Volt on Kubernetes: stateful workloads that scale and upgrade in service

    What you’ll learn

  • Stateful workloads on Kubernetes usually end up outside the cluster on separate infrastructure or inside it as a fixed-size StatefulSet that Kubernetes cannot manage.
  • A K-safe cluster must resize in steps of K+1 nodes, while the Horizontal Pod Autoscaler adds or removes one pod at a time based on pod CPU or memory.
  • The Volt Operator autoscales on transactions per second, CPU, memory footprint, and partition idle time, with a ten-minute default stabilization window that prevents flapping.
  • In-service software upgrades (ISSU) let a Volt cluster on Kubernetes upgrade one node at a time while it keeps processing transactions.
  • Active(N) replication runs Volt active-active across separate Kubernetes clusters, regions, and clouds, including three-site deployments where every site serves live traffic.

Platform teams adopt Kubernetes to get one way to deploy, scale, patch, and observe everything they run. Stateless services get there easily, but the systems of record behind real-time charging, policy control, fraud detection, and mediation usually don’t. They end up in one of two compromises: running outside the cluster on virtual machines or bare metal with their own tooling and on-call rotation, or running inside it as a fixed-size StatefulSet that Kubernetes schedules but can’t really manage. Either way, every capacity change or operating system patch on the data tier becomes a project with its own change window.

Where built-in Kubernetes tooling falls short

Closing that gap starts with knowing exactly where the Horizontal Pod Autoscaler, rolling restarts, and node drains break down for a strongly consistent, partitioned data platform. The HPA adds one pod at a time, while a K-safe cluster, which keeps K replicas of every partition, has to grow and shrink in steps of K+1 nodes. When the HPA scales a StatefulSet down, it deletes the highest-numbered pod regardless of the data that pod holds. Pod CPU also says little about a transactional system’s real headroom, and rolling upgrades run in pod order rather than an order that keeps every partition available.

How an Operator makes stateful workloads Kubernetes-native

Kubernetes answers this class of problem with an Operator, a controller that encodes the domain knowledge Kubernetes itself lacks. You’ll get a clear picture of how the Volt Operator orders every structural change so K-safety holds, from a single custom resource and Helm chart to online scaling in K+1 steps that moves partitions off departing nodes before they leave. Autoscaling gets easier to reason about once it runs on transactions per second, CPU, memory footprint, and partition idle time, with a stabilization window of ten minutes by default to prevent flapping. In-service software upgrades, zone-aware placement groups, and Active(N) replication between Kubernetes clusters complete the model.

Who it’s for

If you’re a platform engineer, SRE, or architect deciding whether your data tier can live on the same Kubernetes platform as everything else, you’ll come away knowing what to ask of any stateful system before it joins the cluster. You’ll be able to tell a data platform that Kubernetes merely schedules from one it can actually scale, upgrade, and replicate, and to spot where a second platform with its own DBAs, runbooks, and maintenance windows is still costing you. Real deployment patterns from Tier-1 communications service providers on OpenShift, EKS, AKS, and vendor distributions give you something concrete to measure your own setup against. Start with the gap between pod order and partition order, and the rest of the model follows.