search

cubed-dev / cubed

Bounded-memory serverless distributed N-dimensional array processing
https://cubed-dev.github.io/cubed/
Apache License 2.0
122 stars 14 forks source link

Write to Icechunk as intermediate store? #601

Open TomNicholas opened 3 weeks ago

TomNicholas commented 3 weeks ago

Maybe useful? Could simply roll back to state before failed stage. Also then it's Icechunk 's problem to worry about atomic writes... Idea from:

A reason to not run backup tasks is if the filesystem does not support atomic writes. Cloud object stores generally are atomic (see https://cubed-dev.github.io/cubed/user-guide/reliability.html#stragglers), but local filesystems are not.

Discussing with @applio, we thought this PR should be changed so that the default was based on the store rather than the executor. Backup tasks would be off by default, except if the store was a well-known cloud store like S3 or GCS.

Originally posted by @tomwhite in https://github.com/cubed-dev/cubed/pull/600#issuecomment-2442268954

tomwhite commented 3 weeks ago

The interesting case is HPC, where there are multiple nodes (and hence the possibility of stragglers), but an intermediate store that uses a shared filesystem that does not support atomic writes. Icechunk might be useful to provide atomicity - but perhaps there are other ways too?

Also, I'd like to move to general blob stores (not just Zarr stores) for the intermediate store, so we have more control over the chunk access pattern (to control memory usage). Work like #464 will enable this.