Old Revisions need to be scaled down to 0 about cloudstate HOT 1 OPEN

I've been thinking about this actually - when using CDRTs, we don't want them scaled down to zero immediately, because we need to give the new nodes enough time to sync their replicators with the old nodes. How long this will take all depends on the amount of data, in many cases it will be very fast, but if there's a lot of CRDTs, eg, if the data is in the GBs, then it could take minutes I guess. A typical rolling update might be faster than this, which would mean after an update, if the old nodes were shutdown immediately, data would be lost.

Also I'm not sure how ddata syncs data to new nodes, if it's completely sequential across all CRDTs, or if it's somewhat random. If random, then having a higher minScale should alleviate the issue somewhat, since a complete sync is not necessary, all that's necessary is that each CRDT gets sent by at least one old node to at least one new node. That said, if it's random that would leave it up to chance, so probably still good to leave them running for at least enough time to guarantee syncing, or to do a sync based handoff.

Finally, this isn't an issue if using durable CRDTs with a shared durable store (though I don't know if that's possible).

from cloudstate.