When we dockerized a service the p95 time in testing notched up a noticeable amount. I was already juggling so much other work at that point that for shits and giggles I tried vertically scaling - reducing the cluster size by half and the cores per server by 2x. Zeroed out the p95 delta.

OPS gave me shit about it, and I was like kiss my ass, the cluster costs EXACTLY the same and deployments are 25% faster.

I think people forget that in the cloud, bigger servers don't really cost more until you get crazy about it.