The path between developing and serving machine learning models.
summary
Deploys a Kubernetes cluster (kops on AWS), installs PipelineIO components — Spark, HDFS, JupyterHub, a Java PMML server plus a Python prediction service — and shows pushing models with the Pio CLI (it zips the folder and uploads versions). 2,000 requests give a 99th‑percentile ≈300 ms (90th ≈1 ms) for a demo model. Demonstrates A/B via separate model URLs and recommends building Docker images + CI (Airflow/Jenkins). Several modules are buggy, docs thin, and TensorFlow prediction support is incomplete.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 2:26Serving ML models in production
- 5:04Pipeline I/O overview
- 7:06DevOps and Kubernetes basics
- 11:00Pipeline I/O on Kubernetes
- 21:25Kubernetes cluster setup
- 24:38Kubernetes dashboard and Weave Scope
- 27:18Deploying Pipeline I/O
- 29:06Model versioning and A/B testing
- 30:54Serializing models with PMML and Python
- 39:56Reliability, scalability, and deployment
- 43:20Custom models with PMML and Python
- 48:21TensorFlow models with Pipeline I/O
- 51:11Scaling prediction services
- 53:20Enterprise considerations and limits
- 55:34Real-time data access
related talks
33mDeploying and Scaling Spark ML and Tensorflow AI Models
48mDeploying Machine Learning Models in Practice
46mDeploying scalable Machine Learning models with Kubernetes
28mProductionizing Machine Learning Models at Scale with Kubernetes
46mAccelerating Machine Learning DevOps with Kubeflow
21mMastering the DevOps Machine Learning Pipeline
57mDevOps for Machine Learning
43mAI Pipelines With OPEA: Best Practices for Cloud Native ML Operati...