BUG: Raise recursion error during dimension discovery - #17786

Well, this is confusing. The test doesn't fail in the Python 3.6 (windows) tests. May have to dig out an older Python version :/.

This isn't working.

Yeah, I wasn't yet able to figure out why python 3.6 (or whatever it is) is behaving differently, it is pretty confusing. Should have marked as draft.

@eric-wieser just in case you have a thought on this. By now I think this whole thing may be complicated by a Python bug somewhere. On a linux 3.6, I get no failure at all even with changes that I think should ensure all but attribute errors being raised.
However, if I add prints to __new__ or __array__, I do seem to run into a RecursionError reliably (so it definitely should be recursive, but for some reason, I don't see anything about it?).

I tried adding a Py_EnterRecursiveCall() before calling obj.__array__ to no avail. Printing out a lot of things, it does recursive calls at least during dimension discovery (always finding the list of length one and entering it, then finding the array like and doing a dimension discovery again). Until suddenly all of those do return an object array including self successfully.

Possibly the recursion triggers a MemoryError as well? In principle if dimension discovery succeeds finishes successfully for whatever reason, the __new__ can conceivably succeed.

After poking this for almost a few hours now, I must have been flipping the wrong stones (unless there is a Python bug involved), though. And whether to do with this or not, the machine (which is at BIDS) that I used to try things out on Python 3.6 is not accessible anymore.

Well, now at least the test suit reproduces the stack overflow that Eric is seeing :(.

OK, so the issue is pretty definitely that we are ignoring recursion errors, no idea if the new code has the same problems (it does a similiar thing for the sequence check (but the order is different, so you likely need to try harder if it is problematic).

But... there are more PyErr_clear(), even if that part of the code probably had particularly many, so assuming tests succeed now, I am not sure its watertight.

Ok, as said on the issue. I can clean this up and remove the test if you want. I would need convincing to dig deeper into trying to actually fix it. Maybe its just another place where exceptions get caught that needs to also check for recursion error. But, it seems fine on master (the code on master is much less likely to misbehave, since it will never try to call __array__ on an object after deciding that it wasn't an array-like before – which should always happen if the RecursionError is cleared).

This is still marked as "draft", @eric-wieser are you sure you meant to approve it?

I think the change is a good idea even if we end up ripping out the test that still segfaults. When I approved I hadn't noticed the segfault though.

OK, I updated gh-17817 (the master version) for now to remove the tricky test. I would do the same update here then (with the exception of not raising all errors, but always checking only for MemoryError and RecursionError. Just to be a bit pedantic about trying to be minimal in a bug-fix release. (That is if we don't want to stick to master to begin with.)

I put a 1.19.5 milestone on this, but it still needs finishing.

OK, updated and ported the tests that are in master. Unlike master, I have chosen to still clear all errors except MemoryError and RecursionError here (some deep recursions can also cause MemoryError if I saw the cpython source correctly; so it seemed right to me to consider both fatal errors.

Master is identical now, but the only place where it still clears these errors is for the equivalent code in discover_dimensions here.

Thanks Sebastian.