From 5 PB of Daily Logs to Thousands of ML Models #bbuzz
summary
Criteo's pipeline reads 5 PB of daily logs to train and deploy several thousand machine-learning models per day that predict which ads to show, supporting 3.5 billion ad impressions and about 500,000 predictions per second. The talk describes the engineering challenges and solutions using HDFS, YARN, Spark plus in-house systems for ML algorithms, scheduling, testing and monitoring to continuously train, improve and serve models.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:20Data lifecycle & ML pipeline
- 3:57User behavior modeling
- 5:22Offline to online pipeline
- 7:14Feature transformation challenges
- 9:43Model lifecycle engineering
- 11:45Tennis training example
- 16:11Model implementation
- 18:22Distributed computing
- 20:59Performance optimization
- 23:51Memory debugging
- 26:36Memory optimization
- 28:29Business models
- 30:23Online deployment
- 31:54Feature vectors & automation
- 33:23Open source tools & monitoring
- 34:26Q&A
related talks
32mMaking Models Work: How an optimized ML lifecycle can drive better ROI
41mChallenges of Machine Learning at Scale
24mImprove your marketing reach using ...
33mMachine Learning to moderate ads
53mReal-Time Big Data Analytics with R
36mFrom Machine Learning Startup to Big Data Company
40mMachine Learning
29mScaling to Millions of ML Models to Solve the Problems of SRE and Security