Building Portable ML Pipelines for Development and Production
summary
Describes a concrete portable MLOps pattern: containerize training and prediction, compile a Kubeflow pipeline to a YAML, then run the identical YAML on Vertex AI (managed) and a local Q flow UI to verify parity; demo shows the same pipeline printing a Vertex run and a local run (messages like "hello next 24" and "hello local next4"). Compares managed Vertex AI vs open-source tradeoffs, recommends model monitoring + continuous training, notes TPU/XLA/JAX give hardware-agnostic speed (TPU ~2x claim).
AI-written summary · about
chapters (11)
- 0:00Introduction and agenda
- 1:02MLOps vs DevOps workflow
- 3:28Pipelines in MLOps
- 5:37Portability in ML solutions
- 7:03Model framework choices
- 9:32Containerizing ML pipelines
- 11:00Vertex AI overview
- 16:23Demo setup
- 19:14Running the hello pipeline
- 22:26Portability demonstration
- 22:47Key takeaways and conclusion
related talks
45mVertex AI: Pipelines for your MLOps workflows
32mMLOps at Renault: A Generic Pipeline for Scalable Deployment
40mVertex AI: Pipelines for your MLOps workflows
33mBuilding Resilient (ML) Pipelines for MLOps
31mVertex AI: Pipelines for your MLOps workflows
29mTeam Teardown: MLOps in the Cloud for IoT
25mBuilding Scalable MLOps Pipelines That Actually Work
12mBuild end-to-end solutions with Vertex AI