The S3 store extension caches the full objects in a local cache directory, both when flushing them to S3, or when downloading them from S3. This cache was designed to follow what DIR store does. We can switch to range reads without caching now. Caching will eventually be done at a different layer and we have a project for that.

The azure and GCP extensions already do range reads of the objects instead of doing any caching. S3 store should do the same.