Serverless containers: Nodeless Kubernetes and vertical pod autoscaling

Description

The pricing model layer includes spot, on-demand, and reserved instances (RIs). Spot instances are spare compute capacity on AWS and one of three ways that Amazon sells its compute capacity—the other two being on-demand instances and reserved instances. In terms of the servers themselves, there’s no difference between the three. Spot instances are typically recommended for stateless and non-mission-critical workloads since they can be interrupted at a moment’s notice, while RIs and on demand are more recommended for stateful and long-running applications. Byron Berrisford explores what happens if you mix and match all three business models within the same cluster and run stateless and web services containers on spot instances while running databases on on-demand and RIs. He introduces some of the best practices on how to achieve this automatically and efficiently without risking your applications’ high availability.

Different containers and applications come in different forms and sizes and have different requirements. Instead of provisioning infrastructure and then deploying the containers that might have the right resources to run on and might not, Byron introduces a new concept of pod-driven autoscaling, where the size and type of the servers in the cluster are being determined in real time by pods and deployment parameters in order to meet pod, container, and application needs and makes sure that all pods in the cluster have a place and the capacity to run. Byron also outlines how to use the containers-utilization layer to change in real time the limits and resource requests of the containers based on their consumption and utilization.