El 2/5/25 a las 20:20, Étienne Mollier escribió:
Santiago, about the errors you witnessed with pyranges on random occurrences, do you have build logs somewhere? The package has a history of heavy ressource usage at test time, and I suspect bad parallelism conditions would have caused e.g. out of memory conditions. I've had 0% failure rate on my end after many tries, but also lowish cores count combined to stupidly large amount of swap along my RAM. There is also the possibility of having run into race conditions, but the rate of failures you hit does not match rare conditions (unless you threw several dozens of cores at the problem). The package has been a problem with ressource consumption to other contributors in the past, so I think I will drop parallelism anyway.
I've now filed a proper bug for the failures I get on amd64. My guess is that parallelism is not directly related with the test failures. My feeling is that those tests run some kind of Monte-Carlo simulation, and the data on which they act is different each time, because the AWS machines I use for archive rebuilds are always in a limited set of machine types (i.e. the test suite does not seem to be deterministic because otherwise I would never get any failure, or I would always get a failure). Thanks.
