> I don't think it's true that "a time series CRDT is just an append-only log" You are correct. Append-only log as an implementation detail doesn't converge an

> I don't think it's true that "a time series CRDT is just an append-only log" You are correct. Append-only log as an implementation detail doesn't converge and cannot be a CRDT. Although semantically logging can and should be a CRDT, as it doesn't need to be append-only at all. > As long as you ensure that time is unique and monotonic (usually by including things like node ID and a per-node monotonic counter) As other comment noted, it's not enough. You need some causality here, at least a lamport timestamp (which is a rule of thumb on timestamps anyway). Because even with synced clocks some will always be faster and always win writes.