Why are Python Enums slow?

Question

Today I tried to outsource some of my "final" values (members of my class) from a class to an Enum. I ran the unittests I wrote and noticed, that the tests needed much longer to complete than before. When I put these "final" values back into the class, everything was back to old speed again. Here an example of how I accessed them when they were in the Enum:

class SpecialCharacters(Enum):

    TONE_NUMBERS = ["0", "1", "2", "3", "4"]

Accessing it like this:

SpecialCharacters.TONE_NUMBERS.value

Example of accessing the values when they are in the class:

self.TONE_NUMBERS

So I wonder why my tests take 3x (!) as much time when I put the values in the Enum. It should be a simple call to another class' members, but that wouldn't make such a big difference I guess.

(The goal of outsourcing them was that they're available for other classes and those other classes don't have to access a class with which they have nothing to do.)

My python version is 3.4.

Answer 1

This is a known bug in Python 3.4's enum: https://bugs.python.org/issue23486

It's been "fixed" in Python 3.5, such that enum attribute lookup is only 3x slower than normal, instead of 20x.

Answer 2

To answer the question Why?:

The original design for Enum was to have members be virtual attributes. This means they were not kept in the Enum class dictionary, which meant that every lookup possible had to first fail, after which __getattr__ would be called on the class and the member found.

So basically lots of busy work.

The fix was to put the members in the class dictionary when possible (which is most of the time).