using a separate image for every component makes it much harder to add a third-party component (e.g. tensorflow, or a Python library) to a k8s installation. There's really no difference between the images except for the command that is executed.

Using a single base image for all components means that an extended image can be created by anyone needing additional components, without needing to either repeat that process 10 times or know details of how our containers are built.