data durability

Durability is a measure of how likely data is to be lost on a storage system. It’s commonly measured in “nines:” for example, a storage system may offer eleven nines of durability.¹

Nines of durability is largely a marketing metric because of how it is defined. Amazon originally defined it as:²

the probability that the object will remain intact and accessible after a period of one year

This makes “eleven nines” sound good when it really may not be. Consider:

• If you have a billion objects ($10^9$) and eleven nines of durability ($P_{\text{loss}}=1-0.99999999999=10^{-11}$ ), you wind up with a 1% chance of losing data a year ($1-e^{10^9\times 10^{-11}}\approx 0.01$). Not so good.
• If what you think of as an object is actually sharded over multiple parts, you might have more than a billion objects, and therefore a higher probability of data loss.
• This durability says nothing about the size of each object.

Footnotes

1. For example, “LRS provides at least 99.999999999% (11 9s) durability of objects over a given year.” (from Data redundancy - Azure Storage | Microsoft Learn) ↩

2. New: Amazon S3 Reduced Redundancy Storage (RRS) | AWS News Blog ↩