Rate Limiting in Tyk

In the realm of API management, rate limiting is one of the fundamental aspects of managing traffic to your APIs. Rate limiting can easily become one of the easiest and most efficient ways to control traffic to your APIs.

Rate limiting can help with API overuse caused by accidental issues within client code which results in the API being overwhelmed with requests. On the malicious side, a denial of service attack meant to overwhelm the API resources can also be easily executed without rate limits in place.

What is rate limiting and how does it work?

Rate limits are calculated in Requests Per Second (RPS). For example, let’s say a developer only wants to allow a client to call the API a maximum of 10 times per minute. In this case the developer would apply a rate limit to their API expressed as “10 requests per 60 seconds”. This means that the client will be able to successfully call the API up to 10 times within any 60 second interval and after that the user will get an error stating their rate limit has been exceeded if they call it an 11th time within that time frame.

Types Of Rate Limiting

Tyk offers 2 different rate limiting modes:

Distributed Rate Limiter. Most performant, not 100% accurate. Recommended for most use cases. Uses the leaky bucket algorithm.
Redis Rate Limiter. Less performant, 100% perfect accuracy. Uses the sliding window algorithm.

Distributed Rate Limiter (DRL)

This is the default rate limiter in Tyk. It is the most performant, and the trade-off is that the limit is approximate, not exact. To use a less performant, exact rate limiter, review the Redis rate limiter below.

The DRL rate limiter will be used automatically unless one of the other rate limit algorithms are explicitly enabled via configuration.

With the DRL, the gateways divide the rate limit evenly across all the gateways in the cluster (a cluster of gateway shares the same Redis.) These gateways store the running rate in memory and return 429 (Rate Limit Exceeded) when their share is used up.

This relies on having a fair load balancer since it assumes a well distributed load between all the gateways.

It also uses what’s called a leaky bucket algorithm. In this case if the request rate is higher than the rate limit it will attempt to let through requests at the specified rate limit. It’s important to note that this is the only rate limit method that uses this algorithm and that it will yield approximate results.

Redis rate limiter

This uses redis to track the rate of incoming API calls. It’s important to note that it blocks access to the API when the rate exceeds the rate limit. Unlike the leaky bucket algorithm, it doesn’t let API calls through until the rate drops below the specified rate limit, it acts like a cool-down period. For example, if the rate limit is 3000/minute the call rate would have to be reduced below 3000 for a whole minute before the 429s stop.

This algorithm can be managed using the following configuration option enable_redis_rolling_limiter.

Sentinel Rate Limiter

The sentinel-based rate limiter delivers a smoother performance curve as rate-limit calculations happen off-thread, but a stricter time-out based cool-down for clients. For example, when a throttling action is triggered, they are required to cool-down for the period of the rate limit. The default behaviour is for the rate-limit calculations to happen on-thread and which offers a staggered cool-down and a smoother rate-limit experience for the client. For example, you can slow your connection throughput to regain entry into your rate limit. This is more of a “throttle” than a “block”. The standard rate limiter offers similar performance as the sentinel-based limiter. This is disabled by default.

This option can be enabled using the following configuration option enable_sentinel_rate_limiter.

Performance

The Redis limiter is indeed slower than the DRL, but that performance can be improved by enabling the enable_non_transactional_rate_limiter. This leverages Redis Pipelining to enhance the performance of the Redis operations. Here are the Redis documentation for more information.

DRL Threshold

TYK_GW_DRLTHRESHOLD

Optionally, you can use both rate limit options simultaneously. This is suitable for hard-syncing rate limits for lower thresholds, ie for more expensive APIs, and using the more performant Rate Limiter for the higher traffic APIs.

Tyk switches between these two modes using the drl_threshold. If the rate limit is more than the drl_threshold (per gateway) then the DRL is used. If it’s below the DRL threshold the redis rate limiter is used.

Rate limiting levels

Tyk has two approaches to rate limiting:

Key-level rate limiting

Key-level rate limiting is more focused on controlling traffic from individual sources and making sure that users are staying within their prescribed limits. This approach to rate limiting allows you to configure a policy to rate limit in two possible ways: limiting the rate of calls the user of a key can make to all available APIs, another form of global rate limiting just from one specific user, and limiting the rate of calls to specific individual APIs, also known as a “per API rate limit”.

API-level rate limiting

API-level rate limiting assesses all traffic coming into an API from all sources and ensures that the overall rate limit is not exceeded. Overwhelming an endpoint with traffic is an easy and efficient way to execute a denial of service attack. By using a global rate limit you can easily ensure that all incoming requests are within a specific limit. This limit may be calculated by something as simple as having a good idea of the maximum amount of requests you could expect from users of your API. It may also be something more scientific and precise like the amount of requests your system can handle while still performing at a high-level. This may be easily uncovered with some performance testing in order to establish this threshold.

When rate limiting measures are put in place, they are assessed in this order (if applied):

API-level global rate limit
Key-level global rate limit
Key-level per-API rate limit

When might you want to use rate limiting?

For key-level rate limiting you will be aiming to ensure that one particular user or system accessing the API is not exceeding a determined rate. This makes sense in a scenario such as APIs which are associated with a monetisation scheme where you may allow so many requests per second based on the tier in which that consumer is subscribed or paying for.

An API-level global rate limit may be used as an extra line of defence around attempted denial of service attacks. For instance, if you have load tested your current system and established a performance threshold that you would not want to exceed to ensure system availability and/or performance then you may want to set a global rate limit as a defence to make sure that it is not exceeded.

Of course, there are plenty of other scenarios where applying a rate limit may be beneficial to your APIs and the systems that your APIs leverage behind the scenes. The simplest way to figure out which type of rate limiting you’d like to apply can be determined by asking a few questions:

Do you want to protect against denial of service attacks or overwhelming amounts of traffic from all users of the API? You’ll want to use an API-level global rate limit!

Do you want to limit the number of requests a specific user can make to all APIs they have access to? You’ll want to use a key-level global rate limit!

Do you want to limit the number of requests a specific user can make to specific APIs they have access to? You’ll want to use a key-level per-API rate limit.

Improve this page

Return to Top