I am wrapping up helping a client with a build out of a drupal site to ec2. The site itself is a pretty standard CMS implementation–custom content types, etc. The site is an extension to an existing brand, and exists to collect email addresses and send out email newsletters. It was a team of three technical people (there were some designers and other folks involved, but I was pretty much insulated from them by my client) and I was lucky enough to do a lot of the infrastructure work, which is where a lot of the challenge, exploration and experimentation was.
The biggest attraction of the cloud was the ability to spin up and spin down extra servers as the expected traffic on the site increased or decreased. We choose Amazon’s EC2 for hosting. They seem a bit like the IBM of the cloud–no one ever got fired, etc. They have a rich set of offerings and great documentation.
Below are some lessons I learned from this project about EC2. While it was a drupal project, I believe many of these lessons are applicable to anyone who is building a similar system in the cloud. If you are building an video processing super computer, maybe not so much.
Fork your AMI
Amazon EC2 running instances are instantiations of a machine image (AMI). Anyone can create a machine image and make it available for others to use. If you start an instance off an image, and then the owner of the image deletes the image (or otherwise removes it), your instance continues to run happily, but, if you ever need to spin up a second instance off the same AMI, you can’t. In this case, we were leveraging some of the work done by Chapter Three called Project Mercury. This was an evolving project that released several times while we were developing with it. Each time, there was a bit of suspense to see if what we’d done on top of it worked with the new release.
This was suboptimal, of course, but the solution is easy. Once you find an AMI that works, you can start up an instance, and then create your own AMI from the running instance. Then, you use that AMI as a foundation for all your instances. You can control your upgrade cycle. Unless you are running against a very generic AMI that is unlikely to go away, forking is highly recommended.
Use Capistrano
For remote deployment, I haven’t seen or heard of anything that compares to Capistrano. Even if you do have to learn a new scripting language (Ruby), the power you get from ‘cap’ is fantastic. There’s pretty good EC2 integration, though you’ll want to have the EC2 response XML documentation close by when you’re trying to parse responses. There’s also some hassle involved in getting cap to run on EC2. Mostly it involves making sure the right set of ssh keys is in the correct place. But once you’ve got it up and running, you’ll be happy. Trust me.
There’s also a direct capistrano/EC2 integration project, but I didn’t use that. It might be worth a look too.
Use EBS
If you are doing any kind of database driven website, there’s really no substitute for persistent storage. Amazon’s Elastic Block Storage (EBS) is relatively cheap. Here’s an article explaining setting up MySQL on EBS. I do have a friend who is using EC2 in a different manner that is very write intensive, that is having some performance issues with his database on EBS, but for a write seldom, read often website, like this one, EBS seems plenty fast.
EC2 Persistence
Some of the reasons to use Capistrano are that it forces you to script everything, and makes it easy to keep everything in version control. The primary reason to do that is that EC2 instances aren’t guaranteed to be persistent. While there is an SLA around overall EC2 availability, individual instances don’t have any such assurances. That’s why you should use EBS. But, surprisingly, the EC2 instances that we are using for the website haven’t bounced at all. I’m not sure what I was expecting, but they (between three and eight instances) have been up and running for over 30 days, and we haven’t seen a single failure.
Use ElasticFox
This is a FireFox extension that lets you do every workaday task, and almost every conceivable operation, to your EC2 instances. Don’t delay, use this today.
Consider CloudFront
For distributed images, CloudFront is a natural fit. Each instance can then reference the image, without you needing to sync files across instances. You could use this for other files as well.
Use Internal Network Addressing where possible
When you start an EC2 instance, Amazon assigns it two IP addresses–an external name that can be used to access it from the internet, and an internal name. For most contexts, the external name is more useful, but when you are communicating within the cloud (pushing files around, or a database connection), prefer the internal DNS. It looks like there are some performance benefits, but there are definitely pricing benefits. “Always use the internal address when you are communicating between Amazon EC2 instances. This ensures that your network traffic follows the highest bandwidth, lowest cost, and lowest latency path through our network.” We actually used the internal DNS, but it makes more sense to use the IP address, as you don’t get any abstraction benefits from the internal DNS, which you don’t control–that takes a bit of mental adjustment for me.
Consider reserved instances
If you are planning to use Amazon for hosting, make sure you explore reserved instance pricing. For an upfront cost, you get significant savings on your runtime costs.
On Flexibility
You have a lot of flexibility with EC2–AMIs are essentially yours to customize as you want, starting up another node takes about 5 minutes, you control your own DNS, etc. However, there are some things that are set at startup time. Make sure you spend some time thinking about security groups (built in firewall rules)–they fall into this category. Switching between AMIs requires starting up a new instance. Right now we’re using DNS round robin to distribute load across multiple nodes, but we are planning to use elastic IPs which allow you to remap a routable ip address to a new instance without waiting for DNS timeouts. EBS volumes and instances they attach to must be in the same availability zone. None of these are groundbreaking news, it’s really just a matter of reading all the documentation, especially the FAQs.
Documentation
Be aware that there are a ton of documentation, one set for each API release, for EC2 and the other web services that Amazon provides. Rather than starting with Google, which often leads you to an outdated version of documentation, you should probably start at the AWS documentation center. This is especially true if you’re working with any of the systems that are newer with perhaps not as stable an API.
In the end
Remember that, apart from new tools and a few catches, using EC2 is not that different than using a managed server where you don’t have access to the hardware. The best document I found on deploying drupal to EC2 doesn’t talk about EC2 at all–it focuses on the architecture of drupal (drupal 5 at that) and how to best scale that with additional servers.
[tags]ec2,amazon web services,capistrano rocks[/tags]