Out of curiosity, done HN use any CDN or other way of DDOS protection? dang?
I host a dedicated server there (running https://www.circuitlab.com/) and when I traceroute/ping news.ycombinator.com, it's two hops (and 0.175 ms) away :)
Holy crap I am thinking either there is some magic or everything we are doing in the modern web are wrong.
Edit: The number is from Dang [1]
>These days around 5.5M page views daily and something like 5M unique readers a month, depending on how you try to count them.
Just need good stable code and server side caching.
Spin up an apache installation and see how many requests you can serve per second if you're just serving static files off of an SSD. It's a lot.
edit: I see that there are already a bunch of other comments to this effect. I think you're comment is really going to bring out the old timers, haha. From my perspective, the "modern web" is absolutely insane.
That is great ! :D
>It's a lot.
Well yes, but HN isn't really static though. Fairly Dynamics with Huge number of users and comments. But still, I think I need to rethink lots of assumption in terms of speed, scale and complexity.
I've been at my work laptop (not logged in) and found something I wanted to reply to, so I pulled out my phone and did so. For a good 10 seconds afterwards, I could refresh my phone and see my comment, but refresh the laptop and not see it.
Serving the same content several times in a row requires very few resources - remember, reads far outnumber writes, so even dynamic comment pages will be served many times in between changes. 5.5 million page views a day is only 64 views a second, which isn't that hard to serve.
As for the writes, as long as significant serialization is avoided, it is a non-issue.
(The vast majority of websites could easily be designed to be as efficient.)
Agreed.
I was brought up as a computer systems engineer... So, not a scientist, but I always worked with the basic premise of keep it simple. I've worked on projects where we built all the fangled clustering and master/slave (sorry to the PC crowd, but that's what it was called) stuff but never once needed it in practice. Our stuff could easily handle saturated gigabit networks as the 2 core cpu only running at 40%. We had cpu spare and could always add more network cards before we needed to split the server. It was less maintenance, for sure. It also had self healing so that some packets could be dropped if the client config allowed it, if the server decided it wanted to (but only ever did on the odd dodgey client connection)
That said, I was always impressed by the map-reduce of for search results (yes, I know they've moved on) which showed how massive systems can be fast too. It seemed that the rest of the world wanted to become like Google, and the complexity grew for the std software shop, when it didn't need to imho.
I jumped ship at that point and went embedded, which was a whole lot more fun for me.
Sincerely, old timer
You know, it should be even better than it was in the past, because a lot of heavy lifting is now done on the client. If we properly optimized our stuff, we could potentially request tiny pieces of information from servers, as opposed to rendering the whole thing.
Kinda like native apps can do(if the backend protocols are not too bloated)
We were serving around that traffic off a single dual pentium 3 in 2002 quite happily off IIS/SQL Server/ASP. The amount of information presented has not grown either.
That little box had some top tier brand main corporate web sites on it too and was pumping out 30-40 requests a second peak. There was no CDN.
Not a one trick marketoid pony.
90 to 99% of those are logged-out users, so fully cacheable.
Only a handful of dynamic requests each second remain.
It's ergonomic in a very lispy way but perfectly reasonably so from the POV of that aesthetic.
Even a heavyweight and badly written web server can hit 100 QPS per core, and cores are a dime a dozen these days, and storage devices that can hit a million ops per second don't cost anything anymore, either.
HN is written in a lisp variant and most the stack is built in-house , it is not difficult to imagine efficiency improvements when many abstraction layers have been removed from your stack .
What I do remember, is that it was a social data collection experiment for a couple of social scientists, that never originally expected that many people would actually find ways to find each other and hook up using it.
I miss their old findings reports about how weird humans are and what they lie about. Now, it's just plain boring with no insights released to the public.
Nick carver from SO also one mentioned that they could run SO if a single server , while it wasn’t fun it was doable and had happened some time .
I don't know what the server's specs are but I'm sure it must be quite beefy and have quite a few cores, so let's say that it runs about 10 billions instructions per second. That means a budget of about one million instructions per page load in this pessimistic estimate.
The original PlayStation's CPU ran at 33MHz and most games ran at 30fps, so about 1million cycles per fully rendered frame. The CPU was also MIPS and had 4KiB of cache, so it did a lot less with a single cycle than a modern server would. Meanwhile the HN servers has the same instruction budget to generate some HTML (most of which can be cached) and send it to the client.
A middle of the line modern desktop CPU can nowadays emulate the entire PlayStation console on a single core in real time, CPU, GPU and everything else, without even breaking a sweat.
>Holy crap I am thinking either there is some magic or everything we are doing in the modern web are wrong.
Magic, clearly.
It doesn't need to be crazy.
A static site on DigitalOcean's $5 / month plan using nginx will happily serve that type of traffic.
The author of https://gorails.com hosts his entire Rails video platform app on a $20 / month DO server. The average CPU load is 2% and half the memory on the server is free.
The idea that you need some globally redundant Kubernetes cluster with auto fail-over capabilities seems to be popular but in practice it's totally not necessary in so many cases. This outage is also an unfortunate reminder that you can have the fanciest infrastructure set up ever and you're still going down due to DNS.
True, but this is why it shouldn't be bashed either. When you need it, you need it (cue very complex enterprise applications with SLA requirements).
To support this, look at how many people criticize Kubernetes as being too focused on what huge companies need instead of what their small company needs. Kubernetes still has its place, but some peoples expectations may be misplaced.
For a side project, or anything low traffic with low reliability requirements, a simple VPS or share hosting suffices. Wordpress and PHP are still massively popular despite React and Node.js existing. Someone who runs a site off of shared hosting with Wordpress can have a very different vision about what their business/sideproject/etc will accomplish compared to someone who writes a custom application with a "modern" stack.
It's easily served by a simple server.
i used to host a wordpress site that has 5M pageviews a month on a $10 (and later $20) digitalocean instance.
that's wordpress and a shared vps. I imagine it could be a lot higher if I have dedicated server and use self-written software.
I've been working on something where the DB is also part of the application layer. The performance you can get on one machine is insane, since you spend minimal time on marshalling structures and moving things around.
They are still using Cloudflare. Unlike CF, M5 does not require SNI.
curl --resolve news.ycombinator.com:443:104.20.43.44 https://news.ycombinator.comTheir IP belongs to AS21581; which is registered with a company called 'M5 Computer hosting' out of west-coast USA.
m5hosting.com
The last hop is Santa Barbera.
Definitely does not fall in the AWS ranges.