Running and monitoring distributed ML systems can be challenging. The need to manage multiple servers, and the fact that those servers emit different logs, means that there can be a lot of overhead involved in scaling up a distributed ML system. Fortunately, Ray makes parallelizing Python processes easy, and the open source whylogs enables users to monitor ML models in production, even if those models are running in a distributed environment.

Ray is an exciting project that allows you to parallelize pretty much anything written in Python. One of the advantages of the whylogs architecture is that it operates on mergeable profiles that can be easily generated in distributed systems and collected into a single profile downstream for analysis, enabling monitoring for distributed systems. This post will review some options that Ray users have for integrating whylogs into their architectures as a monitoring solution.

Continue reading on the Anyscale Ray Blog