Flask API 适用于 GET 但不适用于 POST 方法（heroku）

Question

I run a very spartan app for machine learning inference that's structured as such:

app.py:

from flask import Flask, request, jsonify
import pandas as pd
from joblib import load

app = Flask(__name__)

@app.route('/predict', methods=['POST'])
def predict():
    df = pd.DataFrame(request.json)
    df = df.reindex(columns=feature_columns, fill_value=0)
    prediction = list(ada.predict(df))
    return jsonify({"prediction": int(prediction[0])})

@app.route('/test', methods=['GET'])
def test():
    return "get successful"

if __name__ == '__main__':
    ada = load("./model/adaboost_model")
    feature_columns = load("./model/feature_columns")
    app.run()

The requirements.txt :

Flask
gunicorn
pandas
joblib

Procfile :

web: gunicorn -b:$PORT app:app

and a Model/ folder which contains the two pickled Adaboost models.

I deployed this on Heroku, the GET on /test endpoint works fine, the app works fine, returns 404 for random requests, and recognizes that /predict only has a POST method. However when I run POST I get an Internal Server error, I'm sure it's not the actual code of the POST method that's causing problems, because a simple POST method I wrote for testing also fails and it doesn't do anything code related.

Locally this runs perfectly fine, tested with Postman, the inference is correct and all. What could be going wrong with the POST requests when deployed on Heroku?

Answer 1

So the problem seems to be with gunicorn and the code that runs before app.run() , transferring the loading of the model and feature columns into the actual POST predict() method works perfectly now.

However this isn't optimal, because every time a POST request is made you need to load the model again, but it works and isn't so bad because I will only use this POST method once every hour or so.