Episode 9. Tail latency. Day 9 of your 12 day sprint.

Scenario: "p50 is fine but p99 doubled. Where do you look?" Not at the average. At the histogram and the wait sources.

Say: "Averages lie. I read p50, p99, and p999 plus the histogram, then split query time from everything else."

Say the first-data line: "EXPLAIN ANALYZE for the query tail, pg_stat_activity plus wait events for locks and IO, then OS and trace narration."

Say the non-query list: checkpoints, autovacuum, compaction, GC pauses, CPU steal from noisy neighbors, retransmits, replica lag.

Lab: run pgbench with a latency limit, read the histogram, pick the dominant p99 source, and write the mitigation with before-and-after numbers.

Depth tag: kernel tracing is Recognize. Say: "I describe what the perf or eBPF trace would prove. I do not claim kernel authorship."

Drill: name three non-query p99 causes from memory tonight.
Say: checkpoints, autovacuum, compaction, GC pauses, steal, retransmits, lag.
Say the numbers rule: no fix without before-and-after p99 quoted.
Done-check: your mitigation plan names the dominant source plus its proof.
Repeat the spine line: averages lie, the histogram tells the truth.

Teaser: tomorrow, Episode 10, behavioral plus design. The marketplace whiteboard and the stories with metrics.