The big cloud providers use something like erasure coding across a collection of tapes (or hard drives) to preserve information. If you've ever downloaded something from usenet and restored it with quickpar, this is basically the same idea. Every physical cartridge gets a chunk of the file. Depending on parity settings you would have to lose many tapes distributed across an arbitrary # of locations to actually lose any data.

FAcebook had a tape team, however it never really went into prod, as far as I'm aware it was shut down in 2024.

I think the issue with tape at scale is that there is a lot of mechanical stuff that needs to be overcome. You will need a custom robot to allow for the management of hundreds of thousands of slots. I think also you'll need a custom tool for identifying the individual tapes (LTO does have a few ways of ID-ing the individual tapes)

Then you'll need make sure that your data stream don't mix. that seems easy, but you are optimising for filling tapes, and minimum retrievals per object. One advantage is that ibjecst should be encrypted before they even enter the tape storage system so the risk oulling the wrong file becomes one of correctly decrpyting, rather than fine management of file paths.

Note that the hardware the big cloud providers use for tape archives is very different from the devices the GP poster appears to have used, and usually have robust verification systems.

In addition the need for cloud providers to comply with userdata erasure means they have to frequently test their ability to read and write their backups - which should increase your confidence in their correctness. Yes problems can and do still happen, but IME far less often than the more traditional tape setups GP describes.

[flagged]