Bug #65658
Updated by Patrick Donnelly 23 days ago
There was a lot of discussion surrounding this in
https://github.com/ceph/ceph/pull/26004/
but circling back we have since seen evidence this is causing significant problems: after a long up:replay recovery, the MDS can be flooded with metrics messages by clients and the lock contention in fast_dispatch is preventing the MDS from sending beacons to the monitors. This then leads to undesirable MDS failovers.
We should convert this to using regular dispatch (and optimize later if needed).