New file system snapshots should be visible on the clients faster than currently. We aim for a ~5 mn delay from the point when the publishing finishes. Several points along the distribution chain should be tackled
- Speedup of the publishing process itself. This is largely completed by René's recent work.
- Notification service to start replication as part of the publish process (
- Caching and reuse of "object packs" developed as part of the distributed backend. This will drastically reduce the replication of small files.
- Stream change sets concurrently to authoritative storage and stratum 1s.
- Reduction (including validation) of the .cvmfspublished TTL. It can probably be reduced to values as low as 30 seconds.
- An interface to change the catalog default TTL. Possibly adjusting the default TTL from 15 minutes to 5 minutes.
- Investigate if the fuse-maintained kernel cache on clients can be actively purged to avoid the 1 minute delay to drain the caches.