Glossary

Scale-out storage

Scale-out storage grows by adding servers, called nodes, each bringing its own drives, processing and network connections to one system. Capacity and performance grow together, and older hardware retires gradually, with data rebalancing in the background, so there is no forklift migration project or planned outage.

What separates scale-up from scale-out?

A traditional array scales up. Drive shelves are added behind a fixed pair of controllers until the controllers become the ceiling, either for performance or for how much they can address. The next step is a new array and a migration. A scale-out system adds complete nodes one at a time while the system stays online. Every node contributes processing and network capacity, so the system does not have a single controller to outgrow.

How growth and retirement work

  1. A new node joins the cluster and announces itself.
  2. Data rebalances in the background, with some existing data moving to the new node.
  3. Capacity and throughput rise as soon as the node takes load.
  4. When a node ages out, its data drains to other nodes and it leaves with no outage.

The effect on operations is that hardware generations can run side by side, and replacement is a continuous routine, not a project every five years. This pattern is what makes exabyte-scale systems feasible, since no migration of that size is practical.

Where are the real limits?

Drives are rarely the limit. Two other resources are. The network between nodes carries rebalancing, repair and client traffic, and a cluster that grows past its network design slows down however many drives it has. Metadata is the second: tracking where every object or block lives gets harder as the cluster grows, and a design that keeps its index on a few nodes will bottleneck there. These constraints are why a scale-out claim is tested by asking how the system behaves at ten times its current size, not by counting nodes.

The coupling cost

A further question is whether compute and capacity can grow independently. If every node pairs fixed CPU with a fixed number of drives, adding space means buying processors that are not needed, which is the pattern that makes hyper-converged infrastructure inefficient for data that grows on its own. Designs that let storage nodes and processing scale separately avoid it. The approach also underlies most modern software-defined storage and distributed storage.

Frequently asked questions

Is scale-out storage the same as distributed storage?

They overlap. Distributed storage describes spreading data across multiple machines. Scale-out describes the growth model of adding nodes. Most scale-out systems are distributed, but a distributed system can be fixed in size.

Does scale-out storage need identical hardware?

Not necessarily. Well-designed systems accept nodes of different generations and sizes in one cluster, which is what allows gradual refresh instead of replacing everything at once.

What happens to performance as nodes are added?

Throughput generally rises with node count because more servers handle requests in parallel. Small-object and metadata-heavy workloads depend on how the index scales, so results vary by design.