Build Data Apps by Deploying ML Models as API Services
summary
Demonstration converts a trained scikit-learn KNN on iris into a pickled artifact, wraps it in a Tornado async API that uses a thread pool to run blocking predict calls, and returns JSON predictions. A simple Flask webapp queries that API. Deployment uses cookiecutter templates, Conda, supervisor and an EC2 instance so the service runs on port 80. Optional pieces discussed: authentication, Redis caching, and that the pipeline actually returns correct predictions end-to-end.
AI-written summary · about
chapters (10)
- 0:00Deploying ML models as APIs
- 1:25Tutorial goals
- 3:31Development environment setup
- 13:18Building API with Tornado
- 27:03Querying API from Jupyter
- 31:51Why Tornado for async
- 48:17DIY vs managed services
- 55:06Flask web app for API
- 1:04:45Deploying on AWS EC2
- 1:15:31Final demo and Q&A
related talks
01h 45mBuilding a Model Prediction Server
37mBuild a Data-Driven Web App That Everyone Can Use
01h 19mDatascience on the web
29mBuilding an ML Platform from Scratch
15mA Fiery approach to R web services
45mData Science Apps: Beyond Notebooks
39mFrom data science to scalable NLU and vision cloud service.
42mStreamlit: A Faster Way to Build and Share Data Apps