Mastering Kubernetes - Second Edition

Measuring end-to-end pod startup time

One of the most important performance characteristics of a large dynamic cluster is end-to-end pod startup time. Kubernetes creates, destroys, and shuffles pods around all the time. You could say that the primary function of Kubernetes is to schedule pods.

In the following diagram, you can see that pod startup time is less volatile than API calls. This makes sense since there is a lot of work that needs to be done, such as launching a new instance of a runtime that doesn't depend on cluster size. With Kubernetes 1.2 on a 1,000-node cluster, the 99th percentile end-to-end time to launch a pod was less than 3 seconds. With Kubernetes 1.3, the 99th percentile end-to-end time to launch a pod was a little ...