mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-10-14 12:15:25 +02:00
Elior Nataf LackritzandGitHub
39c523eb0a
`bulk_update_state` stores each update's writes on the checkpoint it
updates, under a task id. An update whose node has no pending task to
reuse got `uuid5(checkpoint_id, INTERRUPT)`, so every such update in one
superstep shared an id. Savers keep one write per `(task_id, idx)`, so
all but the first update's writes were dropped.
Plain channels don't notice, since their value is stored in the new
checkpoint. A `DeltaChannel` rebuilds its value from those writes, so it
lost every update after the first, on memory, sqlite and postgres:
```python
graph.bulk_update_state(config, [[
StateUpdate({"d": ["u1"], "p": ["u1"]}, as_node="a"),
StateUpdate({"d": ["u2"], "p": ["u2"]}, as_node="m"),
]])
# plain: [..., 'u1', 'u2']
# delta: [..., 'u1']
```
The existing multi-update test only passed because it gave each update
an explicit `task_id`, and its docstring said so.
## Fix
The `i`th update gets `uuid5(checkpoint_id, f"{INTERRUPT}:{i}")`. The
first keeps the old id, so a single `update_state` stores exactly what
it did before. Updates that reuse a pending task's id, or pass their
own, are unchanged.
Each update also gets the task path `(__interrupt__, i)`, stored with
its writes. Live execution applies a superstep's updates in the order
given, and a saver that replays a checkpoint's writes by task path
(#8544) now gives them back in that order instead of by task id. Savers
whose `put_writes` takes no `task_path` get the old call.
## Limits
- On savers that order a checkpoint's writes by task id only (the
released ones, until #8544), several updates in one superstep still
replay in task id order.
- langgraphjs has the same shared id in `bulkUpdateState` and loses the
same update; fixed in langchain-ai/langgraphjs#2923 with the same ids.
## Tests
`test_delta_channel_update_state.py` adds sync, async and
next-to-a-pending-task cases across the checkpointer fixtures, with no
explicit task ids, on a thread that already has history (a fresh thread
snapshots every delta channel and hides the bug). All 19 fail on main.
Two more (sync and async) give six updates in one superstep and check
that `get_state` returns them in that order, on a saver that replays by
task path; they fail without the path.