Open dcherian opened 6 months ago
With size-1 chunks, e.g.|0|0|0|1|1|1|1
, I'm not sure there's any benefit. It would be a regression for when num_blocks_per_groups > split_every
Note that using blockwise
is still correct since we rechunk so that all group members are in a single block.
I'm thinking of
groupby("time.year")
, which looks like a resample, and also resampling type aggregations,In general, if
by
is sorted, thenblockwise
is potentially the best option.We could gate this on
by
I still prefer the idea of using
preferred_chunks
on xarray's new Grouper objects to rechunk it intentionally but a heuristic might be effective in the intermediate.method="blockwise"
automatically rechunks, so this would be step towards the above.