270 front end servers sounds absurd for the TripAdvisor model.

700k requests per minute is ~12k requests per second.

I've had a sustained 4k-5k on a single un-optimised instance of nginx.

TripAdvisor pages are obviously hugely more complex but a lot of the processing will be happening in the middle tier / back end servers. Most of these front-end requests should be trivially cacheable - static content served to anonymous users, images, CSS and the like.

I used to be a TA engineer. Their pages are not static; almost everything is dynamic. Moreover, their engineering practices are...just awful.

TA java code (used to at least) store every ___location in a giant ___location tree in each server. "Location" includes things like continents and countries and towns and ice cream stores and hotels. Just loading in that much data and creating that many objects took ~10 minutes in part because they insisted on writing their own serialization code but didn't bother to do a good job.

Is it really fair to take a cursory look at someone's architecture and call them out like that? It's especially unfair to throw out something like "I've had a sustained 4k-5k on a single un-optimised instance of nginx." Serving what? Is your problem in any way similar to theirs? Do you have a full understanding of their problem?

I used to work at TripAdvisor, and I don't think he's wrong. TA is a fun company with a lot of really sharp engineers on the payroll--it was a privilege to work with the folks I worked with. But the business focus is on time-to-market, and 10+ years of that can lead to technical debt, which can often present in ways like "requires enough AWS instances to calculate every digit of pi to not fall over." That focus is a conscious decision on their part, though, and their success speaks for itself.

I'm not saying he's wrong about their efficiency. I'm saying it sucks to come behind someone and make blanket statements about their efficiency of utilization without knowing any details. If you worked there, you're in a much better place to comment.

My suspicion is what they're calling "front-end" servers aren't serving static HTTP requests, and their "back-end" servers handle background/offline tasks.

And now that I look again, I realize "they" aren't even connected with TA or Pintrest, just a person who has a somewhat shinier AWS cost calculator.

Not really a user of TripAdvisor but damn are they using these servers efficiently? I have difficulty buying these numbers. Especially considering each one has 6.5 compute units with 17.1GiB of ra... and this is their frontend stuff. Not sure what kinda tasks/db they're running in the backend to require another 70 machines considering they're using 32 memcache machines. Can you imagine even say 30 mongodb sharded instances at 17.1Gib each?

I wonder how may developers 1.7 million per year buys you? A team of 15? Even less if you count office space, managerial costs, etc.

Not sure how that is relevant unless you plan on the engineers handling all the web requests.

I assume it's a reference to the cost of hiring engineers to make the code more efficient, versus just throwing hardware at the problem.

Less again when you include taxes.

The rule of thumb in the UK, at least, is to double an employee's salary to get the total cost of employing them.