Designing rate limits that protect the service and inform the client

Goal

Keep one client from consuming the capacity of all, and tell well-behaved clients exactly when to try again.

Prerequisites

A verified client identity per request: the authenticated account, or a network identifier derived from the address behind a trusted proxy.

Steps

1. Choose budgets per action class: reads, content writes, registrations, expensive queries; cheap actions get large budgets, creation of durable objects small ones.
2. Store counters where all workers see them (a database row per identity and window with an atomic upsert, or a shared store); in-memory counters reset on restart and diverge between workers.
3. Use fixed windows for simplicity or sliding windows for smoothness; document which.
4. Reject with 429 Too Many Requests (RFC 6585) and Retry-After in seconds; keep the body machine-readable.
5. Add a global ceiling so that many identities together cannot overload the service.
6. Publish effective limits in discovery metadata and make them administratively adjustable without a deployment.

Expected result

Bursts from one identity are rejected predictably, others are unaffected, and clients back off precisely.

Limits and test basis

Network identifiers are coarse behind carrier-grade NAT and can be shared by many users; combine with account limits where possible. Rate limits mitigate, they do not prevent, distributed abuse. The design mirrors this wiki's persistent quota implementation.