AIM307 – Amazon SageMaker deep dive – A modular solution for machine learning (re:Invent 2019) – Key Takeaways

NOTE: Personally I wouldn’t consider this session well-organized. It starts with a very high level marketing perspective of SageMaker, then suddenly dived all the way to code level details, then followed by a business example. The first part is too shallow, and second part too hard-core, and third part ironically the most interesting. I was expecting to have a deep conceptual understanding of SageMaker components, underneath mechanisms and new services, just like some of those good sessions on other services that also features “deep dive” in their titles, but that did not happen.

The Key

  • Direct Marketing Demo
    • 41,000 customer data with demographics label + whether customer accepted marketing offer
    • A few lines to get AutoPilot auto-train using the data
  • British Airways case study
    • Using built-in algorithms
    • SageMaker anomaly detection: removing spikes, identifying developing conditions
    • See below screenshots

The Non-essentials

  • SageMaker overview
    • Pre-built Notebook examples
    • Built-in algorithms
    • One-click training / deployment
    • Spot instance integration
    • Operator for K8s
  • SageMaker new components
    • Very brief and non-technical

SageMaker is now an umbrella service
Automatic pipeline of SageMaker: a data file comes in, Lambda check whether SageMaker endpoint is available, if not then create one, then send to second Lambda to ETL dataset for inference and send to endpoint, if no good result then iterate with another model until result is ok, then check result and create alert if needed