Rate limits and quotas

Reads and writes are limited separately, because a write is more expensive and more dangerous to retry blindly. Exceeding a limit returns 429 with a Retry-After header stating how long to wait.

A client that retries without honouring that interval makes things worse for itself and everyone else. Exponential backoff with jitter is the correct behaviour.

Concurrent runs, concurrent endpoints and stored bytes are all capped by plan. Exceeding one returns a conflict with a specific code rather than queueing the work, because silently queueing converts a clear error into a mysterious delay.

A quota bounds concurrency and footprint. A spend limit bounds money. An account can hit either without approaching the other — three cheap runs can exhaust a concurrency quota while barely touching a budget.