[eloff]

Availability is the % of times you try to access your data that you get it back. So 52.5 minutes of downtime a year is still within SLA.

Durability is the % of your data that doesn't die. Eleven 9s means that if you store 1TB on AWS S3 you can expect to lose 10 bytes and still be within SLA.

[jacquesm]

No, it means that if you store your data there that there is a .000000001% chance that you will lose all of it.

[hanoz]

For those wondering .000000001% per what? The answer apparently is per object year.

i.e. you could expect to lose 10 bytes of your 1TB every year if your stored it as a trillion one byte objects, but if you stored it as a single object you could expect to lose the whole thing once every hundred billion years, but none of it the rest of the time.

[speedplane]

Is that true? How can they possibly measure a probability event so small? If every human in the world was their customer, then .05 humans would lost heir data?

[jontas]

I don't know much about actuarial math but I think this number is for insurance policies more than anything else. It could be based on something like the rate of hardware failures they experience now amortized over a long period and many customers, and then adjusted to account for redundancy.

As a very simplified example, imagine they are expecting to lose 2 servers every day, this percentage might be the probability of those two servers storing the same exact object (and thus, losing it irretrievably).

[eloff]

It doesn't mean that either. It's just an SLA. Could have been a number pulled out of the air. Likely loss in real life would be granular at the object level.