As with so many things, the right answer depends on the expected ratio and complexity of of writes, queries and updates. Something that I've come to believe is

As with so many things, the right answer depends on the expected ratio and complexity of of writes, queries and updates. Something that I've come to believe is that developers don't spend nearly enough time gathering process data that would let them make explicit forecasts (even order of magnitude). Once your team has aligned on those forecasts, a bunch of design debates just fall away. If you really want to have fun, start a prediction market - though that might end up creating its own set of debates :)