Quota Management

Rate limiting controls speed: "no more than 100 requests per minute." Quota management controls volume: "no more than 10,000 requests this month."

Rate limits protect infrastructure from overload. Quotas protect your business model. Free-tier users get 1,000 API calls per month, pro users get 100,000, enterprise gets unlimited. Rate limits reset every second or minute. Quotas reset daily, monthly, or per billing cycle.

OpenAI gives you a monthly token quota based on your tier. Tier 1 gets 100,000 tokens per minute and 500,000 tokens per day. Tier 5 gets 10,000,000 per minute. Google Maps gives you $200 of free API calls per month, then charges per call. AWS has service quotas (formerly "limits") on everything from Lambda concurrent executions (1,000 default) to S3 bucket counts (100 per account).

The distinction matters for HTTP responses too. Rate limit exceeded returns 429 Too Many Requests. Quota exceeded returns 403 Forbidden. The semantics are different: 429 says "slow down, try again soon." 403 says "you have used your allowance for this period."

Quotas are typically tracked in a database (not just Redis) because they span long periods and the numbers are billable:

CREATE TABLE api_quotas (
 api_key_id UUID REFERENCES api_keys(id),
 period_start TIMESTAMP NOT NULL,
 period_end TIMESTAMP NOT NULL,
 usage_count BIGINT DEFAULT 0,
 quota_limit BIGINT NOT NULL,
 PRIMARY KEY (api_key_id, period_start)
);

Per-request flow:

1. Check cached quota status (Redis counter, ~0.3ms)
2. If within limits: allow, atomic INCR on the Redis counter
3. Background job syncs Redis counter to Postgres every 60 seconds
4. When quota is exceeded: return 403 with quota headers

The hot path never touches the database. Redis INCR is atomic and costs one round trip. A SQL update would need a row lock and a transaction. At ten thousand requests per second, only the Redis path keeps up. But the billing system reads from Postgres, not Redis. Redis is for gating. Postgres is for billing.

Quota headers (every response):

X-Quota-Limit: 10000
X-Quota-Remaining: 3456
X-Quota-Reset: 2024-02-01T00:00:00Z

Overage handling strategies:

• Hard cutoff: Reject all requests once quota is reached (most common for free tiers)
• Soft limit with overage billing: Allow requests beyond the quota but charge extra, typically $0.001 per extra call (common for pay-as-you-go plans like AWS, GCP, Twilio)
• Degraded service: Switch to cached data or reduced rate (Spotify does this for free-tier API partners)
• Grace period: Allow 10% overage, then hard cutoff (gives time to upgrade)