An API should allow an average of 10 requests per second per key, but let clients make short bursts of up to 50 requests. Which algorithm fits most naturally?

A token bucket can store up to 50 unused tokens. So a client that was idle can send a burst of 50 requests. The refill rate keeps the long-run average at 10 per second. "Leaky bucket draining at 10 per second" is tempting, but it smooths output to a constant 10 per second. A burst is queued or dropped, not served right away.