Connection pooling only helps if you have a limited # of clients running on machines you control. In a maximum-scaled-out situation, you're gonna have eg hundre

Connection pooling only helps if you have a limited # of clients running on machines you control. In a maximum-scaled-out situation, you're gonna have eg hundreds of physical app servers so connection pooling is gonna make a difference, but only in individual client latency. In the CouchDB case your clients might even be remote web-browsers and while some of them might use keep-alive, with real traffic most clients will be requesting just a couple of resources anyway, so again, connection-pooling / keep-alive is only going to make a small difference. Not saying that you shouldn't try another benchmark with more efficient resource utilization, but at some point (hopefully) these databases all become disk-bound.