hesela.devCeph deep scrub
Ceph deep scrub reads placement-group object data and checks it using checksums, beyond the metadata comparisons of a light scrub.
In Ceph Squid, a light scrub compares object sizes and attributes; deep scrub additionally reads data for checksum verification. The configured interval is a scheduling input, not evidence that a scan completed. [1]
The active state permits requests and clean describes the required replication count. Neither establishes when a deep scrub last completed. The deep state describes checking in progress, not a successful result. [2]
For replicated pools, pg repair can mark an inconsistent copy missing so recovery can reconstruct it; repair and recovery are distinct steps. For the documented erasure-coded and BlueStore cases, automatic scrub repair is disabled by default in Squid and, when enabled, is limited by a configured damaged-object threshold. [3] [1]
A read_error in a shard report calls for investigation of the underlying storage. Preserve the affected PG and object, shard errors, scan completion and subsequent recovery outcome before concluding that an inconsistency was resolved. [3]
Historically, the 2007 RADOS design kept a log of expected object versions even when replicas were missing. Peering reconciled that history before recovery; this is architectural context, not a description of Squid's scrub implementation. [4]
Relations
Limitations
- Scrubbing competes with client work, so increasing its frequency can affect cluster performance. [1]
Citations
- Ceph Squid: OSD Config Reference, light/deep scrubbing and automatic repair controls
- Ceph Squid: Placement Group States, active, clean, scrubbing, deep and repair
- Ceph Squid: Troubleshooting PGs, inconsistent shards and More Information on PG Repair
- Weil et al. (PDSW 2007): RADOS, sections 3.4 and 3.4.1
Further reading
Sources & caveats
- Version-scoped documentation review of Ceph Squid, not a running-cluster experiment or a measured reduction in data loss.
- The distinction between active+clean and completed checksum verification follows from the separately documented PG states. [2]