The last month has seen a lot of work on the storage cluster, fixing recovery related bugs, improving threading, and working out a mechanism for online scrubbing. In this case, scrubbing is basically a low-level fsck of the object storage layer. For each PG being scrubbed, the primary and any replica nodes generate a catalog of all objects in the PG and compare them to ensure that no objects are missing or mismatched (currently we check size and attributes; soon, we’ll pull the checksums out of btrfs to ensure the object contents match too). Assuming the replicas all match up, one OSD does a final semantic sweep to ensure that all of the snapshot-related object metadata is consistent. Errors are reported to a (new) central system log.
An administrator can tell the system to scrub the entire storage cluster, a single OSD, or a single placement group. Eventually, we’ll probably want to have the system automatically schedule a slow background scrub when the system is idle.
This is only one piece of the overall ‘fsck’ problem–the file system metadata is more complicated and also needs to be verified.