Feature #24917
open
Gracefully deal with upgrades when bluestore skipping of data_digest becomes active
Added by David Zafman almost 6 years ago.
Updated over 5 years ago.
Description
Once the data_digest is no longer being used, but is still set from an earlier version, we can get EIO from read but deep-scrub doesn't report errors.
- Related to Bug #23871: luminous->mimic: missing primary copy of xxx, wil try copies on 3, then full-object read crc mismatch added
We need to wait to turn off data_digest once all OSDs are running bluestore AND we must disallow a filestore OSD to join the cluster once the switch is flipped. The flag must be a monitor maintained bit.
Once this code is in place the osd_skip_data_digest could be phased out in a future release.
- Assignee set to David Zafman
- Source set to Development
- Assignee changed from David Zafman to Josh Durgin
Let's include this with any other feature bit addition.
- Tracker changed from Bug to Feature
Josh, this code needs to be written. It needs a feature bit AND a mon flag that can only be set when all OSDs are using bluestore. When the feature bit is set on all OSDs AND bluestore is running on all OSDs with the mon flag set, then all OSDs can skip data_digest processing. The config value osd_skip_data_digest should also be phased out.
Also available in: Atom
PDF