h2oai / typesentry Goto Github PK

Python library for run-time type checking for type-annotated functions. It supports both Python-3 annotations (as defined in PEP 484), and legacy Python-2 annotations via decorators. The library also provides function is_type analogous to builtin isinstance to allow type checking in other scenarios.

The goal of this library is to provide performat, convenient, configurable module that can be used in a variety of different contexts.

Python is a dynamically typed language; it is the source of its strength and of its appeal. Still, every object in the language does have its own type, and every function has an implied signature — the types of arguments that it can operate on. Passing arguments of incorrect types leads either to unexpected results, or more likely to an exception.

Thus, PEP-484 was introduced, which mentions in its introduction that

This PEP aims to provide a standard syntax for type annotations, opening up Python code to easier static analysis and refactoring, potential runtime type checking ...

Of course, it is each developer's choice whether or not to use runtime type checking. Obviously, there is a certain amount of overhead to check the type of each argument (especially when it's a container type). However we have found that not doing so itself causes problems:

• Exceptions that are raised are either only tangentially related to the actual problem, or outright cryptic. You may see errors about unsupported operations on a type, unavailable attributes, and worse. But very rarely you see a clear message such as "Argument x expected to be an integer, but instead received None".
• A function may not do anything to its argument other than storing it for later use. If the argument had invalid type, it will manifest itself some time later in the program, at which point it would already be impossible to know how this bad argument came out to be.

In short, runtime type checks slow down the program, whereas their absence slow down the programmer. We also recognize that Python code may be executed in different environments, which would lead to a different choice in this tradeoff:

• Production environment, where the code runs against a set of predefined and carefully vetted inputs. Perhaps in this case runtime checks can be forfeited.
• Testing environment, where speed is much less of an importance but ability to quickly identify errors is more crucial. Runtime checks should be always on.
• Interactive programming, via a console or Jupyter Notebook, where the software engineer writes the code in a trial-and-error loop. Again, runtime checks are extremely useful here, since speed of debugging the problems dominates the speed of code execution.

The typesentry library is intended to be used from other packages. The minimal example of how to do it looks like this:

import typesentry
tc1 = typesentry.Config()
typed = tc1.typed        # decorator to check function arguments at runtime
is_typed = tc1.is_typed  # equivalent of isinstance()

Typically you would put this into a separate file in your project, and then import symbols typed and is_typed (or however you want to name them) from that file.

The Config object here allows you to specify how exactly the type checking should behave. In particular you can enable/disable type checking, pass custom exception class that will be used when a type error is detected, etc. More settings are expected to be added in the future.

You can create more than one Config object, with different settings, allowing you to have typecheck-decorators with different degrees of persistence.

Once you this file (let's assume it's utils/types.py for concreteness), then the annotation @typed and function is_typed() can be used as follows (in both Python 2 and 3):

from .utils.types import typed
from typing import List, Union

@typed(x=Union[List[int], List[str]])
def together(x):
    if not x:
        return None
    elif isinstance(x[0], int):
        return sum(x)
    else:
        return "".join(x)

or using Python 3 type annotations:

from .utils.types import typed, is_type
from typing import List, Union

@typed()
def together(x: Union[List[int], List[str]]):
    if not x:
        return None
    elif isinstance(x[0], int):
        return sum(x)
    else:
        return "".join(x)

Now let's try calling this function with various arguments:

>>> together([])
>>> together([1, 5, -2])
4
>>> together(["hello", ",", " ", "world", "!"])
'hello, world!'
>>> # Notice how in 2 examples below the error message is different depending
>>> # on whether the argument looks more like List[int] or List[str]
>>> together(["hello", ",", " ", "world", 1])
TypeError: Parameter x expects type List[str] but received a list where 5th element is 1 of type int
           File <stdin>, line 1, in
                together(x)
           File <stdin>, line 1, in <module>()
>>> together(["hello", 2, 9, 11, 1])
TypeError: Parameter x expects type List[int] but received a list where 1st element is 'hello' of type str
           File <stdin>, line 1, in
                together(x)
           File <stdin>, line 1, in <module>()
>>> # If it doesn't look like either, then a more generic message is displayed
>>> together([False, True])
TypeError: Parameter x of type Union[List[int], List[str]] received value [False, True] of type list
           File <stdin>, line 1, in
                together(x)
           File <stdin>, line 1, in <module>()
>>> # Also note that we treat booleans as types distinct from int:
>>> isinstance(True, int), isinstance(True, bool)
(True, True)
>>> is_type(True, int), is_type(True, bool)
(False, True)

In addition to trying to generate helpful messages to the user upon seeing a type mismatch, this module also advocates for the use of "kind" (or "soft") exceptions. It applies in the context of interactive programming from within a console or a Jupyter notebook.

The idea is that when a user makes a small innocent mistake, such as a typo in a parameter's name, or providing wrong parameter value — then throwing back at them exceptions with long intimidating stack traces is rather rude. The error message should not attempt to overwhelm the user, but rather help them correct the problem.

Hence, the notion of "soft exceptions". They can be turned on/off using Config``s parameter ``soft_exceptions (which is True by default). In this mode typesentry installs an exception hook (via sys.excepthook) such that whenever an exception exposing a _handle_() method propagates to the outer level, then instead of printing the default stack trace, this _handle_() method would be invoked. Thus, "soft exceptions" are just regular exceptions for all intents and purposes, except with respect to how they appear in the console.

The default exception class used by typesentry implements custom _handle_() method which prints the error message at the top, then the signature of the function where type mismatch has occurred, and finally the compactified stack trace. It also uses colors to accentuate the most important parts of the error message.

The user may override this behavior by either specifying turning off soft exceptions in the Config, providing their own exception class which may or may not implement _handle_(), or submitting a Pull Request ;)

We intend to further improve and refine this functionality (for example, currently support for Jupyter Notebooks is missing).

You can extend functionality of this module by declaring custom types as classes deriving from typesentry.MagicType. At a minimum, you would override methods check(self, var) which should return True iff a variable var matches the type; and method name(self) which returns string description of your new type (to be used in error messages).

In addition there are also methods fuzzy_check(self, var) returning a float value from 0 to 1 indicating how well var matches the type; and get_error_msg(self, param, var) which should return an error message about parameter param = var not matching the type. These two methods are advanced and need not be implemented. However they are useful if you want to provide smarter-than-usual feedback to the user.

For example, suppose you have a set of functions that work with rectangular matrices, i.e. objects of the type List[List[float]]. At some point you realize that this is insufficient: you need to guarantee that all internal arrays have the same dimensions, otherwise it's not really a matrix. The code to implement such type may look like this:

from typesentry import MagicType

class MatrixT(MagicType):
    def check(self, var) -> bool:
        if not isinstance(var, list) or not var: return False
        for elem in var:
            if not isinstance(elem, list): return False
            if len(elem) != len(var[0]): return False
            if not all(isinstance(x, float) for x in elem): return False
        return True

    def name(self) -> str:
        return "Matrix"

pip install typesentry

• PEP 484 — Python standard for declaring type annotations.
• MyPy — static type analyzer (i.e. at compile time).
• typeguard — alternative runtime type checker.
• enforce — another runtime type checker.
• runtime_typecheck — yet another one.