We have seen that a node restarting in the cluster can have very negative effects on request latency.

A reproduction test was added in #95161

Allowing Lease Transfers
Extremely High P99 SQL Latency

Blocking Lease Transfers
No impact on P99 SQL Latency

This issue tracks the improvements to crdb in 23.1 to prevent a restarted node impacting end latency. The improvements are narrowly scoped but have been proven to have a significant effect in keeping request latency normal.

• storepool: consider stores suspect on join #97532
• kvserver: consider io overload for lease transfers #96508

Jira issue: CRDB-24164