Set poll, response, and execution timeouts on every worker task definition. Retry only idempotent or compensated work, use exponential backoff for transient remote failures, and route resource-bound work to domains.
Workers may receive a task more than once. Persist idempotency keys before external
side effects and use TaskInProgress or lease extension for long-running work.
Verify behavior with failure-path tests and inspect task reasons before retrying.