This document discusses building a machine learning orchestration framework on Mesos. It describes challenges with machine learning pipelines like heterogeneous environments and separate orchestration/execution. It then introduces Meson, a general purpose workflow orchestration system that delegates execution to resource managers like Mesos. Meson is optimized for machine learning pipelines and visualization. The document outlines how Mesos is used for executors, custom executors, executor caching, executor cleanup, framework messages, multi-tenancy and cluster management.