The S3 backend retried only a 5xx answer: a dropped connection or a
timeout on the way to the object store failed the read or write at
once, and the first retry waited a whole second. A transport error and
a 429 now retry the same way, and the waits run 100 ms, 200 ms, 400 ms
... capped at 5 s, so a Garage node restarting costs a request tenths
of a second rather than failing it.