A taste of CloudFormation.

By: on March 20, 2017

So, we’ve recently had cause to move one of our internal applications to the cloud; which has largely been an excuse for me to get some experience in some relatively modern operations technologies. Amazon’s CloudFormation is designed so that you can declaratively specify the infrastructure resources (eg: virtual machines, load balancers, container configuration, &c) that your application needs, and apply changes relatively atomically.

For someone who’s spent a goodly deal of their career dealing with physical machines that live in a datacenter (so-called ‘Pets‘), this is something of a revelation. For a start, because everything is specified in terms of textual configuration (usually JSON, but we’ll get to that); it can be checked into version control. This might not seem like a big deal, but it certainly hugely more effective than having to manage a fleet of hardware, with all of the cabling, inventorying, and fiddling that involves.

What this buys you principally is leverage. For example, it’s common to have staging environments for validating applications on before they’re deployed into production, and one issue with that is that it can be difficult to maintain parity between those staging, and your production environments. For one, it’s relatively common to have a simplified network topology in staging (eg: skipping firewalls, &c) to save on cost; but that can bite you say, when you introduce new communication pathways between services. If you haven’t told networking that service A needs to talk to service B, then you may suddenly find that what works on stage will mysteriously fail in production. By being able to spin up an exact copy of your running infrastructure on demand, you minimise the risk of any unpleasant surprises.

It’s quite common to have different teams responsible for different parts of the stack–eg: at one place I worked, we had a hardware team who would fly out to far-flung datacenters, and deploy and manage the physical kit and networking, wheras the sysadmin team would manage the operating systems and everything up. So, in a sense, each team presents an abstraction; the hardware team presents running boxes, and the sysadmins manage and monitor, as well as providing means to bootstrap a new deployment.

So, in the same way, we end up with a split of applications in the cloud. Systems like Kubernetes, Rancher, &c provide a platform, assuming that all you care about is providing some code artifacts and having them run; and that the underlying infrastructure can be abstracted away, and worried about by someone else.

CloudFormation can provide this; but by necessity ends up providing a lower level of abstraction. So, it’s probably fairer to say that you can build a platform on top of CloudFormation’s services; but you’ll likely need to provide your own abstraction on top for this to be workable long term.

So, wanting to avoid working with the CloudFormation JSON syntax directly, Tom found a library named Troposphere, which allows you to express configurations as python code, and provides some degree of configuration linting, too. Expressing the configuration as an embedded DSL in python allows you to take advantage of the structuring features of the host language, so in python, you can use classes to represent sub-groupings of resources (eg: a cluster configuration for ECS), and those can be consumed by say, an application instance without it needing to worry about how that container system was configured.

You can replicate this in plain CloudFormation, by having a number of separate stacks; and importing references from them; but you’ll end up needing to specify the names of each stack somehow. You could potentially end up Greenspunning a module system from scratch, though, using these imported references and string concatenation, so using a language that provides modularity feels more natural.

However, because the python ecosystem provides a packaging mechanism, it’s entirely possible to publish a python module providing say, 90% of what you need for a typical application stack, and have the consuming application inject references to source trees or deployable artifacts, and provide a deployable configuration. It’s not quite as convenient as a fully plumbed solution like Kubernetes since you’ll have to integrate that with your build / deploy mechanisms, but it’s still a pretty good solution when you need the flexibility.

So; when would I want to use CloudFormation, vs. something higher level? Well, as mentioned above, the principle use of this is going to be building infrastructure for others to consume, or for more complex projects where you need relatively finely grained control over say, database usage or placement for real-time services and batch jobs.

Conversely though, because CloudFormation provides relatively thin abstractions; this can potentially make it easier to debug or trace faults than systems like Kubernetes if you don’t have an expert team managing it, simply because the more fine-grained control can make it easier to trace what process is running on which server; even if the servers themselves are ephemeral, the roles they serve may not be.

Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>