Scalable Rate Limiter

Overview

Scalable-rate-limiter is a plugin for Kong built on top of Rate-limiting plugin. It adds batch updates of rate-limiting counters and also adds support for clustered redis.

Issues in the bundled Rate Limiting plugin

The Rate Limiting plugin bundled with Kong works fine upto a certain throughput, after which cassandra and redis policy become hard to scale. This is mainly due to the following problems:

Problem 1: At high throughputs, updating the rate-limiting counters on each request can increase the load on the database causing Hot-Key problem.
Solution: We created a new policy batch-redis which maintains a counter in local memory and updates the rate-limiting counters in DB in a batch. Hot keys are not an issue anymore as the number of requests to Redis go down by a factor of batch size.

Problem 2: When rate-limiting counters data grows, we cannot shard them with a redis-cluster since the plugin did not have support for redis-cluster.
Solution: The lua-resty-redis client does not support clustered-redis so we replaced it with resty-redis-cluster. We faced some issues during stress tests when one of the Redis shard in a cluster went down. To make it more fault tolerant, we made some tweaks to resty-redis-cluster library and added the modified version to our plugin's code.

Changes made

Major changes made to the plugin code:

1. Added batch-redis policy
2. Added support for redis cluster
3. Made plugin fault tolerant by default i.e. requests will continue to be served if there is some problem with code or redis
4. Removed local, cluster (cassandra/postgres) policies

How does it work?

The plugin uses fixed time windows to maintain rate-limiting counters (similar to the bundled Rate Limiting plugin)

redis policy

The rate limiting counters are updated on each request. Recommended for low throughput API's.

batch-redis policy

Instead of updating the global counter in redis after each request, the request counts are maintained at a (local) node level in the shared cache (amongst nginx workers) which is synchronized with the (global) redis counter whenever a batch is complete. Consider this scenario:

batch-size = 500, and throughput = 1 million RPM

Each nginx worker updates the local shared cache after serving a request. Once the local_counter % 500 == 0 , the global redis counter is incremented (by batch size). This global count which is a representative of total number of requests is then used to update the local cache as well. And this process is repeated until the API limit is reached or the time period expires.
Batching reduces the effective writes made on global redis counter by a factor of batch_size. It therefore makes the rate limiting scalable and faster as network calls are avoided on each request.

Installation

If you're using luarocks execute the following:

luarocks install scalable-rate-limiter

You will also need to enable this plugin by adding it to the list of enabled plugins using KONG_PLUGINS environment variable or the plugins key in kong.conf

export KONG_PLUGINS=scalable-rate-limiter

OR

plugins=scalable-rate-limiter

Parameters

Parameter	Type	Default	Required	description
second	integer		false	Maximum number of requests allowed in 1 second
minute	integer		false	Maximum number of requests allowed in 1 minute
hour	integer		false	Maximum number of requests allowed in 1 hour
day	integer		false	Maximum number of requests allowed in 1 day
limit_by	string	service	true	Property to limit the requests on (service / header)
header_name	string		true (limit_by: header)	The header name by which to limit the requests if limit_by is selected as header
policy	string	redis	true	Update redis at each request (redis) or in batches (batch-redis)
batch_size	integer	10	true (when policy: batch-redis)	Redis counters will be updated in batches of this size
redis_host	string		true	Redis host
redis_port	integer	6379	true	Redis port
redis_password	string		false	Redis password
redis_connect_timeout(ms)	integer	200	true	Redis connect timeout
redis_send_timeout(ms)	integer	100	true	Redis send timeout
redis_read_timeout(ms)	integer	100	true	Redis read timeout
redis_max_connection_attempts	integer	2	true	Total attempts to connect to a Redis node
redis_keepalive_timeout(ms)	integer	60000	true	Keepalive timeout for Redis connections
redis_max_redirection	integer	2	true	Number of times a keys is tried when MOVED or ASK response is received from redis server
redis_pool_size	integer	4	true	Pool size of redis connection pool
redis_backlog	integer	1	true	Size of redis connection pool backlog queue
error_message	string	API rate limit exceeded	true	Error message sent when rate limit is exhausted

Caveats

1. Batching introduces an error of upto batch_size * (number of kong instances) since the local count is used to check if request should be allowed or not and it can be outdated (in the worst case) by batch_size * (number of kong instances) giving the above error margin.
   This can be mitigated by reducing the batch size to the smallest possible value (which the database can support). In our tests we found this error margin to be too minuscule (<0.5%). This error margin can also be accounted for while deciding the rate limit (if rate limit is 100, set it to 99 accounting for upto 1% error).
2. This plugin only works with a redis cluster.

Roadmap

1. Add support for sliding windows.
2. Add support for custom window size (eg. 2hrs, 3hrs, 10 minutes, 3 days)