[POC] Distributed checkpointing API for dask collections #8483
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR includes a rough POC for a distributed checkpointing API.
Background
I have come across several RAPIDS + Dask projects that are forced to perform a global "checkpoint" in the middle of their workflow by writing a Parquet dataset to shared storage. Due to extreme memory pressure, it is not practical for these workflows to execute their end-to-end data pipeline as a single task graph. P2p shuffling can sometimes help a subset of these workflows. However, explicit IO tends to be much faster and more stable when fast/shared storage is available. The problem with this naive checkpointing approach is that "shared" storage is not always fast and/or available. It is much more common to have fast node-local storage.
We are currently exploring the possibility of adding an API for distributed checkpointing in RAPIDS. By "distributed", I mean that workers must be able to write to and read from node-local storage. The idea is similar to persisting a collection and forcing workers to spill all partitions. However, this API would also allow the user to completely restart the cluster after checkpointing.
I'm not sure if this kind of API is generally desired. However, I wanted to share the idea early in case it is.
Example Usage