Rate-limit extensions for RESTful bucketed endpoints.

Provides implementations for the complex rate limiting mechanisms that Discord requires for rate limit handling that conforms to the passed bucket headers correctly.

This was initially a bit of a headache for me to understand, personally, since there is a lot of “implicit detail” that is easy to miss from the documentation.

In an attempt to make this somewhat understandable by anyone else, I have tried to document the theory of how this is handled here.

What is the theory behind this implementation?#

In this module, we refer to a hikari.internal.routes.CompiledRoute as a definition of a route with specific major parameter values included (e.g. POST /channels/123/messages), and a hikari.internal.routes.Route as a definition of a route without specific parameter values included (e.g. POST /channels/{channel}/messages). We can compile a hikari.internal.routes.CompiledRoute from a hikari.internal.routes.Route by providing the corresponding parameters as kwargs, as you may already know.

In this module, a “bucket” is an internal data structure that tracks and enforces the rate limit state for a specific hikari.internal.routes.CompiledRoute, and can manage delaying tasks in the event that we begin to get rate limited. It also supports providing in-order execution of queued tasks.

Discord allocates types of buckets to routes. If you are making a request and there is a valid rate limit on the route you hit, you should receive an X-RateLimit-Bucket header from the server in your response. This is a hash that identifies a route based on internal criteria that does not include major parameters. This X-RateLimitBucket is known in this module as an “bucket hash”.

This means that generally, the route POST /channels/123/messages and POST /channels/456/messages will usually sit in the same bucket, but GET /channels/123/messages/789 and PATCH /channels/123/messages/789 will usually not share the same bucket. Discord may or may not change this at any time, so hard coding this logic is not a useful thing to be doing.

Rate limits, on the other hand, apply to a bucket and are specific to the major parameters of the compiled route. This means that POST /channels/123/messages and POST /channels/456/messages do not share the same real bucket, despite Discord providing the same bucket hash. A real bucket hash is the str hash of the bucket that Discord sends us in a response concatenated to the corresponding major parameters. This is used for quick bucket indexing internally in this module.

One issue that occurs from this is that we cannot effectively hash a hikari.internal.routes.CompiledRoute that has not yet been hit, meaning that until we receive a response from this endpoint, we have no idea what our rate limits could be, nor the bucket that they sit in. This is usually not problematic, as the first request to an endpoint should never be rate limited unless you are hitting it from elsewhere in the same time window outside your hikari.applications. To manage this situation, unknown endpoints are allocated to a special unlimited bucket until they have an initial bucket hash code allocated from a response. Once this happens, the route is reallocated a dedicated bucket. Unknown buckets have a hardcoded initial hash code internally.

Initially acquiring time on a bucket#

Each time you BaseRateLimiter.acquire() a request timeslice for a given hikari.internal.routes.Route, several things happen. The first is that we attempt to find the existing bucket for that route, if there is one, or get an unknown bucket otherwise. This is done by creating a real bucket hash from the compiled route. The initial hash is calculated using a lookup table that maps hikari.internal.routes.CompiledRoute objects to their corresponding initial hash codes, or to the unknown bucket hash code if not yet known. This initial hash is processed by the hikari.internal.routes.CompiledRoute to provide the real bucket hash we need to get the route’s bucket object internally.

The BaseRateLimiter.acquire() method will take the bucket and acquire a new timeslice on it. This takes the form of a asyncio.Future which should be awaited by the caller and will complete once the caller is allowed to make a request. Most of the time, this is done instantly, but if the bucket has an active rate limit preventing requests being sent, then the future will be paused until the rate limit is over. This may be longer than the rate limit period if you have queued a large number of requests during this limit, as it is first-come-first-served.

Acquiring a rate limited bucket will start a bucket-wide task (if not already running) that will wait until the rate limit has completed before allowing more futures to complete. This is done while observing the rate limits again, so can easily begin to re-ratelimit itself if needed. Once the task is complete, it tidies itself up and disposes of itself. This task will complete once the queue becomes empty.

The result of RESTBucketManager.acquire() is a tuple of a asyncio.Future to await on which completes when you are allowed to proceed with making a request, and a real bucket hash which should be stored temporarily. This will be explained in the next section.

Handling the rate limit headers of a response#

Once you have received your response, you are expected to extract the values of the vital rate limit headers manually and parse them to the correct data types. These headers are:

  • X-RateLimit-Limit:

    an int describing the max requests in the bucket from empty to being rate limited.

  • X-RateLimit-Remaining:

    an int describing the remaining number of requests before rate limiting occurs in the current window.

  • X-RateLimit-Bucket:

    a str containing the initial bucket hash.

  • X-RateLimit-Reset-After:

    a float containing the number of seconds when the current rate limit bucket will reset with decimal millisecond precision.

Each of the above values should be passed to the update_rate_limits method to ensure that the bucket you acquired time from is correctly updated should Discord decide to alter their ratelimits on the fly without warning (including timings and the bucket).

This method will manage creating new buckets as needed and resetting vital information in each bucket you use.

Tidying up#

To prevent unused buckets cluttering up memory, each RESTBucketManager instance spins up a asyncio.Task that periodically locks the bucket list (not threadsafe, only using the concept of asyncio not yielding in regular functions) and disposes of any clearly stale buckets that are no longer needed. These will be recreated again in the future if they are needed.

When shutting down an application, one must remember to close() the RESTBucketManager that has been used. This will ensure the garbage collection task is stopped, and will also ensure any remaining futures in any bucket queues have an asyncio.CancelledError set on them to prevent deadlocking ratelimited calls that may be waiting to be unlocked.

Body-field-specific rate limiting#

As of the start of June, 2020, Discord appears to be enforcing another layer of rate limiting logic to their HTTP APIs which is field-specific. This means that special rate limits will also exist on some endpoints that limit based on what attributes you send in a JSON or form data payload.

No information is sent in headers about these specific limits. You will only be made aware that they exist once you get ratelimited. In the 429 ratelimited response, you will have the "global" attribute set to False, and a "reset_after" attribute that differs entirely to the X-RateLimit-Reset-After header. Thus, it is important to not assume the value in the 429 response for the reset time is the same as the one in the bucket headers. Hikari’s hikari.api.rest.RESTClient implementation specifically uses the value furthest in the future when working out which bucket to adhere to.

It is worth remembering that there is an API limit to the number of 401s, 403s, and 429s you receive, which is around 10,000 per 15 minutes. Passing this limit results in a soft ban of your account.

At the time of writing, the only example of this appears to be on the PATCH /channels/{channel_id} endpoint. This has a limit of two changes per 10 minutes. More details about how this is implemented have yet to be released or documented…

Module Contents#

class hikari.impl.buckets.RESTBucket(name, compiled_route, global_ratelimit, max_rate_limit)[source]#

Bases: hikari.impl.rate_limits.WindowedBurstRateLimiter

Represents a rate limit for an HTTP endpoint.

Component to represent an active rate limit bucket on a specific HTTP route with a specific major parameter combo.

This is somewhat similar to the WindowedBurstRateLimiter in how it works.

This algorithm will use fixed-period time windows that have a given limit (capacity). Each time a task requests processing time, it will drip another unit into the bucket. Once the bucket has reached its limit, nothing can drip and new tasks will be queued until the time window finishes.

Once the time window finishes, the bucket will empty, returning the current capacity to zero, and tasks that are queued will start being able to drip again.

Additional logic is provided by the RESTBucket.update_rate_limit call which allows dynamically changing the enforced rate limits at any time.

property is_unknown: bool[source]#

Return True if the bucket represents an UNKNOWN bucket.

async acquire()[source]#

Acquire time and the lock on this bucket.


You should afterwards invoke RESTBucket.update_rate_limit to update any rate limit information you are made aware of and RESTBucket.release to release the lock.


If the rate limit is longer than max_rate_limit.


Release the lock on the bucket.


Resolve an unknown bucket.


The real bucket hash for this bucket.


If the hash of the bucket is already known.

update_rate_limit(remaining, limit, reset_at)[source]#

Update the rate limit information.


The reset_at epoch is expected to be a time.monotonic monotonic epoch, rather than a time.time date-based epoch.


The calls remaining in this time window.


The total calls allowed in this time window.


The epoch at which to reset the limit.

class hikari.impl.buckets.RESTBucketManager(max_rate_limit)[source]#

The main rate limiter implementation for HTTP clients.

This is designed to provide bucketed rate limiting for Discord HTTP endpoints that respects the X-RateLimit-Bucket rate limit header. To do this, it makes the assumption that any limit can change at any time.


The max number of seconds to backoff for when rate limited. Anything greater than this will instead raise an error.

property is_alive: bool[source]#

Whether the component is alive.

acquire_bucket(compiled_route, authentication)[source]#

Acquire a bucket for the given route.


You MUST keep the context manager acquired during the full duration of the request: from making the request until calling update_rate_limits.


The route to get the bucket for.


The authentication that will be used in the request.


The context manager to use during the duration of the request.

async close()[source]#

Close the garbage collector and kill any tasks waiting on ratelimits.

start(poll_period=20.0, expire_after=10.0)[source]#

Start this ratelimiter up.

This spins up internal garbage collection logic in the background to keep memory usage to an optimal level as old routes and bucket hashes get discarded and replaced.


Period to poll the garbage collector at in seconds. Defaults to 20 seconds.


Time after which the last reset_at was hit for a bucket to remove it. Higher values will retain unneeded ratelimit info for longer, but may produce more effective rate-limiting logic as a result. Using 0 will make the bucket get garbage collected as soon as the rate limit has reset. Defaults to 10 seconds.


Throttle the global ratelimit for the buckets.


How long to throttle for.

update_rate_limits(compiled_route, authentication, bucket_header, remaining_header, limit_header, reset_after)[source]#

Update the rate limits for a bucket using info from a response.


The compiled route to get the bucket for.


The authentication that was used in the request.


The X-RateLimit-Bucket header that was provided in the response.


The X-RateLimit-Remaining header cast to an int.


The X-RateLimit-Limit header cast to an int.


The X-RateLimit-Reset-After header cast to a float.

hikari.impl.buckets.UNKNOWN_HASH: Final[str] = 'UNKNOWN'[source]#

The hash used for an unknown bucket that has not yet been resolved.