Morreski/timed_cache.py

Morreski · 2020-03-27T20:21:18Z

I think it should be next_update = datetime.utcnow() + update_delta but in fact it does not change the correctness of the solution since if will force a flush on the first call. It's just not needed and if copy pasted to another context it could be wrong.

from datetime import datetime, timedelta
import functools


def timed_cache(**timedelta_kwargs):                                              
                                                                                  
    def _wrapper(f):                                                              
        update_delta = timedelta(**timedelta_kwargs)                              
        next_update = datetime.utcnow() + update_delta                            
        # Apply @lru_cache to f with no cache size limit                          
        f = functools.lru_cache(None)(f)                                          
                                                                                                                      
        @functools.wraps(f)                                                       
        def _wrapped(*args, **kwargs):                                            
            nonlocal next_update                                                  
            now = datetime.utcnow()                                               
            if now >= next_update:                                                
                f.cache_clear()                                                   
                next_update = now + update_delta                                
            return f(*args, **kwargs)                                             
        return _wrapped                                                           
    return _wrapper

Hi ! You're 100% right. I updated the gist with your fixed version. Thanks !

svpino · 2020-06-03T19:55:57Z

Thanks for this! Very helpful.

I used this function in one of my projects but modified it a little bit before using it.

def cache(seconds: int, maxsize: int = 128, typed: bool = False):
    def wrapper_cache(func):
        func = functools.lru_cache(maxsize=maxsize, typed=typed)(func)
        func.delta = timedelta(seconds=seconds)
        func.expiration = datetime.utcnow() + func.delta

        @functools.wraps(func)
        def wrapped_func(*args, **kwargs):
            if datetime.utcnow() >= func.expiration:
                func.cache_clear()
                func.expiration = datetime.utcnow() + func.delta

            return func(*args, **kwargs)

        return wrapped_func

    return wrapper_cache

Here are some notes about this version:

The @cache decorator simply expects the number of seconds instead of the full list of arguments expected by timedelta. This avoids leaking timedelta's interface outside of the implementation of @cache. Having the number of seconds should be flexible enough to invalidate the cache at any interval.
maxsize and typed can now be explicitly declared as part of the arguments expected by @cache.
By adding the delta and expiration variables to the func we don't have to use the nonlocal variables, which makes for more readable and compact code.

Also, here is a pytest test case:

def test_cache():
    count = 0

    @cache(seconds=1)
    def test(arg1):
        nonlocal count
        count += 1
        return count

    assert test(1) == 1, "Function should be called the first time we invoke it"
    assert test(1) == 1, "Function should not be called because it is already cached"

    # Let's now wait for the cache to expire
    time.sleep(1)

    assert test(1) == 2, "Function should be called because the cache already expired"

MioYvo · 2020-08-05T08:21:07Z

Thanks for this! Very helpful.

I used this function in one of my projects but modified it a little bit before using it.
def cache(seconds: int, maxsize: int = 128, typed: bool = False):
    def wrapper_cache(func):
        func = functools.lru_cache(maxsize=maxsize, typed=typed)(func)
        func.delta = timedelta(seconds=seconds)
        func.expiration = datetime.utcnow() + func.delta

        @functools.wraps(func)
        def wrapped_func(*args, **kwargs):
            if datetime.utcnow() >= func.expiration:
                func.cache_clear()
                func.expiration = datetime.utcnow() + func.delta

            return func(*args, **kwargs)

        return wrapped_func

    return wrapper_cache
Here are some notes about this version:

The @cache decorator simply expects the number of seconds instead of the full list of arguments expected by timedelta. This avoids leaking timedelta's interface outside of the implementation of @cache. Having the number of seconds should be flexible enough to invalidate the cache at any interval.

maxsize and typed can now be explicitly declared as part of the arguments expected by @cache.

By adding the delta and expiration variables to the func we don't have to use the nonlocal variables, which makes for more readable and compact code.

Also, here is a pytest test case:
def test_cache():
    count = 0

    @cache(seconds=1)
    def test(arg1):
        nonlocal count
        count += 1
        return count

    assert test(1) == 1, "Function should be called the first time we invoke it"
    assert test(1) == 1, "Function should not be called because it is already cached"

    # Let's now wait for the cache to expire
    time.sleep(1)

    assert test(1) == 2, "Function should be called because the cache already expired"

Thanks your share, it's very good!

I add some test and info about test_cache for some people's doubts.

def test_cache():
    count = 0
    count2 = 0

    @cache(seconds=1)
    def test(arg1):
        nonlocal count
        count += 1
        return count

    @cache(seconds=10)
    def test_another(arg2):
        nonlocal count2
        count2 += 1
        return count2

    assert test(1) == 1, "Function test with arg 1 should be called the first time we invoke it"
    assert test(1) == 1, "Function test with arg 1 should not be called because it is already cached"

    assert test(-1) == 2, "Function test with arg -1 should be called the first time we invoke it"
    assert test(-1) == 2, "Function test with arg -1 should not be called because it is already cached"

    assert test_another(1) == 1, "Function test_another with arg 1 should be called the first time we invoke it"
    assert test_another(1) == 1, "Function test_another with arg 1 should not be called because it is already cached"

    # Let's now wait for the cache to expire
    time.sleep(1)

    assert test(1) == 3, "Function test with arg 1 should be called because the cache already expired"
    assert test(-1) == 4, "Function test with arg -1 should be called because the cache already expired"

    # func.cache_clear clear func's cache, not all lru cache 
    assert test_another(1) == 1, "Function test_another with arg 1 should not be called because the cache NOT expired yet"

fdemmer · 2020-08-31T21:06:45Z

many thanks to everybody sharing here! to further pile on to this gist, here are my suggested changes to @svpino's version:

def lru_cache(timeout: int, maxsize: int = 128, typed: bool = False):
    def wrapper_cache(func):
        func = functools.lru_cache(maxsize=maxsize, typed=typed)(func)
        func.delta = timeout * 10 ** 9
        func.expiration = time.monotonic_ns() + func.delta

        @functools.wraps(func)
        def wrapped_func(*args, **kwargs):
            if time.monotonic_ns() >= func.expiration:
                func.cache_clear()
                func.expiration = time.monotonic_ns() + func.delta
            return func(*args, **kwargs)

        wrapped_func.cache_info = func.cache_info
        wrapped_func.cache_clear = func.cache_clear
        return wrapped_func
    return wrapper_cache

renamed the decorator to lru_cache and the timeout parameter to timeout ;)
using time.monotonic_ns avoids expensive conversion to and from datetime/timedelta and prevents possible issues with system clocks drifting or changing
attaching the original lru_cache's cache_info and cache_clear methods to our wrapped_func

svpino · 2020-08-31T21:42:05Z

Solid update, @fdemmer.

jianshen92 · 2020-11-09T17:17:11Z

Further tidying up from @fdemmer version, a fully working snippet

from functools import lru_cache, wraps
from time import monotonic_ns


def timed_lru_cache(
    _func=None, *, seconds: int = 600, maxsize: int = 128, typed: bool = False
):
    """Extension of functools lru_cache with a timeout

    Parameters:
    seconds (int): Timeout in seconds to clear the WHOLE cache, default = 10 minutes
    maxsize (int): Maximum Size of the Cache
    typed (bool): Same value of different type will be a different entry

    """

    def wrapper_cache(f):
        f = lru_cache(maxsize=maxsize, typed=typed)(f)
        f.delta = seconds * 10 ** 9
        f.expiration = monotonic_ns() + f.delta

        @wraps(f)
        def wrapped_f(*args, **kwargs):
            if monotonic_ns() >= f.expiration:
                f.cache_clear()
                f.expiration = monotonic_ns() + f.delta
            return f(*args, **kwargs)

        wrapped_f.cache_info = f.cache_info
        wrapped_f.cache_clear = f.cache_clear
        return wrapped_f

    # To allow decorator to be used without arguments
    if _func is None:
        return wrapper_cache
    else:
        return wrapper_cache(_func)

With documentations, imports, and allow decorators to be called without arguments and paratheses

Alcheri-zz · 2021-09-21T05:41:05Z

@jianshen92 👌💪

valentinbrasov · 2022-07-05T10:01:47Z

The implementation has a big problem: if you have a function that you can call with different values and you obviously want the result cached with TTL for each calling value, then when the TTL is reached for one calling value, the cache is cleared of ALL CACHED RESULTS, that is FOR ALL CALLING VALUES.

Sample code:

import time
import random

@timed_cache(seconds=10)
def expensive_operation(a: int):
    return random.randint(1, 1 + a)
    
def ex_op_wrapper(a: int):
    return f'{time.time()}: {expensive_operation(a)}'

Calling in reply with a 6 secs pause between the first and second call:

ex_op_wrapper(1000)
'1657014039.3334417: 762'
ex_op_wrapper(100)
'1657014045.5532942: 4'
ex_op_wrapper(1000)
'1657014047.3158472: 762'
ex_op_wrapper(100)
'1657014048.6246898: 4'
ex_op_wrapper(1000)
'1657014049.7079725: 847'
ex_op_wrapper(100)
'1657014050.7649162: 70'

You can see that the first cached result for calling with 100 was 4 at '1657014045.5532942', then that was changed at '1657014050.7649162' to 70, so only 5 secs after the first caching of 4, instead of 10.

The problem in the above code is that f.cache_clear() clears the cache for all calling values, not just for the expired one.

Naelpuissant · 2022-10-13T13:10:12Z

Thanks guys ! Btw it can leads to a TypeError: unhashable type: 'list' if you have list args.
A fix could be to cast those args to tuple (more info : https://stackoverflow.com/a/49210802)
This piece of code will fix this :

for arg, value in kwargs.items():
    kwargs[arg] = tuple(value) if type(value) == list else value

giordano91 · 2023-01-13T11:05:55Z

Thanks guys ! Btw it can leads to a TypeError: unhashable type: 'list' if you have list args. A fix could be to cast those args to tuple (more info : https://stackoverflow.com/a/49210802) This piece of code will fix this :
for arg, value in kwargs.items():
    kwargs[arg] = tuple(value) if type(value) == list else value

The behavior remains the same but I would suggest to use isinstance() instead of type()

for arg, value in kwargs.items():
    kwargs[arg] = tuple(value) if isinstance(value, list) else value

HansBambel · 2023-08-07T06:46:30Z

Thanks for the implementations! Really helpful!

Something I noticed is that neither of these implementations work with pytest-antilru. This is likely due to the lru_cache which is monkeypatched is not patched early enough: ipwnponies/pytest-antilru#28.

guschnwg · 2023-09-14T17:30:55Z

args = [tuple(v) if isinstance(v, list) else v for v in args]

for args too

Thanks guys ! Btw it can leads to a TypeError: unhashable type: 'list' if you have list args. A fix could be to cast those args to tuple (more info : https://stackoverflow.com/a/49210802) This piece of code will fix this :
for arg, value in kwargs.items():
    kwargs[arg] = tuple(value) if type(value) == list else value
The behavior remains the same but I would suggest to use isinstance() instead of type()
for arg, value in kwargs.items():
    kwargs[arg] = tuple(value) if isinstance(value, list) else value

Morreski/timed_cache.py

Morreski commented Mar 27, 2020

svpino commented Jun 3, 2020 •

edited

Loading

MioYvo commented Aug 5, 2020

fdemmer commented Aug 31, 2020 •

edited

Loading

svpino commented Aug 31, 2020

jianshen92 commented Nov 9, 2020 •

edited

Loading

Alcheri-zz commented Sep 21, 2021 •

edited

Loading

valentinbrasov commented Jul 5, 2022

Naelpuissant commented Oct 13, 2022 •

edited

Loading

giordano91 commented Jan 13, 2023 •

edited

Loading

HansBambel commented Aug 7, 2023

guschnwg commented Sep 14, 2023

	from datetime import datetime, timedelta
	import functools


	def timed_cache(**timedelta_kwargs):

	def _wrapper(f):
	update_delta = timedelta(**timedelta_kwargs)
	next_update = datetime.utcnow() + update_delta
	# Apply @lru_cache to f with no cache size limit
	f = functools.lru_cache(None)(f)

	@functools.wraps(f)
	def _wrapped(args, *kwargs):
	nonlocal next_update
	now = datetime.utcnow()
	if now >= next_update:
	f.cache_clear()
	next_update = now + update_delta
	return f(args, *kwargs)
	return _wrapped
	return _wrapper

Morreski/timed_cache.py

Morreski commented Mar 27, 2020

svpino commented Jun 3, 2020 • edited Loading

MioYvo commented Aug 5, 2020

fdemmer commented Aug 31, 2020 • edited Loading

svpino commented Aug 31, 2020

jianshen92 commented Nov 9, 2020 • edited Loading

Alcheri-zz commented Sep 21, 2021 • edited Loading

valentinbrasov commented Jul 5, 2022

Naelpuissant commented Oct 13, 2022 • edited Loading

giordano91 commented Jan 13, 2023 • edited Loading

HansBambel commented Aug 7, 2023

guschnwg commented Sep 14, 2023

svpino commented Jun 3, 2020 •

edited

Loading

fdemmer commented Aug 31, 2020 •

edited

Loading

jianshen92 commented Nov 9, 2020 •

edited

Loading

Alcheri-zz commented Sep 21, 2021 •

edited

Loading

Naelpuissant commented Oct 13, 2022 •

edited

Loading

giordano91 commented Jan 13, 2023 •

edited

Loading