TLS session cache miss spike

After a coordinated full restart of your TLS-terminating edge fleet (a fast rolling restart that cycled almost all nodes within a couple minutes), you see a few minutes of elevated p99, raised CPU on the edge nodes, and a spike of client handshake errors/timeouts, even though traffic volume is unchanged. Metrics show TLS session-resumption RATE collapsed to near zero right after the restart and then climbed back over ~5 minutes, exactly tracking the error/CPU spike as it subsided. New full handshakes per second jumped enormously during the window. How do you triage and mitigate?