Correct logging setup for LangGraph - Getting I/O errors

Product Help LangGraph
1

What’s the correct way to set up logging in LangGraph? I’m getting frequent I/O errors when the application tries to log during async operations:

Traceback (most recent call last):
  File "/usr/local/lib/python3.11/logging/__init__.py", line 1113, in emit
    stream.write(msg + self.terminator)
ValueError: I/O operation on closed file.
...
  File "/usr/local/lib/python3.11/site-packages/httpx/_client.py", line 1740, in _send_single_request
    logger.info(
Message: 'HTTP Request: %s %s "%s %d %s"'
Arguments: ('POST', URL('https://api.openai.com/v1/chat/completions'), 'HTTP/1.1', 200, 'OK')

This happens when httpx tries to log OpenAI API responses within LangGraph workflows.

What’s the recommended approach for configuring loggers in LangGraph to avoid these issues?

2

I’m assuming you’re talking about langgraph platform? Since langgraph is just an OSS lib.

If the former, you can use stdlib or structlog regularly:

import logging

logger = logging.getLogger(__name__)
logger.info(“hi”)

or with structlog

import structlog
logger = structlog.stdlib.get_logger(__name__)
logger.info(“hi”, some_key=some_val)

3

I’m using both the LangGraph platform and local development. I’ve implemented the logger as suggested with an additional SafeStreamHandler to handle stream issues, but the error persists:

python

import logging
import sys

class SafeStreamHandler(logging.StreamHandler):
    def emit(self, record):
        try:
            super().emit(record)
        except (ValueError, OSError):
            pass

logger = logging.getLogger(__name__)
logger.handlers.clear()
handler = SafeStreamHandler(sys.stdout)
logger.addHandler(handler)

Any other suggestions for resolving this logging issue?

4

Interesting discussion. I’ve been experimenting with a slightly different angle on logging in LangGraph.

Instead of trying to capture everything (which can get noisy and brittle), I’ve been testing “collapse-point logging” — only recording when reasoning fails (contradictions, skipped steps, unknowns).

This produces a compact audit trail and reduces I/O churn, since you’re not writing every event — just the meaningful failure points.

Has anyone here tried a similar “failure-only” logging pattern with LangGraph?