Audience: production workflow and worker owners.
Works with: OSS and Orkes.
Conductor delivers tasks at least once. A worker must tolerate retries, timeouts, and process replacement without duplicating business effects.
Set pollTimeoutSeconds to detect a task no worker claims, responseTimeoutSeconds to recover a task after a worker disappears, and timeoutSeconds to bound total execution. Set a workflow timeout as a separate guardrail. Choose values from observed latency plus realistic failure recovery, not defaults.
Configure retry count, backoff, and delay on task definitions. Use FAILED_WITH_TERMINAL_ERROR for invalid input or an irreversible business rejection; allow transient network or provider failures to use the retry policy. Design each external action with an idempotency key or a durable application record before enabling retries.
Use task domains for dedicated worker pools, regions, tenants, or high-risk integrations. A domain is a routing boundary, not an authorization boundary; still use scoped credentials and network controls.
Exercise success, transient failure, terminal failure, duplicate delivery, and timeout paths in an integration test. The maintained workflow test harness runs against a real local server.
Next: deployment and scaling, security, and debugging.