Skip to main content

Stateful Workloads at Edge: Storage that Fits the Talos Paradigm

October 15, 2026

Our CTO and co-founder, Michael Schmidt, presented “Stateful Workloads at Edge: Storage that Fits the Talos Paradigm” at TalosCon 2026 in Amsterdam on October 15.

Talos Linux runs every node from an immutable image and manages it only through its API. There is no SSH, no hand-tuning, and the same node image runs at an edge site and in the data center. Stateful workloads test that model. Their data has to survive a failed node, and a node rebooting into a new image, at every site you run.

The talk compared three ways to store that data with storage built for the Talos model.

Three ways to store data at the edge, and what each costs

Local disks

Local NVMe is fast, but it pins every volume to the node it sits on. When that node goes down, its volumes go down with it. Pods and VMs cannot restart on another node with their data, and capacity is planned node by node, so free space on one node is stranded for the rest.

Top: with local disks, a node goes down and its volume is unavailable. Bottom: with simplyblock, a node goes down and the volume stays available on the other nodes
Figure 1: Local disk pins the volume to one node. Pooled storage keeps it available when a node fails.

Storage arrays

An array puts storage in a separate box, connected to the nodes over a storage fabric. At the edge, that means an array in every store, plant or cell site, each with its own controllers and fabric to run next to Kubernetes. The array also manages LUNs, not the containers and VMs that use them.

Top: Kubernetes nodes connected to a separate storage array over a storage fabric. Bottom: simplyblock running on the Kubernetes nodes themselves
Figure 2: An array is a separate box. Simplyblock runs on the servers that already run Kubernetes.

Typical software-defined storage

Software-defined storage runs on standard servers, which suits the edge. The usual price is speed and isolation. A kernel I/O path and three-way replication burn CPU on every I/O, so IOPS flatten as nodes are added. Extra network hops and protocol translation add latency over local NVMe. One busy volume slows every other volume in the pool, and tenant keys and access control are added late, if at all.

Storage that fits the Talos model

Simplyblock is software-defined, distributed block storage for Kubernetes that closes those gaps.

  • Volumes are served over NVMe/TCP end to end, from a user-space SPDK engine that bypasses the kernel I/O stack. There is no translation layer.
  • Erasure coding spreads data and parity chunks across nodes. Data survives node loss with less overhead than three copies, and no triple write traffic.
  • Per-volume IOPS and bandwidth limits stop one busy volume from slowing the rest, with pools and QoS classes per tenant.
  • Each volume can be encrypted, with tenant keys in an external KMS, FIPS 140-3 modules and NVMe-oF authentication.

On Talos, simplyblock is set up the way Talos sets up everything else:

Talos Linux providesSimplyblock adds
Immutable, minimal OS, managed only by APIOperator and CSI driver deployed as pods
Host settings as declarative machine configHugepages, nvme-tcp and core isolation as config patches
Image-based, rolling node upgradesVolumes stay available while a node reboots
One node image from edge site to data centerOne storage stack from edge sites to large clusters

Rolling upgrades are where this shows. Talos upgrades a cluster by rebooting one node at a time into the new image. Because the data is spread across the cluster, every volume stays available while each node reboots.

A node rebooting during a rolling upgrade while the volumes on the cluster stay available
Figure 3: One node reboots during an upgrade. The volumes stay available.

The same software runs hyperconverged on the worker nodes, disaggregated on storage nodes, or both, on standard rack servers. You add NVMe drives to servers you already own, and capacity and performance grow with every node. Platform engineers install, upgrade and monitor it through Kubernetes, and because simplyblock knows which volumes belong to which container or VM, migration, consistency groups and disaster recovery act on the application, not on LUNs.

Virtual machines on Talos

KubeVirt VMs on Talos get the same block storage as containers, with the data services VMs expect. A VMware® exit moves vSphere VMs onto NVMe block volumes with snapshots, clones and disaster recovery already in place. Copy-on-write clones boot many VMs from one golden image without copying data, for VDI, CI and labs. Databases that cannot be containerised get NVMe-class I/O and per-volume QoS.

Shared RWX block volumes let VMs live-migrate between nodes for maintenance and Talos upgrades. For recovery, VM disks and definitions replicate to a second cluster with failover and failback per application, and immutable snapshot generations in S3 let you restore VMs on a clean cluster from a chosen point.

One declarative stack, from node to volume

Talos made the node declarative. The talk closed on doing the same for storage: one declarative stack, from node to volume, that stays available through upgrades and failures and runs the same from a single edge site to a large cluster.

Since not everybody could join the session, you can download the presentation. To try it on your own nodes, Sidero Labs, the team behind Talos, has a step-by-step guide to running simplyblock on Talos in its docs.

If you run Talos, the next step is a proof of concept on your own cluster.

Related reading: simplyblock + Talos and storage for virtual machines in Kubernetes.