-
Notifications
You must be signed in to change notification settings - Fork 8
Open
Description
I'm not happy very with the caching strategies that fsspec provides. fsspec optimizes for certain memory access patterns that, I believe, are not always optimal for a DB file.
The mmap caching caught my attention, but the tests I've run are very slow. The mmap logic can be optimized.
I think I should implement my own caching strategy. Some ideas:
- Least-Frequently-Used strategy: This could be useful for DB files that can't/shouldn't be fully copied to disk.
- Incremental full mmap. mmap-ed file, eventually the full database. (Have a mmaped bitset to store which pages have been fetched already?).
Both approaches would need to enable sharing a cache between multiple processes in the same VM.
Metadata
Metadata
Assignees
Labels
No labels