But other signs is what we wanted:
:compaction_daemon<0.620.0>:compaction_daemon:handle_info:309]Finished compaction iteration.
:compaction_daemon<0.620.0>:compaction_daemon:schedule_next_compaction:1199]Finished compaction too soon. Next run will be in 30s
:system_stats_collector<0.611.0>:system_stats_collector:handle_info:133]lost 2 ticks
:mb_master<0.555.0>:mb_master:handle_info:204]Skipped 12 heartbeats
:mb_master<0.555.0>:mb_master:handle_info:219]Haven't heard from a higher priority node or a master, so I'm taking over.
So looks like CPU was taken away from mb_master for about 24 seconds. Or there's some issue with erlang timers. It's still a bit mystery why we're not seeing this issue elsewhere. But leading guess is master is both latency sensitive _and_ is constantly exercised. Perhaps a bit more often than say ns_config or nodes_disco