FYI dask has async compute but it seems hard to work in here :)

https://distributed.dask.org/en/stable/asynchronous.html

If this parallels test_load then lets keep this on the base class and use pytest.skip() on the netCDF subclasses. That way it's easy to keep the two in sync. Let's add a comment requesting future contributors to keep the two in sync

Yeah I did think about doing this but skipping an inherited method seemed messy.

I made it work in cf1d127, though. Note that the MRO becomes important to get the correctly overridden test to be inherited (so it can be skipped). I think the cleaner solution here would be if we could simply ask the backends whether or not they support async indexing, which is an idea we also discussed for #10579 (comment).

This is fine, but we could combine the sync and async checks in one async function and just use that everywhere in this file.

Done in 4f40792, though now I'm a little worried that this pattern

    async def check_orthogonal_indexing(self, v):
        expected = self.d[[8, 3]][:, [2, 1]]

        result = v.isel(x=[8, 3], y=[2, 1])
        assert np.allclose(result, expected)

        result = await v.isel(x=[8, 3], y=[2, 1]).load_async()
        assert np.allclose(result, expected)

might be automatically passing the second assertion by still being in-memory after the first assert?

You could parametrize the test over async and non async calls

@pytest.mark.parametrize("use_async", [True, False])
async def check_orthogonal_indexing(self, v, use_async):
        expected = self.d[[8, 3]][:, [2, 1]]

        if use_async:
            result = await v.isel(x=[8, 3], y=[2, 1]).load_async()
        else:
            result = v.isel(x=[8, 3], y=[2, 1])
        assert np.allclose(result, expected)

Could use assert not v._in_memory to be really sure.

I went with the parametrization idea, though the syntax has to be a little messier because you can't parametrize normal functions in pytest. a074a25

👏🏾 👏🏾 👏🏾