CoreOS Fest 2016

Fleet is a distributed init system that we use to schedule both infrastructure as well as user services on our clusters. Fleet is one of the main building blocks of our architecture that we use to schedule both infrastructure as well as user services on our microservices infrastructures. Fleet has seen production use for some time and is largely considered stable. The versatility and simplicity of fleet offers a distributed orchestration system, on which services in the form of systemd units (optionally running containers), can easily be deployed on the cluster. However, fleet can run into scalability issues that grow when using it in large-scale infrastructures. In particular, we realized of these scalability issues when fleet has to coordinate many nodes and/or thousands of deployed units in a cluster. Thus, we set out to improve the scalability of this distributed init system. Through the (optional) use of gRPC for communication between the fleet nodes as well as reducing etcd load through storing transient data in-memory, we could improve fleet performance and stability significantly. This talk goes into details of our design decisions and the implementation of the performance-improving changes, which we are currently working on merging upstream.