Big increase in latency around the time the new server joins the cluster and starts serving queries

The following image is from a server getting killed, then restarting and rejoining the cluster a minute later.

The graph shows QPS


and here's latency:

Latency spikes to timeouts around the time the server dies and then once again once the server starts serving queries.

Start by reproducing the reported server kill, restart, and cluster rejoin sequence while comparing the QPS and latency graphs attached to the issue. Done means identifying the cause of the timeout spikes and verifying a fix; the issue names no source file, entry point, or test.

Written by the indexing model from the issue text.