Use when designing rate limiting for an API, choosing between token bucket / sliding window / leaky bucket / fixed window, implementing it in Redis, deciding edge (Cloudflare/Upstash) vs origin enforcement, sizing per-user vs per-IP vs per-endpoint quotas, returning the right 429 response with Retry-After, or fixing the boundary-burst bug in fixed-window limiters. Triggers: 429 too many requests, INCR + EXPIRE, ZADD + ZREMRANGEBYSCORE + ZCARD, X-RateLimit-Remaining header, Cloudflare WAF rate limiting rules, Upstash @upstash/ratelimit, leaky bucket shaping vs policing, distributed rate limiter consistency. NOT for DDoS mitigation specifically (different scale), CAPTCHA / bot management, full WAF design, or per-user quota billing.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Use when designing rate limiting for an API, choosing between token bucket / sliding window / leaky bucket / fixed window, implementing it in Redis, deciding edge (Cloudflare/Upstash) vs origin enforcement, sizing per-user vs per-IP vs per-endpoint quotas, returning the right 429 response with Retry-After, or fixing the boundary-burst bug in fixed-window limiters. Triggers: 429 too many requests, INCR + EXPIRE, ZADD + ZREMRANGEBYSCORE + ZCARD, X-RateLimit-Remaining header, Cloudflare WAF rate limiting rules, Upstash @upstash/ratelimit, leaky bucket shaping vs policing, distributed rate limiter consistency. NOT for DDoS mitigation specifically (different scale), CAPTCHA / bot management, full WAF design, or per-user quota billing.
allowed-tools
Read,Grep,Glob,Edit,Write,Bash
metadata
{"category":"Backend & Infrastructure","tags":["rate-limiting","redis","api-design","cloudflare","throttling","infrastructure"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"pairs-with":[{"skill":"reverse-proxy-for-agents","reason":"Owns the Nginx/Caddy/Cloudflare gateway layer where the edge half of the edge-plus-origin limiter architecture is enforced."},{"skill":"error-handling-patterns","reason":"The client-side retry/backoff and circuit-breaker discipline that consumes the 429 + Retry-After contract designed here."},{"skill":"logging-observability","reason":"The fail-open metrics, denied-rate alerting, and OTel span attributes this skill's quality gates require."}],"io-contract":{"kind":"deliverable","consumes":["[Truncated]","[Truncated]"],"produces":["[Truncated]","[Truncated]"]}}
Rate Limiting Strategy
Three decisions before any code: which algorithm, where in the stack, what key. Get those right and the implementation is 30 lines of Lua. Get any of them wrong and you'll either let an attacker through or wreck legitimate clients on traffic spikes.
The widely-recommended starting position, repeated across Redis's own tutorial, API7's algorithm guide, and Cloudflare's deployment data, is sliding window counter as a general-purpose default — it has the accuracy of a sliding window log without the O(n) memory (Redis — Build 5 Rate Limiters, API7 — Rate Limiting Guide). Cloudflare's published numbers: only 0.003% misclassification across 400M requests with sliding window counter (API7).
algorithm → sliding window counter (general); token bucket (allow bursts); leaky bucket (smooth output)
where → edge (Cloudflare WAF / Upstash) for unauth; origin (Redis) for per-user
key → authenticated user ID > IP+UA > IP > global
The Redis tutorial's compressed comparison: (redis-ratelimit)
"Start with Sliding Window Counter — it balances accuracy, simplicity, and memory efficiency for most production APIs."
Fixed window's boundary trap
The simplest algorithm has a real bug worth understanding before you ship it:
limit = 10 requests per 10 seconds, fixed windows aligned to wall clock.
Window A: [00:00:00 – 00:00:09]
Window B: [00:00:10 – 00:00:19]
Client sends 10 requests at 00:00:09.5 → all in window A, all allowed.
Client sends 10 requests at 00:00:10.5 → all in window B, all allowed.
Net: 20 requests in 1 second within a "10 per 10 seconds" limit.
That's the boundary burst. Sliding-window counter fixes it by interpolating across windows. Sliding-window log fixes it exactly at O(n) memory cost.
Cheapest — drops before origin pays. Globally distributed.
Coarse keys (IP, header, JWT subject).
Unauthenticated traffic, abuse, DDoS-adjacent.
API gateway (Kong, Envoy, Traefik)
Per-route, easy to standardize.
Single point if it's a single instance.
Per-tier quotas, internal mesh.
Application + Redis
Full context (user_id, plan, endpoint cost).
Each request hits origin.
Per-user quota, plan-based limits, billing.
In-process (@nestjs/throttler without Redis)
Zero infrastructure.
Per-instance only — N pods = N× the limit.
Single-replica internal tools. Rarely correct in prod.
The right architecture is usually edge + origin in series: edge drops obvious abuse cheaply; origin enforces per-user quotas with full context. Don't try to do per-user quota at the edge with imperfect identification, and don't try to do DDoS at the origin.
Cloudflare offers WAF Rate Limiting rules that key on common headers and Bot Score; Upstash offers @upstash/ratelimit for Workers/edge runtimes — both are sliding-window-counter-based. (Upstash docs)
Choosing the key
The key determines who shares the limit. Bad keys are how you DOS your own users.
Tier
Key
Notes
1
Authenticated user ID
Best for product limits. Requires auth to be settled before the limiter.
2
API key / token hash
For B2B; pair with plan tier.
3
IP + user-agent + path
For unauthenticated; UA reduces NAT collisions.
4
IP only
Coarsest; corporate NATs and CGNAT collide hundreds of users.
5
Global (per-endpoint)
Use only as a backstop ("never more than 5k/sec to /search").
For login: rate-limit by (IP, username) AND (IP) AND (username) simultaneously. An attacker brute-forcing one user's password rotates IPs; an attacker spraying weak passwords across many users uses one IP; both must be blocked. Limit each tuple separately.
Redis recipes
The single common requirement: atomic read-modify-write. That means a Lua script via EVAL (or EVALSHA cached). Without atomicity you have race-induced over-allows under contention.
-- KEYS[1] = bucket key, ARGV = now (ms), windowMs, limit, memberlocal now = tonumber(ARGV[1])
local windowMs = tonumber(ARGV[2])
local limit = tonumber(ARGV[3])
local member = ARGV[4]
redis.call('ZREMRANGEBYSCORE', KEYS[1], '-inf', now - windowMs)
local count = redis.call('ZCARD', KEYS[1])
if count < limit then
redis.call('ZADD', KEYS[1], now, member)
redis.call('PEXPIRE', KEYS[1], windowMs + 1000)
return {1, limit - count - 1} -- allowed, remainingendreturn {0, 0} -- denied
member should be now .. ':' .. randomBytes so duplicates don't collide. Keep limit ≤ ~10k; sorted set memory grows linearly with traffic.
Sliding window counter (general default)
Approximate sliding window using two adjacent fixed windows and weighted interpolation:
-- ARGV: now_ms, windowMs, limitlocal now = tonumber(ARGV[1])
local windowMs = tonumber(ARGV[2])
local limit = tonumber(ARGV[3])
local currentWindow = math.floor(now / windowMs)
local currentKey = KEYS[1] .. ':' .. currentWindow
local prevKey = KEYS[1] .. ':' .. (currentWindow - 1)
local current = tonumber(redis.call('GET', currentKey) or'0')
local previous = tonumber(redis.call('GET', prevKey) or'0')
local elapsedInCurrent = (now % windowMs) / windowMs
local weighted = previous * (1 - elapsedInCurrent) + current
if weighted + 1 <= limit then
redis.call('INCR', currentKey)
redis.call('PEXPIRE', currentKey, windowMs * 2)
return {1, math.floor(limit - weighted - 1)}
endreturn {0, 0}
O(1) memory per principal. Cloudflare's deployment shows ~0.003% error on 400M requests. (api7-ratelimit)
Token bucket (allow bursts to capacity)
-- ARGV: now_ms, capacity, refill_per_sec, costlocal key = KEYS[1]
local now = tonumber(ARGV[1])
local capacity = tonumber(ARGV[2])
local refill = tonumber(ARGV[3])
local cost = tonumber(ARGV[4])
local s = redis.call('HMGET', key, 'tokens', 'ts')
local tokens = tonumber(s[1]) or capacity
local lastTs = tonumber(s[2]) or now
local elapsedSec = (now - lastTs) / 1000
tokens = math.min(capacity, tokens + elapsedSec * refill)
if tokens >= cost then
tokens = tokens - cost
redis.call('HMSET', key, 'tokens', tokens, 'ts', now)
redis.call('EXPIRE', key, math.ceil(capacity / refill) + 60)
return {1, math.floor(tokens)}
end
redis.call('HMSET', key, 'tokens', tokens, 'ts', now)
return {0, 0}
cost per request lets you charge expensive endpoints more (a vector embedding might cost 10 tokens, a metadata read 1).
Fixed window (when you must keep it simple)
local count = redis.call('INCR', KEYS[1])
if count == 1then
redis.call('EXPIRE', KEYS[1], tonumber(ARGV[1]))
endreturn count
Cheap, simple, has the boundary-burst bug. Acceptable for "no more than N writes per minute to a low-stakes internal endpoint" but not for anything customer-facing.
HTTP/1.1 429 Too Many Requests
Content-Type: application/json
Retry-After: 30
RateLimit-Limit: 100
RateLimit-Remaining: 0
RateLimit-Reset: 1714592400
{
"error": "rate_limit_exceeded",
"message": "100 requests per minute exceeded.",
"retry_after": 30
}
Retry-After: integer seconds (or HTTP date). Required for the client to back off correctly.
RateLimit-* headers (IETF draft): Limit, Remaining, Reset (epoch seconds). Some implementations still use the legacy X-RateLimit-*; emit both during transition.
Body should be machine-parseable; clients automate around it.
Don't return 429 from auth endpoints with helpful detail — "4 attempts left" becomes a brute-force aid. Keep the message generic.
Distributed gotchas
One Redis, multiple app pods: fine, atomic via Lua. This is the common case.
Redis Cluster: keys for one principal MUST hash to the same slot. Use {user:42} as part of the key so the surrounding key components don't break atomicity.
Multi-region with separate Redises: the limiter is per-region. A client switching regions doubles their limit. Either accept it (most APIs) or use a global counter (separate concern, much higher latency).
Failure mode: if Redis is down, what happens? Open question. Common answers:
Fail-open (allow everything) — preserves availability but lets abuse through.
Fail-closed (deny everything) — preserves protection but tanks availability.
Local fallback (per-pod in-memory limiter at higher allowance) — degraded but still some protection.
OTel span attributes per request: ratelimit.key, ratelimit.allowed, ratelimit.remaining (see opentelemetry-instrumentation).
Anti-patterns
Per-pod in-memory limiter
Symptom: Limit is "100/min" but production allows 100×N (N pods) per minute. Scaling out silently raises the effective limit.
Diagnosis: No shared state.
Fix: Redis (or equivalent) backing.
Fixed window for customer-facing API
Symptom: Reports of "I'm getting 200 requests through in two seconds" despite the 100/min limit.
Diagnosis: Boundary burst.
Fix: Sliding window counter (or sliding window log if exactness required).
Keying on IP only behind a proxy
Symptom: All requests appear to come from the load balancer's internal IP; one user's spam blocks everyone.
Diagnosis: App reads req.connection.remoteAddress, not X-Forwarded-For.
Fix: Read the trusted forwarded IP. Validate the proxy chain. Better yet, key on the authenticated user where possible.
No Retry-After header
Symptom: Clients hammer the API in tight loops after 429s; outages look worse than they are.
Diagnosis: No explicit backoff signal.
Fix: Always emit Retry-After. Document the header. Pair with RateLimit-*.
Same limit on auth and read endpoints
Symptom: Brute-force succeeds because a 1000/min limit on /login is way too generous, OR legitimate clients get blocked because a 5/min limit also applies to /health.
Diagnosis: One global limit policy.
Fix: Differentiate by endpoint cost / risk: auth ~5–10/min per IP+username; reads ~1000/min per user; expensive (LLM, search) by token-bucket cost.
Race conditions without Lua
Symptom: Counter occasionally allows a few above the limit under concurrency.
Diagnosis: Sequence of GET/SET that isn't atomic.
Fix: All limiter logic in a single EVAL (Lua). Use EVALSHA after the script is loaded.
Cluster keys not co-located
Symptom:MOVED errors, slot mismatch, or atomicity broken.
Diagnosis: Lua script touches multiple keys that hash to different slots.
Fix: Use {tag} syntax: rl:{user:42}:current and rl:{user:42}:prev hash to the same slot.
Failing open without telemetry
Symptom: Abuse goes through during a Redis blip; nobody notices.
Diagnosis: Limiter caught the error, allowed the request, didn't log/alert.
Fix: Emit a metric on every fail-open; alert if rate > N per minute.
Returning detailed quotas in error messages
Symptom: A scraper learns the exact limit and stays just under it.
Diagnosis:"You have 4/5 attempts remaining" is reconnaissance.
Fix: Generic 429 message. Quota status only on authenticated, scoped endpoints.
Quality gates
Test: load test asserts the limit is enforced in aggregate (Nx pods, K concurrent clients) — not just per-pod.
Test: boundary test for sliding-window correctness — assert no 2× burst at window flips.
Test: Redis-down failure mode is exercised (chaos toggle); behavior matches the documented decision (fail-open / fail-closed / local-fallback).
Algorithm choice is documented and matched to use case (sliding window counter is the default).
All limiter logic runs in a single EVAL/EVALSHA Lua script — no race in the read-modify-write.
Cluster-safe keys ({tag} hash co-location) if using Redis Cluster.
Per-endpoint and per-key-tier limits configured (auth vs read vs expensive); not a single global limit.
Retry-After AND RateLimit-Limit/Remaining/Reset headers on every 429.
Login rate-limited by (IP, username), (IP), and (username) tuples simultaneously.
Edge-layer rate limiting in place for unauthenticated traffic (Cloudflare WAF, Upstash, AWS WAF).
OTel span attributes recorded per request (ratelimit.key, ratelimit.allowed, ratelimit.remaining) — see opentelemetry-instrumentation.
Deterministic Audit
Before shipping a limiter design (or reviewing another agent's), write the three decisions —
algorithm, placement, key — as a JSON rate-limit-plan matching
schemas/rate-limiting-strategy-plan.schema.json and run it through the deterministic
auditor:
auditRateLimitingStrategy(plan) (in scripts/rate_limiting_strategy_audit.mjs) turns this
skill's decision tables and Quality Gates into machine-checkable rules over structured
fields — no keyword matching: fixed window on a customer-facing API (the boundary burst),
an in-process limiter multiplied by replica count, Redis read-modify-write without Lua,
Cluster keys without hash tags, a missing Retry-After, an undecided Redis failure mode,
fail-open with no telemetry, uncovered auth endpoints or an incomplete login-tuple set,
and IP-only keying for customer traffic. It returns
{ pass, score, findings, recommendations }; examples/sample-input.json is a
sliding-window-counter, edge+origin, user-keyed design that audits pass: true. Version
history: CHANGELOG.md.
NOT for
DDoS mitigation at scale — different problem; needs anycast, scrubbing, BGP. → CDN + WAF.
CAPTCHA / bot management — behavioral signals, not rate-based.
Full WAF design — much bigger surface (rules, signatures, anomalies). No dedicated skill yet.
Per-user billing / quota for a metered product — rate limiting is the enforcement layer; billing is a different system.
Network-layer SYN-flood / packet rate limiting — kernel-level, different toolset.
Rate-limiting jobs in a queue (concurrency caps on workers) — different shape. → background-job-queue-design.
Sources
Redis — Build 5 Rate Limiters with Redis: Algorithm Comparison Guide (Lua scripts for all four algorithms with pros/cons). redis.io/tutorials/howtos/ratelimiting/
