การจำกัดอัตราและการป้องกันการใช้งานโดยบอต

A single FastAPI worker holding a rate-limit counter in process memory breaks the moment you scale horizontally. With N Uvicorn workers behind a load balancer, an abusive client gets N times the allowance because each worker counts independently.

• Goal: one shared counter that every worker and every pod sees.
• Tool: Redis, an atomic in-memory store that all instances connect to.
• Targets: absorb legitimate traffic spikes while throttling or blocking abusive bots.

Throughout this lesson we build the algorithms in plain Python first, then wire them into FastAPI dependencies.

The simplest algorithm: divide time into fixed windows (e.g. 60-second buckets) and increment a counter per client per window. When the counter exceeds the limit, reject the request.

Below is a pure-Python simulation so you can see the mechanics without Redis. Notice the boundary problem: a client can send the full quota at the end of one window and again at the start of the next, briefly doubling the effective rate.

To kill the boundary problem we track a log of timestamps per client and count only the events that fall within the trailing window relative to now. This is exact but memory-heavy: one entry per request.

The pattern below is exactly what we will translate to a Redis sorted set, where the score is the timestamp.

Fixed and sliding windows are strict. To absorb bursts while enforcing a long-run average, the token bucket is the standard choice.

• Each client owns a bucket with a capacity (max burst) and a refill rate (tokens per second).
• Each request consumes one token; if the bucket is empty, the request is throttled.
• A burst of up to capacity requests passes instantly, then traffic is smoothed to the refill rate.

This is the algorithm we recommend for public APIs facing spiky-but-legitimate traffic.

Across many workers the read-modify-write of a counter is a classic race condition: two workers read count=4, both increment, both write 5, and a request that should have been blocked slips through.

Redis solves this because each command is atomic, but a rate-limit decision usually needs several commands (increment, set expiry, compare). The robust solution is a Lua script executed by EVAL: Redis runs the whole script atomically, with no other client interleaving.

The next scenes show the Redis-backed fixed window and token bucket using this approach.

The cheapest distributed limiter: a Redis key per client per window. INCR returns the new count atomically; on the first hit we set a TTL equal to the window so the key self-cleans.

We wrap both commands in a small Lua script so the increment and the expiry are one atomic unit. This is FastAPI-side glue code, not a standalone program.

For burst absorption we port the token bucket to a Lua script. State lives in a Redis hash holding tokens and ts (last refill time). The script refills based on elapsed time, tries to consume one token, and writes the state back, all atomically.

Returning 1 means allowed, 0 means throttled. We also set a TTL so idle clients free their memory.

We expose the limiter as a dependency so any route can opt in. The dependency derives the client identity, runs the limiter, and raises HTTPException(429) when the quota is exhausted.

Prefer an authenticated identity (API key, user id) over raw IP, because IPs are shared behind NAT and trivially rotated by bots. Fall back to IP only for anonymous traffic.

A correct limiter is also a polite one. Well-behaved clients respect standard signals; sending them reduces retries and support tickets.

• 429 Too Many Requests is the only correct status; never use 403 or 503 for throttling.
• Retry-After tells the client how many seconds to wait.
• X-RateLimit-Limit, X-RateLimit-Remaining, and X-RateLimit-Reset let clients self-pace.

Have your Lua script also return the remaining count and reset time so you can populate these headers without extra round-trips.

One global limit is blunt. Production systems layer several:

• Per-route limits: a cheap GET tolerates far more traffic than an expensive search or login endpoint.
• Tiered identities: anonymous IPs get a tight quota, authenticated users more, paid plans the most.
• Escalation: when a client repeatedly hits 429, write it to a Redis denylist with an exponential TTL so abusive bots get progressively longer bans.

The helper below computes a doubling ban duration capped at one hour.

Rate limiting caps volume, but sophisticated bots stay just under the limit. Combine throttling with cheap behavioral signals:

• Missing or junk headers: absent User-Agent, or a UA on a known bad list.
• Login failure ratio: a high failed-to-successful auth ratio per IP signals credential stuffing.
• Path entropy: rapid hits across many unrelated endpoints suggest scraping.
• Proof of work / CAPTCHA as a gate when a score crosses a threshold, rather than an outright block.

Feed these into a per-client risk score in Redis and tighten the token-bucket capacity dynamically for high-risk clients.

Your public FastAPI API runs across many Uvicorn workers and several pods. Legitimate clients sometimes send short legitimate bursts, but you must enforce a steady long-run average and keep the decision consistent across all instances. Which design fits best?

You can now build distributed, abuse-resistant throttling for FastAPI:

• Centralize state in Redis so every worker and pod shares one counter.
• Pick the algorithm by need: fixed window for simple caps, sliding log for exactness, token bucket to absorb bursts with an enforced average.
• Guarantee atomicity with Lua scripts via EVAL to avoid read-modify-write races.
• Expose limiters as FastAPI dependencies keyed on authenticated identity first, IP as fallback.
• Respond honestly with 429, Retry-After, and X-RateLimit-* headers.
• Layer defenses: per-route and per-tier limits, exponential-backoff denylists, and behavioral bot scoring beyond raw counting.

Always fail open carefully: if Redis is unreachable, decide deliberately whether to allow or block, and alert on it.