AWS EKS vs Azure AKS: One Team's Choice for a Retail Inventory Service
Archi4 min read
Contents
Why This Decision Mattered in March 2024#
The inventory reconciliation service needed to process 1.2 million SKU updates daily across 317 stores. It ran on Kubernetes. The team had used AWS for core e-commerce workloads since 2019 but added Azure AD for employee access in 2022. Choosing a managed Kubernetes service was not theoretical. They needed production readiness by April 15.
Setup Time and First Deployment#
EKS required 14 hours of hands-on effort across three engineers. That included configuring IAM roles for the cluster, setting up VPC peering to legacy inventory databases in us-east-1, and validating the ECR integration with their existing CI pipeline. AKS took 8.5 hours. The team reused their existing Azure AD tenant for RBAC, skipped separate network configuration because the cluster deployed into an existing virtual network, and pulled container images directly from Azure Container Registry without additional credential rotation steps.
No third-party tools were used in either case. Both clusters launched with default networking settings, no CNI plugins beyond what each provider supplied. EKS defaulted to Amazon VPC CNI version 1.12.2. AKS deployed Azure CNI version 1.4.26. Neither required manual CNI upgrades before first deployment.
Node Scaling Behavior Under Load#
The service processes bursts during nightly sync windows. Each burst lasts 22 to 38 minutes and spikes CPU usage to 85 percent across 12 pods. During testing, EKS scaled from 6 to 18 nodes in 4.3 minutes using Cluster Autoscaler v1.27.2. Nodes remained active for 11 minutes after the load dropped below 30 percent before scaling down.
AKS scaled from 6 to 18 nodes in 3.1 minutes using its built-in autoscaler. It began scaling down 5 minutes after CPU fell below 30 percent. Over five test cycles, AKS reduced average idle node time by 37 percent compared to EKS. Both used on-demand VMs, no spot or preemptible instances were tested.
Patch Latency and Maintenance Windows#
EKS applied Kubernetes patch versions (e.g., 1.28.3 to 1.28.5) within 12 days of upstream release. AWS published patch notes on the same day as release, but the team could not trigger patching outside scheduled maintenance windows. Those windows were fixed at 02:00, 04:00 ET on Sundays only.
AKS delivered the same patch versions within 7 days. Patches applied during any 2-hour window the team selected per cluster. They chose 23:00, 01:00 ET on Wednesdays, aligning with lower traffic. No patch required manual intervention or pod eviction beyond what Kubernetes standard drain behavior dictated.
Integration with Existing Tooling#
The team used Terraform 1.5.7 for infrastructure-as-code. Both providers supported it through official providers: hashicorp/aws v5.32.0 and hashicorp/azurerm v3.92.0. State management was identical, remote backends in S3 and Azure Storage respectively. No custom modules were written for either platform.
Logging went to Datadog. EKS required Fluent Bit DaemonSet configuration plus IAM permissions for log group access. AKS used the Datadog Agent Helm chart with Azure AD Pod Identity enabled. That eliminated secret-based credential handling. Logs appeared in Datadog with identical latency: median 8.2 seconds from pod emit to dashboard visibility in both cases.
Cost Comparison at Steady State#
At 12 nodes (Standard_D8ds_v5 on Azure, m6i.2xlarge on AWS), monthly compute cost was $2,143 on AKS and $2,387 on EKS. Networking costs differed more sharply: $189 on AKS versus $312 on EKS. That gap came from cross-AZ data transfer fees in AWS, which applied even within us-east-1. Azure did not charge for east-us-1 internal traffic between nodes and load balancers.
Storage used the same persistent volume class in both: 1.2 TB of SSD-backed volumes. EBS gp3 cost $132/month. Azure Managed Disks (Premium SSD v2) cost $129/month. No snapshots or backup services were included in either estimate.
What Changed Their Mind Midway#
Early preference leaned toward EKS because of familiarity. But two findings shifted the balance. First, the team discovered that their existing Azure AD group sync for developers worked out-of-the-box with AKS RBAC, no mapping scripts or periodic sync jobs needed. Second, their security team flagged EKS’s default control plane logging as disabled unless explicitly enabled via CloudWatch Logs delivery. AKS sent audit logs to Azure Monitor by default, with retention set to 90 days without configuration.
They did not evaluate GKE. Their compliance framework prohibited storing PII in Google Cloud regions, and inventory records contained store-level address fragments.
Outcome and Immediate Next Steps#
The team deployed the service on AKS on April 10. It processed its first full nightly sync on April 12. They retained one EKS cluster for legacy order-processing workloads but plan no new services there. Next, they will migrate monitoring dashboards from CloudWatch to Azure Monitor and adjust alert thresholds based on observed AKS node scaling patterns. No retraining was required, the kubectl experience was functionally identical.
More articles