This is surprising since 1) Facebook is most likely using Hack and 2) it could save a lot of money by moving to Go.
> Facebook’s entire site runs on HHVM (desktop, API and mobile)
Isn't Facebook running mostly on React? I'm guessing HHVM only really powers the API, business logic, etc.
That's why they invented Hack and HHVM in the first place -- it was a cheap-enough compromise that didn't require rewriting the whole code base.
Wouldn't most of the pressure be on whatever the data layer is?
PHP is the presentation layer. Much of the actual logic is in C++, Java, and maybe even Go by now.
The VAST majority of online serving code at Facebook is written in Hack (PHP) and is part of a gigantic codebase called fbwww. Backend services that do heavy lifting (mostly search, feed ranking, ad matching, TAO, various caches, proxies, etc) are fairly light on business logic and written in C++.
It isn't a microservices architecture at all: they only move stuff to C++ for latency and efficiency reasons. For instance most of the ad stack is written in Hack, but certain key pieces use C++ for better performance. I'm not aware of any Java or Go code that is part of the online serving stack.
Its just an observation of having worked on systems at scale and I dont see the microservices / containers used.
I am curious if anyone else has.
The more I am around containers and the network complexity/issues that are created the less I am convinced it is the most optimal way and I am looking for opinions that differ from mine so I can learn
Microservices are probably a bad idea (for now) if you aren't large enough for such a team.
But you've constructed your question to exclude the companies that work this way, which I think any reasonable person would call "at scale."
This is like the guy in these threads saying $400k a year for a software engineer is "market rate" and/or "easily achievable."
Another fun bit from that deck was a newsfeed-related PHP function with something like over 70 arguments across a dozen lines -- basically template data to render that had organically grown over time. It earned its own slide. There was an audible gasp from the room when it was revealed in all its glory, and they talked about the work it takes to refactor stuff like that (which is extensive).
[citation needed]
The majority (all?) of the front end and mobile apps are powered by React and React Native, respectively.
So "only" everything except UI (which is run by the browser, not FB's servers), then. And that's only for browsers that use it at all, as they still need to render a lot of stuff from the servers both for speed and for people who turn off js or use older or limited browsers.
And this is a weird thing to say anyway, nothing runs "on" react; both the framework(s) and the business logic run on something else.
Go outperforms HHVM by a wider margin than 20%. I've seen benchmarks as high as 50%. Sure, benchmarks aren't real world applications, but it's safe to say that generally speaking Go is significantly faster than HHVM.