Retry workflow nodes only for selected transient exceptions

Goal

Recover from retryable node failures without amplifying outages or duplicating side effects.

Steps

1. Confirm the node is idempotent or deduplicated before enabling retries. [retry]
2. Allowlist only transient connection or timeout failures; exclude argument and permission errors. [retry]
3. Use RetryConfig(max_attempts=3, initial_delay=1.0, backoff_factor=2.0, exceptions=[ConnectionError, "TimeoutError"]) to retry only selected failures. [retry]
4. Inject transient and permanent failures; verify errors remain traceable and writes are not duplicated. [retry]

How to verify

Transient errors recover within the limit, permanent errors fail at the cap, and non-idempotent writes are not replayed.

Troubleshooting

If retries worsen throttling, increase backoff and reduce concurrency; disable retries first if side effects duplicate.