If modular is the best way to scale, why didn't we going for that approach off the bat? 

It seems like it's more a case of "we wanted to scale the L1 but it doesn't work well enough for ethereum so we went down the modularity route" 

and maybe it will turn to be the best approach generally, we'll see.

Enthusiastic about web3 domains. Imported a web2 DNS address into ENS and voila, got my grail domain, Jack.top

Scaling by execution layer sharding was the original plan. Dropped cos takes too long to implement on L1, but most importantly: the modular approach allows maximal experimentation/customisation/bazaar style design of diverse execution layers for diverse usecases, none of which the base layer could possibly anticipate