mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-10-16 13:15:32 +02:00
Elior Nataf LackritzandGitHub
9d92f33cca
A fork from a checkpoint with pending writes to a DeltaChannel snapshots
the channel on the fork's first checkpoint, so the fork doesn't replay
the other branch's writes. If the channel had never been written there,
the snapshot also gives it its first version, and `_mark_bumps_seen`
only marked the bump seen for nodes that already have a `versions_seen`
entry. A node subscribed to the channel that never ran has none, so the
bump started it, and it read the empty value before the node that writes
the channel had run:
```python
graph.invoke("go", config) # writer writes [1] to d, reader reads [1]
fork = graph.update_state(first_checkpoint, {"a": "go"}, as_node="__input__")
graph.get_state(fork).next # ("writer", "reader"): resuming runs reader on [] first
```
A replay from that checkpoint saves the same kind of fork from the loop.
A channel bumped from no version was never written, so
`create_checkpoint` now also marks the bump seen for the nodes the
channel triggers, from the graph's `trigger_to_nodes`, at every call
that can bump. Only raw `Pregel` graphs hit this: `StateGraph` nodes
trigger on their `branch:to:*` channels, not on state keys. An
`as_node="__copy__"` fork still starts such a node, with a plain list
channel as well, so that case isn't DeltaChannel-specific and isn't
changed here.
Raised in review of langchain-ai/langgraphjs#2974. JS twin:
langchain-ai/langgraphjs#2983
## Tests
`test_delta_channel_seal_subscribers.py`: an update as input (sync and
async) and a replay from the checkpoint before a DeltaChannel's first
write leave only the writer next, and the reader then reads the written
value once. All three fail on main. The `langgraph` suite passes
(without Redis), and `lint_package` and `lint_tests` are clean.