A few weeks ago Arte and Mario asked me to swing by to chat with folks participating in the Momentum accelerator to talk about scaling technology. While we were talking I pointed folks to a few posts and videos of talks I consider to be some of the root nodes of a lot of other conversations. I’m not sure I’ve ever pulled this together before.
- Allspaw/Hammond - 10+ Deploys Per Day
- Randy Shoup - Ebay’s Architectural Principles
- AKF Partners - Splitting Applications or Services for Scale
- High Scalability
- Netflix Tech Blog
That Allspaw/Hammond presentation is the most primary in my opinion. It’s the talk that lots of us consider to be the inflection point of the devops movement, when the practice started to really gain visibility. What had previously been a random collection of techniques we didn’t have a great way to refer to became “devops”. There’s a ton of useful information in that presentation. Not just the overall concepts of dev and ops working together with a different kind of coupling, but some specific tools to make that happen like shared metrics, feature flags, and shipping dark.
The presentation from Randy Shoup has also withstood the test of time fantastically well. The four architectural principles are great top level items: partition everything, async everywhere, automate everything, remember everything fails. And the presentation from AKF Partners dives into a bunch more detail about how to partition services so that you can just throw hardware at problems. That model also provides a lot of the basis for most of the conversations I end up having about geographic distribution of services (which actually sounds exactly like a discussion about how to design the I/O layer of an operating system, you just swap terminology).
The High Scalability and Netflix blogs provide great concrete examples of what some of the problems and solutions look like. Principles are great, but like always there’s more than one way to do it. So it’s great to have some concrete examples to poke through. Plus Netflix being one of the pioneers of the static AMI methodology of infrastructure automation it makes their approaches particularly strongly opinionated.
The platforms themselves have changed the level of abstraction you need to work at when putting this stuff into practice. Thanks to MongoDB and Riak we don’t have to wrapper tons of instances of MySQL or Postgres to make a datastore that scales horizontally. And thanks to services like Elastic Beanstalk, Heroku, and EngineYard we don’t have to pay as much attention to how to distribute load across multiple application servers. But even though you can count on your platform to remove a bunch of the details in some cases, it’s a good idea to understand the principles. First of all cause it’s possible to misuse the platform and end up not getting the full advantage you could. But also because platforms can only help out so far. Eventually you need to crack open the abstractions and either continue with an application optimized line of evolution, or figure out how to wrapper additional layers of capability around what the base platform provide. In either case at least being aware of the principles your underlying platforms are working on will pay back hugely.