Designing a scalable ML Ops Pipeline: Insights and best practices
summary
Scans an end-to-end fraud detection build that processes billions of transactions, trains with big query data, and gates deployment on a hard threshold (example given: 98% accuracy) so only models that pass automated evaluation go live. Implements three automated pipelines: data pipeline for ingest/validation/cleaning, a model pipeline that triggers automl training and creates endpoints, and a code pipeline that runs a model deploy op, plus logging, monitoring and optional notifications via a vertex notification email op.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:50How ML models work
- 3:14Fraud detection example
- 5:21Model decay challenges
- 6:24What is MLOps
- 7:30Building MLOps pipelines
- 9:17MLOps tools
- 10:43Vertex AI for MLOps
- 11:26Fraud detection pipeline demo
- 17:29Summary
related talks
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
25mBuilding Scalable MLOps Pipelines That Actually Work
18mAutomate the boring ML stuff with pipelines
13mMLOps at Scale
10mFrom Model to Production: Building Scalable MLOps Pipelines
56mDevOps for Machine Learning
01h 21mBuilding a ML Model Observability pipeline
33mBuilding Resilient (ML) Pipelines for MLOps