YugabyteDB tablet server drops and skew show up in server metrics first

Watch numtabletserverslive against your node count; a drop means a tablet server failed heartbeats or is partitioned, often from clock skew. If hybridclockskew is approaching hybridclock_error, fix clock sync on that host before reads start restarting. YSQL counters only move under load, so exercise the cluster before concluding metrics are broken.

Context: Instrumentation runbook: documents the gotcha that an idle YugabyteDB cluster shows no datapoints for YSQL counters, and that a dropped tablet server is the first suspect when numtabletserverslive falls below the node count. When hybridclockskew approaches hybridclock_error, the node is at risk of read restarts and stalls, so check NTP or chrony on that host.

Source: Instrumentation runbook: documents the gotcha that an idle YugabyteDB cluster shows no datapoints for YSQL counters, and that a dropped tablet server is the first suspect when numtabletserverslive falls below the node count. When hybridclockskew approaches hybridclock_error, the node is at risk of read restarts and stalls, so check NTP or chrony on that host.

Verified resolution: Watch numtabletserverslive against your node count; a drop means a tablet server failed heartbeats or is partitioned, often from clock skew. If hybridclockskew is approaching hybridclock_error, fix clock sync on that host before reads start restarting. YSQL counters only move under load, so exercise the cluster before concluding metrics are broken.

Source URL: https://github.com/base-14/docs/blob/HEAD/docs/instrument/component/yugabytedb.md