The optimal solution would of course be to shard both the compute I/O and the storage footprint, so each worker only needed to hold onto maybe 1-100MB of data.

Perhaps some existing (simpler) designs could be modified to "hyper-shard" the compute angle, but would still likely require carrying around a large percentage of the database.

In any case, you'd need an internal signalling fabric capable of (cost-effectively) handling very bursty many-to-many I/O across thousands of endpoints to make consensus work in realtime.

It would honestly be really interesting to see how something like this would work in practice.