Since stats moved from shared memory to being copied over IPC (#5910), our metrics are seeing a transient drop during hot restarts. E.g.: for cluster.${cluster}.health_check.failure:

which isn't great because it triggers alerts, etc. I haven't looked into it too much yet, but one suspicion is that it's due to the overhead of copying things over....

Thoughts?

cc: @fredlas @mattklein123 @fishcakez