I am seeing that this conflicts with #20389 so maybe we should get to a conclusion. For me, it is better for lib._dataloader.open to raise the corresponding exceptions instead of handling them repeatedly in genfromtxt and loadtxt.

@WarrenWeckesser @andrekwr @mattip

#20389 was merged. If you still want to pursue this as a simplification, please rebase/rework on current master.

#20389 was merged. If you still want to pursue this as a simplification, please rebase/rework on current master.

Done

@mattip ping

@seberg will this cause issues for #20580? It does seem like the change here is for the better.

... except for the failing tests :)

No, should be fine for the loadtxt to C move. I am not touching that datasource stuff after all. It might be that we need to tweak a test checking for the exact error being returned for bad encoding, but that's it.

Not sure why this had stalled, probably due to the whole C-loadtxt changes.

This still seems right, and merge-conflict aside probably good to go.

The one thing I would like to check is whether we currently accept byte-strings, in which case it may be good to deprecate that support (or simply keep it).