PyrateLimiter#

The request rate limiter using Leaky-bucket Algorithm.

Full project documentation can be found at pyratelimiter.readthedocs.io.

PyPI version PyPI - Python Versions codecov Maintenance PyPI license


Features#

  • Tracks any number of rate limits and intervals you want to define

  • Independently tracks rate limits for multiple services or resources

  • Handles exceeded rate limits by either raising errors or adding delays

  • Several usage options including a normal function call, a decorator

  • Out-of-the-box workable with both sync & async

  • Includes optional SQLite and Redis backends, which can be used to persist limit tracking across multiple threads, or application restarts

Installation#

PyrateLimiter supports python ^3.8

Install using pip:

pip install pyrate-limiter

Or using conda:

conda install --channel conda-forge pyrate-limiter

Quickstart#

Let’s say you want to limit 5 requests over 2 seconds, and raise an exception if the limit is exceeded:

from pyrate_limiter import Duration, Rate, Limiter, BucketFullException

rate = Rate(5, Duration.SECOND * 2)
limiter = Limiter(rate)

# Or you can pass multiple rates
# rates = [Rate(5, Duration.SECOND * 2), Rate(10, Duration.MINUTE)]
# limiter = Limiter(rates)

for request in range(6):
    try:
        limiter.try_acquire(request)
    except BucketFullException as err:
        print(err)
        print(err.meta_info)
# Bucket for item=5 with Rate limit=5/2.0s is already full
# {'error': 'Bucket for item=5 with Rate limit=5/2.0s is already full', 'name': 5, 'weight': 1, 'rate': 'limit=5/2.0s'}

Basic Usage#

Key concepts#

Clock#

  • Timestamp items

Bucket#

  • Hold items with timestamps.

  • Behave like a FIFO queue

  • It can leak - popping items that are no longer relevant out of the queue

BucketFactory#

  • BucketFactory keeps references to buckets & clocks: determine the exact time that items arrive then route them to their corresponding buckets

  • Help schedule background tasks to run buckets’ leak periodically to make sure buckets will not explode from containing too many items

  • Where user define his own logic: routing, condition-checking, timing etc…

Limiter#

The Limiter’s most important responsibility is to make user’s life as easiest as possible:

  • Sums up all the underlying logic to a simple, intuitive API to work with

  • Handles async/sync context seamlessly (everything just works by adding/removing async/await keyword to the user’s code)

  • Provides different ways of interacting with the underlying BucketFactory (plain method call, decorator, context-manager (TBA))

  • Provides thread-safety using RLock

Defining rate limits and buckets#

Consider some public API (like LinkedIn, GitHub, etc.) that has rate limits like the following:

- 500 requests per hour
- 1000 requests per day
- 10000 requests per month

You can define these rates using the Rate class. Rate class has 2 properties only: limit and interval

from pyrate_limiter import Duration, Rate

hourly_rate = Rate(500, Duration.HOUR) # 500 requests per hour
daily_rate = Rate(1000, Duration.DAY) # 1000 requests per day
monthly_rate = Rate(10000, Duration.WEEK * 4) # 10000 requests per month

rates = [hourly_rate, daily_rate, monthly_rate]

Rates must be properly ordered:

  • Rates’ intervals & limits must be ordered from least to greatest

  • Rates’ ratio of limit/interval must be ordered from greatest to least

Existing implementations of Bucket come with rate-validation when init. If you are to use your own implementation, use the validator provided by the lib

from pyrate_limiter import validate_rate_list

assert validate_rate_list(my_rates)

Then, add the rates to the bucket of your choices

from pyrate_limiter import InMemoryBucket, RedisBucket

basic_bucket = InMemoryBucket(rates)

# Or, using redis
from redis import Redis

redis_connection = Redis(host='localhost')
redis_bucket = RedisBucket.init(rates, redis_connection, "my-bucket-name")

# Async Redis would work too!
from redis.asyncio import Redis

redis_connection = Redis(host='localhost')
redis_bucket = await RedisBucket.init(rates, redis_connection, "my-bucket-name")

If you only need a single Bucket for everything, and python’s built-in time() is enough for you, then pass the bucket to Limiter then ready to roll!

from pyrate_limiter import Limiter

# Limiter constructor accepts single bucket as the only parameter,
# the rest are 3 optional parameters with default values as following
# Limiter(bucket, clock=TimeClock(), raise_when_fail=True, max_delay=None)
limiter = Limiter(bucket)

# Limiter is now ready to work!
limiter.try_acquire("hello world")

If you want to have finer grain control with routing & clocks etc, then you should use BucketFactory.

Defining Clock & routing logic with BucketFactory#

If you need more than one type of Bucket, and be able to route items to different buckets based on some condition, you can use BucketFactory to do that.

From the above steps, you already have your buckets. Now it’s time to define what Time is (funny?!). Most of the time (again!?), you can use the existing Clock backend provided by pyrate_limiter.

from pyrate_limiter.clock import TimeClock, MonotonicClock, SQLiteClock

base_clock = TimeClock()

PyrateLimiter makes no assumption about users logic, so to map coming items to their correct buckets, implement your own BucketFactory class! At minimum, there are only 2 methods require implementing

from pyrate_limiter import BucketFactory
from pyrate_limiter import AbstractBucket


class MyBucketFactory(BucketFactory):
    # You can use constructor here,
    # nor it requires to make bucket-factory work!

    def wrap_item(self, name: str, weight: int = 1) -> RateItem:
        """Time-stamping item, return a RateItem"""
        now = clock.now()
        return RateItem(name, now, weight=weight)

    def get(self, _item: RateItem) -> AbstractBucket:
        """For simplicity's sake, all items route to the same, single bucket"""
        return bucket

Creating buckets dynamically#

If more than one bucket is needed, the bucket-routing logic should go to BucketFactory get(..) method.

When creating buckets dynamically, it is needed to schedule leak for each newly created buckets.

To support this, BucketFactory comes with a predefined method call self.create(..). It is meant to create the bucket and schedule that bucket for leaking using the Factory’s clock

def create(
        self,
        clock: AbstractClock,
        bucket_class: Type[AbstractBucket],
        *args,
        **kwargs,
    ) -> AbstractBucket:
        """Creating a bucket dynamically"""
        bucket = bucket_class(*args, **kwargs)
        self.schedule_leak(bucket, clock)
        return bucket

By utilizing this, we can modify the code as following:

class MultiBucketFactory(BucketFactory):
    def __init__(self, clock):
        self.clock = clock
        self.buckets = {}

    def wrap_item(self, name: str, weight: int = 1) -> RateItem:
        """Time-stamping item, return a RateItem"""
        now = clock.now()
        return RateItem(name, now, weight=weight)

    def get(self, item: RateItem) -> AbstractBucket:
        if item.name not in self.buckets:
            # Use `self.create(..)` method to both initialize new bucket and calling `schedule_leak` on that bucket
            # We can create different buckets with different types/classes here as well
            new_bucket = self.create(YourBucketClass, *your-arguments, **your-keyword-arguments)
            self.buckets.update({item.name: new_bucket})

        return self.buckets[item.name]

Wrapping all up with Limiter#

Pass your bucket-factory to Limiter, and ready to roll!

from pyrate_limiter import Limiter

limiter = Limiter(
    bucket_factory,
    raise_when_fail=False,  # Default = True
    max_delay=1000,         # Default = None
)

item = "the-earth"
limiter.try_acquire(item)

heavy_item = "the-sun"
limiter.try_acquire(heavy_item, weight=10000)

If your bucket’s backend is async, well, we got you covered! Passing await to the limiter is enought to make it scream!

await limiter.try_acquire(item)

Alternatively, you can use Limiter.try_acquire as a function decorator. But you have to provide a mapping function that map the wrapped function’s arguments to a proper limiter.try_acquire argument - which is a tuple of (str, int) or just str

my_beautiful_decorator = limiter.as_decorator()

def mapping(some_number: int):
    return str(some_number)

@my_beautiful_decorator(mapping)
def request_function(some_number: int):
    requests.get('https://example.com')

# Async would work too!
@my_beautiful_decorator(mapping)
async def async_request_function(some_number: int):
    requests.get('https://example.com')

Limiter API#

bucket(): get list of all active buckets#

Return list of all active buckets with limiter.buckets()

dispose(bucket: int | BucketObject): dispose/remove/delete the given bucket#

Method signature:

def dispose(self, bucket: Union[int, AbstractBucket]) -> bool:
    """Dispose/Remove a specific bucket,
    using bucket-id or bucket object as param
    """

Example of usage:

active_buckets = limiter.buckets()
assert len(active_buckets) > 0

bucket_to_remove = active_buckets[0]
assert limiter.dispose(bucket_to_remove)

If a bucket is found and get deleted, calling this method will return True, otherwise False. If there is no more buckets in the limiter’s bucket-factory, all the leaking tasks will be stopped.

as_decorator(): use limiter as decorator#

Limiter can be used as decorator, but you have to provide a mapping function that maps the wrapped function’s arguments to limiter.try_acquire function arguments. The mapping function must return either a tuple of (str, int) or just a str

The decorator can work with both sync & async function

decorator = limiter.as_decorator()

def mapping(*args, **kwargs):
    return "demo", 1

@decorator(mapping)
def handle_something(*args, **kwargs):
    """function logic"""

@decorator(mapping)
async def handle_something_async(*args, **kwargs):
    """function logic"""

Weight#

Item can have weight. By default item’s weight = 1, but you can modify the weight before passing to limiter.try_acquire.

Item with weight W > 1 when consumed will be multiplied to (W) items with the same timestamp and weight = 1. Example with a big item with weight W=5, when put to bucket, it will be divided to 5 items with weight=1 + following names

BigItem(weight=5, name="item", timestamp=100) => [
    item(weight=1, name="item", timestamp=100),
    item(weight=1, name="item", timestamp=100),
    item(weight=1, name="item", timestamp=100),
    item(weight=1, name="item", timestamp=100),
    item(weight=1, name="item", timestamp=100),
]

Yet, putting this big, heavy item into bucket is expected to be transactional & atomic - meaning either all 5 items will be consumed or none of them will. This is made possible as bucket put(item) always check for available space before ingesting. All of the Bucket’s implementations provided by PyrateLimiter follows this rule.

Any additional, custom implementation of Bucket are expected to behave alike - as we have unit tests to cover the case.

See Advanced usage options below for more details.

Handling exceeded limits#

When a rate limit is exceeded, you have two options: raise an exception, or add delays.

Bucket analogy#

https://upload.wikimedia.org/wikipedia/commons/c/c4/Leaky_bucket_analogy.JPG

At this point it’s useful to introduce the analogy of “buckets” used for rate-limiting. Here is a quick summary:

  • This library implements the Leaky Bucket algorithm.

  • It is named after the idea of representing some kind of fixed capacity – like a network or service – as a bucket.

  • The bucket “leaks” at a constant rate. For web services, this represents the ideal or permitted request rate.

  • The bucket is “filled” at an intermittent, unpredicatble rate, representing the actual rate of requests.

  • When the bucket is “full”, it will overflow, representing canceled or delayed requests.

  • Item can have weight. Consuming a single item with weight W > 1 is the same as consuming W smaller, unit items - each with weight=1, with the same timestamp and maybe same name (depending on however user choose to implement it)

Rate limit exceptions#

By default, a BucketFullException will be raised when a rate limit is exceeded. The error contains a meta_info attribute with the following information:

  • name: The name of item it received

  • weight: The weight of item it received

  • rate: The specific rate that has been exceeded

Here’s an example that will raise an exception on the 4th request:

rate = Rate(3, Duration.SECOND)
bucket = InMemoryBucket([rate])
clock = TimeClock()


class MyBucketFactory(BucketFactory):

    def wrap_item(self, name: str, weight: int = 1) -> RateItem:
        """Time-stamping item, return a RateItem"""
        now = clock.now()
        return RateItem(name, now, weight=weight)

    def get(self, _item: RateItem) -> AbstractBucket:
        """For simplicity's sake, all items route to the same, single bucket"""
        return bucket


limiter = Limiter(MyBucketFactory())

for _ in range(4):
    try:
        limiter.try_acquire('item', weight=2)
    except BucketFullException as err:
        print(err)
        # Output: Bucket with Rate 3/1.0s is already full
        print(err.meta_info)
        # Output: {'name': 'item', 'weight': 2, 'rate': '3/1.0s', 'error': 'Bucket with Rate 3/1.0s is already full'}

The rate part of the output is constructed as: limit / interval. On the above example, the limit is 3 and the interval is 1, hence the Rate 3/1.

Rate limit delays#

You may want to simply slow down your requests to stay within the rate limits instead of canceling them. In that case you pass the max_delay argument the maximum value of delay (typically in ms when use human-clock).

limiter = Limiter(factory, max_delay=500) # Allow to delay up to 500ms

As max_delay has been passed as a numeric value, when ingesting item, limiter will:

  • First, try to ingest such item using the routed bucket

  • If it fails to put item into the bucket, it will call wait(item) on the bucket to see how much time remains until the bucket can consume the item again?

  • Comparing the wait value to the max_delay.

  • if max_delay >= wait: delay (wait + 50ms as latency-tolerance) using either asyncio.sleep or time.sleep until the bucket can consume again

  • if max_delay < wait: it raises LimiterDelayException if Limiter’s raise_when_fail=True, otherwise silently fail and return False

Example:

from pyrate_limiter import LimiterDelayException

for _ in range(4):
    try:
        limiter.try_acquire('item', weight=2, max_delay=200)
    except LimiterDelayException as err:
        print(err)
        # Output:
        # Actual delay exceeded allowance: actual=500, allowed=200
        # Bucket for 'item' with Rate 3/1.0s is already full
        print(err.meta_info)
        # Output: {'name': 'item', 'weight': 2, 'rate': '3/1.0s', 'max_delay': 200, 'actual_delay': 500}

Backends#

A few different bucket backends are available:

  • InMemoryBucket: using python built-in list as bucket

  • RedisBucket, using err… redis, with both async/sync support

  • PostgresBucket, using psycopg2

  • SQLiteBucket, using sqlite3

InMemoryBucket#

The default bucket is stored in memory, using python list

from pyrate_limiter import InMemoryBucket, Rate, Duration

rates = [Rate(5, Duration.MINUTE * 2)]
bucket = InMemoryBucket(rates)

This bucket only availabe in sync mode. The only constructor argument is List[Rate].

RedisBucket#

RedisBucket uses Sorted-Set to store items with key being item’s name and score item’s timestamp Because it is intended to work with both async & sync, we provide a classmethod init for it

from pyrate_limiter import RedisBucket, Rate, Duration

# Using synchronous redis
from redis import ConnectionPool
from redis import Redis

rates = [Rate(5, Duration.MINUTE * 2)]
pool = ConnectionPool.from_url("redis://localhost:6379")
redis_db = Redis(connection_pool=pool)
bucket_key = "bucket-key"
bucket = RedisBucket.init(rates, redis_db, bucket_key)

# Using asynchronous redis
from redis.asyncio import ConnectionPool as AsyncConnectionPool
from redis.asyncio import Redis as AsyncRedis

pool = AsyncConnectionPool.from_url("redis://localhost:6379")
redis_db = AsyncRedis(connection_pool=pool)
bucket_key = "bucket-key"
bucket = await RedisBucket.init(rates, redis_db, bucket_key)

The API are the same, regardless of sync/async. If AsyncRedis is being used, calling await bucket.method_name(args) would just work!

SQLiteBucket#

If you need to persist the bucket state, a SQLite backend is available. The SQLite bucket works in sync manner.

Manully create a connection to Sqlite and pass it along with the table name to the bucket class:

from pyrate_limiter import SQLiteBucket, Rate, Duration
import sqlite3

rates = [Rate(5, Duration.MINUTE * 2)]
bucket = SQLiteBucket.init_from_file(rates)

You can also pass custom arguments to the init_from_file following its signature:

class SQLiteBucket(AbstractBucket):
    @classmethod
    def init_from_file(
        cls,
        rates: List[Rate],
        table: str = "rate_bucket",
        db_path: Optional[str] = None,
        create_new_table = True
    ) -> "SQLiteBucket":
        ...

If the db_path is not provided, it will create a temporary database in memory as tempdir / "pyrate-limiter.sqlite"

PostgresBucket#

Postgres is supported, but you have to install psycopg[pool] either as an extra or as a separate package. The PostgresBucket currently does not support async.

You can use Postgres’s built-in CURRENT_TIMESTAMP as the time source with PostgresClock, or use an external custom time source.

from pyrate_limiter import PostgresBucket, Rate, PostgresClock
from psycopg_pool import ConnectionPool

connection_pool = ConnectionPool('postgresql://postgres:postgres@localhost:5432')

clock = PostgresClock(connection_pool)
rates = [Rate(3, 1000), Rate(4, 1500)]
bucket = PostgresBucket(connection_pool, "my-bucket-table", rates)

Async or Sync?#

The Limiter is basically made of a Clock backend and a Bucket backend. Depends on how each of these component works async-or-sync wise, PyrateLimiter will change its methods’ signatures to sync or async accordingly.

Here is a simple rule how to know which mode the Limiter is operating on:

If either of the backends is async-based component, the Limiter will be async.

Advanced Usage#

Component level diagram#

Time sources#

Time source can be anything from anywhere: be it python’s built-in time, or monotonic clock, sqliteclock, or crawling from world time server(well we don’t have that, but you can!).

from pyrate_limiter import TimeClock      # use python' time.time()
from pyrate_limiter import MonotonicClock # use python time.monotonic()

Clock’s abstract interface only requires implementing a method now() -> int. And it can be both sync or async.

Leaking#

Typically bucket should not hold items forever. Bucket’s abstract interface requires its implementation must be provided with leak(current_timestamp: Optional[int] = None).

The leak method when called is expected to remove any items considered outdated at that moment. During Limiter lifetime, all the buckets’ leak should be called periodically.

BucketFactory provide a method called schedule_leak to help deal with this matter. Basically, it will run as a background task for all the buckets currently in use, with interval between leak call by default is 10 seconds.

# Runnning a background task (whether it is sync/async - doesnt matter)
# calling the bucket's leak
factory.schedule_leak(bucket, clock)

You can change this calling interval by overriding BucketFactory’s leak_interval property. This interval is in miliseconds.

class MyBucketFactory(BucketFactory):
    def __init__(self, *args):
        self.leak_interval = 300

When dealing with leak using BucketFactory, the author’s suggestion is, we can be pythonic about this by implementing a constructor

class MyBucketFactory(BucketFactory):

    def constructor(self, clock, buckets):
        self.clock = clock
        self.buckets = buckets

        for bucket in buckets:
            self.schedule_leak(bucket, clock)

Concurrency#

Generally, Lock is provided at Limiter’s level, except SQLiteBucket case.

Custom backends#

If these don’t suit your needs, you can also create your own bucket backend by implementing pyrate_limiter.AbstractBucket class.

One of PyrateLimiter design goals is powerful extensibility and maximum ease of development.

It must be not only be a ready-to-use tool, but also a guide-line, or a framework that help implementing new features/bucket free of the most hassles.

Due to the composition nature of the library, it is possbile to write minimum code and validate the result:

  • Fork the repo

  • Implement your bucket with pyrate_limiter.AbstractBucket

  • Add your own create_bucket method in tests/conftest.py and pass it to the create_bucket fixture

  • Run the test suite to validate the result

If the tests pass through, then you are just good to go with your new, fancy bucket!

Reference Documentation#