opentelemetry dependency of pydantic-ai makes AWS lambda crash at import time

[fedexman]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When running pydantic-ai in a AWS lambda (3.12 or 3.13), pydantic ai imports opentelemetry:
from opentelemetry.trace import NoOpTracer, use_span
that has some init of _load_runtime_context at import time
in /opentelemetry/context/__init__.py

this runtime context function double fails 😅

[ERROR] StopIteration
Traceback (most recent call last):
  File "/var/lang/lib/python3.13/importlib/__init__.py", line 88, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 1387, in _gcd_import
  File "<frozen importlib._bootstrap>", line 1360, in _find_and_load
  File "<frozen importlib._bootstrap>", line 1331, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 935, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 1026, in exec_module
  File "<frozen importlib._bootstrap>", line 488, in _call_with_frames_removed
  File "/var/task/handlers/dummy_lambda.py", line 1, in <module>
    import pydantic_ai  # noqa: F401
  File "/opt/python/pydantic_ai/__init__.py", line 3, in <module>
    from .agent import (
  File "/opt/python/pydantic_ai/agent/__init__.py", line 13, in <module>
    from opentelemetry.trace import NoOpTracer, use_span
  File "/opt/python/opentelemetry/trace/__init__.py", line 85, in <module>
    from opentelemetry import context as context_api
  File "/opt/python/opentelemetry/context/__init__.py", line 70, in <module>
    _RUNTIME_CONTEXT = _load_runtime_context()
  File "/opt/python/opentelemetry/context/__init__.py", line 60, in _load_runtime_context
    return next(  # type: ignore

I tried looking into how to fix open-telemetry for lambda, by tinkering the ENV variable or adding opentelemetry lambda layers. I was not able to fix this problem, this bug is not technically from pydantic-ai as it fails in opentelemetry, or maybe a bug of packaging from serverless framework, but I was wondering if this opentelemetry feature is necessary of is it something optional that could be deactivated ?

Example Code

lambda handler

import pydantic_ai  # noqa: F401
from aws_lambda_powertools.event_handler.exceptions import InternalServerError
from aws_lambda_powertools.utilities.typing import LambdaContext


def handler(event: dict, context: LambdaContext):
    LOGGER.info("Received event: %s", event)
    try:
        return {"statusCode": 200, "body": "Hello from Dummy Lambda!"}
    except Exception as err:
        LOGGER.exception("Lambda encountered an error.")
        raise InternalServerError("Lambda encountered an error.") from err

I'm using serverless to deploy the lambda and package the dependencies, serverless config:

  service: test
  frameworkVersion: "3"

  provider:
    name: aws
    runtime: python3.13
    stage: ${opt:stage, 'dev'}
    region: ${opt:region, 'ap-northeast-1'}

  functions:
    dummy-lambda:
      handler: handlers.dummy_lambda.handler
      layers:
        - !Ref PythonRequirementsLambdaLayer

  plugins:
    - serverless-python-requirements

  custom:
    pythonRequirements:
      dockerizePip: true
      layer: true
      slim: true

Python, Pydantic AI & LLM client version

python3.13
pydantic-ai-slim[openai]==1.0.8
(using slim as pydantic-ai is too big for a base lambda)

[gmetzker-4c]

I'm having this issue as well on python 3.12.
I'm using pydantic-ai-slim = {extras = ["bedrock", "openai"], version = "^1.0.10"}

Also using serverless v3, with serverless-python-requirements.
Very similar options on. my serverless config, although not using layers

    dockerizePip: true
    usePoetry: true
    noDeploy: []
    slim: true

[Viicos]

AWS has optimizations with slim: true on package metadata, breaking usage of importlib.metadata. See the related documentation from Pydantic for more details.

[fedexman]

Thanks @Viicos, setting slim: false fixed the issue 🙏 it could be useful to add this also to the pydantic-ai doc, as I didn't find the pydantic doc during my research 🥲

[Viicos]

I think a better option would be to catch the exception and provide a better error message, giving some possible indication on why it happened. But this would have to happen on the OTel SDK side.

[gmetzker-4c]

That worked for me as well.
opentelemetry is looking for an entrypoints.txt in the distinfo folder. slim=true deletes those.

The downside is if I don't use slim=true the package balloons another 30%.
I might try playing with slimPatterns:. Ideally it could slim everything and then white list opentelemetry_api-1.37.0.dist-info/entry_points.txt. Not sure if there is any easy way to do that though.

Another note. You may have to blast your cache to get it to rebuild the packages.

sls requirements clean
sls requirements cleanCache

[gmetzker-4c]

With a little extra glob magic, I was able to get slim to work with a few exceptions. Looks genai_prices and pydantic_ai both use the METADATA in dist-info, and then open telemetry needs entry_points.

    slim: true
    slimPatternsAppendDefaults: false
    slimPatterns:
      - '**/*.py[c|o]'
      - '**/__pycache__*'
      - '**/!(opentelemetry*|genai_prices*|pydantic_ai*).dist-info/**'
      - '**/opentelemetry*.dist-info/!(entry_points.txt)'
      - '**/opentelemetry*.dist-info/**/!(entry_points.txt)'

[alexmojaki]

I think a better option would be to catch the exception and provide a better error message, giving some possible indication on why it happened. But this would have to happen on the OTel SDK side.

No, it could happen when pydantic-ai imports opentelemetry.

Also, this should still be documented.

[Viicos]

@alexmojaki I'm not sure I understand, yes pydantic-ai can catch the exception and display a better message on what's going on, but why not fix this where importlib.metadata is actually used, that is in the Otel SDK?

On a side note, it is very annoying that AWS lambda is doing this kind of space optimization (without really any kind of documentation about this -- at least couldn't find any). importlib.metadata is becoming more and more relied on as part of the CPython's stdlib.

[alexmojaki]

@alexmojaki I'm not sure I understand, yes pydantic-ai can catch the exception and display a better message on what's going on, but why not fix this where importlib.metadata is actually used, that is in the Otel SDK?

That's good too, but an issue should be opened there, and even if that gets resolved there it won't be resolved here until pydantic-ai's only depends on that version or newer.

On a side note, it is very annoying that AWS lambda is doing this kind of space optimization (without really any kind of documentation about this -- at least couldn't find any). importlib.metadata is becoming more and more relied on as part of the CPython's stdlib.

That should be reported to AWS.