ENH: note on the error when out may be partially written - #31965

cc @ngoldbaum @seberg @ikrommyd

Could you do a scan if there are any other functions where partial modification can occur on the out array?

Could you do a scan if there are any other functions where partial modification can occur on the out array?

caller's array can be left partially modified on error in these

• sort / partition (in-place, raising object comparison)
• a[mask] = vals and a[idx] = vals / put_along_axis (failing value cast mid-write)
• np.copyto / dst[...] = src / concatenate(out=) when the casting rule allows an element-wise-failing cast
• np.round(out=, decimals != 0) (leaves intermediate a * 10**decimals values) and np.mean(out=) (leaves undivided sum)
• ufuncs with a matching-dtype out (elementwise, reduce, accumulate, reduceat, object matmul/dot, einsum) the loop writes straight into out
• argmax / argmin with out on object dtype

fyi: take/choose weren't fully shielded by their internal copy either so i added to the pr

@seberg do you think one should add the note to all cases where an exception being raised can cause partial modification to out array here in this PR?

This has ballooned into way to much reviewer work for something that is really not a problem. Please don't continue adding on more followup PRs for this.

This has ballooned into way to much reviewer work for something that is really not a problem. Please don't continue adding on more followup PRs for this.

the scan was a response to @ikrommyd 's question. The other issues are independent of this specific fix :)

I will avoid adding more follow ups from now on. thanks for flagging this!

I was probably not clear enough. I only wanted to identify if there are any other places to put the add_note helper where the user is already visibly experiencing a partial modification in the out array and we can just improve the error they are already getting.

do you think one should add the note to all cases where an exception being raised can cause partial modification to out array here in this PR?

To be clear, just because I think that this is an interesting idea, doesn't mean that I think it is important or even worthwhile right now. In fact, the whole thought of "there are a lot of places" might mean that we should just not do it, because it is general NumPy behavior.

Yeah this makes sense. This was actually why I asked for the locations where this happens. I did not expect an error being raised causing a partial state modification to be super common cause if it is, one can argue is that this is common numpy behavior and it probably doesn’t deserve special treatment.

I'll just close this for now. I don't care about re-opening it, but the motivatio nneeds to be better then us musing about it maybe being nice.