Niloy Gupta, Software Engineer - Machine Learning
- Jan 17, 2018
Growing Cache Friendly Trees In case you missed it, Part I of this blog series outlines how we built a distributed machine learning platform to train gradient boosted tree (GBT) models on large datasets. While we were able to observe significant improvements in offline metrics, the resulting models are too large for standard XGBoost prediction libraries to meet our latency requirements. As result, we were unable to launch the models in production as we needed to serve ads within 50ms (p50) and evaluating these large models caused time-out exceptions. This article will discuss how we compressed and reordered the trees...