Rate limits

Each key, and each signed-in person, has its own bucket of requests: 120 requests of burst, refilled at 2 per second.

Every request spends one, however malformed it is, because the bucket is spent while the credential is checked and before anything else is read.

A request made with an empty bucket is refused:

429 Retry-After: 3
{"error":{"code":"limited","message":"Rate limit exceeded; retry after 3s"}}

Wait the number of seconds Retry-After names, then try again. The mdbrain CLI exits with code 4 on this answer, so a script can tell it from every other refusal.