The Coding Portion of Data Science
summary
Argues that data scientists must code beyond models, showing concrete practices: feature engineering in SQL rather than slow pandas, automated CI/CD with git flow and code reviews, containerized deployment using Docker/Kubernetes, and runtime targets under 0.5 seconds for online predictions. Calls out Jupyter as fine for exploration but unsuitable for production pipelines; recommends unit tests, monitoring for model drift, and attention to cost (stolen AWS keys inflated a client bill). Emphasizes reproducibility, pipeline automation, and scalable integration.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:51WHY DO I CARE ABOUT THE TOPIC?
- 1:53CAN I BE A DATA SCIENTIST WITHOUT CODING?
- 3:01LEARNING DATA SCIENCE
- 7:46WHAT KAGGLE TEACHES US?
- 8:41KAGGLE VS REALITY
- 10:00DECISION MAKING PROCESS
- 12:19DATA IS MESSY
- 14:57DATA SCIENCE IS A TEAMWORK
- 15:58CODE VERSIONING
- 20:07BUILDING A ROBUST AND OPTIMAL SYSTEM
- 25:16PIPELINE EFFICIENCY
related talks
59mDevOps for Data Science
39mRight Code, Right Place, Right Time
41mBetter Code for Data Science
29mData Science Best Practices : From Proof of Concepts to Production
05mProductionalizing Data Science Models
35mApplied Data Science
42mHow to be a 10x Data Scientist
40mThe data scientists are gone. how do I deploy their work to prod?