Chapter 10 · Fault tolerance, retries, and idempotency

Simulate one transient failure and verify a node retry without looping forever on all errors. No real service is contacted.

Mechanics and design choices

RetryPolicy specifies attempt limits and retryable exceptions. Authentication and validation errors should not normally be retried as transient failures. A checkpoint restores graph state, but an external write may have succeeded while its response was lost; retry with the same idempotency key. Current docs also describe 1.2 node timeouts and error handlers; check installed versions before adopting them.

Example

from typing import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.types import RetryPolicy

class TemporaryFailure(Exception):
    pass

class State(TypedDict):
    result: str

attempts = {"count": 0}

def fetch_fixture(state: State):
    attempts["count"] += 1
    if attempts["count"] == 1:
        raise TemporaryFailure("Simulated transient failure")
    return {"result": "fixture loaded"}

builder = StateGraph(State)
builder.add_node("fetch", fetch_fixture, retry_policy=RetryPolicy(
    max_attempts=2, initial_interval=0.01,
    retry_on=TemporaryFailure,
))
builder.add_edge(START, "fetch")
builder.add_edge("fetch", END)
result = builder.compile().invoke({"result": ""})
assert attempts["count"] == 2
assert result["result"] == "fixture loaded"
print(result)

Run and verify

Expect success on the second attempt. attempts is a local test counter, not recoverable business state. Change the error to ValueError and expect immediate failure, verifying the retry policy is narrow.

Pitfalls

Avoid multiplicative retries across model SDK, tool client, and graph node. Set one overall budget and record attempts at each layer together with the final exception.

🧪 随堂测验

点击你认为正确的选项。答错时会展示正确答案与原因解析。

1. Does max_attempts include the initial attempt?

2. Should authentication errors retry indefinitely?

3. Can checkpoints prevent every duplicate payment?

4. What is a risk of layered retries?

Hands-on exercises

1. Always raise TemporaryFailure and assert only two attempts.
2. Add a non-retryable error case and verify one attempt.
3. Design a request_id-based mock write that records repeated requests once.

Official references

• fault-tolerance

Verified on: 2026-09-17.