ARCHIVED: This repo is outdated and has been incorporated in the Strider repo.

TRAPI Throttle

Proxy service that enforces rate limits on TRAPI endpoints using batching.

Local Development

Docker + Management Script

The codebase comes with a zero-dependency python management script that can be used to automate basic local development tasks. Make sure you have docker installed and then run:

./manage.py dev # starts server accessible at 7830
./manage.py test # run tests
./manage.py coverage # run tests, open coverage report in browser

Without Docker

To run locally without Docker, you will need Python 3.9 installed. After that is installed, you can use pip to install the rest of the dependencies:

pip install -r requirements-test-lock.txt

then either pytest or python run.py.

Usage

Before requests can be initiated, a KP has to be registered. A KP registration object looks like this:

{
  "url": "http://kp1/query",
  "request_qty": 3,
  "request_duration": "0:01"
}

This allows 3 requests every minute to the underlying endpoint. Duration can be specified in any of the Pydantic timedelta formats including an ISO8601 string or an integer number of seconds.

After the KP is registered, any requests to /{kp_name}/query endpoint will be forwarded to the KP with the rate limiting and appropriate buffering applied.

Architecture

This codebase makes extensive use of Python's asyncio features to handle rate limiting and batching. It uses asyncio.Queues for buffering requests and responses. This is the general process for how requests are handled:

A request comes in as a TRAPI message to the /query/kp1 endpoint. The request is given a UUID and added to the associated queue. The request is blocked from returning.
When each KP is registered, it sets up a batch processing coroutine. This coroutine wakes up when there is an item available in its queue.
The coroutine reads all requests in the queue, merging them.
The coroutine makes a request to the underlying KP and receives a response.
The response is split into responses for each initial request. These responses are written to the response queues provided with each request.
The kp1 coroutine keeps track of a Theoretical Arrival Time (TAT), when the next request will be allowed based on the rate limit specified. The thread waits for the TAT to elapse. If it is unset the thread simply continues.
The coroutine updates the TAT key. This key is updated based on the formula TAT = now + (kp_duration / kp_qty). This ensures a smooth set of requests. Then, the coroutine continues waiting for another item to be added to the request queue.
The original request coroutine has been waiting for the response queue. Once the batch processing coroutine adds the finished request to this queue, the request coroutine wakes up.
The request coroutine returns the TRAPI message.