`matminer.datasets.utils._validate_dataset()` flaky on Windows?

[janosh]

load_dataset has been inconsistent in pymatviz CI occasionally raising

UserWarning: Error, hash of downloaded file does not match that included in metadata, the data may be corrupt or altered

I think it's only happening on Windows which I don't have access to locally. Was wondering if @sgbaird who uses Windows has also experienced this? Here's minimal code that (sometimes) causes the error:

from matminer.datasets import load_dataset

load_dataset("matbench_glass")

[sgbaird]

I've been getting the same error on some of my CI, although I've been seeing it on Linux CI, too. I think it's also "sometimes" for me too.

materialsproject/matbench#155

[janosh]

Good to know it's not Windows specific. I've never seen the error locally but then I guess that's due to Ci re-downloading the files on every run whereas locally they're cached.

@ardunn Any clues where the intermittency comes from? Maybe Figshare load balancers that point at servers with slightly different files?

[ardunn]

This is weird, though I'm not sure its a windows issue. I think it is something having to do with downloading from the same IP from figshare and the file gets corrupted agnostic of OS. At least, this is what it seems like to be because CI for matbench will randomly fail with that error and then pass on re-runs.

[sgbaird]

Since it seems to be an issue with FigShare, one option might be to implement a "retry" option, i.e. if the first attempt to download fails, wait 10 s and then try downloading again. Not sure if it will fix it, and obviously not ideal, but if it's something weird on the FigShare side of things might be better to notify FigShare and then implement some kind of workaround in the meantime.

[ardunn]

That’s an excellent idea, I will try that! Thanks, Alex

…

On Aug 12, 2022 at 5:07 PM -0700, Sterling Baird ***@***.***>, wrote: Since it seems to be an issue with FigShare, one option might be to implement a "retry" option, i.e. if the first attempt to download fails, wait 10 s and then try downloading again. Not sure if it will fix it, and obviously not ideal, but if it's something weird on the FigShare side of things might be better to notify FigShare and then implement some kind of workaround in the meantime. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>