Closed jrbourbeau closed 1 year ago
https://github.com/dask/distributed/pull/6614 introduced a significant performance regressions for at least one common type of dataframe operations (repartition + align + reduce), see also https://github.com/coiled/coiled-runtime/issues/295
The regression increases average memory usage in our benchmarks by more than a factor of two, i.e. if a workload is affected this can cause critical failure and this should be treated as a release blocker.
https://github.com/dask/distributed/pull/6985 is the potential fix. I think getting this in should probably block the release so we can do more performance testing and tuning before releasing this behavior change.
I suggest to include https://github.com/dask/distributed/pull/6991 if nothing pops up there (tornado 6.2 support)
Mea culpa. I got too excited and merged a PR we preferred to only merge after the release. https://github.com/dask/distributed/pull/6994 for the revert
Alright, thanks all for the input here. All reported blockers have been resolves, so I'll start pushing out the release in a few minutes
2022.9.0 is on PyPI and conda-forge PRs are open:
Packages are on conda-forge and Docker images have been published. Thanks all!
Best effort
Try to close before the release but will not block the release
Blocker
Issues that would cause us to block and postpone the release if not fixed
Comments
cc @ian-r-rose @fjetter @quasiben @jakirkham