Amazon SageMaker

What it is

Amazon SageMaker is a machine learning service from Amazon Web Services (AWS). AWS describes it as a fully managed, end-to-end service for building, training and hosting models at scale ^[1]. "Fully managed" means the provider runs the servers, storage and software underneath, so users work with notebooks, training jobs and endpoints rather than with machines themselves. "End-to-end" means the service covers the whole path from exploring data to serving predictions.

Where it came from

AWS announced SageMaker on 29 November 2017 at its re:Invent conference, in a post by Randall Hunt ^[1]. The announcement presented the service as three main components — authoring, model training and model hosting — which are also the three stages most machine learning work passes through ^[1].

Authoring

The authoring component gives users hosted Jupyter notebook environments that need no setup, for data exploration, cleaning and preprocessing, running on general-purpose or GPU instances ^[1]. A Jupyter notebook is an interactive document that mixes code, its output and written notes, which is why it is widely used in data work; GPUs matter because many machine learning calculations run much faster on them than on ordinary processors.

Training

Model training is a distributed service. Users can pick from built-in supervised and unsupervised algorithms and frameworks, or bring their own training code packaged in Docker containers ^[1]. A container bundles code together with its dependencies so that it behaves the same way on any machine that can run it. Training can scale to tens of instances, reads training data from Amazon S3 and writes the resulting model artefacts back to S3 ^[1].

Hosting

Model hosting provides HTTPS endpoints for real-time inference. The endpoints scale with traffic, and several models can be served at once so that they can be compared in A/B tests ^[1]. A/B testing here means splitting incoming requests between model versions and comparing how they perform, instead of replacing a live model in a single step.

Using one piece at a time

Each of the three components can be used in isolation, so the service can fill gaps in a pipeline that already exists ^[1]. A team with working training code that only lacks a way to serve models can use hosting by itself; a team that mainly needs somewhere to explore and clean data can use authoring without adopting the rest.

Pricing

The free tier covered the first two months: 250 hours a month of t2.medium notebooks, 50 hours of m4.xlarge training and 125 hours of m4.xlarge hosting ^[1]. Beyond the free tier, the service is billed per second of instance usage, per GB of storage and per GB of data transfer ^[1]. Per-second billing suits training jobs, which often run for minutes or hours rather than continuously.

Where it sits

By offering authoring, training and hosting as one service, with training data and model artefacts kept in S3, SageMaker covers steps that otherwise have to be stitched together from separate tools ^[1]. Because each component is optional, teams can adopt it gradually rather than moving an entire workflow at once.