How do I clear the cache from @cached_property decorator?

As you can read in the CPython source, the value for a cached_property in Python 3.8 is stored in an instance variable of the same name. This is not documented, so it may be an implementation detail that you should not rely upon.

But if you just want to get it done without regards to compatibility, you can remove the cache with del instance.value. And who knows, maybe the current behavior will be documented in the future, so it will be safe to use it in any version or interpreter implementation.

(Aditional to @Blckknght answer)

In case that you have a mutable object and you need to refresh all the @cached_property (because the object has been mutated), you could delete the properties that are already cached on the self.__dict__ dictionary (that's where the properties are storaged)

from functools import cached_property

class Test:
    datalist: List[int]

    @cached_property
    def value(self):
        result = None
        # heavy calculate based on datalist
        return result

    def add_element(self, new:int)-> None:
        # restore cache if calculated 
        self.__dict__.pop('value', None) # this will delete the cached val if already cached, otherwise do nothing  
        self.datalist.append(new)

or in case you want to do it more elegant you can directly edit the __setattr__ method

from functools import cached_property

class Test:
    datalist: List[int]

    @cached_property
    def value(self):
        result = None
        # heavy calculate based on datalist
        return result

    def __setattr__(self, name, val):
        self.__dict__[name] = val
        self.__dict__.pop('value', None)

I offer an alternative approach, which might be useful in some cases. If the type of the dataset you need to do the computation on is hashable, you can make use of the regular functools.cache or lru_cache decorator, applied to a static method that takes the dataset as input.

Here is an example of what I mean:

from functools import lru_cache

class MyClass():

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

    @property    
    def slow_attribute(self):
        return self._slow_attribute(self.data)
        
    @staticmethod
    @lru_cache
    def _slow_attribute(data):

        # long computation, using data,
        # here is just an example

        return sum(data)

Here there is no need to concern yourself with when to clear the cache: if the underlying dataset changes, the staticmethod automatically knows it cannot use the cached value anymore.

This has the additional perk that, if the dataset were to be restored to a previously-used state, the lookup may still be able to use a cached value.

Here is a demo of the code above working:

from time import perf_counter_ns

def print_time_and_value_of_computation(c):
    
    t1 = perf_counter_ns()    
    val = c.slow_attribute
    t2 = perf_counter_ns()

    print(f'Time taken: {(t2 - t1)/1000} microseconds')
    print(f'Value: {val}')


c = MyClass(range(10_000))

print_time_and_value_of_computation(c)
print_time_and_value_of_computation(c)

print('Changing the dataset!')
c.data = range(20_000)

print_time_and_value_of_computation(c)
print_time_and_value_of_computation(c)

print('Going back to the original dataset!')
c.data = range(10_000)

print_time_and_value_of_computation(c)

which returns:

Time taken: 162.074 microseconds
Value: 49995000
Time taken: 2.152 microseconds
Value: 49995000
Changing the dataset!
Time taken: 264.121 microseconds
Value: 199990000
Time taken: 1.989 microseconds
Value: 199990000
Going back to the original dataset!
Time taken: 1.144 microseconds
Value: 49995000

I ran across this problem and came across this thread as I was trying to solve it. The data in my case effectively is immutable, except that the setup of this object in some cases involves using the properties, with the properties being out of date after the setup. @Pablo's answer was helpful, but I wanted that process to dynamically reset everything cached.

Here's a generic example:

Setup and broken thing:

from functools import cached_property


class BaseThing:
    def __init__(self, *starting_numbers: int):
        self.numbers = []
        self.numbers.extend(starting_numbers)

    @property
    def numbers_as_strings(self) -> dict[int, str]:
        """This property method will be referenced repeatedly"""

    def process_arbitrary_numbers(self, *arbitrary_numbers: int) -> list[str]:
        return [self.numbers_as_strings.get(number) for number in arbitrary_numbers]

    def extend_numbers(self, *additional_numbers: int):
        self.numbers.extend(additional_numbers)


class BrokenThing(BaseThing):
    @cached_property
    def numbers_as_strings(self) -> dict[int, str]:
        print("Working on:", " ".join(map(str, self.numbers)))
        return {number: str(number) for number in self.numbers}

output:

>>> thing = BrokenThing(1, 2, 3, 4)
>>> thing.process_arbitrary_numbers(1, 3) == ["1", "3"]
Working on: 1 2 3 4
True
>>> thing.extend_numbers(4, 5, 6)
>>> thing.process_arbitrary_numbers(5, 6) == ["5", "6"]
False

@cached_property replaced with @property to make it work, leaving it inefficient:

class InefficientThing(BaseThing):
    @property
    def numbers_as_strings(self) -> dict[int, str]:
        print("Working on:", " ".join(map(str, self.numbers)))
        return {number: str(number) for number in self.numbers}

output:

>>> thing = InefficientThing(1, 2, 3)
>>> thing.process_arbitrary_numbers(1, 3) == ["1", "3"]
Working on: 1 2 3
Working on: 1 2 3
True
>>> thing.extend_numbers(4, 5, 6)
>>> thing.process_arbitrary_numbers(5, 6) == ["5", "6"]
Working on: 1 2 3 4 5 6
Working on: 1 2 3 4 5 6
True

Solution:

class EfficientThing(BaseThing):
    def _clear_cached_properties(self):
        for name in dir(type(self)):
            if isinstance(getattr(type(self), name), cached_property):
                print(f"Clearing self.{name}")
                vars(self).pop(name, None)

    def extend_numbers(self, *additional_numbers: int):
        self._clear_cached_properties()
        return super().extend_numbers(*additional_numbers)

    @cached_property
    def numbers_as_strings(self) -> dict[int, str]:
        print("Working on:", " ".join(map(str, self.numbers)))
        return {number: str(number) for number in self.numbers}

output:

>>> thing = EfficientThing(1, 2, 3, 4)
>>> thing.process_arbitrary_numbers(1, 3) == ["1", "3"]
Working on: 1 2 3 4
True
>>> thing.extend_numbers(4, 5, 6)
Clearing self.numbers_as_strings
>>> thing.process_arbitrary_numbers(5, 6) == ["5", "6"]
Working on: 1 2 3 4 4 5 6
True

This loops through all attributes of the object's parent class. If the value of the attribute is an instance of cached_property, it's most likely a cached_property. The attribute is then popped from the instance dictionary. None is passed to pop in case the property hadn't been cached yet.