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.

[tallaxes]

It may be worth describing the potential interaction of client-side use (and customization) of pricing with provider-specific server-side price-sensitive allocation strategies (like those supported by EC2 Fleet).

[rlindsberg]

Hi all!
Do we have an estimation of when this feature will be publicly available? We have been waiting for it for over a year now..

Our use case is we run Circle CI runner jobs on a EKS cluster. And we have this limit for buildah to work:

resources:
  limits:
    github.com/fuse: 2

But we got the following error:

{
    "level": "ERROR",
    "logger": "controller.provisioner",
    "message": "Could not schedule pod, incompatible with nodepool \"default\", daemonset overhead={\"cpu\":\"180m\",\"memory\":\"120Mi\",\"pods\":\"8\"}, no instance type satisfied resources {\"cpu\":\"2180m\",\"github.com/fuse\":\"1\",\"memory\":\"8312Mi\",\"pods\":\"9\"} and requirements karpenter.k8s.aws/instance-category In [m t], karpenter.k8s.aws/instance-generation Exists >2, karpenter.sh/capacity-type In [on-demand], karpenter.sh/nodepool In [default], kubernetes.io/arch In [amd64], kubernetes.io/os In [linux] (no instance type has enough resources); incompatible with nodepool \"arm64\", daemonset overhead={\"cpu\":\"180m\",\"memory\":\"120Mi\",\"pods\":\"8\"}, no instance type satisfied resources {\"cpu\":\"2180m\",\"github.com/fuse\":\"1\",\"memory\":\"8312Mi\",\"pods\":\"9\"} and requirements karpenter.k8s.aws/instance-category In [c m t], karpenter.k8s.aws/instance-generation Exists >2, karpenter.sh/capacity-type In [on-demand], karpenter.sh/nodepool In [arm64], kubernetes.io/arch In [arm64], kubernetes.io/os In [linux] (no instance type has enough resources)"
}

[wmgroot]

To be clear, the "extended resource support" in this design would only support statically applied extended resources on selected nodes, correct? This would not cover cases where dynamic extended resources are needed, for example setting an extended resource for available network bandwidth which varies by EC2 instance type.

Additionally, karpenter still needs to have its provisioning and deprovisioning logic updated to account for arbitrary extended resources other than the gpu resource, right?

[ellistarn]

This would not cover cases where dynamic extended resources are needed, for example setting an extended resource for available network bandwidth which varies by EC2 instance type.

You could achieve through multiple NodeOverlays, that match different selectors.

e.g. m5.large has 2, m5.2xlarge has 4.

[wmgroot]

I don't think defining a NodeOverlay for each network bandwidth amount is a reasonable solution for my use case. We use a wide variety of instance types and sizes.

Do you think it could be feasible to incorporate some level of templating for values that can be easily retrieved from the EC2 API?

[ellistarn]

For NetworkBandwidth, I'd love to just support this one first class.

We could always try to do something that attempts to model this like:

resourcesPerCPU: 
  network: 1Gi

But it's hard to capture all potential relationships without devolving into something like https://github.com/google/cel-spec

[coveralls]

Pull Request Test Coverage Report for Build 9882748974

Details

• 0 of 60 (0.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.4%) to 77.764%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
pkg/apis/v1alpha1/zz_generated.deepcopy.go	0	60	0.0%

Totals
Change from base Build 9880691959:	-0.4%
Covered Lines:	8750
Relevant Lines:	11252

---

💛 - Coveralls

[daverin]

Hi All

I'm a bit confused about the relationship between kubernetes-sigs/karpenter and aws/karpenter-provider-aws.

The lack of support for custom resource requests/limits is blocking me from autocalling with xilinx devices.

I want to understand how to deploy this branch to EKS.

Any help or understanding would be appreciated.

[github-actions]

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

[sftim]

I want to understand how to deploy this branch to EKS.

This branch adds a “request for comments“. @ellistarn should the RFC and implementation land as one commit? Seems unusual.

[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.