#highscalability

This site is supposed to be my own website; as in some sort of portfolio website and technical blog rolled into one. I did have to debate a little with myself though as to whether to try and develop my own site from scratch (using something like Django or maybe even using a new framework that I've not used before so as to learn something in the process) or use something pre-built and highly customisable such as Wordpress but I realised something important after a while; I'm very lazy.

Well, not just the normal kind of lazy; I'm sort of a dreamer when it comes to computers. I have these fanciful ideas about new technologies and projects, and interesting and creative thoughts about how I could make the website aesthetically pleasing yet rather kooky yet they usually all end up dead in the "cyber-gutter".

Rolling with something like tumblr and having a custom URL (which I'm pretty sure I've sorted out, but I've not yet figured out how to point the non-www version here) seems adequate enough.

The best part about something like this is that you pay zero hosting fees; I was previously hosting my website on linode (good service; no complaints there) but it just felt like overkill. I almost re-rolled with EC2 or another private VPS service - Hetzner, they are a German hosting company that offer incredible value for money, and from what I've heard are very reliable - but once again they were overkill for a site such as this. Also, I've read a few interesting posts about how tumblr scale their architecture1 so no worries there.

So, tumblr, hopefully I am here to stay for a while; or at least until I find something new and shiny.

텀블로의 아키텍처와 각종 통계에 대해 작성된 블로그 글이 있음. 마이크로블로그이지만 수치가 어마어마함. 서비스는 Simple & Beauty 인데, 아키텍처는 페이지뷰가 많다보니 복잡해짐.

통계

하루에 5억 페이지뷰

한달에 150억 페이지뷰

20명 이하의 엔지니어

순간 1초에 최대 4만건 요청.

하루에 1TB 이상이 하둡 클러스터에 쌓임.

수 테라바이트가 MySQL/HBase/Redis/Memcached 에 쌓임.

한달에 30% 씩 늘어남.

서비스 유지를 위해 1000대의 하드웨어 노드들을 유지.

한 엔지니어당 한달에 10억 페이지뷰를 감당함. (15명이라고 치면..)

하루에 50GB 의 글이 올라옴. 팔로워 리스트 업데이트는 하루에 2.7 TB 정도.

대시보드는 1초에 백만번 써지고, 5만번 읽히고, 이 수치는 점점 증가함.

소프트웨어

개발환경, OS X, Linux(Cent OS)

Apache

PHP, Scala, Ruby

Redis, HBase, MySQL

Varnish, HA-Proxy, nginx

Memcached, Gearman, Kafka, Kestrel, Finagle

Thrift, HTTP

Func

Git, Capistrano, Puppet, Jenkins

하드웨어

웹서버 500대

데이터베이스 서버 200대

Memcached 서버 30대

Redis 서버 22대

Varnish 서버 15대

HAProxy 25대

nginx 서버 8대

job queue 서버 14대