Each transport call's milliseconds, in order — for a percentile.
Requests sent again at once on a new connection, because the kept-alive one had
been closed by the server (HttpResponse.reconnected). Not retries: the first
never reached the server.
Calls to the transport: every attempt and every redirect hop.
Of those, attempts after the first of one request.
Why each retry was made (issue #31): a 429/503, a timeout, a failed connection, anything else.
429 and 503 answers: the upstream asking to slow down.
Milliseconds inside the transport, summed.
Milliseconds spent waiting before a request: pacing (--min-host-interval, a source's floor) and retry backoff.
Where an engine's time went, counted as it goes (issue #14): a slow sync could not tell waiting on the upstream from extracting without
ps,iostatandnettop. Read by the pipeline for a run'stimingand progress line. On the engine's own clock, so a test with an injected one gets exact numbers.