mgr/Mgr.cc: clear daemon health metrics instead of removing down/out osd from daemon...

Reverts the change from https://github.com/ceph/ceph/pull/53993
and directly clears daemon health metrics for down and out OSDs.
The former approach of removing down/out OSDs from the daemon
state has undesirable consequences for stat output, including
the prometheus exporter.

Fixes: https://tracker.ceph.com/issues/66168
Signed-off-by: Cory Snyder <csnyder@1111systems.com>
(cherry picked from commit 282558cf40274366360bb3b1ec0fa102fbb592a6)

diff --git a/src/mgr/Mgr.cc b/src/mgr/Mgr.cc
index cb988cf761cc20e7f5b4f9b00fa546f8f1d25fa1..48dcc13ca3d6c3b0258c14108f57727a20272b5f 100644 (file)
--- a/src/mgr/Mgr.cc
+++ b/src/mgr/Mgr.cc
@@ -537,9 +537,16 @@ void Mgr::handle_osd_map()
       if (daemon_state.is_updating(k)) {
         continue;
       }
+        
+      DaemonStatePtr daemon = daemon_state.get(k);
+        
+      if (daemon && osd_map.is_out(osd_id) && osd_map.is_down(osd_id)) {
+        std::lock_guard l(daemon->lock);
+        daemon->daemon_health_metrics.clear();
+      }
 
       bool update_meta = false;
-      if (daemon_state.exists(k)) {
+      if (daemon) {
         if (osd_map.get_up_from(osd_id) == osd_map.get_epoch()) {
           dout(4) << "Mgr::handle_osd_map: osd." << osd_id
                  << " joined cluster at " << "e" << osd_map.get_epoch()