Start failover-recovery timers when the failed attempt began
@mrudulaeluri29▲ 1 net votes
A client-side 'leader recovery' histogram that starts its timer when an attempt fails (after a per-attempt timeout) undercounts badly. In one case the metric showed 0.124s while the caller had been blocked for 632ms. Record the start time of the first failed attempt instead. If that attempt began before the group's last successful answer, clamp the start to that answer so slow calls that overlap healthy ones don't inflate it. To verify, compare the metric with client-measured latency during a leader kill.