Vertical Pod Autoscaling for DaemonSets with Heterogeneous Resource Requirements

[bernot-dev]

Enhancement Description

• One-line enhancement description (can be used as a release note): Improve Vertical Pod Autoscaling for workloads with heterogeneous resource requirements by applying requests/limits per-pod.
• Kubernetes Enhancement Proposal:
• Discussion Link: SIG Autoscaling Oct. 14, 2024
• Primary contact (assignee): @bernot-dev
• Responsible SIGs: sig-autoscaling
• Enhancement target (which target equals to which milestone):
  • Alpha release target (x.y): 1.35
  • Beta release target (x.y): 1.36
  • Stable release target (x.y): 1.38
• Alpha
  • KEP (k/enhancements) update PR(s):
  • Code (k/k) update PR(s):
  • Docs (k/website) update PR(s):

Please keep this description up to date. This will help the Enhancement Team to track the evolution of the enhancement efficiently.

[bernot-dev]

/sig autoscaling

[bernot-dev]

/assign

[mrueg]

fwiw, there's an external controller that might implement some of the features: https://github.com/DataDog/extendeddaemonset?tab=readme-ov-file#overwrite-containers-pod-resources-for-a-specific-node

[alvaroaleman]

@mrueg thanks for the pointer. From a quick look, while that allows to override per-node requests it does not have any sort of per-node vertical autoscaling or did I miss it?

[adrianmoisey]

/retitle Vertical Pod Autoscaling for DaemonSets with Heterogeneous Resource Requirements

[mrueg]

@mrueg thanks for the pointer. From a quick look, while that allows to override per-node requests it does not have any sort of per-node vertical autoscaling or did I miss it?

Sorry for the late response here. You're correct, no autoscaling component. The per-node overrides are the concept that could be interesting for this KEP.

[alvaroaleman]

I think what people want is per-node autoscaling. A lot of daemonsets are some sort of daemon (log collection, metric collection, maybe some security checks) that scale relative to the size of the node, i.E. they will need less resources on smaller and more resources on bigger nodes. Ideally, the VPA would store the historic data by node or maybe node instance type