references/queues/patterns.md
Queues Patterns & Best Practices
Fetch the guide matching the task and adapt its example to the application's delivery and failure requirements.
| Task | Documentation | |------|---------------| | Accept requests and enqueue asynchronous tasks; publish to multiple queues | Publish to a Queue via Workers | | Buffer writes to an external API or defer a job | Batching, retries, and delays | | Handle upstream rate limits and backpressure | Handle rate limits of external APIs and consumer concurrency | | Isolate workloads with different latency or capacity needs | Configure Queues | | Retain exhausted retries for inspection and recovery | Dead Letter Queues | | Process R2 object events | R2 event notifications | | Batch output into R2 | Use Queues to store data in R2 | | Batch writes into D1 | D1 database API | | Start durable multi-step jobs | Trigger Workflows | | Publish from a Durable Object | Use Queues from Durable Objects | | Route consumer work to a Durable Object | Invoke Durable Object methods |
Design side effects for at-least-once delivery. A separate check-then-write deduplication flag is not an atomic guarantee: concurrent delivery or a crash between the side effect and recording completion can repeat work. Prefer idempotency keys or transactional enforcement at the destination.
Acknowledge after the destination confirms success. For fan-out, plan for some sends succeeding before another fails; retries must not duplicate downstream effects. Separate queues can isolate workloads, but do not imply a global priority or ordering guarantee.