[Python-ideas] Proposal: Use mypy syntax for function annotations

[Guido van Rossum]

[There is no TL;DR other than the subject line. Please read the whole thing
before replying. I do have an appendix with some motivations for adding
type annotations at the end.]

Yesterday afternoon I had an inspiring conversation with Bob Ippolito (man
of many trades, author of simplejson) and Jukka Lehtosalo (author of mypy:
http://mypy-lang.org/). Bob gave a talk at EuroPython about what Python can
learn from Haskell (and other languages); yesterday he gave the same talk
at Dropbox. The talk is online (
https://ep2014.europython.eu/en/schedule/sessions/121/) and in broad
strokes comes down to three suggestions:

  (a) Python should adopt mypy's syntax for function annotations
  (b) Python's use of mutabe containers by default is wrong
  (c) Python should adopt some kind of Abstract Data Types

Proposals (b) and (c) don't feel particularly actionable (if you disagree
please start a new thread, I'd be happy to discuss these further if there's
interest) but proposal (a) feels right to me.

So what is mypy?  It is a static type checker for Python written by Jukka
for his Ph.D. thesis. The basic idea is that you add type annotations to
your program using some custom syntax, and when running your program using
the mypy interpreter, type errors will be found during compilation (i.e.,
before the program starts running).

The clever thing here is that the custom syntax is actually valid Python 3,
using (mostly) function annotations: your annotated program will still run
with the regular Python 3 interpreter. In the latter case there will be no
type checking, and no runtime overhead, except to evaluate the function
annotations (which are evaluated at function definition time but don't have
any effect when the function is called).

In fact, it is probably more useful to think of mypy as a heavy-duty linter
than as a compiler or interpreter; leave the type checking to mypy, and the
execution to Python. It is easy to integrate mypy into a continuous
integration setup, for example.

To read up on mypy's annotation syntax, please see the mypy-lang.org
website. Here's just one complete example, to give a flavor:

  from typing import List, Dict

  def word_count(input: List[str]) -> Dict[str, int]:
      result = {}  #type: Dict[str, int]
      for line in input:
          for word in line.split():
              result[word] = result.get(word, 0) + 1
      return result

Note that the #type: comment is part of the mypy syntax; mypy uses comments
to declare types in situations where no syntax is available -- although
this particular line could also be written as follows:

    result = Dict[str, int]()

Either way the entire function is syntactically valid Python 3, and a
suitable implementation of typing.py (containing class definitions for List
and Dict, for example) can be written to make the program run correctly.
One is provided as part of the mypy project.

I should add that many of mypy's syntactic choices aren't actually new. The
basis of many of its ideas go back at least a decade: I blogged about this
topic in 2004 (http://www.artima.com/weblogs/viewpost.jsp?thread=85551 --
see also the two followup posts linked from the top there).

I'll emphasize once more that mypy's type checking happens in a separate
pass: no type checking happens at run time (other than what the interpreter
already does, like raising TypeError on expressions like 1+"1").

There's a lot to this proposal, but I think it's possible to get a PEP
written, accepted and implemented in time for Python 3.5, if people are
supportive. I'll go briefly over some of the action items.

*(1) A change of direction for function annotations*

PEP 3107 <http://legacy.python.org/dev/peps/pep-3107/>, which introduced
function annotations, is intentional non-committal about how function
annotations should be used. It lists a number of use cases, including but
not limited to type checking. It also mentions some rejected proposals that
would have standardized either a syntax for indicating types and/or a way
for multiple frameworks to attach different annotations to the same
function. AFAIK in practice there is little use of function annotations in
mainstream code, and I propose a conscious change of course here by stating
that annotations should be used to indicate types and to propose a standard
notation for them.

(We may have to have some backwards compatibility provision to avoid
breaking code that currently uses annotations for some other purpose.
Fortunately the only issue, at least initially, will be that when running
mypy to type check such code it will produce complaints about the
annotations; it will not affect how such code is executed by the Python
interpreter. Nevertheless, it would be good to deprecate such alternative
uses of annotations.)

*(2) A specification for what to add to Python 3.5*

There needs to be at least a rough consensus on the syntax for annotations,
and the syntax must cover a large enough set of use cases to be useful.
Mypy is still under development, and some of its features are still
evolving (e.g. unions were only added a few weeks ago). It would be
possible to argue endlessly about details of the notation, e.g. whether to
use 'list' or 'List', what either of those means (is a duck-typed list-like
type acceptable?) or how to declare and use type variables, and what to do
with functions that have no annotations at all (mypy currently skips those
completely).

I am proposing that we adopt whatever mypy uses here, keeping discussion of
the details (mostly) out of the PEP. The goal is to make it possible to add
type checking annotations to 3rd party modules (and even to the stdlib)
while allowing unaltered execution of the program by the (unmodified)
Python 3.5 interpreter. The actual type checker will not be integrated with
the Python interpreter, and it will not be checked into the CPython
repository. The only thing that needs to be added to the stdlib is a copy
of mypy's typing.py module. This module defines several dozen new classes
(and a few decorators and other helpers) that can be used in expressing
argument types. If you want to type-check your code you have to download
and install mypy and run it separately.

The curious thing here is that while standardizing a syntax for type
annotations, we technically still won't be adopting standard rules for type
checking. This is intentional. First of all, fully specifying all the type
checking rules would make for a really long and boring PEP (a much better
specification would probably be the mypy source code). Second, I think it's
fine if the type checking algorithm evolves over time, or if variations
emerge. The worst that can happen is that you consider your code correct
but mypy disagrees; your code will still run.

That said, I don't want to *completely* leave out any specification. I want
the contents of the typing.py module to be specified in the PEP, so that it
can be used with confidence. But whether mypy will complain about your
particular form of duck typing doesn't have to be specified by the PEP.
Perhaps as mypy evolves it will take options to tell it how to handle
certain edge cases. Forks of mypy (or entirely different implementations of
type checking based on the same annotation syntax) are also a possibility.
Maybe in the distant future a version of Python will take a different
stance, once we have more experience with how this works out in practice,
but for Python 3.5 I want to restrict the scope of the upheaval.


*Appendix -- Why Add Type Annotations?*
The argument between proponents of static typing and dynamic typing has
been going on for many decades. Neither side is all wrong or all right.
Python has traditionally fallen in the camp of extremely dynamic typing,
and this has worked well for most users, but there are definitely some
areas where adding type annotations would help.

- Editors (IDEs) can benefit from type annotations; they can call out
obvious mistakes (like misspelled method names or inapplicable operations)
and suggest possible method names. Anyone who has used IntelliJ or Xcode
will recognize how powerful these features are, and type annotations will
make such features more useful when editing Python source code.

- Linters are an important tool for teams developing software. A linter
doesn't replace a unittest, but can find certain types of errors better or
quicker. The kind of type checking offered by mypy works much like a
linter, and has similar benefits; but it can find problems that are beyond
the capabilities of most linters.

- Type annotations are useful for the human reader as well! Take the above
word_count() example. How long would it have taken you to figure out the
types of the argument and return value without annotations? Currently most
people put the types in their docstrings; developing a standard notation
for type annotations will reduce the amount of documentation that needs to
be written, and running the type checker might find bugs in the
documentation, too. Once a standard type annotation syntax is introduced,
it should be simple to add support for this notation to documentation
generators like Sphinx.

- Refactoring. Bob's talk has a convincing example of how type annotations
help in (manually) refactoring code. I also expect that certain automatic
refactorings will benefit from type annotations -- imagine a tool like 2to3
(but used for some other transformation) augmented by type annotations, so
it will know whether e.g. x.keys() is referring to the keys of a dictionary
or not.

- Optimizers. I believe this is actually the least important application,
certainly initially. Optimizers like PyPy or Pyston
<https://github.com/dropbox/pyston> wouldn't be able to fully trust the
type annotations, and they are better off using their current strategy of
optimizing code based on the types actually observed at run time. But it's
certainly feasible to imagine a future optimizer also taking type
annotations into account.

-- 
--Guido "I need a new hobby" van Rossum (python.org/~guido)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-ideas/attachments/20140813/2b249ded/attachment-0001.html>