Glossary
Data durability
Data durability is the probability that stored data remains intact and readable over a period, despite drive, server and other hardware failures. It is usually quoted as a percentage per object per year, often as a run of nines, and it comes from a model of failure and repair rates rather than from measurement.
How durability is expressed
A figure of 99.999999999 percent, known as eleven nines, states an annual chance of one in one hundred billion of losing a given object. Across ten million stored objects, that works out to an expected loss of one object every ten thousand years on average. The figure is a model output: an assumed drive failure rate, the amount of redundancy and the time needed to restore it after a failure, combined into the odds that too many failures overlap before repair completes.
Three inputs that move it
| Input | What it means | Effect on durability |
|---|---|---|
| Redundancy | Extra data kept, as full copies through replication or as calculated pieces through erasure coding | More redundancy tolerates more simultaneous failures |
| Independence | How far apart the pieces sit across drives, servers, racks and sites, as in distributed storage | Pieces sharing a chassis or power feed can fail together, which a simple model may not capture |
| Repair speed | How quickly lost redundancy is recreated after a failure | Each hour of repair is an hour in which a further failure can cause loss |
Failure rate is a property of the drives. Repair speed is a property of the design, and it is the input that has shifted most as drives have grown. A system that rebuilds only the data actually written, spread across many servers at once, closes the window far faster than one that refills a whole replacement drive.
Reading a durability figure
Beyond a certain number of nines, extra digits describe events so unlikely that other risks dominate. Two systems quoting the same figure may rest on different assumptions about failure rates and repair time, so the number is only comparable alongside its inputs. How quickly redundancy is restored after a drive or server failure says more than the headline percentage.
Risks outside the model
Durability counts hardware loss only. A deleted bucket, an overwritten object, data encrypted by ransomware, a software defect and an operator error all fall outside it, because the system stores the damaged state as faithfully as any other. Those risks are handled by versioning, retention locks and separate copies such as an immutable backup. Durability also differs from availability: data can be fully durable and still unreachable during an outage.