functools

Q: lru_cache vs cache?

functools.cache unbounded LRU-like simple cache (3.9+). lru_cache controls size and stats.

Q: when partial?

Bind config callbacks - partial(send_email, smtp=cfg) in loops.

Q: wraps necessary?

Yes for public decorators - preserves __name__, __doc__, annotations for tooling.

Q: reduce still pythonic?

Acceptable for functional folds; often clearer total = 1; for x in nums: total *= x for beginners.

Q: singledispatch extension?

Register in other modules with same function name imported - order matters.

Q: total_ordering caveat?

Must define __eq__ and one comparison; others derived - can be slower than manual.

Q: cache_info?

fib.cache_info() hits/misses/maxsize/currsize for tuning.

functools supplies higher-order functions for caching, partial application, reduction, and single-dispatch generics - common building blocks in decorators and APIs.

Recipe

from functools import lru_cache, partial
 
@lru_cache(maxsize=128)
def fib(n: int) -> int:
    if n < 2:
        return n
    return fib(n - 1) + fib(n - 2)
 
double = partial(pow, 2)

When to reach for this:

Memoizing expensive pure functions
Currying configuration into callbacks
Writing decorators that preserve __name__
Type-specific function behavior without if isinstance chains

Working Example

from functools import lru_cache, partial, reduce, singledispatch, wraps
 
def logger(fn):
    @wraps(fn)
    def wrapper(*args, **kwargs):
        print(f"call {fn.__name__}")
        return fn(*args, **kwargs)
    return wrapper
 
@logger
@lru_cache(maxsize=32)
def tokenize(text: str) -> tuple[str, ...]:
    return tuple(text.lower().split())
 
@singledispatch
def serialize(value) -> str:
    return str(value)
 
@serialize.register(int)
def _(value: int) -> str:
    return f"i:{value}"
 
def product(nums: list[int]) -> int:
    return reduce(lambda a, b: a * b, nums, 1)
 
if __name__ == "__main__":
    print(tokenize("Hello World"))
    print(serialize(42))
    print(product([2, 3, 4]))
    add5 = partial(lambda a, b: a + b, 5)
    print(add5(10))

What this demonstrates:

@wraps copies metadata to wrapper for introspection/debugging
lru_cache requires hashable arguments
singledispatch registers per-type handlers
reduce folds iterable with binary function

Deep Dive

How It Works

lru_cache - Dict keyed by arguments; evicts least recently used at maxsize.
cache_clear - Invalidates memoized results when inputs change externally.
partial - partial(func, *args, **kwargs) fixes leading arguments.
singledispatch - Dispatches on type of first argument; register with decorator.
total_ordering - Generates comparison methods from __eq__ and one ordering.

Decorator Stack Order

@outer
@inner
def f(): ...
 
# equals f = outer(inner(f)) - inner applied first

Python Notes

from functools import cached_property  # 3.8+
 
class Config:
    @cached_property
    def heavy(self) -> dict:
        return load()

Gotchas

lru_cache on mutable args - Lists unhashable. Fix: Tuple arguments or disable cache.
Caching impure functions - Stale reads when hidden state changes. Fix: cache_clear or no cache.
Unbounded cache maxsize=None - Memory growth on infinite key space. Fix: Set maxsize.
partial obscures signature - inspect less clear. Fix: wrapper with explicit params for public APIs.
singledispatch not for multiple args - Only first arg type. Fix: multimethod libraries or manual match.

Alternatives

Alternative	Use When	Don't Use When
manual dict cache	Custom eviction	lru_cache suffices
`@cache` (3.9+)	Unbounded small key space	Need LRU eviction
match/case dispatch	Few types inline	Open extension across modules
`itertools.accumulate`	Running fold without reduce import	Need final single value only

FAQs

lru_cache vs cache?

functools.cache unbounded LRU-like simple cache (3.9+). lru_cache controls size and stats.

thread-safe lru_cache?

CPython protects cache dict for single operations - still coordinate if underlying function not thread-safe.

when partial?

Bind config callbacks - partial(send_email, smtp=cfg) in loops.

wraps necessary?

Yes for public decorators - preserves __name__, __doc__, annotations for tooling.

reduce still pythonic?

Acceptable for functional folds; often clearer total = 1; for x in nums: total *= x for beginners.

singledispatch extension?

cached_property vs lru_cache?

cached_property stores on instance once; lru_cache on function arguments globally.

total_ordering caveat?

Must define __eq__ and one comparison; others derived - can be slower than manual.

cache_info?

fib.cache_info() hits/misses/maxsize/currsize for tuning.

functools in hot path?

lru_cache helps expensive pure calls; partial negligible cost.

Writing Decorators - wraps pattern
Closures & Scope Capture - factory pattern
Overloads & Callable Types - typing overloads
itertools - accumulate vs reduce

Stack versions: This page was written for Python 3.14.0 (stable 3.14, maintenance 3.13), FastAPI 0.115+, Django 5.2, Flask 3.1, Pydantic 2, PyTorch 2.6+, pandas 2.2+, Polars 1.x, ruff 0.9+, and uv 0.6+.

Recipe

from functools import lru_cache, partial
 
@lru_cache(maxsize=128)
def fib(n: int) -> int:
    if n < 2:
        return n
    return fib(n - 1) + fib(n - 2)
 
double = partial(pow, 2)

When to reach for this:

Memoizing expensive pure functions
Currying configuration into callbacks
Writing decorators that preserve __name__
Type-specific function behavior without if isinstance chains

Working Example

from functools import lru_cache, partial, reduce, singledispatch, wraps
 
def logger(fn):
    @wraps(fn)
    def wrapper(*args, **kwargs):
        print(f"call {fn.__name__}")
        return fn(*args, **kwargs)
    return wrapper
 
@logger
@lru_cache(maxsize=32)
def tokenize(text: str) -> tuple[str, ...]:
    return tuple(text.lower().split())
 
@singledispatch
def serialize(value) -> str:
    return str(value)
 
@serialize.register(int)
def _(value: int) -> str:
    return f"i:{value}"
 
def product(nums: list[int]) -> int:
    return reduce(lambda a, b: a * b, nums, 1)
 
if __name__ == "__main__":
    print(tokenize("Hello World"))
    print(serialize(42))
    print(product([2, 3, 4]))
    add5 = partial(lambda a, b: a + b, 5)
    print(add5(10))

What this demonstrates:

@wraps copies metadata to wrapper for introspection/debugging
lru_cache requires hashable arguments
singledispatch registers per-type handlers
reduce folds iterable with binary function

Deep Dive

How It Works

lru_cache - Dict keyed by arguments; evicts least recently used at maxsize.
cache_clear - Invalidates memoized results when inputs change externally.
partial - partial(func, *args, **kwargs) fixes leading arguments.
singledispatch - Dispatches on type of first argument; register with decorator.
total_ordering - Generates comparison methods from __eq__ and one ordering.

Decorator Stack Order

@outer
@inner
def f(): ...
 
# equals f = outer(inner(f)) - inner applied first

Python Notes

from functools import cached_property  # 3.8+
 
class Config:
    @cached_property
    def heavy(self) -> dict:
        return load()

Gotchas

lru_cache on mutable args - Lists unhashable. Fix: Tuple arguments or disable cache.
Caching impure functions - Stale reads when hidden state changes. Fix: cache_clear or no cache.
Unbounded cache maxsize=None - Memory growth on infinite key space. Fix: Set maxsize.
partial obscures signature - inspect less clear. Fix: wrapper with explicit params for public APIs.
singledispatch not for multiple args - Only first arg type. Fix: multimethod libraries or manual match.

Alternatives

Alternative	Use When	Don't Use When
manual dict cache	Custom eviction	lru_cache suffices
`@cache` (3.9+)	Unbounded small key space	Need LRU eviction
match/case dispatch	Few types inline	Open extension across modules
`itertools.accumulate`	Running fold without reduce import	Need final single value only

FAQs

lru_cache vs cache?

functools.cache unbounded LRU-like simple cache (3.9+). lru_cache controls size and stats.

thread-safe lru_cache?

CPython protects cache dict for single operations - still coordinate if underlying function not thread-safe.

when partial?

Bind config callbacks - partial(send_email, smtp=cfg) in loops.

wraps necessary?

Yes for public decorators - preserves __name__, __doc__, annotations for tooling.

reduce still pythonic?

Acceptable for functional folds; often clearer total = 1; for x in nums: total *= x for beginners.

singledispatch extension?

cached_property vs lru_cache?

cached_property stores on instance once; lru_cache on function arguments globally.

total_ordering caveat?

Must define __eq__ and one comparison; others derived - can be slower than manual.

cache_info?

fib.cache_info() hits/misses/maxsize/currsize for tuning.

functools in hot path?

lru_cache helps expensive pure calls; partial negligible cost.

Writing Decorators - wraps pattern
Closures & Scope Capture - factory pattern
Overloads & Callable Types - typing overloads
itertools - accumulate vs reduce

Stack versions: This page was written for Python 3.14.0 (stable 3.14, maintenance 3.13), FastAPI 0.115+, Django 5.2, Flask 3.1, Pydantic 2, PyTorch 2.6+, pandas 2.2+, Polars 1.x, ruff 0.9+, and uv 0.6+.

functools

Search Documentation

functools