Troubleshoot trace nesting

When tracing with the LangSmith SDK, LangGraph, and LangChain, tracing should automatically propagate the correct context so that code executed within a parent trace will be rendered in the expected location in the UI.

If you see a child run go to a separate trace (and appear on the top level), it may be caused by one of the following known "edge cases".

Python

The following outlines common causes for "split" traces when building with python.

Context propagation using asyncio

When using async calls (especially with streaming) in Python versions < 3.11, you may encounter issues with trace nesting. This is because Python's asyncio only added full support for passing context in version 3.11.

Why

LangChain and LangSmith SDK use contextvars to propagate tracing information implicitly. In Python 3.11 and above, this works seamlessly. However, in earlier versions (3.8, 3.9, 3.10), asyncio tasks lack proper contextvar support, which can lead to disconnected traces.

To resolve

1. Upgrade Python Version (Recommended) If possible, upgrade to Python 3.11 or later for automatic context propagation.

2. Manual Context Propagation If upgrading isn't an option, you'll need to manually propagate the tracing context. The method varies depending on your setup:

   a) Using LangGraph or LangChain Pass the parent config to the child call:

   import asyncio
   from langchain_core.runnables import RunnableConfig, RunnableLambda
   
   @RunnableLambda
   async def my_child_runnable(
                 inputs: str,
                 # The config arg (present in parent_runnable below) is optional
                 ):
       yield "A"
       yield "response"
   
   @RunnableLambda
   async def parent_runnable(inputs: str, config: RunnableConfig):
       async for chunk in my_child_runnable.astream(inputs, config):
           yield chunk
   
   async def main():
       return [val async for val in parent_runnable.astream("call")]
   
   asyncio.run(main())

   b) Using LangSmith Directly Pass the run tree directly:

   import asyncio
   import langsmith as ls
   
   @ls.traceable
   async def my_child_function(inputs: str):
       yield "A"
       yield "response"
   
   @ls.traceable
   async def parent_function(
                 inputs: str,
                 # The run tree can be auto-populated by the decorator
                 run_tree: ls.RunTree,
             ):
       async for chunk in my_child_function(inputs, langsmith_extra={"parent": run_tree}):
           yield chunk
   
   async def main():
       return [val async for val in parent_function("call")]
   
   asyncio.run(main())

   c) Combining Decorated Code with LangGraph/LangChain Use a combination of techniques for manual handoff:

   import asyncio
   import langsmith as ls
   from langchain_core.runnables import RunnableConfig, RunnableLambda
   
   @RunnableLambda
   async def my_child_runnable(inputs: str):
       yield "A"
       yield "response"
   
   @ls.traceable
   async def my_child_function(inputs: str, run_tree: ls.RunTree):
       with ls.tracing_context(parent=run_tree):
           async for chunk in my_child_runnable.astream(inputs):
               yield chunk
   
   @RunnableLambda
   async def parent_runnable(inputs: str, config: RunnableConfig):
       # @traceable decorated functions can directly accept a RunnableConfig when passed in via "config"
       async for chunk in my_child_function(inputs, langsmith_extra={"config": config}):
           yield chunk
   
   @ls.traceable
   async def parent_function(inputs: str, run_tree: ls.RunTree):
       # You can set the tracing context manually
       with ls.tracing_context(parent=run_tree):
           async for chunk in parent_runnable.astream(inputs):
               yield chunk
   
   async def main():
       return [val async for val in parent_function("call")]
   
   asyncio.run(main())

Context propagation using threading

It's common to start tracing and want to apply some parallelism on child tasks all within a single trace. Python's stdlib ThreadPoolExecutor by default breaks tracing.

Why

Python's contextvars start empty within new threads. Here are two approaches to handle maintain trace contiguity:

To resolve

1. Using LangSmith's ContextThreadPoolExecutor

   LangSmith provides a ContextThreadPoolExecutor that automatically handles context propagation:

   from langsmith.utils import ContextThreadPoolExecutor
   from langsmith import traceable
   
   @traceable
   def outer_func():
       with ContextThreadPoolExecutor() as executor:
           inputs = [1, 2]
           r = list(executor.map(inner_func, inputs))
   
   @traceable
   def inner_func(x):
       print(x)
   
   outer_func()

2. Manually providing the parent run tree

   Alternatively, you can manually pass the parent run tree to the inner function:

   from langsmith import traceable, get_current_run_tree
   from concurrent.futures import ThreadPoolExecutor
   
   @traceable
   def outer_func():
       rt = get_current_run_tree()
       with ThreadPoolExecutor() as executor:
           r = list(
               executor.map(
                   lambda x: inner_func(x, langsmith_extra={"parent": rt}), [1, 2]
               )
           )
   
   @traceable
   def inner_func(x):
       print(x)
   
   outer_func()

   In this approach, we use get_current_run_tree() to obtain the current run tree and pass it to the inner function using the langsmith_extra parameter.

Both methods ensure that the inner function calls are correctly aggregated under the initial trace stack, even when executed in separate threads.