In the middle of an order peak, the shop's AKS cluster adds no more nodes and pods stay Pending

The MasterLinux shop (Raspberry Pi kits, books, training vouchers) runs on the AKS cluster aks-boutique-prod (resource group rg-boutique-prod, region francecentral). The user node pool hosts the commandes API; the cluster autoscaler is enabled on it (3 to 12 nodes) and the HPA adds pods when CPU goes above 70%.

This Tuesday, the “back to school” campaign was sent at 10:45. Since 11:00, traffic has tripled and the HPA asks for 18 replicas instead of 6: about ten pods stay Pending, payment response time is above 9 seconds and customer service is receiving the first complaints.

On the team side: nothing has been deployed to the cluster since Friday. The governance team, for its part, “cleaned up costs” on Monday at the end of the day.

You are the on-call engineer, on your workstation with Azure CLI and kubectl (context aks-boutique-prod). The nodes must arrive, fast, without breaking governance.