1. The encyclopedia and the archive
Isaac Asimov’s Foundation opens with the Encyclopedia Galactica, an attempt to preserve all human knowledge on a single planet. The desktop faces a smaller version of the same problem: where do you put the data so it survives faults and still reaches the people who need it?
Entry 223 sketched the storage attachment. This entry decides where storage lives relative to compute.
2. Topology one: storage on compute boards
The simplest approach is to put SSDs on the same boards that do the processing. This is how many small spacecraft work today. It is compact, low-latency, and requires no extra network.
The problem is coupling. If a compute board fails, its local storage may be offline until the robot replaces the whole board. If storage needs grow faster than compute, you are forced to add compute you do not need. And if a customer wants pure storage service, there is no place to put it.
3. Topology two: dedicated storage cells
A dedicated storage cell is a separate attachment whose only job is to hold data. Compute boards talk to it over the desktop network. This decouples scaling: you can add storage without adding compute, and vice versa.
It also makes servicing cleaner. A failed drive is replaced in a storage cell, not in a running compute board. The robot uses the same rails and latches it already knows.
The downside is extra network traffic and slightly higher latency. For most Earth observation workloads, neither is a problem.
4. Topology three: fully distributed storage
In a fully distributed design, storage is spread across many cells — compute, storage, and perhaps other payloads — and presented as a single federated pool. This is resilient: losing one cell does not lose a dataset if replicas exist elsewhere.
The complexity is software. A distributed filesystem or object store must handle network partitions, slow nodes, radiation-induced bit flips, and robot-triggered hardware changes. This is not impossible, but it is not a first-generation problem.
5. The choice for the first desktop
Start with dedicated storage cells. They give the benefits of separation without the full complexity of a distributed storage layer. Put a small amount of local storage on compute boards only for boot and scratch, and keep the durable data in the storage cells.
Over time, as the desktop grows and the software matures, the storage layer can become more distributed. The first generation does not need to solve that.
What this changes
- Storage should live in dedicated cells on the first desktop, not only on compute boards.
- Compute boards keep local storage limited to boot and scratch.
- Fully distributed storage is a future evolution, not a first-generation requirement.
- The next entry can size the first storage cell and decide how many the desktop needs.