Given how many users threads claims, I suspect threads members must be de facto limited to threads communities. Even if they can, technically, subscribe to regular lemmy communities, how would they discover them? “320 million” threads users vs 65,000? 100,000? lemmy users? And community search is going to be flooded with options from the platform with 2000x more users.
You can configure HA to use an external database, so you could (presumably) config two instances to use the same DB. Not sure how much conflict that would cause for entities that are only attached to one of those instances, but it seems like both should have the same access to state data and history. Could probably even set one instance up with read-only DB access to limit data conflicts, although I imagine HA will complain about that.
Even with an external database, HA still uses its internal DB for some things, so I don’t think you’d ever get identically mirrored instances.
Back in the day, I set up a little cluster to run compute jobs. Configured some spare boxes to netboot off the head-node, figured out PBS (dunno what the trendy scheduler is these days), etc. Worked well enough for my use case - a bunch of individually light simulations with a wide array of starting conditions - and I didn’t even have to have HDs for every system.
These days, with some smart switches, you could probably work up a system to power nodes on/off based on the scheduler demand.