Getting KeyError in pregel/__init__.py -> _prepare_next_tasks() due to packet.node not being in processes

[Glinte]

I don't have a reproducible example yet, still debugging. Just putting it out here if any core devs knows the issue immediately. The dispatcher -> Subgraph edge is a conditional edge that returns a list of Sends. It seems like the error is that the Send is getting propagated inside the subgraph, where there is no longer a node with the name "Subgraph"

 def _prepare_next_tasks(
        checkpoint: Checkpoint,
        processes: Mapping[str, PregelNode],
        channels: Mapping[str, BaseChannel],
        managed: ManagedValueMapping,
        config: RunnableConfig,
        step: int,
        *,
        for_execution: bool,
        get_next_version: Union[None, Callable[[int, BaseChannel], int]] = None,
        manager: Union[None, ParentRunManager, AsyncParentRunManager] = None,
    ) -> tuple[Checkpoint, Union[list[PregelTaskDescription], list[PregelExecutableTask]]]:
        checkpoint = copy_checkpoint(checkpoint)
        tasks: Union[list[PregelTaskDescription], list[PregelExecutableTask]] = []
        # Consume pending packets
        for packet in checkpoint["pending_sends"]:
            if not isinstance(packet, Send):
                logger.warn(f"Ignoring invalid packet type {type(packet)} in pending sends")
                continue
            if for_execution:
>               if node := processes[packet.node].get_node():
E               KeyError: 'Subgraph'

[nfcampos]

@Glinte Is this happening in langgraph-api / LangGraph Cloud?

[Glinte]

@nfcampos Not using LangGraph cloud, this is on stock langgraph from at least 0.1.9 to 0.1.11 using any sort of checkpointer (I checked Memory and AsyncSqlite). This happens only in some very specific situation, my description here is insufficient for reproduction. I haven't got a reproducible example yet because it seems like there is more to the shape of the graph that is causing this problem.

I would be very surprised if LangGraph Cloud fixes this though, since the error comes from the core algortihm having the (apparently incorrect) assumption that all node names can be found in the current graph

[nfcampos]

This should be fixed in the latest version, if you could try to update? Question on LangGraph Cloud was to confirm whether the fix was needed there or in the library.

[Glinte]

For future reference, relevant PR: #1119