AZ imbalance with equal NodePool weights for Spot instances (ARM64)

Summary

I’m trying to have Karpenter evenly launch Spot nodes across two Availability Zones (eu-south-1a and eu-south-1b).
To achieve this, I created two NodePools — one restricted to each AZ — and set the weight to 100 for both.
The idea was to let Karpenter decide freely while keeping the AZ distribution balanced.

However, when scaling up workloads, I noticed Karpenter heavily prefers eu-south-1b (around 90% of nodes), even though both NodePools have the same weight and configuration

Expected Behavior

With equal weights and similar NodePool configurations, I would expect Karpenter to roughly balance node provisioning across the two AZs over time.

Actual Behavior

When scaling a single deployment from 1 to 160 replicas (to trigger node creation), almost all new nodes are provisioned in eu-south-1b.

• The application has no affinity rules, only a nodeSelector for arm64.
• It's a single container per pod.
• I’m not using TopologySpreadConstraints because many of my workloads have only one pod.

Interestingly, if I reduce the weight of the eu-south-1b NodePool to 70 while keeping eu-south-1a at 100, the resulting distribution becomes almost perfectly balanced (around 50% nodes in each AZ).

This makes me wonder:

• Is this caused by a significant Spot capacity imbalance between AZs for ARM64 instances?
• Or am I missing a configuration that would help achieve fair AZ balancing without manually tweaking weights?

Steps to Reproduce

1. Create two NodePools, one restricted to eu-south-1a and one to eu-south-1b, both with:
   • weight: 100
   • capacity-type: spot
   • kubernetes.io/arch: arm64, amd64
2. Deploy a single Deployment with:
   • replicas: 160
   • nodeSelector for arm64
3. Observe node provisioning across AZs.

Environment

• Karpenter version: 1.6.0
• EKS version: 1.32
• AWS region: eu-south-1
• Instance types: ARM64, CPU 4 or 8, instance-generation > 2
• Capacity type: Spot

My config nodepool

# Vedi 'README.md#nodepool' Karpenter per dettagli.
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
name: standard-az-a
spec:
# Aggiunto weight per bilanciare con nodepool-standard-az-b
weight: 100
template:
 metadata:
   labels:
     node-scope: standard-az-a
 spec: 
   requirements:
     - key: kubernetes.io/arch
       operator: In
       values: ["arm64", "amd64"]
     - key: kubernetes.io/os
       operator: In
       values: ["linux"]
     - key: karpenter.sh/capacity-type
       operator: In
       values: ["spot"]
     - key: "karpenter.k8s.aws/instance-cpu"
       operator: In
       values: ["4", "8"]
     - key: karpenter.k8s.aws/instance-generation
       operator: Gt
       values: ["2"]
     - key: "topology.kubernetes.io/zone"
       operator: In
       values: ["eu-south-1a"]

   nodeClassRef:
     group: karpenter.k8s.aws
     kind: EC2NodeClass
     name: default
limits:
 cpu: 1000
disruption:
 consolidationPolicy: WhenEmptyOrUnderutilized
 consolidateAfter: 15s
 expireAfter: 720h

 maxSurge: "20%"
 budgets:
 - nodes: "10%"  # Limita il numero di nodi che possono essere drenati contemporaneamente
   schedule: "* * * * *"  # Cron schedule per quando applicare questo budget
   duration: "1h"  # Per quanto tempo questo budget è valido dopo l'attivazione




# Vedi 'README.md#nodepool' Karpenter per dettagli.
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
name: standard-az-b
spec:
# Aggiunto weight per bilanciare con nodepool-standard-az-a
weight: 100
template:
 metadata:
   labels:
     node-scope: standard-az-b
 spec: 
   requirements:
     - key: kubernetes.io/arch
       operator: In
       values: ["arm64", "amd64"]
     - key: kubernetes.io/os
       operator: In
       values: ["linux"]
     - key: karpenter.sh/capacity-type
       operator: In
       values: ["spot"]
     - key: "karpenter.k8s.aws/instance-cpu"
       operator: In
       values: ["4", "8"]
     - key: karpenter.k8s.aws/instance-generation
       operator: Gt
       values: ["2"]
     - key: "topology.kubernetes.io/zone"
       operator: In
       values: ["eu-south-1b"]
   nodeClassRef:
     group: karpenter.k8s.aws
     kind: EC2NodeClass
     name: default
limits:
 cpu: 1000
disruption:
 consolidationPolicy: WhenEmptyOrUnderutilized
 consolidateAfter: 15s
 expireAfter: 720h
 maxSurge: "20%"
 budgets:
 - nodes: "10%"  # Limita il numero di nodi che possono essere drenati contemporaneamente
   schedule: "* * * * *"  # Cron schedule per quando applicare questo budget
   duration: "1h"  # Per quanto tempo questo budget è valido dopo l'attivazione