redis-func-cache

redis_func_cache

A Python library that provides decorators for caching function results in Redis, supporting multiple serialization formats and caching strategies, as well as asynchronous operations.

Abstract

redis_func_cache is a Python library that provides decorators for caching function results in Redis, similar to the caching functionality offered by the standard library. Like functools module, it includes useful decorators such as lru_cache, which are valuable for implementing memoization.

When we need to cache function return values across multiple processes or machines, Redis can be used as a distributed backend. The purpose of this project is to provide simple and clean decorator functions to use Redis as a cache backend. It implements caches with various eviction/replacement policies such as LRU, FIFO, RR, and LFU. (Refer to Cache Replacement Policies on Wikipedia for more details.)

Here is a simple example:

1. Let's first startup a redis server at 127.0.0.1:6379, eg:

   docker run -it --rm -p 6379:6379 redis:alpine
   

2. Then install the library on your python environment:

   pip install -U redis_func_cache
   

3. Finally run the following Python code:

   from time import sleep, time
   from redis import Redis
   from redis_func_cache import LruTPolicy, RedisFuncCache
   
   # Create a redis client
   redis_client = Redis.from_url("redis://")
   
   # Create an lru cache, it connects Redis by previous created redis client
   lru_cache = RedisFuncCache(__name__, LruTPolicy, redis_client)
   
   @lru_cache # Decorate a function to cache its result
   def a_slow_func():
       sleep(10) # Sleep to simulate a slow operation
       return "OK"
   
   t = time()
   r = a_slow_func()
   print(f"duration={time()-t}, {r=}")
   
   t = time()
   r = a_slow_func()
   print(f"duration={time()-t}, {r=}")
   

The output should be like:

duration=10.002939939498901, r='OK'
duration=0.0008025169372558594, r='OK'

We can see that the second call to a_slow_func() is served from the cache, which is much faster than the first call.

Features

• Based on redis-py, the official Python client for Redis.
• Simple decorator syntax.
• Both asynchronous and synchronous I/O support.
• Redis cluster support.
• Supports multiple cache eviction policies: LRU, FIFO, LFU, RR ...

Install

• install from PyPI:

  pip install -U redis_func_cache
  

• install from source:

  git clone https://github.com/tanbro/redis_func_cache.git
  cd redis_func_cache
  pip install .
  

• Or install from Github directly:

  pip install git+https://github.com/tanbro/redis_func_cache.git@main
  

Data structure

The library combines a pair of Redis data structures to manage cache data:

• The first one is a sorted set, which stores the hash values of the decorated function calls along with a score for each item.

  When the cache reaches its maximum size, the score is used to determine which item to evict.

• The second one is a hash map, which stores the hash values of the function calls and their corresponding return values.

This can be visualized as follows:

The main idea of eviction policy is that the cache keys are stored in a set, and the cache values are stored in a hash map. Eviction is performed by removing the lowest-scoring item from the set, and then deleting the corresponding field and value from the hash map.

Here is an example showing how the LRU cache's eviction policy works(maximum size is 3):

The RedisFuncCache executes a decorated function with specified arguments and cache its result. Here's a breakdown of the steps:

1. Initialize Scripts: Retrieve two Lua script objects for cache hitting and update from policy.lua_scripts.
2. Calculate Keys and Hash: Compute the cache keys using policy.calc_keys, compute the hash value using policy.calc_hash, and compute any additional arguments using policy.calc_ext_args.
3. Attempt Cache Retrieval: Attempt retrieving a cached result. If a cache hit occurs, deserialize and return the cached result.
4. Execute User Function: If no cache hit occurs, execute the decorated function with the provided arguments and keyword arguments.
5. Serialize Result and Cache: Serialize the result of the user function and store it in redis.
6. Return Result: Return the result of the decorated function.

flowchart TD
    A[Start] --> B[Initialize Scripts]
    B --> C{Scripts Valid?}
    C -->|Invalid| D[Raise RuntimeError]
    C -->|Valid| E[Calculate Keys and Hash]
    E --> F[Attempt Cache Retrieval]
    F --> G{Cache Hit?}
    G -->|Yes| H[Deserialize and Return Cached Result]
    G -->|No| I[Execute User Function]
    I --> J[Serialize Result]
    J --> K[Store in Cache]
    K --> L[Return User Function Result]

Basic Usage

First example

Using an LRU cache to decorate a recursive Fibonacci function:

from redis import Redis
from redis_func_cache import RedisFuncCache, LruTPolicy

redis_client = Redis("redis://")

lru_cache = RedisFuncCache("my-first-lru-cache", LruTPolicy, redis_client)

@lru_cache
def fib(n):
    if n <= 1:
        return n
    if n == 2:
        return 1
    return fib(n - 1) + fib(n - 2)

In this example, we first create a Redis client, then create a RedisFuncCache instance with the Redis client and LruTPolicy as its arguments. Next, we use the @lru_cache decorator to decorate the fib function. This way, each computed result is cached, and subsequent calls with the same parameters retrieve the result directly from the cache, thereby improving performance.

It works almost the same as the standard library's functools.lru_cache, except that it uses Redis as the backend instead of the local machine's memory.

If we browse the Redis database, we can find the pair of keys' names look like:

• func-cache:my-first-lru-cache:lru_t:0

  The key (with 0 suffix) is a sorted set that stores the hash of function invoking and their corresponding scores.

• func-cache:my-first-lru-cache:lru_t:1

  The key (with 1 suffix) is a hash map. Each key field in it is the hash value of a function invoking, and the value filed is the return value of the function.

❗ Important:
The name SHOULD be unique for each RedisFuncCache instance. Therefore, we need to choose a unique name carefully using the name argument.

Async functions

To decorate async functions, we shall pass a Async Redis client to RedisFuncCache's client argument:

from redis.asyncio import Redis as AsyncRedis
from redis_func_cache import RedisFuncCache, LruTPolicy

redis_client = AsyncRedis.from_url("redis://")
my_async_cache = RedisFuncCache(__name__, LruTPolicy, redis_client)

@my_async_cache
async def my_async_func(*args, **kwargs):
    ...

❗ Attention:

• When a RedisFuncCache is created with an async Redis client, it can only be used to decorate async functions. These async functions will be decorated with an asynchronous wrapper, and the I/O operations between the Redis client and server will be performed asynchronously.
• Conversely, a synchronous RedisFuncCache can only decorate synchronous functions. These functions will be decorated with a synchronous wrapper, and I/O operations with Redis will be performed synchronously.

Eviction policies

If you want to use other eviction policies, you can specify another policy class as the second argument of RedisFuncCache. For example, we use FifoPolicy to implement a FIFO cache:

from redis import Redis
from redis_func_cache import RedisFuncCache, FifoPolicy

redis_client = Redis.from_url("redis://")
fifo_cache = RedisFuncCache("my-cache-2", FifoPolicy, redis_client)

@fifo_cache
def func1(x):
    ...

Use RrPolicy to implement a random-remove cache:

from redis import Redis
from redis_func_cache import RedisFuncCache, RrPolicy

redis_client = Redis.from_url("redis://")
rr_cache = RedisFuncCache("my-cache-3", RrPolicy, redis_client)

@rr_cache
def func2(x):
    ...

So far, the following cache eviction policies are available:

• LruTPolicy

  💡 Tip:
  LRU-T stands for LRU on timestamp. It is a pseudo-LRU policy that approximates the behavior of LRU but is not as precise. The policy removes items based on their invocation timestamps, which may not always accurately reflect the least recently used item due to potential timestamp inaccuracies.

  Despite this limitation, LRU-T is highly recommended for common use cases. It offers better performance compared to the traditional LRU policy and provides sufficient accuracy for most applications.

• FifoPolicy: first in first out

• LfuPolicy: least frequently used

• LruPolicy: least recently used

• MruPolicy: most recently used

• RrPolicy: random remove

ℹ️ Info:
Explore source codes in directory src/redis_func_cache/policies for more details.

Multiple Redis key pairs

As described above, the cache keys are currently in a paired form, where all decorated functions share the same two keys. However, there may be instances where we want a unique key pair for each decorated function.

One solution is to use different RedisFuncCache instances to decorate different functions.

Another way is to use a policy that stores cache data in different Redis key pairs for each function. There are several policies to do that out of the box. For example, we can use LruTMultiplePolicy for an LRU cache that has multiple different Redis key pairs to store return values of different functions, and each function has a standalone keys pair:

from redis import  Redis
from redis_func_cache import RedisFuncCache, LruTMultiplePolicy

redis_client = Redis.from_url("redis://")
cache = RedisFuncCache("my-cache-4", LruTMultiplePolicy, redis_client)

@cache
def func1(x):
    ...

@cache
def func2(x):
    ...

In the example, LruTMultiplePolicy inherits BaseMultiplePolicy which implements how to store cache keys and values for each function.

When called, we can see such keys in the Redis database:

• key pair for func1:

  • func-cache:my-cache-4:lru_t-m:__main__:func1#<hash1>:0
  • func-cache:my-cache-4:lru_t-m:__main__:func1#<hash1>:1

• key pair for func2:

  • func-cache:my-cache-4:lru_t-m:__main__:func2#<hash2>:0
  • func-cache:my-cache-4:lru_t-m:__main__:func2#<hash2>:1

where <hash1> and <hash2> are the hash values of the definitions of func1 and func2 respectively.

Policies that store cache in multiple Redis key pairs are:

• FifoMultiplePolicy
• LfuMultiplePolicy
• LruMultiplePolicy
• MruMultiplePolicy
• RrMultiplePolicy
• LruTMultiplePolicy

Redis Cluster support

We have already known that the library implements cache algorithms based on a pair of Redis data structures, the two MUST be in a same Redis node, or it will not work correctly.

While a Redis cluster will distribute keys to different nodes based on the hash value, we need to guarantee that two keys are placed on the same node. Several cluster policies are provided to achieve this. These policies use the {...} pattern in key names.

For example, here we use a LruTClusterPolicy to implement a cluster-aware LRU cache:

from redis import Redis
from redis_func_cache import RedisFuncCache, LruTClusterPolicy

redis_client = Redis.from_url("redis://")
cache = RedisFuncCache("my-cluster-cache", LruTClusterPolicy, redis_client)

@cache
def my_func(x):
    ...

Thus, the names of the key pair may be like:

• func-cache:{my-cluster-cache:lru_t-c}:0
• func-cache:{my-cluster-cache:lru_t-c}:1

Notice what is in {...}: the Redis cluster will determine which node to use by the {...} pattern rather than the entire key string.

Therefore, all cached results for the same cache instance will be stored in the same node, irrespective of the functions involved.

Policies that support cluster are:

• FifoClusterPolicy
• LfuClusterPolicy
• LruClusterPolicy
• MruClusterPolicy
• RrClusterPolicy
• LruTClusterPolicy

Redis Cluster support with multiple key pairs

This policy ensures that all cached results for the same function are stored in the same node. Meanwhile, results of different functions may be stored in different nodes.

Policies that support both cluster and store cache in multiple Redis key pairs are:

• FifoClusterMultiplePolicy
• LfuClusterMultiplePolicy
• LruClusterMultiplePolicy
• MruClusterMultiplePolicy
• RrClusterMultiplePolicy
• LruTClusterMultiplePolicy

Max size and expiration time

The RedisFuncCache instance has two arguments to control the maximum size and expiration time of the cache:

• maxsize: the maximum number of items that the cache can hold.

  When the cache reaches its maxsize, adding a new item will cause an existing cached item to be removed according to the eviction policy.

  ℹ️ Note:
  For "multiple" policies, each decorated function has its own standalone data structure, so the value represents the maximum size of each individual data structure.

• ttl: The expiration time (in seconds) for the cache data structure.

  The cache's redis data structure will expire and be released after the specified time. Each time the cache is accessed, the expiration time will be reset.

  ℹ️ Note:
  For "multiple" policies, each decorated function has its own standalone data structure, so the ttl value represents the expiration time of each individual data structure. The expiration time will be reset each time the cache is accessed individually.

Complex return types

The return value's (de)serializer is JSON (json module of std-lib) by default, which does not work with complex objects. However, we can still use pickle. This can be achieved by specifying either the serializers argument of RedisFuncCache, or the serializer and deserializer arguments of the decorator:

💡 Example:

import pickle
from redis import Redis
from redis_func_cache import RedisFuncCache, LruTPolicy

# like this:
my_pickle_cache = RedisFuncCache(
    __name__,
    LruTPolicy,
    lambda: Redis.from_url("redis://"),
    serializer="pickle"
)

# or like this:
my_pickle_cache1 = RedisFuncCache(
    __name__,
    LruTPolicy,
    lambda: Redis.from_url("redis://"),
    serializer=(pickle.dumps, pickle.loads)
)

# or just like this:
cache = RedisFuncCache(__name__, LruTPolicy, lambda: Redis.from_url("redis://"))

@cache(serializer=pickle.loads, deserializer=pickle.dumps)
def my_func_with_complex_return_value(x):
    ...

Other serialization functions also should be workable, such as simplejson, cJSON, msgpack, cloudpickle, etc.

⚠️ Warning:
pickle is considered a security risk, and also cant not be used with runtime/version sensitive data. Use it cautiously and only when necessary. It's a good practice to only cache functions that return JSON serializable simple data.

Advanced Usage

Custom result serializer

The result of the decorated function is serialized by default using JSON (via the json module from the standard library) and then saved to Redis.

To utilize alternative serialization methods, such as msgpack, you have two options:

1. Specify the serializer argument in the constructor of RedisFuncCache, where the argument is a tuple of (serializer, deserializer):

   This method applies globally: all functions decorated by this cache will use the specified serializer.

   For example:

   import msgpack
   from redis import Redis
   from redis_func_cache import RedisFuncCache, LruTPolicy
   
   cache = RedisFuncCache(
       __name__,
       LruTPolicy,
       lambda: Redis.from_url("redis://"),
       serializer=(msgpack.packb, msgpack.unpackb)
    )
   
   @cache
   def my_func(x):
      ...
   

2. Specify the serializer and deserializer arguments directly in the decorator:

   This method applies on a per-function basis: only the decorated function will use the specified serializer.

   For example:

   import msgpack
   from redis import Redis
   from redis_func_cache import RedisFuncCache, LruTPolicy
   
   cache = RedisFuncCache(__name__, LruTPolicy, lambda: Redis.from_url("redis://"))
   
   @cache(serializer=msgpack.packb, deserializer=msgpack.unpackb)
   def my_func(x):
      ...
   

Custom key format

An instance of RedisFuncCache calculate key pair names string by calling method calc_keys of its policy. There are four basic policies that implement respective kinds of key formats:

• BaseSinglePolicy: All functions share the same key pair, Redis cluster is NOT supported.

  The format is: <prefix><name>:<__key__>:<0|1>

• BaseMultiplePolicy: Each function has its own key pair, Redis cluster is NOT supported.

  The format is: <prefix><name>:<__key__>:<function_name>#<function_hash>:<0|1>

• BaseClusterSinglePolicy: All functions share the same key pair, Redis cluster is supported.

  The format is: <prefix>{<name>:<__key__>}:<0|1>

• BaseClusterMultiplePolicy: Each function has its own key pair, and Redis cluster is supported.

  The format is: <prefix><name>:<__key__>:<function_name>#{<function_hash>}:<0|1>

Variables in the format string are defined as follows:

prefix	prefix argument of RedisFuncCache
name	name argument of RedisFuncCache
__key__	__key__ attribute the policy class used in RedisFuncCache
function_name	full name of the decorated function
function_hash	hash value of the decorated function

0 and 1 at the end of the keys are used to distinguish between the two data structures:

• 0: a sorted or unsorted set, used to store the hash value and sorting score of function invoking
• 1: a hash table, used to store the return value of the function invoking

If you want to use a different format, you can subclass AbstractPolicy or any of above policy classes, and implement calc_keys method, then pass the custom policy class to RedisFuncCache.

The following example demonstrates how to custom key format for an LRU policy:

from __future__ import annotations

from typing import TYPE_CHECKING, Any, Callable, Mapping, Sequence, Tuple, override

import redis
from redis_func_cache import RedisFuncCache
from redis_func_cache.policies.abstract import AbstractPolicy
from redis_func_cache.mixins.hash import PickleMd5HashMixin
from redis_func_cache.mixins.policies import LruScriptsMixin

if TYPE_CHECKING:
    from redis.typing import KeyT


def redis_factory():
    return redis.from_url("redis://")


MY_PREFIX = "my_prefix"


class MyPolicy(LruScriptsMixin, PickleMd5HashMixin, AbstractPolicy):
    __key__ = "my_key"

    @override
    def calc_keys(
            self, f: Callable | None = None, args: Sequence | None = None, kwds: Mapping[str, Any] | None = None
    ) -> Tuple[KeyT, KeyT]:
        k = f"{self.cache.prefix}-{self.cache.name}-{f.__name__}-{self.__key__}"
        return f"{k}-set", f"{k}-map"


my_cache = RedisFuncCache(name="my_cache", policy=MyPolicy, client=redis_factory, prefix=MY_PREFIX)


@my_cache
def my_func(*args, **kwargs):
    ...

In the example, we'll get a cache generates redis keys separated by -, instead of :, prefixed by "my-prefix", and suffixed by "set" and "map", rather than "0" and "1". The key pair names could be like my_prefix-my_cache_func-my_key-set and my_prefix-my_cache_func-my_key-map.

LruScriptsMixin tells the policy which lua script to use, and PickleMd5HashMixin tells the policy to use pickle to serialize and md5 to calculate the hash value of the function.

❗ Important:
The calculated key name SHOULD be unique for each RedisFuncCache instance.

BaseSinglePolicy, BaseMultiplePolicy, BaseClusterSinglePolicy, and BaseClusterMultiplePolicy calculate their key names by calling the calc_keys method, which uses its __key__ attribute and the name property of the RedisFuncCache instance. If you subclass any of these classes, you should override the __key__ attribute to ensure that the key names remain unique.

Custom Hash Algorithm

When the library performs a get or put action with redis, the hash value of the function invocation will be used.

For the sorted set data structures, the hash value will be used as the member. For the hash map data structure, the hash value will be used as the hash field.

The algorithm used to calculate the hash value is defined in AbstractHashMixin, it can be described as below:

import hashlib

class AbstractHashMixin:
    __hash_config__ = ...

    ...

    def calc_hash(self, f = None, args = None, kwds = None):
        if not callable(f):
            raise TypeError(f"Can not calculate hash for {f=}")
        conf = self.__hash_config__
        h = hashlib.new(conf.algorithm)
        h.update(f"{f.__module__}:{f.__qualname__}".encode())
        h.update(f.__code__.co_code)
        if args is not None:
            h.update(conf.serializer(args))
        if kwds is not None:
            h.update(conf.serializer(kwds))
        if conf.decoder is None:
            return h.digest()
        return conf.decoder(h)

As the code snippet above, the hash value is calculated by the full name of the function, the bytes code of the function, the arguments and keyword arguments —— they are serialized and hashed, then decoded.

The serializer and decoder are defined in the __hash_config__ attribute of the policy class and are used to serialize arguments and decode the resulting hash. By default, the serializer is pickle and the decoder uses the md5 algorithm. If no decoder is specified, the hash value is returned as bytes.

This configuration can be illustrated as follows:

flowchart TD
    A[Start] --> B{Is f callable?}
    B -->|No| C[Throw TypeError]
    B -->|Yes| D[Get config conf]
    D --> E[Create hash object h]
    E --> F[Update hash: module name and qualified name]
    F --> G[Update hash: function bytecode]
    G --> H{Are args not None?}
    H -->|Yes| I[Update hash: serialize args]
    H -->|No| J{Are kwds not None?}
    I --> J
    J -->|Yes| K[Update hash: serialize kwds]
    J -->|No| L{Is conf.decoder None?}
    K --> L
    L -->|Yes| M[Return digest bytes]
    L -->|No| N[Return decoded digest]

If we want to use a different algorithm, we can select a mixin hash class defined in src/redis_func_cache/mixins/hash.py. For example:

• To serialize the function with JSON, use the SHA1 hash algorithm, store hex string in redis, you can choose the JsonSha1HexHashMixin class.
• To serialize the function with pickle, use the MD5 hash algorithm, store base64 string in redis, you can choose the PickleMd5Base64HashMixin class.

These mixin classes provide alternative hash algorithms and serializers, allowing for flexible customization of the hashing behavior. The following example shows how to use the JsonSha1HexHashMixin class:

from redis import Redis
from redis_func_cache import RedisFuncCache
from redis_func_cache.policies.abstract import AbstractPolicy
from redis_func_cache.mixins.hash import JsonSha1HexHashMixin
from redis_func_cache.mixins.policies import LruScriptsMixin


class MyLruPolicy(LruScriptsMixin, JsonSha1HexHashMixin, AbstractPolicy):
    __key__ = "my-lru"

my_json_sha1_hex_cache = RedisFuncCache(
    name="json_sha1_hex",
    policy=MyLruPolicy,
    client=lambda: Redis.from_url("redis://")
)

Or even write an entire new algorithm. For that, we subclass AbstractHashMixin and override the calc_hash method. For example:

from __future__ import annotations

import hashlib
from typing import TYPE_CHECKING, override, Any, Callable, Mapping, Sequence
import cloudpickle
from redis import Redis
from redis_func_cache import RedisFuncCache
from redis_func_cache.policies.abstract import AbstractPolicy
from redis_func_cache.mixins.hash import AbstractHashMixin
from redis_func_cache.mixins.policies import LruScriptsMixin

if TYPE_CHECKING:  # pragma: no cover
    from redis.typing import KeyT


class MyHashMixin(AbstractHashMixin):
    @override
    def calc_hash(
        self,
        f: Callable | None = None,
        args: Sequence | None = None,
        kwds: Mapping[str, Any] | None = None
    ) -> KeyT:
        assert callable(f)
        dig = hashlib('balck2b')
        dig.update(f.__qualname__.encode())
        dig.update(cloudpickle.dumps(args))
        dig.update(cloudpickle.dumps(kwds))
        return dig.hexdigest()


class MyLruPolicy2(LruScriptsMixin, MyHashMixin, AbstractPolicy):
    __key__ = "my-lru2"


my_custom_hash_cache = RedisFuncCache(
    name=__name__,
    policy=MyLruPolicy2,
    client=redis_client
)

redis_client = Redis.from_url("redis://")


@my_custom_hash_cache
def some_func(*args, **kwargs):
    ...

💡 Tip:
The purpose of the hash algorithm is to ensure the isolation of cached return values for different function invocations. Therefore, you can generate unique key names using any method, not just hashes.

Known Issues

• Cannot decorate a function that has an argument not serializable by pickle or other serialization libraries.

  • For a common method defined inside a class, the class must be serializable; otherwise, the first self argument cannot be serialized.
  • For a class method (decorated by @classmethod), the class type itself, i.e., the first cls argument, must be serializable.

• Compatibility with other decorators is not guaranteed.

• The cache eviction policies are mainly based on Redis sorted set's score ordering. For most policies, the score is a positive integer. Its maximum value is 2^32-1 in Redis, which limits the number of times of eviction replacement. Redis will return an overflow error when the score overflows.

• High concurrency or long-running decorated functions may result in unexpected cache misses and increased I/O operations. This can occur because the result value might not be saved quickly enough before the next call can hit the cache again.

• Generator functions are not supported.

• If there are multiple RedisFuncCache instances with the same name, they may share the same cache data. This may lead to serious errors, so we should avoid using the same name for different instances.

Testing

A Docker Compose file for unit testing is provided in the docker directory to simplify the process. You can run it by executing:

cd docker
docker compose up --abort-on-container-exit