Kubernetes Node Autoscaling

[markmandel]

Paying for infrastructure that is not used doesn’t make sense.

That being said, how we scale up and down Kubernetes nodes that run gameservers is different from stateless applications - Nodes in the cluster must be empty before they are deleted, and depending on if game servers are warm or not, nodes may need to be created preemptively for expected demand.

This should also be built to allow for other cloud providers / on prem solutions to be able to integrate with the solution.

Questions

• How can we keep the buffer of extra resources low, which still ensuring fast scale up times?
• Can we provide direction on how to use smaller VMs to make the scale up/down buffer smaller?
• This also feels like an area in which stat collection and display would be useful.
• May not need a scheduler, if it's assumes developers will fill the cluster with as many game servers as they can.

Research

• https://github.com/markmandel/paddle-soccer/tree/master/server/nodescaler

[markmandel]

Closing this as a duplicate of #368