분산 속도 제한

An in-memory rate limiter only counts requests on one server. With many app instances behind a load balancer, you need a shared view of usage. Redis, being central and atomic, is the natural coordination point.

The simplest algorithm: a counter per time window. INCR the key; set a TTL equal to the window on first increment. Reject when the count exceeds the limit.

Doing INCR then EXPIRE as two commands risks a key without a TTL if the client dies in between. An atomic Lua script fixes this by running both as one operation.

Across many instances, concurrent requests could otherwise read and write counters in interleaved order. Lua scripts run atomically on the server, so the entire check-and-increment happens with no interleaving.

Fixed windows allow bursts at the boundary: a client can send a full window's worth at the end of one window and again at the start of the next, doubling the effective rate.

A sorted set of request timestamps gives a precise sliding window. Drop old entries, count what remains, and add the new request, all in one script.

The token bucket allows controlled bursts. Tokens refill at a fixed rate up to a cap; each request consumes one. Store tokens and last-refill time in a hash and update atomically with Lua.

On each request, compute elapsed time, add elapsed * rate tokens (capped at the bucket size), then allow the request if at least one token remains. The Lua script keeps this consistent across instances.

Fixed window: cheapest, allows boundary bursts. Sliding log: precise but more memory. Token bucket: smooth with controlled bursts, great for APIs.

Tell clients about their limits: return remaining requests and reset time so well-behaved clients can self-throttle.

Decide a fallback if Redis is unreachable: fail open (allow traffic) for availability, or fail closed (deny) for protection. The right choice depends on whether the limit guards cost or correctness.

You built distributed rate limiting on Redis: fixed-window counters, atomic Lua to avoid TTL leaks and races, sliding-window logs with sorted sets, and token buckets for smooth bursts. Centralizing the counters makes limits consistent across all app instances; choose your fail-open vs fail-closed fallback deliberately.