Creating Kubernetes Autoscalers
- 2. Who am I? ● Final year Computer Science student at Queen’s University Belfast. ● Previously worked at IBM for a year. ● Open sourced projects for QRadar. ● Constantly working on side projects, most of them open source.
- 3. Why open source? ● To have a greater chance that someone will find it and use it. ● To get more responsive feedback per release. ● Community led feature suggestion, bug reporting, and use case explanation. ● Kubernetes ecosystem is open source oriented.
- 4. Kubernetes and autoscaling ● Kubernetes allows sharing resources between applications. ● Built on container technology, such as Docker, standardising applications into units with infinite scalability potential. ● Autoscaling is the act of balancing which applications need more or less resources based on criteria and metrics. ● Horizontal Pod Autoscaler inbuilt method for autoscaling, generally used to scale on CPU/Memory.
- 5. The Problem ● The Horizontal Pod Autoscaler has a hard-coded equation for calculating how to scale. ● Custom metrics go some way to help with this, but ultimately still constrained by core equation. ● If more complex scaling logic required (e.g. conditional logic, queues), a custom autoscaler is required. ● Custom autoscalers are difficult and time consuming to create, requiring knowledge of the Kubernetes API and Kubernetes lifecycle.
- 6. The Goal ● Make autoscalers quick and easy to make, with limited Kubernetes knowledge required. ● Abstract away complicated Kubernetes API interactions. ● Autoscaler logic writeable in any language, framework or environment, and should be simple enough to be contained in a single script.
- 7. Custom Pod Autoscaler Framework ● Works by integrating user provided logic, written in any language or framework, with a program to abstract away Kubernetes interactions. ● Allows quick and easy autoscaler development. ● Fast prototyping and experimentation. ● Any logic that can be started with a shell command can be used. ● Runs in the Kubernetes cluster as a pod. ● Fully configurable, supports everything the Horizontal Pod Autoscaler can be configured with. ● Allows configuration without master node access, allowing use on managed providers such as EKS or GKE.
- 8. Custom Pod Autoscaler Framework
- 9. Predictive Horizontal Pod Autoscaler ● Allows use of statistical models coupled with standard autoscaling metrics. ● Can predict future demand based on historical data - preemptively scaling rather than reacting to demand. ● Can see reductions in latency spikes if configured and tuned correctly.
- 10. Any questions? Some (maybe) useful links jamiethompson.me github.com/jthomperoo/custom-pod-autoscaler github.com/jthomperoo/predictive-horizontal-pod-autoscaler