Abstract

The purpose of this document is to outline known considerations, design patterns, and solutions that customers can leverage today when considering hybrid dimensions of the AWS AI/ML stack across the entire machine learning (ML) lifecycle.

Due to the scalability, flexibility, and pricing models enabled by the cloud, we at AWS continue to believe that the majority of ML workloads are better suited to run in the cloud in the long haul.

However, given that less than 5% of overall IT spend is allocated for the cloud, the actual amount of IT spend on-premises is north of 95%.

This tells us that there is a sizeable underserved market.

Change is hard - particularly for enterprises.

The complexity, magnitude, and length of migrations can be a perceived barrier to getting started.

For these customers, we propose hybrid ML patterns as an intermediate step in their cloud and ML journey.

Hybrid ML patterns are those that involve a minimum of two compute environments, typically local compute resources such as personal laptops or corporate data centers, and the cloud.

See the Basics section for a full introduction to the concept of hybrid ML.

We think customers win when they deploy a workload that touches the cloud to get some value, and we at AWS are committed to supporting any customer’s success, even if only a few percentage points of that workload hit the cloud today.

This document is intended for individuals who already have a baseline understanding of machine learning, in addition to Amazon SageMaker.

We will not dive into best practices for Amazon SageMaker per se, nor into best practices for storage or edge services.

Instead, we will focus explicitly on hybrid workloads, and refer readers to resources elsewhere as necessary.