Machine Learning
summary
Booking.com runs a model-serving pipeline that starts with simple materialized predictions and evolves into an in-memory service written in Scala + Akka Spray that loads trained models and serves live traffic; telemetry goes to Graphite/Grafana and batch sinks into Kafka → Hadoop → Cassandra. A self-service portal catalogs models, enforces 50/50 experiments, and shows prediction histograms and feature-distribution monitors; infrastructure scales from ~200 models to 200 models → 400 models and about 2.5 billion predictions/day, while currency bugs, sharding and parallelism fixes force several architectural changes.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 3:312015 challenges and limits
- 6:20Model service creation
- 8:50Model degradation issues
- 11:27Model quality monitoring
- 13:58Prediction visualization
- 17:10Input data monitoring
- 19:20Model scaling problems
- 21:50Model portal and showcase
- 23:59Technical errors and fixes
- 27:53Future plans and improvements
- 28:56Tips and conclusions
- 31:00Q&A
related talks
44mMachine learning @ booking.com
35mHow Booking.com serves Deep Learning Model Predictions
17mLaying the foundation for deep learning
14mPerformance Optimization in High-Volume Transaction Systems
24mChallenges in building Machine Learning models in production
28mScaling Machine Learning Training and Inference
33mDeep Learning at Booking.com
17mKeynote: Scaling Deep Learning Models in Production Using Kubernetes