Data Science API Example

This repository accompanies my talk at PyCon UK 2017, where I gave a tutorial providing some tips for effective deployment of data science APIs using Flask.

Specifically, my talk demonstrated:

• Flask
• JSON-encoding error handlers
• Authentication with API keys and the before_request decorator
• Flask-SQLAlchemy
• Avoiding SQL injection by using SQLAlchmemy's query interpolator
• Worker queues with python-rq
• Deployment with gunicorn

All of these recommendations have been implemented together in this project. In the near future I will write a blog post or series of posts covering the same material as my talk - once completed it will be linked here.

Installation

$ pip install .

To generate the sqlite database:

$ python gendb.py

You will also need to install redis and have it running for the app to function properly.

With homebrew:

$ brew install redis
$ brew tap homebrew/services  # If not done previously
$ brew services start redis

Running the Server

Run the api server with:

$ gunicorn --workers 4 dsapi:app

In a separate terminal(s), run one or more python-rq workers:

rq worker

Also make sure that redis-server is running.

Calling the API

With the server and worker running, the API can be called externally with any HTTP client. From Python, I recommend using the requests library, available from PyPI:

$ pip install requests

Simple Endpoint and Authentication

You can then query the /predict endpoint using one of the valid keys in the api_keys table:

>>> import requests
>>> response = requests.get(
>>>     'http://localhost:8000/predict/feature_a/1.0/feature_b/1.0',
>>>     headers={'X-API-Key': 'PE1tlZti1TXJ9nTIri30OnPcquDUpjVrayieAAzY'}
>>> )
>>> response.status_code
200
>>> response.json()
{'class': 1, 'class_probabilities': [0.0961415144762564, 0.9038584855237436]})

Note how missing or invalid API keys will return a 401 UNAUTHORIZED response:

>>> response = requests.get(
>>>     'http://localhost:8000/predict/feature_a/1.0/feature_b/1.0',
>>>     headers={'X-API-Key': 'invalid-key'}
>>> )
>>> response.status_code
401
>>> response.json()
{'error': True, 'message': 'Invalid API Key'}

Slow Task Endpoints

The /predict-slow endpoint accepts a POST request with the features in the body and returns an ID field that can be redeemed later for the result:

>>> response = requests.post(
>>>     'http://localhost:8000/predict-slow',
>>>     json={'feature_a': 1.3, 'feature_b': -3.2},
>>>     headers={'X-API-Key': 'PE1tlZti1TXJ9nTIri30OnPcquDUpjVrayieAAzY'}
>>> )
>>> response.status_code
202
>>> response.json()
{'id': 'a779fb1a-8811-48f2-b79e-00f223bb9f43'}

To check for the result:

>>> id = response.json()['id']
>>> response = requests.get(
>>>     f'http://localhost:8000/predict-slow/{id}',
>>>     headers={'X-API-Key': 'PE1tlZti1TXJ9nTIri30OnPcquDUpjVrayieAAzY'}
>>> )

Before the result is ready this will yield a 404 NOT FOUND:

>>> response.status_code
404

Once ready, the result will be returned with a 200 OK response:

>>> response.status_code
200
>>> response.json()
{'class': 0}