NVME is not a hard disk (2021)

There are a lot of ideas in this post, but it’s not very cohesive and jumps around a lot without tying any of them together. And I’m not convinced the core idea relies on NVMe at all to begin with. Once you’re using iSCSI, as the article itself notes, the actual write destination is abstracted away. This was just as doable with a smart iSCSI target as it is with a native NVMe (yes, the e canonically should be lowercased, pedantically speaking) one.

Like I said: lots of ideas, so it’s hard to say what the actual point of the article was, but if you assume it’s the concept the article started with regarding storage, I don’t even think this needed to be solved at the storage layer to begin with. Sharding by proximity to endpoint eliminates the need to remotely store, but regardless of that, and multiple nodes in the same DC serving all the different shards eliminate the need to store with redundancy. In fact, it’s standard practice at FAANG-scale (or whatever they call it now) to make individual nodes fully disposable, and just make them fast and cheap instead of reliable.

Good article. I think the first half, where the author briefly discusses other storage and how some components perform their own replication is a must read for people who aren't deeply into storage. His older blog posts tagged with "storage" are littered with small gems many of which I've had the joy (read, "misery") of learning first-hand.

The corner of the landscape that I'm interested in has really shrunk: Datera was scooped up by VMWare and OnDat by Akamai removing them from the market while PortWorx was picked up by PureStorage (and the price jacked up). Ceph is reliable and, relatively easy to deploy with Rook, but in a lot of ways is less than ideal.

Very nice article, though I will note that NVMe still may be a hard disk (though not when it was written) https://nvmexpress.org/wp-content/uploads/Expanding-NVMe-Technology-to-New-Media-Types-Rotational-Media.pdf