LimitLion

A token bucket rate limiting throttle using Redis as the backend. Inspired by Stripe's Scaling your API with rate limiters blog post. Can be used to control processing rates from one to many processes. Potential implementations include protecting databases from high processing rates, orchestrating queue consumer processes, or enforcing HTTP request rate limits.

Following is a simple example of a throttle named test that allows 5 requests per second (RPS) with a burst factor of 2 using a 8 second window and requesting 1 token (default) for each unit of work. Look in the examples directory for more.

redis = redis.Redis('localhost', 6379)
throttle_configure(redis)
while True:
    allowed, tokens, sleep = throttle('test', 5, 2, 8)
    if allowed:
        print ('Do work here')
    else:
        print ('Sleeping {}'.format(sleep))
        time.sleep(sleep)

Design

The rate limiting logic uses a classic token bucket algorithm but is implemented entirely as a Lua Redis script. It leverages the Redis TIME command which ensures fair microsecond resolution across all callers independent of the caller's clock. Note that buckets start and end on whole seconds.

Redis 3.2+ is required because replicate_commands() is used to support using the TIME command in a Lua script.

Configuring

Default values for RPS, burst factor and window size are supplied to the throttle Lua script. The Lua script creates a throttle:[throttle name]:knobs hash with these values if it does not yet exist in Redis. The script then uses the values in that knobs hash for the token bucket calculations. Each call also sets the TTL for the knobs key to 7 days so it will remain in Redis as long as the throttle has been active in the last week.

Since these settings are stored in Redis a separate process can be used to adjust them on the fly. This could simply be manually issuing the Redis command to change the RPS or a more sophisicated process that polls Prometheus metrics to determine the current load on your database and adjust the RPS accordingly.