ML in Production: Why Analysts and Backend Struggle to Agree
summary
The talk describes how Aviasales deployed ML scoring in production to rank flight tickets by purchase probability. It covers practical issues encountered—data drift, multiple heterogeneous data sources, failures to meet nonfunctional requirements, and unexpected increases in latency—and how these impacted analytics and backend integration.
AI-written summary · about
chapters (9)
- 0:00ML in production intro
- 1:55Ticket recommendation problem
- 8:25Analysts vs backend divide
- 12:23Search engine ticket processing
- 17:01Model update challenges
- 21:16Training inference mismatch
- 24:07Prediction performance issues
- 30:08Key lessons
- 31:12Q&A
related talks
44mML in Production
30mDriving a Higher Level of Customer Support with Machine Learning
29mDemystifying Machine Learning in Production: Reasoning about a Large-Scale ML Platform
52mThe ML Lifecycle in Combat Conditions
21mWhy is Production ML so Hard?
54mML Beyond Metrics: Challenges of Using ML Systems in the Real World
24mChallenges in building Machine Learning models in production
22mBefore the Model: How Machine Learning Products Start...