How Booking.com Uses Kubernetes for Machine Learning

Sahil Dua, developer at Booking.com, explained how they have been able to scale machine learning (ML) models for recommending destinations and accommodation to their customers using Kubernetes, at this year's QCon London conference. In particular, he stressed how Kubernetes elasticity and resource starvation avoidance on containers helps them run computationally (and data) intensive, hard to parallelize, machine learning models. Kubernetes isolation (processes not having to compete for resources), elasticity (auto-scaling up or down based on resource consumption), flexibility (being able to quickly try out new libraries or frameworks) and GPU support (albeit Kubernetes support for NVIDIA GPUs is still in alpha, it allows 20x to 50x speed improvements) are key for Booking.com to run a large number of ML models at their scale (around 1.5 million room nights booked daily and 400 million monthly visitors). Each model runs as a stateless app inside a container. The container image does not include the model itself, it is retrieved at startup time from Hadoop.