Florian Jetter

> This is an important question. > > > Do I understand that delayed, bag and submit/map are to disappear, then? Or will the user APIs still work, but without...

> Quick searching suggests hlg/layer direct usage in the following: `dask-sql` is unmaintained and I did, maybe incorrectly, count `dask-histogram` towards `dask-awkward`

No objections. It appears to not be working, though. The test is failing (linter is also off)

Yes, I am aware of this problem. In the past I once wrote some logic that would check (e.g. lookup in a database) whether a certain task already ran before...

Thanks for writing this up > instead of workers using threads by default, I think they should use subprocesses, and zero-copy shared memory between the processes I'm interested in investigating...

This is a bit more complex topic than the community issue makes it look like. Right now, the behavior is one of three different kinds depending on what API and...

TLDR with https://github.com/dask/dask/pull/11736 I am proposing to hard code this to "always concat on cluster"

Your code proves that the individual objects are serialized sequentially and that they are all running on the same process. It doesn't tell us anything about when or how the...

For future reference: There are also several cases where optimization is fundamentally altered when adding a concat step at the very end. Simple example ```python >>> from dask.datasets import timeseries...

See https://github.com/dask/dask/issues/11768#issuecomment-2681904615 and https://docs.dask.org/en/stable/changelog.html#breaking-changes