🐛 Bug Report: TypeError: cannot pickle '_thread.RLock' object during on_tool_start in langchain instrumentor

Which component is this bug for?

Langchain Instrumentation

📜 Description

Hello first of all thank you for the work on the langchain instrumentor ! I have been using it a lot and it is very useful !

I have an issue with the LangchainInstrumentor in certain conditions:
It looks to me that if the langchain chat model can get a Lock attribute if used in some ways. When on_tool_start of the LangchainInstrumentor is triggered, it display an error: TypeError: cannot pickle '_thread.RLock' object

👟 Reproduction steps

Here is a minimal example to reproduce the bug:

import logging
import operator
import os
from typing import Annotated, Any, List

from langchain.agents import AgentState, create_agent
from langchain.messages import HumanMessage
from langchain.tools import ToolRuntime, tool
from langchain_openai import AzureChatOpenAI
from opentelemetry.instrumentation.langchain import LangchainInstrumentor
from pydantic import BaseModel, Field

logging.basicConfig(level=logging.INFO, force=True)

async def test_func():
    class State(AgentState):
        relevant_documents: Annotated[List[dict[str, Any]], operator.add]
        llm: AzureChatOpenAI

    LangchainInstrumentor().instrument()
    llm = AzureChatOpenAI(
            model="gpt-4.1",
            api_version=os.environ["AZURE_API_VERSION"],
            azure_endpoint=os.environ["AZURE_FOUNDRY_AI_ENDPOINT"],
            api_key=os.environ["FOUNDRY_AI_API_KEY"],
        )

    class CalendarEventsInput(BaseModel):
        calendar_options: str = Field(
            description="Options for filtering calendar events.",
        )

    @tool(args_schema=CalendarEventsInput)
    def async_get_calendar(runtime: ToolRuntime, calendar_options: str):
        """Use this tool to search for calendar events, """
        logging.info("tool called")
        return "asyncc"

    agent_executor = create_agent(
            model=llm,
            tools=[async_get_calendar],
            system_prompt="call async calendar",
            state_schema=State,
        )
    async for event in agent_executor.astream(
        input={
            "messages": HumanMessage(content="call async calendar for event in last 7 days"),
            "relevant_documents": [],
            "llm": llm,
        },
        stream_mode=["debug"]
    ):
        logging.info(event)
await test_func()

I get the error:

Traceback (most recent call last):
  File "/home/pvezia/lynxai/LynxeoAI%20API/.venv/lib/python3.12/site-packages/openai/_base_client.py", line 811, in __del__
    if self.is_closed:
       ^^^^^^^^^^^^^^
  File "/home/pvezia/lynxai/LynxeoAI%20API/.venv/lib/python3.12/site-packages/httpx/_client.py", line 228, in is_closed
    return self._state == ClientState.CLOSED
           ^^^^^^^^^^^
AttributeError: 'SyncHttpxClientWrapper' object has no attribute '_state'

When I look at debug logs of the instrumentor, I see:

INFO:opentelemetry.instrumentation.langchain.callback_handler:OpenLLMetry failed to trace in on_tool_start, error: Traceback (most recent call last):
  File "/home/pvezia/lynxai/LynxeoAI%20API/.venv/lib/python3.12/site-packages/opentelemetry/instrumentation/langchain/utils.py", line 68, in wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^  
File "/home/pvezia/lynxai/LynxeoAI%20API/.venv/lib/python3.12/site-packages/opentelemetry/instrumentation/langchain/callback_handler.py", line 680, in on_tool_start
    json.dumps(
  [....]
  File "/home/pvezia/lynxai/LynxeoAI%20API/.venv/lib/python3.12/site-packages/opentelemetry/instrumentation/langchain/utils.py", line 30, in default
    return dataclasses.asdict(o)
           ^^^^^^^^^^^^^^^^^^^^^
  [...]
  File "/usr/lib/python3.12/copy.py", line 151, in deepcopy
    rv = reductor(4)
         ^^^^^^^^^^^
TypeError: cannot pickle '_thread.RLock' object

👍 Expected behavior

It should not display an error when tracing the model

👎 Actual Behavior with Screenshots

,

🤖 Python Version

3.12

📃 Provide any additional context for the Bug.

When we look at the CallbackFilteredJSONEncoder function, we can see that for dataclasses:

if dataclasses.is_dataclass(o):
     return dataclasses.asdict(o)

The issue is that asdict tries to make a deepcopy of the attributes of the dataclass, which cause an error if one of the attributes is a Lock (you can't make a deepcopy of a Lock)

Maybe a good workaround would be to replace those lines by

if dataclasses.is_dataclass(o):
    return {f.name: self.default(getattr(o, f.name)) for f in dataclasses.fields(o)}

It would be a safer way to get the dict from the dataclass in the cases where the dataclass has some attributes that can't be deepcopied

Maybe I am missing something ? Do you have any ideas ?

Thank you !

👀 Have you spent some time to check if this bug has been raised before?

• I checked and didn't find similar issue

Are you willing to submit PR?

Yes I am willing to submit a PR!

Start in packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/utils.py at CallbackFilteredJSONEncoder.default, then inspect the on_tool_start path in callback_handler.py. Run the provided reproduction and verify that tracing dataclasses containing a lock no longer emits the pickle error while preserving JSON serialization.