Performant Python Pattern Matching and Object Validation

Reusable pattern matching for Python, implemented in Cython. I originally developed this system for the Ibis Project but hopefully it can be useful for others as well.

The implementation aims to be as quick as possible, the pure python implementation is already quite fast but taking advantage of Cython allows to mitigate the overhead of the Python interpreter. I have also tried to use PyO3 but it had higher overhead than Cython. The current implementation uses the pure python mode of cython allowing quick iteration and testing, and then it can be cythonized and compiled to an extension module giving a significant speedup. Benchmarks shows more than 2x speedup over pydantic's model validation which is written in Rust.

Installation

The package is published to PyPI, so it can be installed using pip:

pip install koerce

Library components

The library contains three main components which can be used independently or together:

1. Deferred object builders

These allow delayed evaluation of python expressions given a context:

In [1]: from koerce import var, resolve

In [2]: a, b = var("a"), var("b")

In [3]: expr = (a + 1) * b["field"]

In [4]: expr
Out[4]: (($a + 1) * $b['field'])

In [5]: resolve(expr, {"a": 2, "b": {"field": 3}})
Out[5]: 9

The syntax sugar provided by the deferred objects allows the definition of complex object transformations in a concise and natural way.

2. Pattern matchers which operate on various Python objects

Patterns are the heart of the library, they allow searching and replacing specific structures in Python objects. The library provides an extensible yet simple way to define patterns and match values against them.

In [1]: from koerce import match, NoMatch, Anything

In [2]: context = {}

In [3]: match([1, 2, 3, int, "a" @ Anything()], [1, 2, 3, 4, 5], context)
Out[3]: [1, 2, 3, 4, 5]

In [4]: context
Out[4]: {'a': 5}

Note that from koerce import koerce function can be used instead of match() to avoid confusion with the built-in python match.

from dataclasses import dataclass
from koerce import Object, match

@dataclass
class B:
    x: int
    y: int
    z: float

match(Object(B, y=1, z=2), B(1, 1, 2))
# B(x=1, y=1, z=2)

where the Object pattern checks whether the passed object is an instance of B and value.y == 1 and value.z == 2 ignoring the x field.

Patterns are also able to capture values as variables making the matching process more flexible:

from koerce import var

x = var("x")

# `+x` means to capture that object argument as variable `x`
# then the `z` argument must match that captured value
match(Object(B, +x, z=x), B(1, 2, 1))
# it is a match because x and z are equal: B(x=1, y=2, z=1)

match(Object(B, +x, z=x), B(1, 2, 0))
# is is a NoMatch because x and z are unequal

Patterns also suitable for match and replace tasks because they can produce new values:

# >> operator constructs a `Replace` pattern where the right
# hand side is a deferred object
match(Object(B, +x, z=x) >> (x, x + 1), B(1, 2, 1))
# result: (1, 2)

Patterns are also composable and can be freely combined using overloaded operators:

In [1]: from koerce import match, Is, Eq, NoMatch

In [2]: pattern = Is(int) | Is(str)
   ...: assert match(pattern, 1) == 1
   ...: assert match(pattern, "1") == "1"
   ...: assert match(pattern, 3.14) is NoMatch

In [3]: pattern = Is(int) | Eq(1)
   ...: assert match(pattern, 1) == 1
   ...: assert match(pattern, None) is NoMatch

Patterns can also be constructed from python typehints:

In [1]: from koerce import match

In [2]: class Ordinary:
   ...:     def __init__(self, x, y):
   ...:         self.x = x
   ...:         self.y = y
   ...:
   ...:
   ...: class Coercible(Ordinary):
   ...:
   ...:     @classmethod
   ...:     def __coerce__(cls, value):
   ...:         if isinstance(value, tuple):
   ...:             return Coercible(value[0], value[1])
   ...:         else:
   ...:             raise ValueError("Cannot coerce value to Coercible")
   ...:

In [3]: match(Ordinary, Ordinary(1, 2))
Out[3]: <__main__.Ordinary at 0x105194fe0>

In [4]: match(Ordinary, (1, 2))
Out[4]: koerce.patterns.NoMatch

In [5]: match(Coercible, (1, 2))
Out[5]: <__main__.Coercible at 0x109ebb320>

The pattern creation logic also handles generic types by doing lightweight type parameter inference. The implementation is quite compact, available under Pattern.from_typehint().

3. A high-level validation system for dataclass-like objects

This abstraction is similar to what attrs or pydantic provide but there are some differences (TODO listing them).

In [1]: from typing import Optional
   ...: from koerce import Annotable
   ...:
   ...:
   ...: class MyClass(Annotable):
   ...:     x: int
   ...:     y: float
   ...:     z: Optional[list[str]] = None
   ...:

In [2]: MyClass(1, 2.0, ["a", "b"])
Out[2]: MyClass(x=1, y=2.0, z=['a', 'b'])

In [3]: MyClass(1, 2, ["a", "b"])
Out[3]: MyClass(x=1, y=2.0, z=['a', 'b'])

In [4]: MyClass("invalid", 2, ["a", "b"])
Out[4]: # raises validation error

Annotable object are mutable by default, but can be made immutable by passing immutable=True to the Annotable base class. Often it is useful to make immutable objects hashable as well, which can be done by passing hashable=True to the Annotable base class, in this case the hash is precomputed during initialization and stored in the object making the dictionary lookups cheap.

In [1]: from typing import Optional
   ...: from koerce import Annotable
   ...:
   ...:
   ...: class MyClass(Annotable, immutable=True, hashable=True):
   ...:     x: int
   ...:     y: float
   ...:     z: Optional[tuple[str, ...]] = None
   ...:

In [2]: a = MyClass(1, 2.0, ["a", "b"])

In [3]: a
Out[3]: MyClass(x=1, y=2.0, z=('a', 'b'))

In [4]: a.x = 2
AttributeError: Attribute 'x' cannot be assigned to immutable instance of type <class '__main__.MyClass'>

In [5]: {a: 1}
Out[5]: {MyClass(x=1, y=2.0, z=('a', 'b')): 1}

Available Pattern matchers

It is an incompletee list of the matchers, for more details and examples see koerce/patterns.py and koerce/tests/test_patterns.py.

Anything and Nothing

In [1]: from koerce import match, Anything, Nothing

In [2]: match(Anything(), "a")
Out[2]: 'a'

In [3]: match(Anything(), 1)
Out[3]: 1

In [4]: match(Nothing(), 1)
Out[4]: koerce._internal.NoMatch

Eq for equality matching

In [1]: from koerce import Eq, match, var

In [2]: x = var("x")

In [3]: match(Eq(1), 1)
Out[3]: 1

In [4]: match(Eq(1), 2)
Out[4]: koerce._internal.NoMatch

In [5]: match(Eq(x), 2, context={"x": 2})
Out[5]: 2

In [6]: match(Eq(x), 2, context={"x": 3})
Out[6]: koerce._internal.NoMatch

Is for instance matching

Couple simple cases are below:

In [1]: from koerce import match, Is

In [2]: class A: pass

In [3]: match(Is(A), A())
Out[3]: <__main__.A at 0x1061070e0>

In [4]: match(Is(A), "A")
Out[4]: koerce._internal.NoMatch

In [5]: match(Is(int), 1)
Out[5]: 1

In [6]: match(Is(int), 3.14)
Out[6]: koerce._internal.NoMatch

In [7]: from typing import Optional

In [8]: match(Is(Optional[int]), 1)
Out[8]: 1

In [9]: match(Is(Optional[int]), None)

Generic types are also supported by checking types of attributes / properties:

from koerce import match, Is, NoMatch
from typing import Generic, TypeVar, Any
from dataclasses import dataclass


T = TypeVar("T", covariant=True)
S = TypeVar("S", covariant=True)

@dataclass
class My(Generic[T, S]):
    a: T
    b: S
    c: str


MyAlias = My[T, str]

b_int = My(1, 2, "3")
b_float = My(1, 2.0, "3")
b_str = My("1", "2", "3")

# b_int.a must be an instance of int
# b_int.b must be an instance of Any
assert match(My[int, Any], b_int) is b_int

# both b_int.a and b_int.b must be an instance of int
assert match(My[int, int], b_int) is b_int

# b_int.b should be an instance of a float but it isn't
assert match(My[int, float], b_int) is NoMatch

# now b_float.b is actually a float so it is a match
assert match(My[int, float], b_float) is b_float

# type aliases are also supported
assert match(MyAlias[str], b_str) is b_str

As patterns attempting to coerce the value as the given type

from koerce import match, As, NoMatch
from typing import Generic, TypeVar, Any
from dataclasses import dataclass

class MyClass:
    pass

class MyInt(int):
    @classmethod
    def __coerce__(cls, other):
        return MyInt(int(other))


class MyNumber(Generic[T]):
    value: T

    def __init__(self, value):
        self.value = value

    @classmethod
    def __coerce__(cls, other, T):
        return cls(T(other))


assert match(As(int), 1.0) == 1
assert match(As(str), 1.0) == "1.0"
assert match(As(float), 1.0) == 1.0
assert match(As(MyClass), "myclass") is NoMatch

# by implementing the coercible protocol objects can be transparently
# coerced to the given type
assert match(As(MyInt), 3.14) == MyInt(3)

# coercible protocol also supports generic types where the `__coerce__`
# method should be implemented on one of the base classes and the
# type parameters are passed as keyword arguments to `cls.__coerce__()`
assert match(As(MyNumber[float]), 8).value == 8.0

As and Is can be omitted because match() tries to convert its first argument to a pattern using the koerce.pattern() function:

from koerce import pattern, As, Is

assert pattern(int, allow_coercion=False) == Is(int)
assert pattern(int, allow_coercion=True) == As(int)

assert match(int, 1, allow_coercion=False) == 1
assert match(int, 1.1, allow_coercion=False) is NoMatch
# lossy coercion is not allowed
assert match(int, 1.1, allow_coercion=True) is NoMatch

# default is allow_coercion=False
assert match(int, 1.1) is NoMatch

As[typehint] and Is[typehint] can be used to create patterns:

from koerce import Pattern, As, Is

assert match(As[int], '1') == 1
assert match(Is[int], 1) == 1
assert match(Is[int], '1') is NoMatch

If patterns for conditionals

Allows conditional matching based on the value of the object, or other variables in the context:

from koerce import match, If, Is, var, NoMatch, Capture

x = var("x")

pattern = Capture(x) & If(x > 0)
assert match(pattern, 1) == 1
assert match(pattern, -1) is NoMatch

Custom for user defined matching logic

A function passed to either match() or pattern() is treated as a Custom pattern:

from koerce import match, Custom, NoMatch, NoMatchError

def is_even(value):
    if value % 2:
        raise NoMatchError("Value is not even")
    else:
        return value

assert match(is_even, 2) == 2
assert match(is_even, 3) is NoMatch

Capture to record values in the context

A capture pattern can be defined several ways:

from koerce import Capture, Is, var

x = var("x")

Capture("x")  # captures anything as "x" in the context
Capture(x)  # same as above but using a variable
Capture("x", Is(int))  # captures only integers as "x" in the context
Capture("x", Is(int) | Is(float))  # captures integers and floats as "x" in the context
"x" @ Is(int)  # syntax sugar for Capture("x", Is(int))
+x  # syntax sugar for Capture(x, Anything())

from koerce import match, Capture, var

# context is a mutable dictionary passed along the matching process
context = {}
assert match("x" @ Is(int), 1, context) == 1
assert context["x"] == 1

Replace for replacing matched values

Allows replacing matched values with new ones:

from koerce import match, Replace, var

x = var("x")

pattern = Replace(Capture(x), x + 1)
assert match(pattern, 1) == 2
assert match(pattern, 2) == 3

there is a syntax sugar for Replace patterns, the example above can be written as:

from koerce import match, Replace, var

x = var("x")

assert match(+x >> x + 1, 1) == 2
assert match(+x >> x + 1, 2) == 3

replace patterns are especially useful when matching objects:

from dataclasses import dataclass
from koerce import match, Replace, var, namespace

x = var("x")

@dataclass
class A:
    x: int
    y: int

@dataclass
class B:
    x: int
    y: int
    z: float


p, d = namespace(__name__)
x, y = var("x"), var("y")

# if value is an instance of A then capture A.0 as x and A.1 as y
# then construct a new B object with arguments x=x, y=1, z=y
pattern = p.A(+x, +y) >> d.B(x=x, y=1, z=y)
value = A(1, 2)
expected = B(x=1, y=1, z=2)
assert match(pattern, value) == expected

replacemenets can also be used in nested structures:

from koerce import match, Replace, var, namespace, NoMatch

@dataclass
class Foo:
    value: str

@dataclass
class Bar:
    foo: Foo
    value: int

p, d = namespace(__name__)

pattern = p.Bar(p.Foo("a") >> d.Foo("b"))
value = Bar(Foo("a"), 123)
expected = Bar(Foo("b"), 123)

assert match(pattern, value) == expected
assert match(pattern, Bar(Foo("c"), 123)) is NoMatch

SequenceOf / ListOf / TupleOf

from koerce import Is, NoMatch, match, ListOf, TupleOf

pattern = ListOf(str)
assert match(pattern, ["foo", "bar"]) == ["foo", "bar"]
assert match(pattern, [1, 2]) is NoMatch
assert match(pattern, 1) is NoMatch

MappingOf / DictOf / FrozenDictOf

from koerce import DictOf, Is, match

pattern = DictOf(Is(str), Is(int))
assert match(pattern, {"a": 1, "b": 2}) == {"a": 1, "b": 2}
assert match(pattern, {"a": 1, "b": "2"}) is NoMatch

PatternList

from koerce import match, NoMatch, SomeOf, ListOf, pattern

four = [1, 2, 3, 4]
three = [1, 2, 3]

assert match([1, 2, 3, SomeOf(int, at_least=1)], four) == four
assert match([1, 2, 3, SomeOf(int, at_least=1)], three) is NoMatch

integer = pattern(int, allow_coercion=False)
floating = pattern(float, allow_coercion=False)

assert match([1, 2, *floating], [1, 2, 3]) is NoMatch
assert match([1, 2, *floating], [1, 2, 3.0]) == [1, 2, 3.0]
assert match([1, 2, *floating], [1, 2, 3.0, 4.0]) == [1, 2, 3.0, 4.0]

PatternMap

from koerce import match, NoMatch, Is, As

pattern = {
    "a": Is(int),
    "b": As(int),
    "c": Is(str),
    "d": ListOf(As(int)),
}
value = {
    "a": 1,
    "b": 2.0,
    "c": "three",
    "d": (4.0, 5.0, 6.0),
}
assert match(pattern, value) == {
    "a": 1,
    "b": 2,
    "c": "three",
    "d": [4, 5, 6],
}
assert match(pattern, {"a": 1, "b": 2, "c": "three"}) is NoMatch

Annotable objects

Annotable objects are similar to dataclasses but with some differences:

• Annotable objects are mutable by default, but can be made immutable by passing immutable=True to the Annotable base class.
• Annotable objects can be made hashable by passing hashable=True to the Annotable base class, in this case the hash is precomputed during initialization and stored in the object making the dictionary lookups cheap.
• Validation strictness can be controlled by passing allow_coercion=False. When allow_coercion=True the annotations are treated as As patterns allowing the values to be coerced to the given type. When allow_coercion=False the annotations are treated as Is patterns and the values must be exactly of the given type. The default is allow_coercion=True.
• Annotable objects support inheritance, the annotations are inherited from the base classes and the signatures are merged providing a seamless experience.
• Annotable objects can be called with either or both positional and keyword arguments, the positional arguments are matched to the annotations in order and the keyword arguments are matched to the annotations by name.

from typing import Optional
from koerce import Annotable

class MyBase(Annotable):
    x: int
    y: float
    z: Optional[str] = None

class MyClass(MyBase):
    a: str
    b: bytes
    c: tuple[str, ...] = ("a", "b")
    x: int = 1


print(MyClass.__signature__)
# (y: float, a: str, b: bytes, c: tuple = ('a', 'b'), x: int = 1, z: Optional[str] = None)

print(MyClass(2.0, "a", b"b"))
# MyClass(y=2.0, a='a', b=b'b', c=('a', 'b'), x=1, z=None)

print(MyClass(2.0, "a", b"b", c=("c", "d")))
# MyClass(y=2.0, a='a', b=b'b', c=('c', 'd'), x=1, z=None)

print(MyClass(2.0, "a", b"b", c=("c", "d"), x=2))
# MyClass(y=2.0, a='a', b=b'b', c=('c', 'd'), x=2, z=None)

print(MyClass(2.0, "a", b"b", c=("c", "d"), x=2, z="z"))
# MyClass(y=2.0, a='a', b=b'b', c=('c', 'd'), x=2, z='z')

MyClass()
# TypeError: missing a required argument: 'y'

MyClass(2.0, "a", b"b", c=("c", "d"), x=2, z="z", invalid="invalid")
# TypeError: got an unexpected keyword argument 'invalid'

MyClass(2.0, "a", b"b", c=("c", "d"), x=2, z="z", y=3.0)
# TypeError: multiple values for argument 'y'

MyClass("asd", "a", b"b")
# ValidationError

Performance

koerce's performance is at least comparable to pydantic's performance. pydantic-core is written in rust using the PyO3 bindings making it a pretty performant library. There is a quicker validation / serialization library from Jim Crist-Harif called msgspec implemented in hand-crafted C directly using python's C API.

koerce is not exactly like pydantic or msgpec but they are good candidates to benchmark against:

koerce/tests/test_y.py::test_pydantic PASSED
koerce/tests/test_y.py::test_msgspec PASSED
koerce/tests/test_y.py::test_annotated PASSED


------------------------------------------------------------------------------------------- benchmark: 3 tests ------------------------------------------------------------------------------------------
Name (time in ns)            Min                   Max                  Mean              StdDev                Median                IQR            Outliers  OPS (Kops/s)            Rounds  Iterations
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_msgspec            230.2801 (1.0)      6,481.4200 (1.60)       252.1706 (1.0)       97.0572 (1.0)        238.1600 (1.0)       5.0002 (1.0)      485;1616    3,965.5694 (1.0)       20000          50
test_annotated          525.6401 (2.28)     4,038.5600 (1.0)        577.7090 (2.29)     132.9966 (1.37)       553.9799 (2.33)     34.9300 (6.99)      662;671    1,730.9752 (0.44)      20000          50
test_pydantic         1,185.0201 (5.15)     6,027.9400 (1.49)     1,349.1259 (5.35)     320.3790 (3.30)     1,278.5601 (5.37)     75.5100 (15.10)   1071;1424      741.2206 (0.19)      20000          50

I tried to used the most performant API of both msgspec and pydantic receiving the arguments as a dictionary.

I am planning to make more thorough comparisons, but the model-like annotation API of koerce is roughly twice as fast as pydantic but half as fast as msgspec. Considering the implementations it also makes sense, PyO3 possible has a higher overhead than Cython has but neither of those can match the performance of hand crafted python C-API code.

This performance result could be slightly improved but has two huge advantage of the other two libraries:

1. It is implemented in pure python with cython decorators, so it can be used even without compiling it. It could also enable JIT compilers like PyPy or the new copy and patch JIT compiler coming with CPython 3.13 to optimize hot paths better.
2. Development an be done in pure python make it much easier to contribute to. No one needs to learn Rust or python's C API in order to fix bugs or contribute new features.

TODO:

The README is under construction, planning to improve it:

• Example of validating functions by using @annotated decorator
• Explain allow_coercible flag
• Proper error messages for each pattern

Development

• The project uses poetry for dependency management and packaging.
• Python version support follows https://numpy.org/neps/nep-0029-deprecation_policy.html
• The wheels are built using cibuildwheel project.
• The implementation is in pure python with cython annotations.
• The project uses ruff for code formatting.
• The project uses pytest for testing.

More detailed developer guide is coming soon.

References

The project was mostly inspired by the following projects:

• https://github.com/scravy/awesome-pattern-matching
• https://github.com/HPAC/matchpy