There has been a push in dask core to bring “expr” to dask.arrays. Part of this effort includes the following behaviour:
when doing array.compute(), a task will be added to concatenate the parts of the array on a worker before copying to the client.
I don’t know if this is the default, but I feel people should be aware and try out the option on their workflows, to provide feedback to dask maintainers.
Workflows where the client has significantly more memory than workers, and final materialised arrays of a size approaching the worker memory will crash workers and render workflows broken. If this is your situation, please speak up!
PR: Wrap HLGs in an Expr to avoid Client side materialization by fjetter · Pull Request #11736 · dask/dask · GitHub , probably released in dask 2025.4.0