Not sure what platforms you've worked on but "6M requests/day" is peanuts in my neighborhood. (No offense to dang and team which do a fantastic job, but that's not high throughput compared to almost any platform I've worked on over the past decade.)
I have a site with ~ 2.5M req/day running on an m3.medium instance that rarely spikes 10% cpu. NGINX is frontline server probably handling a good % of those request directly before passing others via proxy.
It’s not trivial to take advantage of that but that is also much easier to handle than QPS that are distributed across your dataset (IE Facebook or Twitter where everybody is seeing different things).
Traffic is a bit of text, extremely basic CSS, no images, no tracking, no bullshit means that load times are probably instantaneous even on a 56k modem.
Content is a bit of plain text => easy rendering, the entire database probably would fit into memory, so not much in terms of i/o going on.
What I have utterly no idea is how the karma system works. Updating karma counts and comment sorting only on new submissions would help with performance, but on the other side karma "decays" over time, so there needs to be some sort of external cron?
In other words: they are serving exactly what the people who come here are looking for, and little else. Don't get me wrong, there are cases where images and moderately more complex CSS improve a website. On the other hand, many websites go to an extreme that degrades the user experience.
I disagree that startups with millions of customers handle less traffic. There are probably some exceptions that prove the rule, but once you're into millions of customers you are most likely over 50K rps at peak times.
Product managers are constantly coming with cool ideas that are hard to shard.
At any rate they should be able to serve that easily off one box, so the architecture holds up.
Let's imagine that all the traffic occurs in a single hour (6M per hour) and is Poisson distributed. Even then to reach 10,000 rps would require a second with 6x the average traffic volume for that hour. Admittedly this could theoretically occur when some correlated event causes everyone to browse HN, but historically the site goes down during those.
For comparison our service (also largely US-centric) served 30M views yesterday with a peak volume of ~1000rps.
Yes, 10K is likely an overestimate by some factor, however it's in the realm of possible. We can argue how much of an overestimate, but I don't think this takes away from the main point that I was trying to make above, which is that they are very unlikely to exceed the serving capacity of a single host (I don't know how fast Arc is, but a well-tuned serving system should be able to handle 40K rps for simple to cache content, which HN surely qualifies as).
While 10k rps may be possible it would either require vastly exceeding 6M requests per day (and per hour) or some very non-standard traffic event (this could happen in practice, for instance email asset delivery is increasingly peaky due to modern email clients and can exhibit very large traffic spikes). As HN is primarily visited by human users spanning many organizations such a strongly correlated event seems unlikely.
HN can do an incredible amount of easy caching. Besides a small box in the top right with user profile info, the entire page is completely identical for everyone.