We've had >1 worker per arch on the new infra this week, so this assumption that a build always helps your cache isn't true unless we set up some kind of affinity to workers that tried the same exact JSON before. But I don't think this is the most common case we should be optimising for, and you can get starvation (blocking on a long-running build for ages) if you "mess" with the scheduler like this.

The problem is the Scaleway builders are just pretty slow and I'm worried about falling behind even with a relatively modest change request traffic.

It certainly makes sense to cancel scheduled builds (that aren't even started) however.