EKS InferenceCluster node pools can only use a region's first three Availability Zones

What happened?

An EKS InferenceCluster's VPC only has subnets in the region's first three Availability Zones, so a pool's zones must be among them. In us-east-1, g6.12xlarge was out of capacity in 1a, 1b and 1c, and AWS suggested 1d or 1f, which no pool could use.

compose-eks-cluster derives each subnet's AZ from its index in subnetCidrs, which defaults to three CIDRs:

modelplane/functions/compose-eks-cluster/function/fn.py

Lines 443 to 444 in b6b28f0

    
           for i, cidr in enumerate(self._networking().subnetCidrs or []): 
        
               az = _az(self.xr.spec.region, i)

modelplane/apis/eksclusters/definition.yaml

Lines 67 to 80 in b6b28f0

    
                             subnetCidrs: 
        
                               type: array 
        
                               description: >- 
        
                                 Subnet CIDRs, one per Availability Zone. EKS requires 
        
                                 at least two subnets in different AZs. The function 
        
                                 picks AZs by index from the region — index 0 maps to 
        
                                 the region's first AZ alphabetically (e.g. us-west-2a), 
        
                                 index 1 to the second, etc. 
        
                               minItems: 2 
        
                               maxItems: 6 
        
                               default: ["10.0.0.0/20", "10.0.16.0/20", "10.0.32.0/20"] 
        
                               items: 
        
                                 type: string 
        
                                 maxLength: 18

The EKSCluster XR accepts up to six, but the InferenceCluster doesn't expose subnetCidrs. Even with more subnets, compose-eks-cluster picks their zones by letter, not by name.

How can we reproduce it?

Create an EKS InferenceCluster in us-east-1 with a pool whose zones is [us-east-1d].
Look at the private subnets the EKSCluster composes. There's none in us-east-1d.

What environment did it happen in?

Modelplane version: main plus #542 (e574dcf)

Crossplane v2.4.0
provider-aws-ec2 v2.8.1

Navigation Menu

Repository navigation

EKS InferenceCluster node pools can only use a region's first three Availability Zones #581

What happened?

How can we reproduce it?

What environment did it happen in?

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	for i, cidr in enumerate(self._networking().subnetCidrs or []):
	az = _az(self.xr.spec.region, i)

	subnetCidrs:
	type: array
	description: >-
	Subnet CIDRs, one per Availability Zone. EKS requires
	at least two subnets in different AZs. The function
	picks AZs by index from the region — index 0 maps to
	the region's first AZ alphabetically (e.g. us-west-2a),
	index 1 to the second, etc.
	minItems: 2
	maxItems: 6
	default: ["10.0.0.0/20", "10.0.16.0/20", "10.0.32.0/20"]
	items:
	type: string
	maxLength: 18

EKS InferenceCluster node pools can only use a region's first three Availability Zones #581

Description

What happened?

How can we reproduce it?

What environment did it happen in?

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions