Thread Local Context


You should call some functions with some arguments.

—David Reid

The behavior of copying itself, adding new values, and returning the result is useful for applications that keep somehow their own context using classes or closures. Twisted is a fine example for that. Another possible approach is passing wrapped loggers around or log only within your view where you gather errors and events using return codes and exceptions. If you are willing to do that, you should stick to it because immutable state is a very good thing[*]. Sooner or later, global state and mutable data lead to unpleasant surprises.

However, in the case of conventional web development, we realize that passing loggers around seems rather cumbersome, intrusive, and generally against the mainstream culture. And since it’s more important that people actually use structlog than to be pure and snobby, structlog contains a dirty but convenient trick: thread local context storage which you may already know from Flask:

Thread local storage makes your logger’s context global but only within the current thread[†]. In the case of web frameworks this usually means that your context becomes global to the current request.

The following explanations may sound a bit confusing at first but the Flask example illustrates how simple and elegant this works in practice.

Wrapped Dicts

In order to make your context thread local, structlog ships with a function that can wrap any dict-like class to make it usable for thread local storage: structlog.threadlocal.wrap_dict().

Within one thread, every instance of the returned class will have a common instance of the wrapped dict-like class:

>>> from structlog.threadlocal import wrap_dict
>>> WrappedDictClass = wrap_dict(dict)
>>> d1 = WrappedDictClass({"a": 1})
>>> d2 = WrappedDictClass({"b": 2})
>>> d3 = WrappedDictClass()
>>> d3["c"] = 3
>>> d1 is d3
>>> d1 == d2 == d3 == WrappedDictClass()
>>> d3  
<WrappedDict-...({'a': 1, 'b': 2, 'c': 3})>

To enable thread local context use the generated class as the context class:



Creation of a new BoundLogger initializes the logger’s context as context_class(initial_values), and then adds any values passed via .bind(). As all instances of a wrapped dict-like class share the same data, in the case above, the new logger’s context will contain all previously bound values in addition to the new ones.

structlog.threadlocal.wrap_dict() returns always a completely new wrapped class:

>>> from structlog.threadlocal import wrap_dict
>>> WrappedDictClass = wrap_dict(dict)
>>> AnotherWrappedDictClass = wrap_dict(dict)
>>> WrappedDictClass() != AnotherWrappedDictClass()
>>> WrappedDictClass.__name__  
>>> AnotherWrappedDictClass.__name__   

In order to be able to bind values temporarily to a logger, structlog.threadlocal comes with a context manager: tmp_bind():

>>> log.bind(x=42)  
<BoundLogger(context=<WrappedDict-...({'x': 42})>, ...)>
>>> log.msg("event!")
x=42 event='event!'
>>> with tmp_bind(log, x=23, y="foo") as tmp_log:
...     tmp_log.msg("another event!")
x=23 y='foo' event='another event!'
>>> log.msg("one last event!")
x=42 event='one last event!'

The state before the with statement is saved and restored once it’s left.

If you want to detach a logger from thread local data, there’s structlog.threadlocal.as_immutable().

Downsides & Caveats

The convenience of having a thread local context comes at a price though:


  • If you can’t rule out that your application re-uses threads, you must remember to initialize your thread local context at the start of each request using new() (instead of bind()). Otherwise you may start a new request with the context still filled with data from the request before.

  • Don’t stop assigning the results of your bind()s and new()s!


    log =
    log = log.bind(x=42)


    Although the state is saved in a global data structure, you still need the global wrapped logger produce a real bound logger. Otherwise each log call will result in an instantiation of a temporary BoundLogger. See Configuration for more details.

The general sentiment against thread locals is that they’re hard to test. In this case we feel like this is an acceptable trade-off. You can easily write deterministic tests using a call-capturing processor if you use the API properly (cf. warning above).

This big red box is also what separates immutable local from mutable global data.

[*]In the spirit of Python’s ‘consenting adults’, structlog doesn’t enforce the immutability with technical means. However, if you don’t meddle with undocumented data, the objects can be safely considered immutable.
[†]Special care has been taken to detect and support greenlets properly.