2026-11-18

November add-on upgrades: renamed GPU metrics and a fixed Secrets Store CSI driver

#add-on  #kubernetes  #eks  #update  #upgrade  #component  #observability 

Our November round of add-on upgrades is rolling out: non-production clusters first, production after a soak period. Two changes are worth knowing about: the GPU metrics have new names, and the Secrets Store CSI driver is back on its latest release.

What changed

  • GPU metrics have new names and units. The NVIDIA DCGM exporter 4.9.0 renames every GPU metric, and most units change: utilization is a ratio from 0 to 1 instead of a percentage, memory is in bytes, and clock speeds are in Hz. Our GPU dashboards already use the new names, and the GPU monitoring guide lists every old and new name.
  • The Secrets Store CSI driver is on its latest release again. We had held it on 1.5 because 1.6.0 could delete the Secrets it syncs from AWS Secrets Manager. Version 1.6.2 fixes that.
  • A new node image. EKS node image v20261107 replaces every node once. Karpenter moves your workloads within your node pools’ disruption budgets, as in any node rotation.
  • Atlantis runs OpenTofu 1.13.9. Your atlantis plan and atlantis apply runs use OpenTofu 1.13.9 now, up from 1.13.1.
  • Small behaviour changes.
    • Traefik now rejects BasicAuth passwords longer than 1024 bytes, and more than 500 values in a single request header.
    • The oauth2-proxy in front of internal tools reports itself as not ready while its Redis is down.
    • On clusters with the Mountpoint for S3 CSI driver, opening a file for writing fails with ENOMEM once the limit of files open for writing at the same time is reached.

What you need to do

If you built your own Grafana dashboards or Prometheus alerts on the DCGM_FI_DEV_* GPU metrics, update them to the new names. Check thresholds too: an alert on > 90 percent utilization becomes > 0.9, and it would never fire on the new ratio. Without GPU nodes, there’s nothing to do.

Add-on upgrades

We also picked up the latest EKS build revisions for CoreDNS, kube-proxy and the snapshot controller (packaging updates only, no version change).