Keynote: Scaling Deep Learning Models in Production Using Kubernetes
summary
Booking.com runs a Kubernetes pipeline that trains deep learning models inside Docker containers (TensorFlow, Torch etc.), streams training data from Hadoop or mounts it via persistent volumes, writes logs and final model artifacts back to HDFS, and deploys lightweight stateless images without the model that fetch a model at startup and expose a REST API for predictions. The platform autos-scales per model, supports distributed TensorFlow training, and operates at 1.5 million room nights/day across 1.4M properties, trading latency vs throughput per use case.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:38About me
- 1:27Agenda
- 1:52Scaling Deep Learning
- 4:09Image Tagging
- 8:53Training with Kubernetes
- 10:42Running the training
- 12:03Deployment
- 12:15What we want
- 12:50Containerization
- 13:55Serving Pod
- 14:42Recap
- 15:34Creating a Deployment
- 16:17Summary
related talks
35mHow Booking.com serves Deep Learning Model Predictions
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
18mKeynote: Adapting Kubernetes for Machine Learning Workflows
12mKeynote: Building Enterprise AI/ML Platform on Kubernetes
28mProductionizing Machine Learning Models at Scale with Kubernetes
54mEconomics and Best Practices of Running AI/ML Workloads on Kubernetes
15mKeynote: Machine Learning on Kubernetes Made Easy With Kubeflow
18mKeynote: Containerizing Deep Learning Using Kubeflow on AWS