Skip to main content
Join the official 2019 Python Developers SurveyStart the survey!

Publication helps you maintain public-api-friendly modules by preventing unintentional access to private implementation details via introspection.

Project description

Setting expectations around what APIs you can rely on in a Python library is very difficult.

Publication makes it easy.

The Problem

As Hyrum’s Law somewhat grimly states,

With a sufficient number of users of an API,
it does not matter what you promise in the contract:
all observable behaviors of your system
will be depended on by somebody.

In general, Python famously has a somewhat different philosophical view of this reality. We assume each other to be responsible users of the libraries we consume. Mucking with implementation details might break every time you upgrade, but it’s sufficiently useful for testing, debugging, and experimentation that retaining that ability is worth paying the cost.

But, critical to this assumption is that everybody knows when they’re breaking into the “private” area of the library’s interface. Here, there’s a mismatch of expectations:

  • library authors write documentation, and then think that users sit down and read the documentation, front to back, and learn about what the “public” interface is by doing so. they then assume that users will know that they’ve used private implementation details if they ever deviate from these documented features.
  • library users pip install a thing, open up a REPL, import the module, and discover the library by doing dir() on the module and its contents, assuming that their program is not using any private implementation details as long as they never had to type library._something_private() while doing so. If they ever encounter a traceback they may consult the documentation, briefly, until it is resolved.

Publication makes it possible to align the wildly divergent expectations of these groups, so that users can still get the benefits of being able to use internal details if they want, but they’ll know that they’re doing so. It makes the runtime namespace of your module look like the public documentation of your library.

How does this look in practice?

You, a prospective library author, want to write a library that makes it easy to zorf a sprocket. Great! You do, and it looks like this:

# sprocket_zorfer.py
from sprocket import sprocket_with_name
from zorf import zorfable_thing

def zorf_sprocket_internal(sprocket, zorfulations):
    ...
def compute_zorfulations():
    ...

def zorf_sprocket_named(sprocket_name, how_much):
    sprocket = sprocket_with_name(sprocket_name)
    zorfulations = compute_zorfulations(how_much)
    return zorf_sprocket_internal(sprocket, zorfulations)


__all__ = [
    'zorf_sprocket_named'
]

Your intent here, of course, is that you have exposed a module with a single function: zorf_sprocket_named, and everything else is an implementation detail. You even said so, explicitly, with __all__. Your API documentation says the same.

However, reading reference documentation and cleanly respecting conventions is not how working programmers really figure out how to use stuff. Your users all do stuff like:

  • Load up an interactive python interpreter and call dir() on your module
  • Install Jupyter and tab-complete their way around your module to find what they want
  • use the auto-import function in PyCharm to grab some private implementation detail

and, before you know it, you have thousands of users of your library with code like

from sprocket_zorfer import compute_zorfulations, zorf_sprocket_internal, sprocket_with_name

sprocket = sprocket_with_name(name)
zorf_sprocket_internal(sprocket, compute_zorfulations(7) * 2)

Now you can never change any of your implementation details! Worse yet, sprocket_with_name isn’t even your own code; that’s something you got from a library! But when someone does import sprocket_zorfer; sprocket_zorfer.<tab> in an interactive shell, none of that information comes through.

Underscore Paranoia

The convention in Python is that we use _ to indicate private names. So when we library authors notice this problem starting to happen, a common reaction is to start putting _ in front of everything – class names, function names, module names – and only explicitly export those things that should be public by “moving” them via an import and an entry in a public module’s __all__.

However, this has a bunch of disadvantages:

  • Most code inspection tooling and IDEs won’t see that the public name is “moved”, so code exploration just makes it seem like everything is an implementation detail now, rather than making it seem like nothing is.
  • All your __repr__s now have ugly and inaccurate function and class names in them, at least from the perspective of your users; how are they supposed to know zorf_sprocket_named is actually defined in zorf_sprocket._impl_details.funcs._zorf_sprocket_public now? How are they supposed to find the good, public name once they’re looking at the goofy internal one?
  • You constantly need to remember to put all of your code in these ugly _-prefixed modules, and educate new contributors as to the risks of creating new modules in your package that are not carefully hidden away from public users.

A Better World

What if you could write all your code as if it were just regular public code, and have all your implementation details and imports automatically squirreled away in an underscore namespace so that curious coders won’t accidentally find every module you ever imported and every temporary helper function you ever defined and think they’re part of the permanent public face of your library?

Enter publication.

publication uses the existing convention of __all__ and a little runtime hackery to hide everything that you have not marked as explicitly public, like so:

# sprocket_zorfer.py

from publication import publish

from sprocket import sprocket_with_name
from zorf import zorfable_thing

def zorf_sprocket_internal(sprocket, zorfulations):
    ...
def compute_zorfulations():
    ...

def zorf_sprocket_named(sprocket_name, how_much):
    sprocket = sprocket_with_name(sprocket_name)
    zorfulations = compute_zorfulations(how_much)
    return zorf_sprocket_internal(sprocket, zorfulations)


__all__ = [
    'zorf_sprocket_named'
]

publish()

That’s it! Now, from sprocket_zorfer import zorf_sprocket_named works as intended, but from sprocket_zorfer import compute_zorfulations is an ImportError.

But what about…

Other modules in my package, like tests, that need to peek at implementation details?

Don’t worry, your code didn’t go anywhere. The original module is still available as a special pseudo-module called <your_module>._private. In the example above, sprocket_zorfer.py’s tests can still do:

from sprocket_zorfer._private import compute_zorfulations

def test_compute_zorfulations():
    assert compute_zorfulations(0) > 7

Mypy?

Your types should probably just be part of your published API, if you’re expecting that users will need to know about them. But, if there are cases which need to be type-checked internally in your library, as far as Mypy is concerned, all your private classes are still there. So, in the simple case you can just do this:

from typing import TYPE_CHECKING
if TYPE_CHECKING:
    from something import T

def returns_a() -> "T":
    ...

and in the hopefully very unusual case you need to mix runtime and type-checking access to a different module’s private details,

from typing import TYPE_CHECKING
if TYPE_CHECKING:
    from something import T
else:
    from something._private import T

def returns_a() -> A:
    ...

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for publication, version 0.0.3
Filename, size File type Python version Upload date Hashes
Filename, size publication-0.0.3-py2.py3-none-any.whl (7.7 kB) File type Wheel Python version py2.py3 Upload date Hashes View hashes
Filename, size publication-0.0.3.tar.gz (5.5 kB) File type Source Python version None Upload date Hashes View hashes

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN SignalFx SignalFx Supporter DigiCert DigiCert EV certificate StatusPage StatusPage Status page