Witnesses do not decide ordering of units. Ordering is determined by the algorithm that looks back at the witnesses-authored units in the DAG.
What TPS limit do you expect to see in the real world (order of magnitude)?
You know, there is no architectural limit in the DAGs.
Regarding the practical limits, I don't buy into this race to Visa tps. The most pressing issue of crypto is not tps, it is adoption (which we address in the first place). Tps will come second after the first is solved.
I'm an adopter. I buy into this. What is currect [practical] TPS limit for Byteball?
My tests with a developer-installation showed 25 transactions per second in one VM with 2 cores and 1gb memory. I could add more VMs or full nodes, to get about 20ish more tps, it scales with number of full nodes.
Networking is no bottleneck, bandwidth nor latency, barely used. Only cpu usage was high but not even all cores at once, about 60% cpu load on 1 core. Which tells me the software isnt optimized but can push the hardware much, much more, I wouldnt be surprised if 1 full node can do 200 tps with 2 cores if software is slighly optimized. Even today if we had same amount of full nodes as bitcoin, 1600ish, low ball it as I have fat servers, 10tps, thats 16 000tps.