Tom White

> changing this leads to some kind of cubed error It's quite hard to debug without the code, but it's worth trying to turn off graph optimization (fuse) by passing...

Thanks for the notebook @TomNicholas. I tried running it, and managed to reproduce the problem. I think what's happening is that Dask doesn't care about the dtype or shape of...

I just reached the same conclusion after looking at it for a while!

No problem! It was quite instructive looking through the code. A couple of things I noticed: * One reason that `blockwise` is used rather than `reduction` to apply the initial...

> I think `blockwise` and `map-reduce` can be made to work out of the box, if we adjust the intermediates to be structured arrays. Great! Does Flox need to use...

> Aside from algorithmic changes, isn't this just `reduction` with `chunk=lambda x: x`? Yes

> @tomwhite mentioned that the sgkit file openers / converters are actually about to be deprecated in favour of a new package called `bio2zarr`. Basically their motivation is that the...

No, haven't seen this problem before. > The Haivvreo should be the same with CDH (assuming they didn't make any changes they didn't contribute back, which I don't believe they...

Thanks @honno for providing the wrapper and the test suite. There are likely to be some compatibility issues with using `dask.array` as the Array API namespace. To pick one example,...

Thanks for all the comments. I've filed Dask issues for some of these errors: https://github.com/dask/dask/issues/8667, https://github.com/dask/dask/issues/8668, https://github.com/dask/dask/issues/8669, https://github.com/dask/dask/issues/8670. > With `eye`, probably the Array API needs to skip passing `dtype`...