Troubleshooting

We expect most users of the Python SDK not to run into any of the problems documented here.

Use the information in this page to help answer these questions:

  • "What do I do if scope data is leaking between requests?"
  • "What do I do if my transaction has nested spans when they should be parallel?"

The short answer to those: check if your contextvars work and you clone the hub where needed.

Addressing Concurrency Issues

Python supports several distinct solutions to concurrency, including threads and coroutines.

The Python SDK does its best to figure out how contextual data such as tags set with sentry_sdk.set_tags is supposed to flow along your control flow. In most cases it works perfectly, but in a few situations some special care must be taken. This is specially true when working with a code base doing concurrency outside of the provided framework integrations.

The general recommendation is to have one hub per "concurrency unit" (thread/coroutine/etc). The SDK ensures every thread has an independent hub. If you do concurrency with asyncio coroutines, clone the current hub for use within a block that runs concurrent code:

Copied
with Hub(Hub.current):
    # in this block Hub.current refers to a new clone
    # of the original hub, with the same client and
    # the same initial scope data.

Issues with asyncio have then an easy workaround: every asyncio coroutine that really does run concurrently with other coroutines needs to be made into a task, then the hub needs to be cloned and reassigned.

See the Threading section for a more complete example that involves cloning the current hub.

Context Variables vs Thread Locals

The Python SDK uses thread locals to keep contextual data where it belongs. There are a few situations where this approach fails.

Read along if you cannot figure out why contextual data is leaking across HTTP requests, or data is missing or popping up at the wrong place and time.

Python 2: Thread Locals and gevent

If the SDK is installed on Python 2, there is not much else to use than the aforementioned thread locals, so the SDK will use just that.

Code that uses async libraries such as twisted is not supported in the sense that you will experience context data leaking across tasks/any logical boundaries, at least out of the box.

Code that uses more "magical" async libraries such as gevent or eventlet will work just fine provided those libraries are configured to monkeypatch the stdlib. If you are only using those libraries in the context of running gunicorn that is the case, for example.

Python 3: Context Variables or Thread Locals

Python 3 introduced asyncio, which, just like Twisted, had the problem of not having any concept of attaching contextual data to your control flow. That means in Python 3.6 and lower, the SDK is not able to prevent leaks of contextual data.

Python 3.7 rectified this problem with the contextvars stdlib module which is basically thread locals that also work in asyncio-based code. The SDK will attempt to use that module instead of thread locals if available.

For Python 3.6 and older, install aiocontextvars from PyPI which is a fully-functional backport of contextvars. The SDK will check for this package and use it instead of thread locals.

Context Variables vs gevent/eventlet

If you are using gevent (older than 20.5) or eventlet in your application and have configured it to monkeypatch the stdlib, the SDK will abstain from using contextvars even if it is available.

The reason for that is that both of those libraries will monkeypatch the threading module only, and not the contextvars module.

The real-world usecase where this actually comes up is if you're using Django 3.0 within a gunicorn+gevent worker on Python 3.7. In such a scenario the monkeypatched threading module will honor the control flow of a gunicorn worker while the unpatched contextvars will not.

It gets more complicated if you're using Django Channels in the same app, but a separate server process, as this is a legitimate usage of asyncio for which contextvars behaves more correctly. Make sure that your channels websocket server does not import or use gevent at all (and much less call gevent.monkey.patch_all), and you should be good.

Even then there are still edge cases where this behavior is flat-out broken, such as mixing asyncio code with gevent/eventlet-based code. In that case there is no right, static answer as to which context library to use. Even then gevent's aggressive monkeypatching is likely to interfere in a way that cannot be fixed from within the SDK.

This issue has been fixed with gevent 20.5 but continues to be one for eventlet.