A least recently used (LRU) 2 layer caching mechanism based in part on the Python 2.7 back-port of lru_cache

## Project description

A least recently used (LRU) 2 layer caching mechanism based in part on the Python 2.7 back-port of functools.lru_cache

This was developed by 3Top, Inc. for use with our ranking and recommendation platform, http://www.3top.com.

lru2cache is a decorator that can be used with any user function or method to cache the most recent results in a local cache and using the django cache framework to cache results in a shared cache.

The first layer of caching is stored in a callable that wraps the function or method. As with ‘functools.lru_cache’ a dict is used to store the cached results, therefore positional and keyword arguments must be hashable. Each instance stores up to l1_maxsize results that vary on the arguments. The discarding of the LRU cached values is handled by the decorator.

The second layer of caching requires a shared cache that can make use of Django’s cache framework. In this case it is assumed that any LRU mechanism is handled by the shared cache backend.

This arrangement allows a process that accesses a function multiple times to retrieve the value without the expense of requesting it from a shared cache, while still allowing different processes to access the result from the shared cache.

## Installation & Configuration

The easiest and best way to install this is with pip:

pip install lru2cache

If available this package will use SpookyHash V2 as a hashing mechanism. Spooky is a good fast hashing algorithm that should be suitable for most uses. If it is not available the package will fall back to SHA-256 from the standard hashlib. Because SHA-256 is a proper cryptographic hash it requires more computation than Spooky. To install spooky use pip:

pip install spooky 2

Once lru2cache is installed you will need to configure a shared cache as an l2 cache. If you are using Django your settings file will contain something similar to the following in the settings file:

CACHES = {
'default': {
'BACKEND': 'django.core.cache.backends.memcached.PyLibMCCache',
'LOCATION': '127.0.0.1:11211',
'TIMEOUT': 1200,
},
'l2cache': {
'BACKEND': 'django.core.cache.backends.memcached.PyLibMCCache',
'LOCATION': '127.0.0.1:11211',
'TIMEOUT': None,
},
}

If you do not want to use either default or l2cache you will need to specify the name of the cache.

## Benefits Over functools.lru_cache

Local and Shared Cache - Combining both types of cache is much more effective than either used on it’s own. The local cache eliminates the latency of calls to a shared cache, while the shared cache eliminates the expense of returning the result

The Ability to Not Cache None Results - This may seem like a minor thing but in our environment it has greatly reduced the frequency of cache invalidations.

## Usage

@utils.lru2cache(l1_maxsize=128, none_cache=False, typed=False, l2cache_name='l2cache', inst_attr='id')

Usage is as simple as adding the decorator to a function or method as seen in the below examples from our test cases:

from lru2cache import utils

@utils.lru2cache()
def py_cached_func(x, y):
return 3 * x + y

class TestLRUPy(TestLRU):
module = utils
cached_func = py_cached_func,

@utils.lru2cache()
def cached_meth(self, x, y):
return 3 * x + y

@staticmethod
@utils.lru2cache()
def cached_staticmeth(x, y):
return 3 * x + y

If l1_maxsize is set to None, the LRU feature is disabled and the L1 cache can grow without bound. The LRU feature performs best when maxsize is a power-of-two.

if none_cache is True than None results will be cached, otherwise they will not.

If typed is set to True, function arguments of different types will be cached separately. For example, f(3) and f(3.0) will be treated as distinct calls with distinct results.

If l2cache_name is specified it will be used as the shared cache. Otherwise it will attempt to use a cache named l2cache and if not found fall back to default.

inst_attr is the attribute used to uniquely identify an object when wrapping a method. In Django this will typically be id however if it is not you will need to specify what attribute should be used.

## Cache Management

Since the lru2cache decorator does not provide a timeout for its cache although it provides other mechanisms for programatically managing the cache.

### Cache Statistics

As with lru_cache, one can view the cache statistics via a named tuple (l1_hits, l1_misses, l2_hits, l2_misses, l1_maxsize, l1_currsize), with f.cache_info(). These stats are stored within an instance, and therefore are specific to that instance. Cumulative statistics for the shared cache would need to be obtained from the shared cache.

### Clearing Instance Cache

the cache and statistics associated with a function or method can be cleared with:

f.cache_clear()

### Clearing Shared Cache

A shared cache can easily be cleared with the following:

from django.core import cache

lru2cache_cache = cache.get_cache('l2cache')
lru2cache_cache.clear()

### Invalidating Cached Results

To invalidate the cache for a specific set of arguments, including the instance one can pass the same arguments to invalidate the both L1 and L2 caches:

f.invalidate(*args, **kwargs)

in the case of a method you do need to explicitly pass the instance as in the following:

foo.f.invalidate(foo, a, b)

### Refreshing the Cache

This is not yet implemented as a function but can be accomplished by first calling invalidate and then calling the function

### Accessing the Function without Cache

The un-cached underlying function can always be accessed with f.__wrapped__.

### Background and Development

At 3Top We needed a way to improve performance of slow queries, not just those using the Django ORM, but also for queries to other data stores and services. We started off with a simpler centralized caching solution using Memcached, but even those queries, when called frequently, can start to cause delays. Therefore we sought a means of caching at two layers.

Initially we looked at the possibility of using two different mechanisms but we quickly saw the advantage of maintaining the same set of keys for both caches and decided to create our own mechanism.

We used a backport python 3 functools.lru_cache() decorator as a starting point for developing an in instance cache with LRU capabilities. However we needed to ensure the keys would also be unique enough to use with a shared cache. We leverage Django’s excellent cache framework for managing the layer 2 cache. This allows the use of any shared cache supported by Django.

### Tests

As a starting point I incorporated most of the tests for functools.lru_cache() with minor changes to make them work with python 2.7 and incorporated the l2_cache stats. We will continue to add tests to validate the additional functionality provided by this decorator.

## Project details

Uploaded py2 py3