| 12 Oct 2023 |
 cole-h | I've already tried that (manually), unfortunately. A few years ago, 3-4 ofborg evaluators was enough to chew through the queue. Nowadays, even 9 is not enough, due to eval times blowing up. | 15:59:45 |
| cole-h | Also, I don't know how I feel about marking "errors as errors" (I assume this means "failed builds turn into failed checks"). There could be any number of reasons as why the build failed that may not have anything to do with the derivation itself.
Maybe the machine OOM'd. Maybe networking died. Maybe the kernel panicked. Maybe there was a hardware failure. Maybe....
Something that was decided early on was that things with a red X should not be merged under any circumstance (as always, there are exceptions, but those should be very rare).
If one of those transient (or not so transient) failures happens, but nobody can reproduce it and someone decides to merge it anyways, that cheapens the meaning of a failed CI check.
At least with a "skipped" check, its communicated that something may have gone wrong, but it may not be anyone in particular's fault. | 16:03:02 |
| cole-h | * Also, I don't know how I feel about marking "errors as errors" (I assume this means "failed builds turn into failed checks"). There could be any number of reasons as why the build failed that may not have anything to do with the derivation itself.
Maybe the machine OOM'd. Maybe networking died. Maybe the kernel panicked. Maybe there was a hardware failure. Maybe....
Something that was decided early on was that things with a red X should not be merged under any circumstance (as always, there are exceptions, but those should be very rare).
If one of those transient (or not so transient) failures happens, but nobody can reproduce it and someone decides to merge it anyways, that cheapens the meaning of a failed CI check.
At least with a "skipped" check, it's communicated that something may have gone wrong, but it may not be anyone in particular's fault. | 16:03:05 |
| cole-h | (Not to say I'd block that change, per se, but it'd be nice to be convinced that it's the right thing to do.) | 16:03:49 |
 @delroth:delroth.net | allowing people to retry failed runs and figuring out how to address infra flakiness seem like they'd both help there - fwiw I've rarely seen ofborg failing for the reasons you're listing, and they seem to be all transient conditions | 16:04:44 |
| @delroth:delroth.net | (could even do something like "retries get scheduled on a different runner" if we wanted to be fancy :p) | 16:05:06 |
| @delroth:delroth.net | I agree that we should at the very least try to measure how often these problems happen before making any decision, but I don't think a low rate of false positives necessarily needs to be a blocker - it would still be a massive improvement | 16:06:01 |
| @delroth:delroth.net | (imo) | 16:06:03 |
 Lily Foster | Yeah I definitely want to get measurements for how often those transient failures happen, before changing anything. Stuff like build timeout, OOM, and nix daemon/builder error should be technically distinguishable from derivation-builder-error, though, so ideally we'd have the ability to do both "neutral" and "failure" conditions depending on how it failed
I think there would be value in having at least some scoped conditions surrounding builds get a red X though, because I see a lot of PRs get merged with failing builds or tests because failures of, e.g. build timeouts due to rebuilding llvm or something, are not clearly distinguishable from failures due to the build just not actually working, without diving into the logs on the ofborg website
I'd have to see where the decision was originally made, but I feel like, too, if there's enough of the community really wanting the red X now, we could decide to try it temporarily and roll it back if it does in fact turn out to be worse than status quo (even though that is admittedly hard to measure) | 17:53:51 |
 @adam:robins.wtf | I've definitely had PRs merged that were failing a build, and nobody noticed. It seems like some of the concerns (timeouts, OOM, hardware failure) are common across other CI systems, yet it is also quite common for them to mark builds as failing through the UI. | 18:14:03 |
| @adam:robins.wtf | Is the difference just a lack of ability to retry? | 18:15:55 |
