Extended docs (including API docs) available at memoize.readthedocs.io.
What & Why
What: Caching library for asynchronous Python applications.
Why: Python deserves library that works in async world (for instance handles dog-piling ) and has a proper, extensible API.
Etymology
In computing, memoization or memoisation is an optimization technique used primarily to speed up computer programs by storing the results of expensive function calls and returning the cached result when the same inputs occur again. (…) The term “memoization” was coined by Donald Michie in 1968 and is derived from the Latin word “memorandum” (“to be remembered”), usually truncated as “memo” in the English language, and thus carries the meaning of “turning [the results of] a function into something to be remembered.” ~ Wikipedia
Getting Started
Installation
Basic Installation
To get you up & running all you need is to install:
pip install py-memoize
Installation of Extras
If you are going to use memoize
with tornado add a dependency on extra:
pip install py-memoize[tornado]
To harness the power of ujson (if JSON SerDe is used) install extra:
pip install py-memoize[ujson]
Usage
Provided examples use default configuration to cache results in memory. For configuration options see Configurability.
You can use memoize
with both asyncio
and Tornado - please see the appropriate example:
asyncio
To apply default caching configuration use:
import asyncio
import random
from memoize.wrapper import memoize
@memoize()
async def expensive_computation():
return 'expensive-computation-' + str(random.randint(1, 100))
async def main():
print(await expensive_computation())
print(await expensive_computation())
print(await expensive_computation())
if __name__ == "__main__":
asyncio.get_event_loop().run_until_complete(main())
Tornado
If your project is based on Tornado use:
import random
from tornado import gen
from tornado.ioloop import IOLoop
from memoize.wrapper import memoize
@memoize()
@gen.coroutine
def expensive_computation():
return 'expensive-computation-' + str(random.randint(1, 100))
@gen.coroutine
def main():
result1 = yield expensive_computation()
print(result1)
result2 = yield expensive_computation()
print(result2)
result3 = yield expensive_computation()
print(result3)
if __name__ == "__main__":
IOLoop.current().run_sync(main)
Features
Async-first
Asynchronous programming is often seen as a huge performance boost in python programming. But with all the benefits it brings there are also new concurrency-related caveats like dog-piling.
This library is built async-oriented from the ground-up, what manifests in, for example, in Dog-piling proofness or Async cache storage.
Tornado & asyncio support
No matter what are you using, build-in asyncio or its predecessor Tornado memoize has you covered as you can use it with both. This may come handy if you are planning a migration from Tornado to asyncio.
Under the hood memoize detects if you are using Tornado or asyncio (by checking if Tornado is installed and available to import).
If have Tornado installed but your application uses asyncio IO-loop,
set MEMOIZE_FORCE_ASYNCIO=1
environment variable to force using asyncio and ignore Tornado instalation.
Configurability
With memoize you have under control:
- timeout applied to the cached method;
- key generation strategy (see :class:`memoize.key.KeyExtractor`); already provided strategies use arguments (both positional & keyword) and method name (or reference);
- storage for cached entries/items (see :class:`memoize.storage.CacheStorage`); in-memory storage is already provided; for convenience of implementing new storage adapters some SerDe (:class:`memoize.serde.SerDe`) are provided;
- eviction strategy (see :class:`memoize.eviction.EvictionStrategy`); least-recently-updated strategy is already provided;
- entry builder (see :class:`memoize.entrybuilder.CacheEntryBuilder`)
which has control over
update_after
&expires_after
described in Tunable eviction & async refreshing
All of these elements are open for extension (you can implement and plug-in your own). Please contribute!
Example how to customize default config (everything gets overridden):
from datetime import timedelta
from memoize.configuration import MutableCacheConfiguration, DefaultInMemoryCacheConfiguration
from memoize.entrybuilder import ProvidedLifeSpanCacheEntryBuilder
from memoize.eviction import LeastRecentlyUpdatedEvictionStrategy
from memoize.key import EncodedMethodNameAndArgsKeyExtractor
from memoize.storage import LocalInMemoryCacheStorage
from memoize.wrapper import memoize
@memoize(configuration=MutableCacheConfiguration
.initialized_with(DefaultInMemoryCacheConfiguration())
.set_method_timeout(value=timedelta(minutes=2))
.set_entry_builder(ProvidedLifeSpanCacheEntryBuilder(update_after=timedelta(minutes=2),
expire_after=timedelta(minutes=5)))
.set_eviction_strategy(LeastRecentlyUpdatedEvictionStrategy(capacity=2048))
.set_key_extractor(EncodedMethodNameAndArgsKeyExtractor(skip_first_arg_as_self=False))
.set_storage(LocalInMemoryCacheStorage())
)
async def cached():
return 'dummy'
Still, you can use default configuration which:
- sets timeout for underlying method to 2 minutes;
- uses in-memory storage;
- uses method instance & arguments to infer cache key;
- stores up to 4096 elements in cache and evicts entries according to least recently updated policy;
- refreshes elements after 10 minutes & ignores unrefreshed elements after 30 minutes.
If that satisfies you, just use default config:
from memoize.configuration import DefaultInMemoryCacheConfiguration
from memoize.wrapper import memoize
@memoize(configuration=DefaultInMemoryCacheConfiguration())
async def cached():
return 'dummy'
Tunable eviction & async refreshing
Sometimes caching libraries allow providing TTL only. This may result in a scenario where when the cache entry expires latency is increased as the new value needs to be recomputed. To mitigate this periodic extra latency multiple delays are often used. In the case of memoize there are two (see :class:`memoize.entrybuilder.ProvidedLifeSpanCacheEntryBuilder`):
update_after
defines delay after which background/async update is executed;expire_after
defines delay after which entry is considered outdated and invalid.
This allows refreshing cached value in the background without any observable latency.
Moreover, if some of those background refreshes fail they will be retried still in the background.
Due to this beneficial feature, it is recommended to update_after
be significantly shorter than expire_after
.
Dog-piling proofness
If some resource is accessed asynchronously dog-piling may occur. Caches designed for synchronous python code (like built-in LRU) will allow multiple concurrent tasks to observe a miss for the same resource and will proceed to flood underlying/cached backend with requests for the same resource.
As it breaks the purpose of caching (as backend effectively sometimes is not protected with cache) memoize has built-in dog-piling protection.
Under the hood, concurrent requests for the same resource (cache key) get collapsed to a single request to the backend. When the resource is fetched all requesters obtain the result. On failure, all requesters get an exception (same happens on timeout).
An example of what it all is about:
import asyncio
from datetime import timedelta
from aiocache import cached, SimpleMemoryCache # version 0.11.1 (latest) used as example of other cache implementation
from memoize.configuration import MutableCacheConfiguration, DefaultInMemoryCacheConfiguration
from memoize.entrybuilder import ProvidedLifeSpanCacheEntryBuilder
from memoize.wrapper import memoize
# scenario configuration
concurrent_requests = 5
request_batches_execution_count = 50
cached_value_ttl_millis = 200
delay_between_request_batches_millis = 70
# results/statistics
unique_calls_under_memoize = 0
unique_calls_under_different_cache = 0
@memoize(configuration=MutableCacheConfiguration
.initialized_with(DefaultInMemoryCacheConfiguration())
.set_entry_builder(
ProvidedLifeSpanCacheEntryBuilder(update_after=timedelta(milliseconds=cached_value_ttl_millis))
))
async def cached_with_memoize():
global unique_calls_under_memoize
unique_calls_under_memoize += 1
await asyncio.sleep(0.01)
return unique_calls_under_memoize
@cached(ttl=cached_value_ttl_millis / 1000, cache=SimpleMemoryCache)
async def cached_with_different_cache():
global unique_calls_under_different_cache
unique_calls_under_different_cache += 1
await asyncio.sleep(0.01)
return unique_calls_under_different_cache
async def main():
for i in range(request_batches_execution_count):
await asyncio.gather(*[x() for x in [cached_with_memoize] * concurrent_requests])
await asyncio.gather(*[x() for x in [cached_with_different_cache] * concurrent_requests])
await asyncio.sleep(delay_between_request_batches_millis / 1000)
print("Memoize generated {} unique backend calls".format(unique_calls_under_memoize))
print("Other cache generated {} unique backend calls".format(unique_calls_under_different_cache))
predicted = (delay_between_request_batches_millis * request_batches_execution_count) // cached_value_ttl_millis
print("Predicted (according to TTL) {} unique backend calls".format(predicted))
# Printed:
# Memoize generated 17 unique backend calls
# Other cache generated 85 unique backend calls
# Predicted (according to TTL) 17 unique backend calls
if __name__ == "__main__":
asyncio.get_event_loop().run_until_complete(main())
Async cache storage
Interface for cache storage allows you to fully harness benefits of asynchronous programming (see interface of :class:`memoize.storage.CacheStorage`).
Currently memoize provides only in-memory storage for cache values (internally at RASP we have others). If you want (for instance) Redis integration, you need to implement one (please contribute!) but memoize will optimally use your async implementation from the start.