If I understand your comment correctly, it isn't the case that the ISR is fixed. The ISR has a minimum size, but it can change over time, so brokers can be removed from the ISR and they can rejoin later once they catch up.
One can drive latency almost arbitrarily low if one is willing to give up exact knowledge of a where the data is.
A simple example is N out of M writes (N and M can be arbitrary with N <= M). M is a set of machines, N is number of replicas we want to have. Now at any given time we write to the N machines that respond fastest. Data is now arbitrarily sprayed over the M machines, but as long as the data itself has ordering information the right state can be reassembled by talking to the M machines.
The "spray" can now be controlled by favouring some nodes from M up to a timeout (i.e. we put more uncertainty in time). We can reduce the reassembly work by using learners and hence increase the likelihood that one machine has all state.
So while providing a lot of flexibility and decreasing the communication overhead when using a lot of nodes, it may introduce performance degradation and latency spikes when nodes are slow. I'd probably prefer a design where some kind of reduced majority is required, but not a full list of nodes, so that outliers would be filtered out.
Are there good percentile numbers over there to check metadata writes over the, say, 99,99%-99,999% percentile, to check how bad this may become?