A coworker has sent me three interesting articles from High Scalability, a site I still didn’t knew but which I’ve already added to my Google Reader list. :) The articles talk about the design and computer/network architecture decisions taken at YouTube, Google and GTalk in order to handle the big load their services face. They also comment the current architecture in each site and their evolution over time:
Some lessons to learn from these articles:
- Don’t try to fix everything with one single architecture or tool. Divide the problem, see if each sub-problem is CPU-, bandwidth- or IO-bound, and optimize it. Specialize server for each task and coordinate their work.
- Cache content whenever possible. Pre-generate content whenever possible. Make good use of HTTP’s cache-control directives. Use squid as a reverse proxy to leverage your application servers’ load.
- Think about externalizing some things, like hosting images or videos off-site. These elements may need more bandwidth that you currently have, and moving them off-site can be a good idea, even if it’s just a temporary measure while you manage to get more bandwidth. The service must run at all times.
- Simplicity. Will let you make changes and evolve your architecture without screwing up.
- Commodity-PC based clusters. They maximize the power/price ratio. Have a redundancy system in place so that when one node goes down or needs maintainence, the system keeps working without it. Have a system to easily install/change a node, also without affecting the service. And start planning the power and cooling problems ahead. ;)
- Programming today is much about libraries and frameworks. Don’t reinvent the wheel. Use a common framework in all your developments, homegrown or not. This way novel programmers will be able to start writting code faster, will be able to switch projects easily, won’t have to code the same things over and over again, and a system upgrade will benefit all your applications.
- Think about the architecture you’ll need from the start. I’m sadly used to developers not caring about what their code runs on, or if their code will lead to CPU, IO or bandwidth problems. Google seems to face every new development looking at the architecture they’ll need to handle the service, and then develop the code arount that architecture. This is what settles Google appart from the rest.
Últimos comentarios
RSS