Why wouldn't that be handled already by the drives themselves?

I wonder about this, too. Since access latency is the primary concern, a neat solution would be to create some dynamic "fastup" storage scheme. For large files, loosen the always-on requirement; to fulfill a request the first N bytes are served from the always-on pod, and the tail is stored and served from pods that dynamically power up drives only at access time. The size of N is computed such that the IO time it takes to make it through the fragment stored on the always-on pod is > the power-on time for the pods where the tail is stored.

Brian from Backblaze here.

> Have you ever computed how much power you would save if you turned them off when not needed?

Electrical power is a gigantic part of our datacenter bill (like 60%). However, we cannot really ever turn off the hard drives.

If they were powered down, it would take too long to retrieve data. It would be like Amazon Glacier where it can take 15 minutes to get a single image back.

> is every pod constantly in use?

The way we built it, yes. If one user deletes a backup or unsubscribes from the Backblaze service, it frees a little bit of space on pods all over the datacenter. So we allow some other customer to use that space up as soon as we can.

We also use ALL the pods in the datacenter to give us "breadth" to accept data faster. This is particularly helpful on surges of new customers due to a press event, or if a competitor like CrashPlan decides to exit the market.