Wheezy Caching comes with the following cache implementations:
Wheezy Caching provides integration with:
It introduces idea of cache dependency that let effectively invalidate dependent cache items.
All cache implementations and integrations provide the same contract. That means caches can be swapped without a need to modify the code. However there do exist challenge: some caches are singletons and correctly provide inter-thread synchronization (thread safe), while others require an instance per thread (not thread safe), some sort of pooling is required. This challenge is transparently resolved.
Here is an example how to configure pylibmc - memcached client written in C):
from wheezy.core.pooling import EagerPool
from wheezy.caching.pylibmc import MemcachedClient
from wheezy.caching.pylibmc import client_factory
# Cache Pool
pool = EagerPool(lambda: client_factory(['/tmp/memcached.sock']), size=10)
# Factory
cache = MemcachedClient(pool)
# Client code
cache.set(...)
The client code remains unchanged even some cache implementations require pooling to remain thread safe.
CacheClient serves mediator purpose between a single entry point that implements Cache and one or many namespaces targeted to cache factories.
CacheClient let partition application cache by namespaces effectively hiding details from client code.
CacheClient accepts the following arguments:
In the example below we partition application cache into three (default, membership and funds):
from wheezy.caching import ClientCache
from wheezy.caching import MemoryCache
from wheezy.caching import NullCache
default_cache = MemoryCache()
membership_cache = MemoryCache()
funds_cache = NullCache()
cache = ClientCache({
'default': default_cache,
'membership': membership_cache,
'funds': funds_cache,
}, default_namespace='default')
Application code is designed to work with a single cache by specifying namespace to use:
cache.add('x1', 1, namespace='default')
At some point of time we might change our partitioning scheme so all namespaces reside in a single cache:
default_cache = MemoryCache()
cachey = ClientCache({
'default': default_cache,
'membership': default_cache,
'funds': default_cache
}, default_namespace='default')
What happened with no changes to application code? Just configuration settings.
MemoryCache is effective, high performance in-memory cache implementation. There is no background routine to invalidate expired items in the cache, instead they are checked on each get operation.
In order to effectively manage invalidation of expired items (those that are not actively requested) each item being added to cache is assigned to time bucket. Each time bucket has a number associated with a point in time. So if incoming store operation relates to time bucket N, all items from that bucket are being checked and expired items removed.
You control a number of buckets during initialization of MemoryCache. Here are attributes that are accepted:
Interval set by bucket_interval shows how often items in cache will be checked for expiration. So if it set to 15 means that every 15 seconds cache will choose a bucket related to that point in time and all items in bucket will be checked for expiration. Since there are 60 buckets in the cache that means only 1/60 part of cache items are locked. This lock does not impact items requested by get/get_multi operations. Taking into account this lock happens only once per 15 seconds it cause minor impact on overall cache performance.
NullCache is a cache implementation that actually does not do anything but silently performs cache operations that result no change to state.
python-memcached is a pure Python memcached client. You can install this package via easy_install:
$ env/bin/easy_install python-memcached
Here is a typical use case:
from wheezy.caching.memcache import MemcachedClient
cache = MemcachedClient(['unix:/tmp/memcached.sock'])
You can specify key encoding function by passing key_encode argument that must be a callable that does key encoding. By default string_encode() is applied.
All arguments passed to MemcachedClient() are the same as to original Client from python-memcache. Note, python-memcached Client implementation is thread local object.
pylibmc is a quick and small memcached client for Python written in C. Since this package is an interface to libmemcached you need development version of this library so it can be compiled. If you are using Debian:
apt-get install libmemcached-dev
Now, you can install this package via easy_install:
$ env/bin/easy_install pylibmc
Here is a typical use case:
from wheezy.core.pooling import EagerPool
from wheezy.caching.pylibmc import MemcachedClient
from wheezy.caching.pylibmc import client_factory
pool = EagerPool(lambda: client_factory(['/tmp/memcached.sock']), size=10)
cache = MemcachedClient(pool)
You can specify key encoding function by passing key_encode argument that must be a callable that does key encoding. By default string_encode() is applied.
All arguments passed to client_factory() are the same as to original Client from pylibmc. Default client factory configures pylibmc Client to use binary protocol, tcp_nodelay and ketama algorithm.
Since pylibmc implementation is not thread safe it requires pooling, so we do here. EagerPool holds a number of pylibmc instances.
Memcached has some restrictions concerning keys used. Text protocol requires a valid key contain only ASCII characters except space (0x20), carriage return (0x0d), line feed (0x0a) since these characters are meaningful in text protocol. Key length is restricted to 250.
There is general purpose function:
You can specify key encoding function by passing key_encode argument to memcache and/or pylibmc cache factory.
CacheDependency introduces a wire between cache items so they can be invalidated via a single operation, thus simplifying code necessary to manage dependencies in cache.
CacheDependency is not related to any particular cache implementation.
CacheDependency can be used to invalidate items across different cache partitions (namespaces). Note that delete must be performed for each namespace and/or cache.
It is important to avoid key collision for master key due to a way how dependency keys are built. The dependency keys are built by adding a suffix with incremental number to master key, e.g. if master key is ‘key’ than dependent keys used by CacheDependency will be ‘key1’, ‘key2’, ‘key3’, etc. The master key stores a number of dependent keys thus this number is incremented each time you add something to dependency.
If a master key is composed as a concatenation with some id it must be suffixed with a delimiter (a symbol that is not part of the id) to avoid key collision. In the example below id is a number so choosing ‘:’ as a delimiter suites our needs:
def master_key_order(id):
return 'mk:order:' + str(id) + ':'
For order id 100 the master key is ‘mk:order:100:’ and dependent keys take space ‘mk:order:100:1’ for first item added, ‘mk:order:100:2’ for the second, etc. If we add 2 items to cache dependency the value stored by master key is 2.
Let demostrate this by example. We establish dependency between keys k1, k2 and k3 for 600 seconds. Please note that dependency does not need to be passed between various parts of application. You can create it in one place, than in other, etc. CacheDependency stores it state in cache:
# this is sample from module a.
dependency = CacheDependency('master-key', time=600)
dependency.add_multi(cache, ['k1', 'k2', 'k3'])
# this is sample from module b.
dependency = CacheDependency('master-key', time=600)
dependency.add(cache, 'k4')
Note that module b have no idea about keys used in module a. Instead they share virtually cache dependency.
Once we need invalidate items related to cache dependency this is what we do:
dependency = CacheDependency('master-key')
dependency.delete(cache)
delete operation must be repeated for each namespace (it doesn’t manage namespace dependency) and/or cache:
# Using namespaces
dependency = CacheDependency('master-key')
dependency.delete(cache, namespace='membership')
dependency.delete(cache, namespace='funds')
# Using caches
dependency = CacheDependency('master-key')
dependency.delete(membership_cache)
dependency.delete(funds_cache)
Cache dependency is an effective way to reduce coupling between modules in terms of cache items invalidation.