Redundancy

Redundancy means: something is intentionally present multiple times, even though one instance would mathematically suffice. In engineering, this refers to components or entire systems that can do the same thing. As long as everything runs smoothly, this looks like waste. But as soon as one part breaks, the spare part immediately takes over the work. An airplane with two engines can keep flying with just one, and that is exactly what the second engine is for. In information technology, the principle works identically, except there it concerns hard drives, servers, power lines, and internet connections.

Why failures are more expensive than spare parts

Every technical component eventually breaks. This is not a question of whether, but of when. A hard drive might last five years, a power supply burns out, an excavator severs a fiber optic cable. In a large data center with tens of thousands of hard drives, statistically at least one fails every day. Anyone hoping that nothing will happen doesn’t have a plan, just luck.

For companies, downtime is directly measurable in money. If an online shop is offline for an hour, that hour’s revenue is lost. For a bank or a payment service provider, the amounts quickly reach the millions. On top of that comes the loss of trust, which is harder to quantify. That’s why providers guarantee a certain level of availability in contracts, such as 99.99 percent of the time.

This number sounds like splitting hairs, but it is crucial. 99 percent availability allows for about 3.65 days of downtime per year. 99.99 percent allows for only about 53 minutes. Each additional nine costs significantly more money, because it requires more duplicate hardware. Companies therefore calculate precisely how much redundancy is worthwhile.

From the second power supply to the second data center

Redundancy is built up on several levels. At the smallest level, a server has two power supplies connected to two different circuits. For hard drives, RAID systems are used, in which data is distributed across multiple drives and additionally protected. If one drive fails, its data can be reconstructed from the others. The technician replaces it while the system keeps running.

A level higher are entire data centers. Large cloud providers operate so-called availability zones, meaning multiple locations within a region. These locations have separate power supplies and separate network connections, but are close enough to each other for fast connections. A fire or power outage at one location does not paralyze the others.

Crucial here is the term single point of failure, meaning the one single spot on which everything depends. Two servers are of little use if both are connected to the same power cable and the same network switch. This exact mistake happens often in practice. True redundancy requires that the backup systems can also fail independently of one another.

Redundancy in AI data centers and in the news

When training large AI models, thousands of graphics chips compute together on a single task for weeks. If one of them fails, the entire run effectively grinds to a halt. That’s why operators regularly save intermediate states, so-called checkpoints, and keep spare chips ready. The computation then restarts from the last saved point instead of from the beginning. Without this safeguard, training runs spanning several weeks would be practically impossible.

In business news, the term usually comes up when redundancy was lacking. Major cloud outages regularly paralyze streaming services, banking apps, and online shops simultaneously. The reason is often that many companies do use multiple servers, but all with the same provider and in the same region. In response, some companies rely on multiple providers in parallel, which experts call multi-cloud.

A common misconception is equating redundancy with backup. A backup is a copy saved yesterday, which is restored after damage occurs. That takes time and costs time. Redundancy, on the other hand, steps in immediately, without anyone having to intervene. However, it does not help against accidentally deleted data, since the deletion is dutifully propagated to all copies. Good systems therefore need both.