You're going to have to show us the work for anything you're saying to even remotely make sense. Otherwise this is just drive-by thread crapping.
I worked at Amazon for 7 years and personally witnessed teams investing in operational efficiencies and cost reductions. They regularly celebrate big wins in that. There's no reason to believe they're being wasteful here, and you bring not a single shred of evidence that they are.
the cpu , ram, ssd are capable of serving over 30k qps or more (assuming a roughly 500kb web page response). The nic is likely the bottleneck, but between 3k - 30k qps . With compression, 2x .
Balanced across instances, they are achieving roughly 10-30 qps currently (LBs, app servers, internal LBs, DBs , etc). Possibly even closer to 5qps.
30 qps / 30k qps = 1/1000
Where does it go? mostly encode / decode & IO wait due to human centered architectures that made legacy development scalable. No longer necessary.
so yes, an 80% reduction is a very conservative and plausible reduction. The theoretical possibility would be a 99.9% reduction ( 1/1000)
You’re vastly underestimating the amount of mechanical (and useful) work that goes into making Amazon’s complete retail shopping experience. You’re sizing it as though it were a 3-tier architecture designed in the late 1990s/early 2000s. It’s much more complicated than that. Maybe you’re implicitly arguing that it could be less complicated, and perhaps there’s some merit to that, but you’re not sitting in a knowledgeable place to contend that there are 3 orders of magnitude of resource waste to fulfill the company’s stated product requirements.
Doctors don’t diagnose patients they haven’t seen.
