Companion

One caller waiting for a connection: the pool the queries share with the appends is used up.

The queries of a store run one at a time at the least: a limit of 0 would refuse every query.

One second: short enough to give way within a few seconds of a slow write, long enough to sample it.

Half again the baseline: the period-to-period spread of a store's p90 under a steady load stays within it, so a period above it is slower than the store's normal latency.

The p90 presses only this far above its baseline, whatever the factor: a factor alone is relative, and a store whose appends take a millisecond (PostgreSQL on a near network) would press on noise of a millisecond or two, as one of a few milliseconds would on the jitter of a busy pool. Twenty milliseconds is about one more round trip queued on a pool under load, a delay the writes feel.

Two periods in a row: one slow second is no trend.