Yiling-J/cacheme

Asyncio cache framework for Python

Cacheme

Asyncio cache framework with multiple cache storages.

• Organize cache better: Cache configuration with node, you can apply different strategies on different nodes.
• Multiple cache storages: in-memory/redis/mongodb/postgres..., also support chain storages.
• Multiple serializers: Pickle/Json/Msgpack serializers.
• Thundering herd protection: Simultaneously requests to same key are blocked by asyncio Event and only load from source once.
• Cache stats API: Stats of each node and colected automatically.
• Performance: See Benchemarks section.

Related projects:

• High performance in-memory cache: https://github.com/Yiling-J/theine

Table of Contents

• Installation
• Add Node
• Register Storage
• Cacheme API
• Cache Node
  • Key
  • Meta Class
  • Serializers
  • DoorKeeper
• Cache Storage
  • Local Storage
  • Redis Storage
  • MongoDB Storage
  • Sqlite Storage
  • PostgreSQL Storage
  • MySQL Storage
• How Thundering Herd Protection Works
• Benchmarks
  • continuous benchmark
  • 200k concurrent requests
  • 20k concurrent batch requests

Requirements

Python 3.7+

Installation

pip install cacheme

Multiple storages are supported by drivers. You can install the required drivers with:

pip install cacheme[redis]
pip install cacheme[aiomysql]
pip install cacheme[motor]
pip install cacheme[asyncpg]

Add Node

Node is the core part of cache. Each node has its own key function, load function and storage options. Stats of each node are collected independently. You can place all node definations into one package/module, so everyone knows exactly what is cached and how they are cached. All cacheme API are based on node.

Each node contains:

• Key attritubes and key method, which are used to generate cache key. Here the UserInfoNode is a dataclass, so __init__ method is generated automatically.
• Async load method, which will be called to load data from data source on cache missing. This method can be omitted if you use Memoize decorator only.
• Meta class, node cache configurations. See Cache Node

import cacheme
from dataclasses import dataclass
from cacheme.serializer import MsgPackSerializer

@dataclass
class UserInfoNode(cacheme.Node):
    user_id: int

    def key(self) -> str:
        return f"user:{self.user_id}:info"

    async def load(self) -> Dict:
        user = get_user_from_db(self.user_id)
        return serialize(user)

    class Meta(cacheme.Node.Meta):
        version = "v1"
        caches = [cacheme.Cache(storage="my-redis", ttl=None)]
        serializer = MsgPackSerializer()

This simple example use a cache storage called "my-redis", which will be registered next step. Also we use MsgPackSerializer here to dump and load data from redis. See [Cache Node] for more details.

Register Storage

Register a redis storage called "my-redis", which you can use in node meta data. The register_storage is asynchronous and will try to establish connection to cache store. See [Cache Storage] for more details.

import cacheme

await cacheme.register_storage("my-redis", cacheme.Storage(url="redis://localhost:6379"))

Cacheme API

get: get data from single node.

user = await cacheme.get(UserInfoNode(user_id=1))

get_all: get data from multiple nodes, same node type.

users = await cacheme.get_all([UserInfoNode(user_id=1), UserInfoNode(user_id=2)])

invalidate: invalidate a node, remove data from cache.

await cacheme.invalidate(UserInfoNode(user_id=1))

refresh: reload node data using load method.

await cacheme.refresh(UserInfoNode(user_id=1))

Memoize: memoize function with this decorator.

Decorate your function with cacheme.Memoize decorator and cache node. Cacheme will load data using the decorated function and ignore load method. Because your function may contain variable number of args/kwargs, we need one more step to map between args/kwargs to node. The decorated map function should have same input signature as memoized function, and return a cache node.

@cacheme.Memoize(UserInfoNode)
async def get_user_info(user_id: int) -> Dict:
    return {}

# function name is not important, so just use _ here
@get_user_info.to_node
def _(user_id: int) -> UserInfoNode:
    return UserInfoNode(user_id=user_id)

nodes: list all nodes.

nodes = cacheme.nodes()

stats: get node stats.

metrics = cacheme.stats(UserInfoNode)

metrics.request_count() # total request count
metrics.hit_count() # total hit count
metrics.hit_rate() # hit_count/request_count
metrics.miss_count() # (request_count - hit_count)/request_count
metrics.miss_rate() # miss_count/request_count
metric.load_success_count() # total load success count
metrics.load_failure_count() # total load fail count
metrics.load_failure_rate() # load_failure_count/load_count
metrics.load_count() # total load count
metrics.total_load_time() # total load time in nanoseconds
metrics.average_load_time() # total_load_time/load_count

set_prefix: set prefix for all keys. Default prefix is cacheme. Change prefix will invalid all keys, because prefix is part of the key.

cacheme.set_prefix("mycache")

Cache Node

Key

Generated cache key will be: {prefix}:{key()}:{Meta.version}. So change version will invalid all keys automatically.

Meta Class

• version[str]: Version of node, will be used as suffix of cache key.
• caches[List[Cache]]: Caches for node. Each Cache has 2 attributes, storage[str] and ttl[Optional[timedelta]]. storage is the name you registered with register_storage and ttl is how long this cache will live. Cacheme will try to get data from each cache from left to right. In most cases, use single cache or [local, remote] combination.
• serializer[Optional[Serializer]]: Serializer used to dump/load data. If storage type is local, serializer is ignored. See Serializers.
• doorkeeper[Optional[DoorKeeper]]: See DoorKeeper.

Multiple caches example. Local cache is not synchronized, so set a much shorter ttl compared to redis one. Then we don't need to worry too much about stale data.

import cacheme
from dataclasses import dataclass
from datetime import timedelta
from cacheme.serializer import MsgPackSerializer

@dataclass
class UserInfoNode(cacheme.Node):
    user_id: int

    def key(self) -> str:
        return f"user:{self.user_id}:info"

    async def load(self) -> Dict:
        user = get_user_from_db(self.user_id)
        return serialize(user)

    class Meta(cacheme.Node.Meta):
        version = "v1"
        caches = [
            cacheme.Cache(storage="local", ttl=timedelta(seconds=30)),
            cacheme.Cache(storage="my-redis", ttl=timedelta(days=10))
        ]
        serializer = MsgPackSerializer()

Cacheme also support creating Node dynamically, you can use this together with Memoize decorator:

@Memoize(cacheme.build_node("TestNodeDynamic", "v1", [Cache(storage="local", ttl=None)]))
async def fn(a: int) -> int:
    return 1


@fn.to_node
def _(a: int) -> cacheme.DynamicNode:
    return DynamicNode(key=f"bar:{a}")

Here we use DynamicNode, which only support one param: key

Serializers

Cacheme provides serveral builtin serializers, you can also write your own serializer.

• PickleSerializer: All python objects.
• JSONSerializer: Use pydantic_encoder and json, support python primitive types, dataclass, pydantic model. See pydantic types.
• MsgPackSerializer: Use pydantic_encoder and msgpack, support python primitive types, dataclass, pydantic model. See pydantic types.

serializer with compression, use zlib level-3

• CompressedPickleSerializer
• CompressedJSONSerializer
• CompressedMsgPackSerializer

DoorKeeper

Idea from TinyLfu paper.

The Doorkeeper is a regular Bloom filter placed in front of the cahce. Upon item arrival, we first check if the item is contained in the Doorkeeper. If it is not contained in the Doorkeeper (as is expected with first timers and tail items), the item is inserted to the Doorkeeper and otherwise, it is inserted to the cache.

from cacheme import BloomFilter

@dataclass
class UserInfoNode(cacheme.Node):

    class Meta(cacheme.Node.Meta):
        # size 100000, false positive probability 0.01
        doorkeeper = BloomFilter(100000, 0.01)

BloomFilter is cleared automatically when requests count == size.

Cache Storage

Local Storage

Local storage use the state-of-the-art library Theine to store data. If your use case in simple, also consider using Theine directly, which will have the best performance.

# lru policy
Storage(url="local://lru", size=10000)

# w-tinylfu policy
Storage(url="local://tlfu", size=10000)

Parameters:

• url: local://{policy}. 2 policies are currently supported:

  • lru
  • tlfu: W-TinyLfu policy

• size: size of the storage. Policy will be used to evict key when cache is full.

Redis Storage

Storage(url="redis://localhost:6379")

# cluster
Storage(url="redis://localhost:6379", cluster=True)

Parameters:

• url: redis connection url.
• cluster: bool, cluster or not, default False.
• pool_size: connection pool size, default 100.

MongoDB Storage

To use mongodb storage, create index first. See mongo.js

Storage(url="mongodb://test:password@localhost:27017",database="test",collection="cache")

Parameters:

• url: mongodb connection url.
• database: mongodb database name.
• collection: mongodb collection name.
• pool_size: connection pool size, default 50.

Sqlite Storage

To use sqlite storage, create table and index first. See sqlite.sql

Storage(url="sqlite:///test", table="cache")

Parameters:

• url: sqlite connection url.
• table: cache table name.
• pool_size: connection pool size, default 50.

PostgreSQL Storage

To use postgres storage, create table and index first. See postgresql.sql

Storage(url="postgresql://username:password@127.0.0.1:5432/test", table="cache")

Parameters:

• url: postgres connection url.
• table: cache table name.
• pool_size: connection pool size, default 50.

MySQL Storage

To use mysql storage, create table and index first. See mysql.sql

Storage("mysql://username:password@localhost:3306/test", table="cache")

Parameters:

• url: mysql connection url.
• table: cache table name.
• pool_size: connection pool size, default 50.

How Thundering Herd Protection Works

If you are familar with Go singleflight, you may have an idea how Cacheme works. Cacheme group concurrent requests to same resource(node) into a singleflight with asyncio Event, which will load from remote cache OR data source only once. That's why in next Benchmarks section, you will find Cacheme even reduce total redis GET command count under high concurrency.

Benchmarks

continuous benchmark

https://github.com/Yiling-J/cacheme-benchmark

200k concurrent requests

aiocache: https://github.com/aio-libs/aiocache

cashews: https://github.com/Krukov/cashews

source code: https://github.com/Yiling-J/cacheme/blob/master/benchmarks/trace.py

How this benchmark run:

1. Initialize Cacheme/Aiocache/Cashews with Redis backend, use Redis blocking pool and set pool size to 100.
2. Decorate Aiocache/Cashews/Cacheme with a function which accept a number and sleep 0.1s. This function also record how many times it is called.
3. Register Redis response callback, so we can know how many times GET command are called.
4. Create 200k coroutines use a zipf generator and put them in async queue(around 50k-60k unique numbers).
5. Run coroutines in queue with N concurrent workers.
6. Collect results.

Identifier:

• Cacheme: Cacheme redis storage
• Aiocahce: Aiocahce cached decorator
• Cashews: Cashews cache decorate
• Cacheme-2: Cacheme use cache chain [local, redis]
• Aiocache-2: Aiocache cached_stampede decorator
• Cashews-2: Cashews decorator with lock=True

Result:

• Time: How long it takes to finish bench.
• Redis GET: How many times Redis GET command are called, use this to evaluate pressure to remote cache server.
• Load Hits: How many times the load function(which sleep 0.1s) are called, use this to evaluate pressure to load source(database or something else).

1k concurrency

	Time	Redis GET	Load Hits
Cacheme	25 s	166454	55579
Cacheme-2	20 s	90681	55632
Aiocache	46 s	200000	56367
Aiocache-2	63 s	256492	55417
Cashews	51 s	200000	56920
Cashews-2	134 s	200000	55450

10k concurrency

	Time	Redis GET	Load Hits
Cacheme	25 s	123704	56736
Cacheme-2	19 s	83750	56635
Aiocache	67 s	200000	62568
Aiocache-2	113 s	263195	55507
Cashews	68 s	200000	66036
Cashews-2	175 s	200000	55709

100k concurrency

	Time	Redis GET	Load Hits
Cacheme	24 s	60990	56782
Cacheme-2	22 s	55762	55588
Aiocache	80 s	200000	125085
Aiocache-2	178 s	326417	65598
Cashews	88 s	200000	87894
Cashews-2	236 s	200000	55647

20k concurrent batch requests

source code: https://github.com/Yiling-J/cacheme/blob/master/benchmarks/trace.py

How this benchmark run:

1. Initialize Cacheme with Redis backend, use Redis blocking pool and set pool size to 100.
2. Decorate Cacheme with a function which accept a number and sleep 0.1s. This function also record how many times it is called.
3. Register Redis response callback, so we can know how many times MGET command are called.
4. Create 20k get_all coroutines use a zipf generator and put them in async queue(around 50k-60k unique numbers). Each get_all request will get 20 unique numbers in batch. So totally 400k numbers.
5. Run coroutines in queue with N concurrent workers.
6. Collect results.

Result:

• Time: How long it takes to finish bench.
• Redis MGET: How many times Redis MGET command are called, use this to evaluate pressure to remote cache server.
• Load Hits: How many times the load function(which sleep 0.1s) are called, use this to evaluate pressure to load source(database or something else).

1k concurrency

	Time	Redis MGET	Load Hits
Cacheme	12 s	9996	55902

10k concurrency

	Time	Redis MGET	Load Hits
Cacheme	11 s	9908	42894