LangGraph Execution Freezes Intermittently Without Errors – Possible Deadlock?

[harshitgupta412]

Hi all,

I've been facing an issue with LangGraph where the graph execution intermittently stops without any error messages or warnings.

Background
I'm building a search application using LangGraph. My setup includes:

• A simple LangGraph flow:
  • Generates search queries
  • Executes searches in parallel
  • Aggregates results using an LLM
    The graph uses the Postgres checkpointer to track graph execution. Since this is a custom service with other components, I am hosting it on a private server using flask+gunicorn+gevent workers.

Issue
After executing a few queries successfully, some nodes stop executing (generally the final step to aggregate results but this has happened with other nodes as well). No errors or logs are emitted—execution just doesn't proceed so it seems like a deadlock may be occurring.
Further, when sending new requests to the graph, some nodes do run but execution hangs at later stages.

The only workaround so far is restarting the server. The issue appears intermittently, making it hard to consistently reproduce. And because there's no indication when this happens, many queries end up getting stuck.

Any help would be appreciated! This has been a blocker for me, and I’d love to collaborate to find a solution or workaround.

Thanks in advance!

[hinthornw]

This is in dev?

[harshitgupta412]

I am using the official pypi release of langgraph (v0.2.60)

[hinthornw]

Are you seeing the issue when langgraph dev? Please provide code & additional context so we can reproduce this issue and/or provide additional guidance.

[harshitgupta412]

I dont use the langgraph commands to run the graph as I mentioned before. I am attaching a minimal version of my server script. The server is running using the command:

gunicorn --worker-class gevent --workers 2 --bind 0.0.0.0:5000 "server:app"

Server code:

from gevent import monkey
monkey.patch_all() 
import gevent
from flask import Flask, request, Response
from flask_cors import CORS
import uuid
from queue import Queue
from langgraph.graph.state import CompiledStateGraph

app = Flask(__name__)
CORS(app)

def stream_from_queue(queue: Queue[dict | None]):
    while True:
        response = queue.get()
        yield f"data: {response}\n\n"
        if response is None:
            break

# Dummy function to load a graph based on configuration.
def load_graph(graph_config: dict) -> CompiledStateGraph:
    pass

# Load graphs from configuration. Dummy implementation.
graphs: dict[str, CompiledStateGraph] = {
    "search_agent": load_graph({"name": "search_agent"}),
    "agent": load_graph({"name": "agent"}),
}

@app.route('/run_graph', methods=['POST'])
def run_graph():
    data = request.json
    stream_queue: Queue[dict | None] = Queue()
    graph = graphs[data.get('graph_type')]

    # Run the graph in a separate thread/greenlet
    def run_graph_with_queue():
        try:
            graph.invoke(
                input={
                    "input": data.get('input', {}),
                    "queue": stream_queue
                },
                config={"configurable": {"thread_id": str(uuid.uuid4())}}
            )
            stream_queue.put(None)
        except Exception as e:
            stream_queue.put({"error": str(e)})
            stream_queue.put(None)
    
    gevent.spawn(run_graph_with_queue)
    return Response(stream_from_queue(stream_queue), mimetype="text/event-stream")

if __name__ == '__main__':
    app.run(host='0.0.0.0', port=80, debug=True)

[Ayush170-Future]

I'm having the same issue. @harshitgupta412 were there any updates or fixes?

[Mgczacki]

This is a problem I was having previously and just started having again. The usage pattern that caused this originally was invoking graphs within a parent graph's nodes.
At the time, removing this pattern (rather adding the subgraphs themselves as nodes within the parent graph) fixed the deadlock problem.
As of a few days ago, using langgraph 0.4.8, a small refactor brought the issue back without having any graph invocations within nodes. This is not fixed as of 0.5.0rc1, but rather it happens more.

One consumer facing product in my company is made with langgraph, and this has become a major blocker to releasing updates. If it helps, we are running everything async, and we have a corroutine that should emit events after a certain time. These never actually trigger when we face the deadlock, so there must be an issue that blocks the main thread somewhere.

I would love to know if this problem is in your team's radar @hinthornw

[Mgczacki]

I have tracked down the deadlock to when calling Langchain ainvokes. Within a node given a specific sequence of nodes, the langchain ainvoke will just halt all processing and become irreparably stuck. Interestingly, it seems to be a specific interaction between Langgraph and Langchain for this case, but I have seen it happen in other cases as well.

[hinthornw]

Do you have a pointer to the lock you are seeing?

[Mgczacki]

I will try and recreate it with a minimal example, our codebase cannot be shared and has too many internal dependencies to make sense outside of our organization. This issue has, as far as I know, existed for more than a year and:

1. Always happens given the same sequence of nodes. If the sequence changes, it may not happen at all.
2. Happens when there is lots of graph nesting going on.

I will try and get a working example to recreate this without any proprietary code.

[Ankit3794]

facing the same issue. do we have any workaround?

[pavelsavva]

Our graph looks very similar to yours 🤔

[GaelCangySpoon]

I don't know if it is relevent but I am a windows user. I got wind of a few of my colleagues using linux and never hitting that issue.

So I tested my theory on WSL and to my suprise i don't hit that limit anymore. I am not sure if it is an underlying issue with windows that blocks certain python or langchain processes but switching to linux fixed it for me.

You can try it and revert to me if you got any similar results? @pavelsavva @Ankit3794

[pavelsavva]

I don't know if it is relevent but I am a windows user. I got wind of a few of my colleagues using linux and never hitting that issue.

So I tested my theory on WSL and to my suprise i don't hit that limit anymore. I am not sure if it is an underlying issue with windows that blocks certain python or langchain processes but switching to linux fixed it for me.

You can try it and revert to me if you got any similar results? @pavelsavva @Ankit3794

Hmm for me this never happens on Mac, locally, but happens in prod when running in a Linux Docker container

[Ankit3794]

@GaelCangySpoon, I am using https://github.com/langchain-ai/open_deep_research Workflow agent. I am facing issue even on Mac and Linux. not sure, but when I put this behind FastAPI and run it in background job it mostly happens.

[roysupriyo10]

I don't know if it is relevent but I am a windows user. I got wind of a few of my colleagues using linux and never hitting that issue.

So I tested my theory on WSL and to my suprise i don't hit that limit anymore. I am not sure if it is an underlying issue with windows that blocks certain python or langchain processes but switching to linux fixed it for me.

You can try it and revert to me if you got any similar results? @pavelsavva @Ankit3794

Hmm for me this never happens on Mac, locally, but happens in prod when running in a Linux Docker container

Same issue here. I though its an AWS throttling issue since I was using t2, but even after changing to m5 family (non throttling). I am facing this issue although much less times than before. Only in the prod vm

Have you found any solution?

[GaelCangySpoon]

Currently having the same issue too, was anything found?
It tends to happen when there are a lot of nodes and subgraphs being created because of heavy rocessing.

[shivampkumar]

Any resolutions found? Any version I can roll back to?

[pavelsavva]

Seeing a similar issue

[zbn1120]

I also got this issue with langgraph v0.5.2. Any resolution?

[javierquin]

seeing a similar issue using https://github.com/langchain-ai/open_deep_research

[OnlyBrinTy]

One thing I noticed is that the exact same issue occurs with PydanticAI's OpenAIModel.

[jcelmeta14]

It also happened to me when using it with google gen ai

[alitaghub]

I see the same problem, random freezing. any comment how to fix? version 0.6.6