kids.cache

kids.cache is a Python library providing a cache decorator. It's part of 'Kids' (for Keep It Dead Simple) library. It has no dependency to any python library.

Its main concern is to offer a very simple default usage scheme, without forgetting to offer full power inside when needed.

Maturity

This code is around ~100 lines of python, and it has a 100% test coverage.

However it still considered beta stage currently.

Compatibility

It is small and simple and should work anywhere.

To put it in longer details: the current code is simple enough that it use a common subset of python that is compatible with any platform on python 2.7 and python >= 3... and this without any specific modification.

Even then, You'll be happy to know that, this code is tested for compatibility at each commit with python 2.7, 3.4, 3.5, 3.6 on linux and windows platform.

Features

• Use one simple call to @cache, and a majority of all hidden complexity will vanish.
  • works out of the box everywhere you can stick a decorator (function, methods, property, classes...).
  • support to be called before or after common decorators as @property, @classmethod, @staticmethod.
• With @cache several design pattern can be achieved:
  • memoization when used on function with arguments.
  • lazy evaluation when placed on properties.
  • singleton patterns when placed on classes.
• Full customization at disposition:
  • cache clearing or cache stats functionality.
  • support of any cache store mecanism from cachetools package.
  • support of custom key function which allows:
    • support of your exotic unhashable objects
    • fine tune which function calls can be considered identic
    • hand pick function dependencies in object (for method)

Basic Usage

Function

This cache decorator is quite straightforward to use:

>>> from kids.cache import cache

>>> @cache
... def answer_to_everything():
...     print("many insightfull calculation")
...     return 42

Then the function answer_to_everything would only do the calculation the first time called, and would save the result, and directly return it the next calls:

>>> answer_to_everything()
many insightfull calculation
42

>>> answer_to_everything()
42

The body of the function was not executed anymore and the cache value was used.

It'll work with arguments:

>>> @cache
... def mysum(*args):
...     print("calculating...")
...     return sum(args)

>>> mysum(2, 2, 3)
calculating...
7
>>> mysum(1, 1, 1, 1)
calculating...
4
>>> mysum(2, 2, 3)
7
>>> mysum(1, 1, 1, 1)
4

And notice that by default, object are not typed, thus:

>>> mysum(1.0, 1, 1, 1)
4

Did trigger the cache, despite the first argument is a float and not an integer.

Methods

With methods:

>>> class MyObject(object):
...    def __init__(self, a, b):
...        self.a, self.b = a, b
...
...    @cache
...    def total(self):
...        print("calculating...")
...        return self.a + self.b

>>> xx = MyObject(2, 3)
>>> xx.total()
calculating...
5
>>> xx.total()
5

Cache is not shared between instances:

>>> yy = MyObject(2, 3)
>>> yy.total()
calculating...
5

Of course, if you change the inner values of the instance, this will NOT be detected by the caching method:

>>> xx.a = 5
>>> xx.total()
5

Look at advanced usages to see how to changes some of these behaviors.

Property

You can use the cache decorator with properties, and provides a good way to have lazy evaluated attributes:

>>> class WithProperty(MyObject):
...
...    @property
...    @cache
...    def total(self):
...        print("evaluating...")
...        return self.a + self.b

>>> xx = WithProperty(1, 1)
>>> xx.total
evaluating...
2
>>> xx.total
2

You can use @cache decorator before or after @property decorator:

>>> class WithProperty(MyObject):
...
...    @cache
...    @property
...    def total(self):
...        print("evaluating...")
...        return self.a + self.b

>>> xx = WithProperty(2, 2)
>>> xx.total
evaluating...
4
>>> xx.total
4

classmethod

You can use the cache decorator with classmethods, and provides a good way to share cache between instances:

>>> class WithClassMethod(MyObject):
...
...    a = 2
...    b = 3
...
...    @classmethod
...    @cache
...    def total(cls):
...        print("evaluating...")
...        return cls.a + cls.b

>>> WithClassMethod.total()
evaluating...
5
>>> WithClassMethod.total()
5

You can use @cache decorator before or after @property decorator:

>>> class WithClassMethod(MyObject):
...
...    a = 1
...    b = 6
...
...    @cache
...    @classmethod
...    def total(cls):
...        print("evaluating...")
...        return cls.a + cls.b

>>> WithClassMethod.total()
evaluating...
7
>>> WithClassMethod.total()
7

staticmethod

You can use the cache decorator with staticmethods:

>>> class WithStaticMethod(MyObject):
...
...    @staticmethod
...    @cache
...    def total(a, b):
...        print("evaluating...")
...        return a + b

>>> WithStaticMethod.total(1, 3)
evaluating...
4
>>> WithStaticMethod.total(1, 3)
4

You can use @cache decorator before or after @property decorator:

>>> class WithStaticMethod(MyObject):
...
...    @cache
...    @staticmethod
...    def total(a, b):
...        print("evaluating...")
...        return a + b

>>> WithStaticMethod.total(2, 6)
evaluating...
8
>>> WithStaticMethod.total(2, 6)
8

class

Using cache with classes will allow variations around the notion of singletons. A singleton shares the same id in memory, so this shows a classical non-singleton behavior:

>>> a, b = object(), object()
>>> id(a) == id(b)
False

Factory based singleton

You can use the cache decorator with classes, effectively implementing a factory pattern for creating singleton:

>>> @cache
... class MySingleton(MyObject):
...     def __new__(cls):
...         print("instanciating...")
...         return MyObject.__new__(cls)
...     def __init__(self):
...         print("initializing...")

>>> a, b = MySingleton(), MySingleton()
instanciating...
initializing...
>>> id(a) == id(b)
True

Notice that both instance are the same object, so it was only instanciated and initialized once.

But be warned: this is not anymore a class:

>>> MySingleton
<function MySingleton at ...>

Instanciation based singletons

Slightly different, the class singleton pattern can be achieved by caching __new__:

>>> class MySingleton(MyObject):
...     @cache
...     def __new__(cls):
...         print("instanciating...")
...         return MyObject.__new__(cls)
...     def __init__(self):
...         print("initializing...")

>>> a, b = MySingleton(), MySingleton()
instanciating...
initializing...
initializing...
>>> id(a) == id(b)
True

Notice that both instance are the same object, so it was only instanciated once. But the __init__ was called both times. This is sometimes perfectly valid, but you might want to avoid this also.

So if you don't want this, you should cache also __init__ method:

>>> class MySingleton(MyObject):
...     @cache
...     def __new__(cls):
...         print("instanciating...")
...         return MyObject.__new__(cls)
...     @cache
...     def __init__(self):
...         print("initializing...")

>>> a, b = MySingleton(), MySingleton()
instanciating...
initializing...
>>> id(a) == id(b)
True

For both cases you'll keep your full object untouched of course:

>>> MySingleton
<class 'MySingleton'>

Singleton with arguments

Actually, these are only singletons if you call them successively with the same arguments.

Or to be more precise, you can share your classes when their instanciation's arguments are the same:

>>> @cache
... class MySingleton(MyObject):
...     def __init__(self, a):
...         self.a = a
...         print("evaluating...")

>>> a, b = MySingleton(1), MySingleton(2)
evaluating...
evaluating...
>>> id(a) == id(b)
False

But:

>>> c = MySingleton(1)
>>> id(a) == id(c)
True

If you want a singleton that give you the same instance even if your successive calls differs, you should check the advanced usage section and the key argument.

Advanced Usage

Most of the advanced usage implies to call the @cache decorator with arguments. Please notice that:

>>> @cache
... def mysum1(*args):
...     print("calculating...")
...     return sum(args)

Or:

>>> @cache()
... def mysum2(*args):
...     print("calculating...")
...     return sum(args)

is equivalent:

>>> mysum1(1,1)
calculating...
2
>>> mysum1(1,1)
2

>>> mysum2(1,1)
calculating...
2
>>> mysum2(1,1)
2

Provide a key function

Providing a key function can be extremely powerfull and will allow to fine tune when the cache should be recalculated.

hashing functions will receive exactly the same arguments than the main function called. It must return an hashable structure (combination of tuples, int, string... avoid list, dicts and sets). This will identify uniquely the result.

For example you could:

>>> class WithKey(MyObject):
...    @cache(key=lambda s: (id(s), s.a, s.b))
...    def total(self):
...        print("calculating...")
...        return self.a + self.b

>>> xx = WithKey(2, 3)
>>> xx.total()
calculating...
5
>>> xx.total()
5

It should detect changes of the given values of the instance:

>>> xx.a = 5
>>> xx.total()
calculating...
8

Without bothering to recalculate when other values change:

>>> xx.c = 7
>>> xx.total()
8

But it should still make a difference between instances:

>>> yy = WithKey(2, 3)
>>> yy.total()
calculating...
5

This last example is important as you could have wanted to share the cache between all instances. You could have done this easily by avoiding returning id(s) in the key function.

Typed key functions

You could ask for typed argument to NOT be treated the same:

>>> @cache(typed=True)
... def mysum(*args):
...     print("calculating...")
...     return sum(args)
>>> mysum(1, 1)
calculating...
2

>>> mysum(1.0, 1)
calculating...
2.0

default key functions

The default key function if not provided is a bold try to make list and dict, set also keyable despite these not being hashable.

The name of the key function is called hippie_hashing, and this is the default value for the key argument:

>>> from kids.cache import hippie_hashing

>>> @cache(key=hippie_hashing)
... def mylength(obj):
...     return len(obj)

This allows you to use the function with list, dict or combination of these:

>>> mylength([set([3]), 2, {1: 2}])
3

Even your objects could be used as key, as long as they are hashable:

>>> class MyObj(object):  ## object subclasses have a default hash
...     length = 5
...     def __len__(self, ):
...         print('calculating...')
...         return self.length

>>> myobj = MyObj()
>>> mylength(myobj)
calculating...
5

>>> mylength(myobj)
5

Be assured that hash collision (they happen!) won't generate cache collisions:

>>> class MyCollidingHashObj(MyObj):
...     def __init__(self, length):
...          self.length = length
...     def __hash__(self):
...          return 1

>>> hash_collide1 = MyCollidingHashObj(6)
>>> hash_collide2 = MyCollidingHashObj(7)

>>> mylength(hash_collide1)
calculating...
6
>>> mylength(hash_collide2)
calculating...
7

But try to avoid them for performance's sake !! And you should probably be aware that if your object compare equal, then THERE WILL BE a cache collision (but at this point, this is probably what you wanted, heh ?):

>>> class MyEqCollidingHashObj(MyCollidingHashObj):
...     def __eq__(self, value):
...          return True
...     def __hash__(self):
...          return 1

>>> eq_and_hash_collide1 = MyEqCollidingHashObj(8)
>>> eq_and_hash_collide2 = MyEqCollidingHashObj(9)

>>> mylength(eq_and_hash_collide1)
calculating...
8
>>> mylength(eq_and_hash_collide2)
8

Huh oh. This is not what was probably expected in this example, but you really had to work hard to make this happen. And most of the time, you'll probably find this convenient and will use it at you advantage. It's a little bit like an extension of the key mecanism that is the objects responsability.

Note

Please verify also that if your object compares the same, their hash HAS TO BE the same. For this very reason, in Python3, when you define the __eq__ method, it'll remove the default __hash__ from objects.

Of course, hippie_hashing will fail on special unhashable object:

>>> class Unhashable(object):
...    def __hash__(self):
...        raise ValueError("unhashable!")

>>> hippie_hashing(Unhashable())  ## doctest: +ELLIPSIS
Traceback (most recent call last):
...
ValueError: <Unhashable ...> can not be hashed. Try providing a custom key function.

If you are not a hippie, you should consider using strict=True and a much more limited method will be used to make a key from your arguments:

>>> @cache(strict=True)
... def mylength(obj):
...     return len(obj)

>>> mylength("hello")
5

But then, don't be surprised if it fails with dict, list, or set arguments:

>>> mylength([set([3]), 2, {1: 2}])
Traceback (most recent call last):
...
TypeError: unhashable type: 'list'

And typed=True can be used in combination with strict=True:

>>> @cache(strict=True, typed=True)
... def mysum(*args):
...     print("calculating...")
...     return sum(args)
>>> mysum(1, 1)
calculating...
2

>>> mysum(1.0, 1)
calculating...
2.0

A good key function can:

• make some cache timeout (but you should then look at cache store section to limit the size of the cache)
• finely select which argument are pertinent to the method to avoid re-evaluating the function when it is non-necessary.
• allow you to cache callables that have very special arguments that can't be hashed properly.

Cleaning Cache

kids.cache uses some lru_cache ideas of python 3 implementation, and each function cached received a cache_clear method:

>>> @cache
... def mysum(*args):
...     print("calculate...")
...     return sum(args)

>>> mysum(1,1)
calculate...
2
>>> mysum(1,1)
2

By calling cache_clear method, we flush all previous cached value:

>>> mysum.cache_clear()
>>> mysum(1,1)
calculate...
2

Cache stats

kids.cache uses some lru_cache ideas of python 3 implementation, and each function cached received a cache_info method:

>>> @cache
... def mysum(*args):
...     print("calculate...")
...     return sum(args)

>>> mysum(1,1)
calculate...
2
>>> mysum(1,1)
2

>>> mysum.cache_info()
CacheInfo(type='dict', hits=1, misses=1, maxsize=None, currsize=1)

Cache Store

kids.cache can use any dict-like structure as a cache store. This means you can provide some more clever cache stores. For example, you can use cachetools caches under the hood to manage the caching store.

Keep in mind that the default cache store is... a dict ! which is not a good idea if your program will run for a long time and you have cached function calls that will be different throughout the running time: the cache store will then grow for each new call making the memory usage of your process grow... perhaps out of bounds.

In these scenario, you must think about using managed cache stores that will clean and remove old unused cache entries. There are many cache store provided in cachetools and kids.cache supports them all.

So if you need any caching store from cachetools you can provide it:

>>> from cachetools import LRUCache

LRU stands for Least Recent Used...

>>> @cache(use=LRUCache(maxsize=2))
... def mysum(*args):
...     print("calculate...")
...     return sum(args)

>>> mysum(1, 1)
calculate...
2
>>> mysum(1, 2)
calculate...
3
>>> mysum(1, 3)
calculate...
4

We have exceeded the cache memory and the least recent used have been tossed away:

>>> mysum(1, 1)
calculate...
2

But we still have this one in memory:

>>> mysum(1, 3)
4

Contributing

Any suggestion or issue is welcome. Push request are very welcome, please check out the guidelines.

Push Request Guidelines

You can send any code. I'll look at it and will integrate it myself in the code base and leave you as the author. This process can take time and it'll take less time if you follow the following guidelines:

• check your code with PEP8 or pylint. Try to stick to 80 columns wide.
• separate your commits per smallest concern.
• each commit should pass the tests (to allow easy bisect)
• each functionality/bugfix commit should contain the code, tests, and doc.
• prior minor commit with typographic or code cosmetic changes are very welcome. These should be tagged in their commit summary with !minor.
• the commit message should follow gitchangelog rules (check the git log to get examples)
• if the commit fixes an issue or finished the implementation of a feature, please mention it in the summary.

If you have some questions about guidelines which is not answered here, please check the current git log, you might find previous commit that would show you how to deal with your issue.

License

Copyright (c) 2017 Valentin Lab.

Licensed under the BSD License.