Ceph » CephFS

The MDS currently does not support rolling upgrades. Normally we recommend upgrading all MDS at the same time for this reason. We also have an unfortunately complex upgrade procedure [1] to deal with the CompatSet changing in the FSMap. Old MDSs will kill themselves if they see a CompatSet they are incompatible with. Another issue is that any bump or difference to the CEPH_MDS_PROTOCOL will cause MDSs to not be able to talk to each other.

Now that MDS messages are now versioned (#12107), this puts us closer to supporting rolling upgrades. The next steps are to:

- Guarantee (via code) that there are no unversioned inter-MDS messages: #41565
- Use some kind of feature bit-vector (maybe feature_bitset_t) to gate the use of new features by a newer MDS until all MDS have the same bit-vector. This is similar to what we do with OSDs and require_osd_release. This should prevent the version upgrade of on-disk metadata and newer versions of inter-MDS messages.
- Testing for rolling upgrades in the fs suite.

[1] https://docs.ceph.com/docs/master/cephfs/upgrading/