It looks like this change is causing bunch of other tests to fail. I think some of them are not to hard to fix, because they are marked as xfail due to the problem this PR is solving. But I'm a bit confused about certain cases where it seems we are ok, having nan chunks ? For example

the test dask/array/tests/test_slicing.py::test_getitem_avoids_large_chunks_missing it seems that works when the first element in chunks is doesn't include nans.

Is this the expected behavior? because if that's the case then the fix we added here breaks this. As any appearance of nan in self.chunks will raise a ValueError now.

We have multiple situations where we were able to check for

X = np.arange(24).reshape((4, 6))
x = da.from_array(X, chunks=(2, 2))
x._chunks = ((2, 2), (np.nan, np.nan, np.nan))

assert_eq(x[0], X[0]) #where here x has unknown chunks

but now this check will fail with the fix on this PR. So test like the one below fail, because the ValueError raises one step before the pytest.raises .

test_index_array_with_array_1d

I think we might need to discuss how to fix all these tests, and what we expect to work or not. Basically with this change any time that there is an unknown chunk, it'll raise as soon as we do any operation that involves __getitem__ it seems this might be very restrictive, as I see that in certain situations we can work with unkown chunks. For example, the case of test_index_array_with_array_2d

After chatting with Julia for a bit, it seems that finding the right place where to raise the error is not quite possible, or at least we weren't able to sort this out. We have multiple situations where dask can do its magic even though there are nan values in the chunks but not always. It seems the best situation, in this case, is to catch the exception in the compute, and give a more informative error there.

Because I don't have push access to this PR, I continued the work in #9285

I continued the work in #9285

Thank you, @ncclementi! I'll close this one then. :)