dask troubles #598
Description
Here are things that have changed since Dec 2024, which are problematic for us.
- tuple tasks became a class hierarchy. This uses more memory and time, and can confuse the garbage collector (so I hear)
- dask-postcompute now becomes an operation in the graph and gets carried out on a worker. This complicates graph optimisation (this last layer behaves differently), and leads to OOM for workflows designed with big clients in mind.
- because only expr is sent on the wire now, HLGs get materialised to LLGs on the client (why isn't one layer-> one node??)
- get() fundamentally changed, so that we can't call get_sync() any more. Did this break dask-on-ray, taskvine and others?