Is __init__.py not required for packages in Python 3.3+

Overview

@Mike's answer is correct but too imprecise. It is true that Python 3.3+ supports Implicit Namespace Packages that allows it to create a package without an __init__.py file. This is called a namespace package in contrast to a regular package which does have an __init__.py file (empty or not empty).

However, creating a namespace package should ONLY be done if there is a need for it. For most use cases and developers out there, this doesn't apply so you should stick with EMPTY __init__.py files regardless.

Namespace package use case

To demonstrate the difference between the two types of python packages, lets look at the following example:

google_pubsub/              <- Package 1
    google/                 <- Namespace package (there is no __init__.py)
        cloud/              <- Namespace package (there is no __init__.py)
            pubsub/         <- Regular package (with __init__.py)
                __init__.py <- Required to make the package a regular package
                foo.py

google_storage/             <- Package 2
    google/                 <- Namespace package (there is no __init__.py)
        cloud/              <- Namespace package (there is no __init__.py)
            storage/        <- Regular package (with __init__.py)
                __init__.py <- Required to make the package a regular package
                bar.py

google_pubsub and google_storage are separate packages but they share the same namespace google/cloud. In order to share the same namespace, it is required to make each directory of the common path a namespace package, i.e. google/ and cloud/. This should be the only use case for creating namespace packages, otherwise, there is no need for it.

It's crucial that there are no __init__py files in the google and google/cloud directories so that both directories can be interpreted as namespace packages. In Python 3.3+ any directory on the sys.path with a name that matches the package name being looked for will be recognized as contributing modules and subpackages to that package. As a result, when you import both from google_pubsub and google_storage, the Python interpreter will be able to find them.

This is different from regular packages which are self-contained meaning all parts live in the same directory hierarchy. When importing a package and the Python interpreter encounters a subdirectory on the sys.path with an __init__.py file, then it will create a single directory package containing only modules from that directory, rather than finding all appropriately named subdirectories outside that directory. This is perfectly fine for packages that don't want to share a namespace. I highly recommend taking a look at Traps for the Unwary in Python’s Import System to get a better understanding of how Python importing behaves with regular and namespace package and what __init__.py traps to watch out for.

Summary

  • Only skip __init__.py files if you want to create namespace packages. Only create namespace packages if you have different libraries that reside in different locations and you want them each to contribute a subpackage to the parent package, i.e. the namespace package.
  • Keep on adding empty __init__.py to your directories because 99% of the time you just want to create regular packages. Also, Python tools out there such as mypy and pytest require empty __init__.py files to interpret the code structure accordingly. This can lead to weird errors if not done with care.

Resources

My answer only touches the surface of how regular packages and namespace packages work, so take a look at the following resources for further information:

  • PEP 420 -- Implicit Namespace Packages
  • The import system - Regular packages
  • The import system - Namespace packages
  • Traps for the Unwary in Python’s Import System

Python 3.3+ has Implicit Namespace Packages that allow it to create a packages without an __init__.py file.

Allowing implicit namespace packages means that the requirement to provide an __init__.py file can be dropped completely, and affected ... .

The old way with __init__.py files still works as in Python 2.