If your failure modes are uncorrelated (i.e., spread across datacenter facilities with separate power supplies), you might be happy knowing a majority has accepted the write in memory, even though none of them have stored it yet (because that's slower if you're on spinning rust).