
Hyperconverged infrastructure: when three servers become one cluster
With hyperconverged infrastructure, compute and storage run together on a handful of standard servers, and software turns their local disks into one shared pool. That replaces separate servers with a SAN behind them. Whether it suits you depends less on the brand than on four design questions: how many nodes, how much usable space, which network and which licence.
Hyperconverged infrastructure puts compute and storage in the same servers and manages them as one cluster. You grow by adding a node rather than replacing a storage array.
Design for the day one node fails: three nodes survive one failure, but not while a second is in maintenance. And plan in usable capacity, because data sits in the cluster more than once.
HCI is not a backup. Copies inside the cluster protect against a failed disk, not against ransomware or a deleted folder.
What hyperconverged infrastructure is, and what it replaces
The classic setup has three tiers: servers running the virtual machines, a central storage system (a SAN or NAS) holding their disks, and a separate network in between. Each of the three has its own management, its own firmware and often its own supplier.
Hyperconverged infrastructure, HCI for short, removes the middle tier. Each server has its own disks, and software spreads the data across every server in the cluster so that the whole behaves like one shared store. If a server fails, its virtual machines start on another one, using data that was already there.
What you get back is mainly simplicity: one management console, one kind of hardware, and growth in steps of one node. What you give up is flexibility: compute and storage grow together, even when you only need one of them.
When HCI fits, and when it does not
It fits when part of your estate has to stay on your own hardware (production software that must sit close to machines, data that may not leave the building, a connection too slow for the cloud), when you run more than one server can carry and losing one server must not mean downtime, or when an ageing SAN is due for replacement and you would rather not buy that tier again. A branch office with a few local systems is a good case too, with a small two-node cluster.
It does not fit when one good server with a sound backup is enough, when most of the work already lives in Microsoft 365 and other cloud services, or when you mainly need storage without compute: CCTV footage, an archive, large file libraries. Then every expansion pays for processors you do not use, and a separate storage system is cheaper to run.
The prior question is whether anything should still run on your own kit at all. That is worked through in our piece on own servers or the cloud.
Two, three or four nodes
The number of nodes decides how much failure the cluster can absorb, and it is the most important design decision.
- Two nodes are possible on most platforms, but only with a witness: a small component in a third location, often a virtual machine elsewhere or a cloud endpoint, that decides which side carries on if the link between the two breaks. Suitable for a branch, not for the core.
- Three nodes is the usual minimum. One may fail. But if one is in maintenance for an update and a second then fails, you have no margin left.
- Four or more gives you that margin: maintenance and a failure at the same time, without the cluster stopping.
Size the cluster with one node missing. If everything only runs when every node is present, the first failure is an outage, and you have paid for a cluster without getting the reason for it.
Raw and usable capacity
An HCI cluster keeps every block of data on more than one node, or a failure would cost data. With two copies the usable space is roughly half of the disks installed; with three copies, a third. Some platforms offer a more economical form of spreading (erasure coding), but it needs more nodes and costs compute when writing.
Leave free space on top of that. When a node fails, the cluster rebuilds the missing copies on the nodes that remain, and there has to be room for them. Vendors state a free margin in their design guidance; do not plan beyond it.
So always ask for a quote in usable capacity at the chosen protection level, not in raw terabytes. Two proposals with the same raw figure can differ by a factor of two in what you can actually store.
The network is now your storage connection
In a classic setup, storage traffic runs over its own network to the SAN. With HCI it runs between the nodes, across the switches. Every write also goes to a second node, so a slow or flaky switch means slow or flaky storage.
- Two switches, every node connected to both. One switch is one fault away from a cluster that sees itself split in pieces.
- Fast ports. The platforms set minimum speeds between nodes, well above what an office switch offers. Check that before the switches are ordered.
- Separated traffic. Storage, management and user traffic in their own VLANs, so a busy backup or a migration does not crowd out storage traffic.
This is the part most often underestimated, because in a quote the switches look small next to the servers.
Platform and licence: the choice that stays longest
You replace the hardware after a few years; the platform usually stays longer, because switching means moving every virtual machine to another hypervisor. The best-known options:
- Nutanix, with its own hypervisor included, or on VMware.
- VMware vSAN, which since the Broadcom acquisition is sold only as a subscription within bundles, no longer separately or as a perpetual licence. If you run VMware, renewal will look different from the original purchase.
- Microsoft Azure Local, the successor to Azure Stack HCI, with Hyper-V and a link to Azure for management.
- Proxmox VE with Ceph, open source, with paid support as an option.
What drives cost over time is not the server but the licensing model: per processor core, per node or per amount of storage, and what happens at renewal. Ask as well whether the hardware is on the platform’s list of supported configurations; a cluster outside that list runs fine until you need the vendor.
HCI is not a backup
The copies inside a cluster protect against a failed disk or a lost node. They do not protect against a deleted folder, a bad update or ransomware: those are written to every copy just as faithfully as your normal data.
Snapshots are not a backup either. They sit on the same storage, so they go down with it if the cluster is the problem, and an attacker with admin rights deletes them first.
So keep the backup outside the cluster, with at least one copy that cannot be altered and one off site. How to set that up is in our article on the 3-2-1 backup rule. And test restoring a whole virtual machine, not just a file.
Where our work stops
What we do. The design (number of nodes, protection, usable capacity, the network underneath), supplying servers and switches, building the cluster, migrating the existing virtual machines and managing it afterwards, including updates in an order that keeps the cluster up, and the backup beside it. We deliver worldwide and have our own local engineers on site in every country.
What we do not do. We do not make the platform: the software, support for that software and the licence terms come from the vendor, who also decides what changes at renewal. We are not building contractors: the electrical circuits and cooling of the room the cluster sits in are work for an installer. And we do not promise availability the design cannot carry; a three-node cluster has the limits described above, even when we manage it.
Questions we get about this
The five that come up most often.
What is hyperconverged infrastructure?
A setup in which compute and storage sit together in a few standard servers, and software turns their local disks into one shared pool. The whole is managed as one cluster. It replaces the classic three tiers of servers, a separate storage system and the network between them.
What is the difference between HCI and a SAN?
With a SAN, storage sits centrally in a separate system that every server connects to. With HCI, storage sits in the servers themselves and software spreads the data across the cluster. HCI is simpler to manage and grows per node; a SAN lets compute and storage grow independently.
How many nodes do you need for a hyperconverged cluster?
Three nodes is the usual minimum, and then one may fail. Two is possible on most platforms, but only with a witness in a third location that decides which side carries on during a fault. If you want to do maintenance while something else breaks, take four.
Does HCI replace backup?
No. The copies in the cluster protect against a failed disk or server, not against deleted files, a failed update or ransomware. Snapshots sit on the same storage, so they are not a backup either. Keep a backup outside the cluster, with a copy that cannot be altered.
Do you also supply the licences and vendor support?
We supply the hardware and can purchase the licences for you, but the licence terms and the support on the software come from the vendor, who also decides what changes at renewal. We design, build and manage the cluster, and bring in the vendor when the problem is in their software.
Further reading
What hyperconverged infrastructure connects to.
The brands above, one page each
What the brand makes, which lines we supply and what we do around it.
Shall we size your cluster with one node missing?
Send us what runs today and what will be added over three years. We work out how many nodes, what usable space and which network that calls for.
Practical IT knowledge in your inbox
New guides on management, security and the workplace, written by the people doing the work. No sales talk, and you can unsubscribe in one click.