Durability modes

Durability controls when checkpoint writes finish relative to graph execution. It matters only when the graph has a checkpointer.¹

Three choices

Mode	When persistence happens	Trade-off
`"sync"`	Finish the write before the next step starts	Strongest boundary, more latency
`"async"`	Write while the next step runs	Default; lower latency, small crash window
`"exit"`	Persist only when the graph exits	Fastest intermediate steps, no intermediate recovery point

Tiny example

result = graph.invoke(
    {"case_id": "104"},
    config={"configurable": {"thread_id": "case-104"}},
    durability="sync",
)

Use sync when the next step must never start before the previous checkpoint is stored. Use exit only when replaying the whole run is acceptable.

Failure note

Durability does not make external effects transactional. A checkpoint and a payment API are two separate systems; use an idempotency key and verify the remote result on retry.

LangChain, Durability reference and invoke reference. ↩