Retry a Failed Node From Its Checkpoint
With a checkpointer, a node exception leaves the last successful node checkpoint available. In a SQLite-backed minimal graph, get_state(config).next named the failed node, and a second invoke(None, same_config) retried that node while retaining prior state. Use this for transient failures; make any side effects inside the failed node idempotent because the node can run again from its beginning.
RetryPolicy follow-up reproduced three times on LangGraph 1.2.22 / checkpoint-sqlite 3.1.3: default RetryPolicy(max_attempts=3) did not retry RuntimeError, so the first invoke made one attempt and left the node failed/pending. With retry_on=RuntimeError and max_attempts=2, two attempts exhausted per invocation; after reopening SQLite, invoke(None, same_config) started a fresh two-attempt budget and succeeded on overall attempt four. Persist a separate total-attempt counter if it must span manual resumes.
On LangGraph 1.2.22 / checkpoint-sqlite 3.1.3, an async node with timeout=0.05 was cancelled and raised NodeTimeoutError. Its finally block ran; with RetryPolicy(max_attempts=1), the live and reopened SQLite snapshots retained the prior node state and the timed-out node remained failed/pending. Cleanup inside the timed-out node must be safe to repeat if a retry re-enters it. Reproduced three times.