Armin Ronacher's Thoughts and Writings

Revenge of the Types

written on Sunday, August 24, 2014

This is part two about "The Python I Would Like To See" and it explores a bit the type system of Python in light of recent discussions. Some of this references the earlier post about slots. Like the earlier post this is a bit of a diving into the CPython interpreter, the Python language with some food for thoughts for future language designers.

As a Python programmer, types are a bit suspicious to you. They clearly exist and they interact in different ways with each other, but for the most part you really only notice their existence when you fail and an exception tells you that a type does not behave like you think it does.

Python was very proud of its approach to typing. I remember reading the language FAQ many years ago and it had a section about how cool duck typing is. To be fair: in practical terms duck typing is a good solution. Because there is basically no type system that fights against you, you are unrestricted in what you can do, which allows you to implement very nice APIs. Especially the common things are super simple in Python.

Almost all the APIs I designed for Python do not work in other languages. Even such simple things such as click's general interface just does not work in other languages. The largest reason for that is that you constantly fight against types.

Recently there have been discussions about adding static typing to Python and I wholeheartedly believe that train long left the station and will never come back. So for the interested, here my thoughts on why I hope Python will not adapt explicit typing.

What's a Type System?

A type system are the rules of how types interact with each other. There is actually a whole part of computer science that seems to be exclusively concerned with types which is pretty impressive by itself. But even if you are not particularly interested in theoretical computer science, type systems are hard to ignore.

I don't want to go too much into type systems for two reasons. The first one is that I barely understand them at all myself. The second one is that they are really not all that important to understand in order to "feel" the consequences of them. For me the way types behave is important because it influences how APIs are designed. So consider this basic introduction more influenced by my obsession with nice APIs than with a correct introduction to types.

Type systems can have many properties but the most important one that sets them all apart is the amount of information they provide when you try to reason with them.

As an example you can take Python. Python has types. There is the number 42 and when you ask the number what type it is, it will reply that it is an integer type. That makes a lot of sense and it allows the interpreter to define the rules of how integers interact with other integers.

However there is one thing Python does not have, and that is composite types. All Python types are primitive. That means that you basically can only work with one of them at the time. The opposite of that is composite types. You do see them in Python every once in a while in other contexts.

The most straightforward composite type that most programming languages have are structs. Python does not have them directly, but there are many situations where libraries need to define their own structs. For instance a Django or SQLAlchemy ORM model is essentially a struct. Each database column is represented through a Python descriptor which in this case, corresponds directly to a field in a struct. So when you say the primary key is called id and it's an IntegerField() you are defining your model as composite type.

Composite types are not limited to structs. When you want to work with more than one integer for instance you would use a collection like an array. In Python you have lists and each item in the list can be of an arbitrary type. This is in contrast with lists defined to be specific to a type (like list of integer).

"List of integer" always says more than list. While you could argue that by iterating over the list you can figure out of which type it is, the empty list causes a problem. When you are given a list in Python without elements in it, you cannot know the type.

The exact same problem is caused by the null reference (None) in Python. When you pass a user to a function and that user might become None you all the sudden do not know that it could be a user object.

So what's the solution? Not having null references and having explicitly typed arrays. Haskell obviously is the language that everybody knows that does this, but there are others which look less hostile. For instance Rust is a language that looks much more like C++ and as such more familiar but brings a very powerful type system to the table.

So how do you express "no user present" if there are no null references? The answer in Rust for instance are option types. Option<User> means there is either Some(user) or None. The former is a tagged enum which wraps a value (a specific user). Because now the variable can be either some value or nothing, all code that deals with it needs to explicitly handle the None case of it will not even compile.

The Future is Gray

In the past the world was very clearly divided between interpreted languages with dynamic typing and ahead of time compiled languages with static typing. This is changing as new trends emerge.

The first indication that we're moving into some unexplored territory was C#. It's a statically compiled language and when it started it was very similar to Java in how the language operated. As the language was improved many new type system related features landed. The most important was the introduction of generics which allowed non compiler provided collections like lists and dictionaries to be strongly typed. After that they also went into the opposite direction of allowing sections of code to opt out of static typing on a variable by variable basis. This is ridiculously useful, especially on the context of working with data provided by webservices (JSON, XML etc.) where you just do some potentially unsafe processing and catch down any type system related exceptions to inform the user about bad input data.

Today C#'s type system is very powerful supporing generics with covariance and contravariance specifications. Not only that, it also grew a lot of language level support to deal with nullable types. For instance the null-coalescing operator (??) was introduced to provide default values for objects represented as null. While C# already went down too far to get rid of null they are controlling the damage that can be done.

At the same time other languages that are traditionally ahead of time compiled and statically typed also explore new areas. While C++ will always be statically typed, it started to explore with type inference on many levels. The days of MyType<X, Y>::const_iterator iter are gone. Today you can in almost all situations replace the type with a mere auto and the compiler will fill in the type for you.

Rust as a language has also excellent support for type inference which lets you write statically typed programs that are entirely void of any type declarations:

use std::collections::HashMap;

fn main() {
    let mut m = HashMap::new();
    m.insert("foo", vec!["some", "tags", "here"]);
    m.insert("bar", vec!["more", "here"]);

    for (key, values) in m.iter() {
        println!("{} = {}", key, values.connect("; "));
    }
}

I believe we're moving in a future with powerful type systems. I do not believe that this will be the end of dynamic typing but there appears to be a noticable trend of embracing powerful static typing with local type inference.

Python and Explicit Typing

So not long ago someone apparently convinced someone else at a conference that static typing is awesome and should be a language feature. I'm not exactly sure how that discussion went but the end result was that mypy's type module in combination with Python 3's annotation syntax were declared to be the gold standard of typing in Python.

In case you have not seen the proposal yet, it advocates something like this:

from typing import List

def print_all_usernames(users: List[User]) -> None:
    for user in users:
        print(user.username)

I honestly believe that this is not exactly a good decision for many reasons, the largest being that Python is already suffering having a not exactly good type system. The language actually has different semantics depending on how you look at it.

For static typing to make sense the type system needs to be good. A type system where you take two types and you can figure out how they relate to each other. Python doesn't have that.

Python's Type Semantics

If you have read the previous post about the slot system you might remember that Python has different semantics depending on if a type is implemented in C or in Python. This is a very unique feature of the language and is usually not found in many other places. While it is true that many languages for bootstrapping purposes have types implemented on the interpreter level, they are typically fundamental types and as such special cased.

In Python there are no real "fundamental" types. There are however a whole bunch of types that are implemented in C. These are not at all limited to primitives and fundamental types, they can appear everywhere and without any logic. For instance collections.OrderedDict is a type implemented in Python whereas collections.defaultdict from the same module is implemented in C.

This is actually causing quite a few problems for PyPy which has to emulate the original types as good as possible to achieve a similar enough API that these differences are not noticeable. It is very important to understand what this general difference between C level interpreter code and the rest of the language means.

As an example I want to point out the re module up to Python 2.7. (This behavior has ultimately been changed in the re module, but the general problem of the interpreter working different than the language are still present.)

The re module provides a function (compile) to compile a regular expression into a regular expression pattern. It takes a string and returns a pattern object. Looks roughly like this:

>>> re.compile('foobar')
<_sre.SRE_Pattern object at 0x1089926b8>

As you can see this pattern object comes from the _sre module which is a bit internal but generally available:

>>> type(re.compile('foobar'))
<type '_sre.SRE_Pattern'>

Unfortunately it's a bit of a lie, because the _sre module does not actually contain that type:

>>> import _sre
>>> _sre.SRE_Pattern
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'module' object has no attribute 'SRE_Pattern'
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'module' object has no attribute 'SRE_Pattern'

Alright, fair enough, would not be the first time that a type lied about its location and it's an internal type anyways. So moving on. We know the type of the pattern, it's an _sre.SRE_Pattern type. As such a subclass of object:

>>> isinstance(re.compile(''), object)
True

And all objects implement some very common methods as we know. For instance all objects implement __repr__:

>>> re.compile('').__repr__()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: __repr__
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: __repr__

Oh. What happened here? Well, the answer is pretty bizarre. Internally the SRE pattern object for reasons unknown to me, until Python 2.7, had a custom tp_getattr slot. In this slot there was a custom attribute lookup which provided access to some custom methods and attributes. When you actually inspect the object with dir() you will notice that lots of stuff is missing:

>>> dir(re.compile(''))
['__copy__', '__deepcopy__', 'findall', 'finditer', 'match',
 'scanner', 'search', 'split', 'sub', 'subn']

In fact, this leads you down to a really bizarre adventure of how this type actually functions. Here is what's happening:

Type type claims that it's a subclass of object. This is true for the CPython interpreter world, but not true for Python the language. That these are not the same things is disappointing but generally the case. The type does not corresponds to the interface of object on the Python layer. Every call that goes through the interpreter works, every call that goes through the Python language fails. So type(x) succeeds, whereas x.__class__ fails.

What's a Subclass

The above example shows that you can have a class in Python that is a subclass of another thing, that disagrees with the behavior of the baseclass. This is especially a problem if you talk about static typing. In Python 3 for instance you cannot implement the interface of the dict type unless you write the type in C. The reason for this is that the type guarantees a certain behavior of the view objects that just simply cannot be implemented. It's impossible.

So when you would statically annotate that the function takes a dictionary with string keys and integer objects, it would not be clear at all if it takes a dict, a dict like object or if it would permit a dictionary subclass.

Undefined Behavior

The bizarre behavior of the pattern objects was changed in Python 2.7, but the core issue remains. As mentioned with the behavior of dicts for instance, the language has different behavior depending on how the code was written and the exact semantics of the type system are completely impossible to understand.

A super bizarre case of these interpreter internals are for instance type comparisons in Python 2. This particular case does not exist like that on Python 3 because the interfaces were changed, but the fundamental problem can be found on many levels.

Let's take sorting of sets as an example. Sets in Python are useful types, but they have very bizarre comparison behavior. In Python 2 we have this function called cmp() which given two types will return a numeric value that indicates which side is larger. A return value smaller than zero means that the first argument is smaller than the second, a return value of zero means that they are equal and any positive number means the second value is larger than the first.

Here is what happens if you compare sets:

>>> cmp(set(), set())
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: cannot compare sets using cmp()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: cannot compare sets using cmp()

Why is that? Not exactly sure to be honest. Probably because of how the comparison operators are actually set subsets and they could not make that work with cmp(). However for instance frozensets compare just fine:

>>> cmp(frozenset(), frozenset())
0

Except when one of the sets is not empty it will fail. Why? The answer to this is that this is not a language feature, but an optimization in the CPython interpreter. A frozenset interns common values. The empty frozenset is always the same value (as it is immutable and you cannot add to it), so any empty frozenset is the same object. When two objects have the same pointer address, then cmp will generally return 0. Why exactly I could not figure out quickly due to how complex the comparison logic in Python 2 is, but there are multiple code paths in the comparison routines which might produce this result.

The point is not so much that there is a bug, but that Python does not actually have proper semantics for how types interact with each other. Instead the type system's behavior for a really long time has been "whatever CPython does".

You can find countless of changesets in PyPy where they tried to reconstruct behavior in CPython. Given that PyPy is written in Python, it becomes quite an interesting problem for the language. If the Python language was defined purely like the actual Python part of the language is, PyPy would have a lot less problems.

Instance Level Behavior

Now let's assume there would be a hypothetical version of Python that fixes all of the problems mentioned, static types would still not be something that would fit into Python well. A big reason for this is that on the Python language level, types traditionally had very little meaning in regards to how objects interact.

For instance datetime objects are generally comparable with other things, but datetime objects are only comparable to other datetime objects if their timezone awareness is compatible. Similarly the result of many operations is not clear until you look at the object at hand. Adding two strings together in Python 2 can either construct a unicode or a bytestring object. APIs like decoding or encoding from the codecs system can return any object.

Python as a language is too dynamic for annotations to work well. Just consider how important generators are for the language, yet generators could perform different type conversions on every single iteration.

Type annotations would be spotty at best but they might even have negative impact on API design. At the very least they will make things slower unless they are removed at runtime. They could never implement a language that compiles efficiently statically without making Python something it is not.

Baggage and Semantics

I think my personal takeaway from Python the language is that it got ridiculously complex. Python is a language that suffers from not having a language specification and already such complex interactions between different types, that we will probably never end up with one. There are so many quirks and odd little behaviors that the only thing a language specification would ever produce, is a textual description of the CPython interpreter.

On this foundation it makes very little sense in my mind to put type annotations.

I think if someone would want to develop another predominantly dynamically typed language in the future, they should probably go the extra mile to clearly define how types should work. JavaScript does a pretty good job at that. All semantics of builtin types are clearly defined even if they are bizarre. I think this generally is a good thing. Once you have clearly defined how the semantics work, you are open to optimize or later put optional static typing on top.

Keeping a language lean and well defined seems to be very much worth the troubles. Future language designers definitely should not make the mistake that PHP, Python and Ruby did, where the language's behavior ends up being "whatever the interpreter does".

I think for Python this is very unlikely to ever change at this point, because the time and work required to clean up language and interpreter outweighs the benefits.

This entry was tagged python, rust and thoughts