Rendezvous Architectures

Definitions

Rendezvous architectures are a class of streaming microservice architecture designed to manage multiple implementations of a streaming function so that new implementations can be deployed easily and accurately, input data can be archived precisely, and past operations can be audited while maintaining strict service-level guarantees. Although suitable for more general application, rendezvous architectures are particularly useful for the special case of managing machine learning models. The discussion here follows the trend of associating rendezvous architecture with machine learning and is largely limited to that context.

Historical Background

The management of logistics in machine learning systems has always been notoriously difficult, particularly when there are multiple data sources and, as is almost always the case, multiple models being iteratively developed, evaluated, and deployed at the same time. Requirements include stability and reliability in production along...