RFC: NodeOverlay

[ellistarn]

Fixes:

• Account for AWS Savings Plan when choosing to deploy on-demand or spot aws/karpenter-provider-aws#3860
• feat: Adding capacityType discount as a percentage aws/karpenter-provider-aws#4697
• Mega Issue: Karpenter doesnt support custom resources requests/limit #751
• karpenter doesn't scaleout when using time-slicing for GPUs #729
• Discover Instance Type Capacity Memory Overhead Instead of vmMemoryOverheadPercent aws/karpenter-provider-aws#5161

Description

---
# Reduce on-demand prices to 90%
# https://github.com/aws/karpenter-provider-aws/issues/3860
# https://github.com/aws/karpenter-provider-aws/pull/4697
kind: NodeOverlay
metadata:
  name: discount
spec:
  selector:
    matchLabels:
      karpenter.sh/capacity-type: on-demand
  pricePercent: 90
---
# Support for extended resource types (e.g. smarter-devices/fuse)
# https://github.com/kubernetes-sigs/karpenter/issues/751
# https://github.com/kubernetes-sigs/karpenter/issues/729
kind: NodeOverlay
metadata:
  name: discount
spec:
  selector:
    matchLabels:
      karpenter.sh/capacity-type: on-demand
    matchExpressions:
    - key: node.kubernetes.io/instance-type
      operator: In
      values:
      - m5.large
      - m5.2xlarge
      - m5.4xlarge
      - m5.8xlarge
      - m5.12xlarge
  capacity:
    smarter-devices/fuse: 1
---
# Add memory overhead of 10Mi to all instances with 2Gi memory
# https://github.com/aws/karpenter-provider-aws/issues/5161
kind: NodeOverlay
metadata:
  name: discount
spec:
  selector:
    matchLabels:
      karpenter.k8s.aws/instance-memory: 2048
  overhead:
    memory: 10Mi
---

How was this change tested?

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: ellistarn

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Needs approval from an approver in each of these files:

• OWNERS [ellistarn]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[sftim]

(nit) In the PR description, the NodeOverlays all have the same .metadata.name

[sftim]

Is this an overlay for Nodes, or an overlay for NodePools (maybe more than one NodePool, if we support matching via a selector)?

[jwcesign]

In a real scenario with a saving plan, there is a commitment to a consistent amount of usage. Once the committed usage is exhausted, the price will revert to the on-demand rate. Does this PR take that into consideration?

What I expect is:

1. When there is remaining committed usage, Karpenter should continue to select on-demand instances(instance type) under the saving plan.
2. When the committed usage is exhausted, Karpenter should choose spot instances.

[daimaxiaxie]

@ellistarn We redefined the value of topology.kubernetes.io/zone. However, the current zone value is provided by the CloudProvider. Is it possible to support custom topology.kubernetes.io/zone and other WellKnownLabel values?

[sftim]

@ellistarn We redefined the value of topology.kubernetes.io/zone. However, the current zone value is provided by the CloudProvider. Is it possible to support custom topology.kubernetes.io/zone and other WellKnownLabel values?

It sounds like you don't agree with the zone naming used in an existing integration. You can:

• use a different label
  • this is the easy and recommended option
• replace the cloud provider integration code with your own
  (Karpenter, cloud controller manager, etc)

I don't think this PR would need to change to accommodate the ask.

[GnatorX]

Thanks for trying to tackle this!

I wonder if there is a cleaner way to represent the overlays to reduce sprawl.

Thinking more generally, each dimension can be additive, adding a new custom resource, and/or adjustment, price changes or resource modification. (Assuming we aren't supporting deleting properties at this time).

Price is a slight outlier from resources so we can separate out. Would it make sense to have at the highest level price and resources. Then subsection of adjustment and additions. Adjustment would support +/- percent or values and additions would just be a dict of new custom resources and values?

[daimaxiaxie]

@sftim I mean nodeoverlay can provide more overlay options, not just resources. For example, Lables.

[github-actions]

This PR has been inactive for 14 days. StaleBot will close this stale PR after 14 more days of inactivity.

[zaafar]

i think this shouldn't be closed.

[njtran]

/lifecycle frozen

[k8s-ci-robot]

@njtran: The lifecycle/frozen label cannot be applied to Pull Requests.

In response to this:

/lifecycle frozen

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[daimaxiaxie]

What is the progress now? Why does it seems not to be updated.

[github-actions]

This PR has been inactive for 14 days. StaleBot will close this stale PR after 14 more days of inactivity.

[zaafar]

commenting to make it not stale. this is a very important feature that will take Karpenter to the next level.

[abbas-square]

Is there any update on this?

[github-actions]

This PR has been inactive for 14 days. StaleBot will close this stale PR after 14 more days of inactivity.

[devusb]

Is this still planned for a future Karpenter release? Working on deploying in some clusters where I use hugepages, bumping up against this issue -- just trying to get a feel for whether I'll need a short-term workaround or if this is planned to be a long-term limitation.

[rschalo]

/reopen

[k8s-ci-robot]

@rschalo: Reopened this PR.

In response to this:

/reopen

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[zarko-a]

Also wondering if this is still planned to be merged.

[likeabbas]

@ellistarn could you provide and update on when this PR will be merged?