Machine learning @ booking.com
summary
Built a lightweight Scala prediction service that logs input vectors and predictions into Hadoop and aggregates results into databases, enabling a self‑service portal to deploy models; system serves 200k predictions/sec across 200+ models with <30 ms latency and uses histogram monitoring of prediction distributions to detect degradation and choose thresholds. Early choices (heavy parallelization, in‑memory model replication) cause memory and overhead issues that required sharding and revisiting the execution model.
AI-written summary · about
chapters (12)
- 0:00ML introduction at Booking.com
- 2:55ML adoption challenges 2015
- 5:24Building prediction service
- 7:09Using predictions in product
- 11:31Model monitoring and degradation
- 16:50Input data issues
- 17:11Self-service for models
- 18:36ML showcase
- 20:00Scaling and performance
- 23:29Future plans
- 25:37Lessons and tips
- 28:17Q&A
related talks
40mMachine Learning
35mHow Booking.com serves Deep Learning Model Predictions
39mMachine Learning with Scala on Spark
24mChallenges in building Machine Learning models in production
28mScaling Machine Learning Training and Inference
34mMachine Learning to know if you R coming
29mMachine Learning in Production: What to do after you’ve trained the models?
22mReal-world applications of machine learning: Beyond the hype