Video: An Immutable Approach to Deployment
In this unit, we're going to talk about the processes of deployment and release. Most of this live lesson deals with all different kinds of software, whether that's embedded, or software as a service, or user-installed. In this module we'll be particularly dealing with web-based systems.
In the case where you're going live with your service, in many ways if we've been practicing continuous delivery correctly, we means that we've been deploying all the time throughout the process of building software to acceptance testing environments, staging environments and so forth, release to production is just another deployment.
Well, almost. There are some differences. When you're actually going into production you have to prepare a remediation plan, what are the known risks when we're deploying to production? What are the things that could go wrong? How are we going to remediate these different things? We need to talk to support, marketing, sales. We need to give potentially training to users, and then we'll typically have go and no go meetings and coordination meetings of all kinds. Communication is much more important when you're going to production. If you have change advisory boards, part of the role of change advisory boards is making sure that all these different kinds of things are in place before you go into prod, and that is an important role of change advisory boards, communicating between all the different people who should be informed or involved in the process of going to production.
Finally, we would hopefully rehearse that release roll back process, or the roll forward process if we're using roll forward instead of roll back in order to push out emergency fixes. These are all things that you need to be aware of that are different from just pushing changes to a staging environment.
In this section we're going to present one of the key patterns behind low risk deployments to production, and that is that low risk releases are incremental. There's a whole series of patterns that we have making sure that low risk releases are incremental, which we'll investigate throughout the course of this unit.
Typically we want to avoid deploying the whole thing at once, and that's for a couple of reasons. Firstly that's a lot of moving parts, and secondly it makes it hard to roll back. One of the ideas that has come to fore recently is the idea of immutable architecture, and immutable might be the wrong metaphor, because obviously you are moving data around and your systems are changing state. Nevertheless the idea that at the API level you're not changing existing objects or the schema of existing objects in production, I think is very important.
This is sometimes called the expand/contract pattern. This is what Mike Nygard calls it in his book, Release It. The idea here is instead of overwriting the static content, you put in the new static content for the new version of the app side-by-side with the old version. You basically version it. You could have just different directories for every release that you're putting out, so you don't override the old stuff. You can put the new content out before you do the release, and if you need to roll back you don't need to change it, you just point the application to the old version and that will work fine.
You can do a similar thing for upstream services. AWS is a great example of this. I can access old version of the AWS API if I want. When they upgrade AWS to new versions of EC2, NS3, or whatever, the old APIs are still available. What that means is that clients of those services aren't forced to upgrade at the same time as the service. We can actually put out changes to upstream services before the versions of the applications that depend on those new versions of the services, and make sure everything's fine. Then the clients of those services can upgrade in their own time.
This adds complexity to your system, and it means that you have to potentially maintain multiple versions of those services in production at any one time. It does mean you need to think carefully about API versioning for those services, so there is trade off here and some extra complexity involved. The benefit is simplicity of deployment, and not needing to orchestrate these deployments. You can independently put the new version of the service in there, and then when that's bedded down and we're sure it's fine, then we can upgrade the application.
We can apply a similar pattern to database changes, and the pattern here is again we don't touch existing objects, we add new objects side-by-side the old ones. To give a very simple example, we want to change our address column to be address line one, address line two. We don't delete the old column and add the new ones. Instead what we do is we add the new columns to the schema side-by-side with the olds ones. Those columns will initially be null. Again we can add those new columns, those new objects side-by-side with the old ones before we release the version of the application that depends on them.
Then in the application we add some extra code which looks to see if those new columns are there and if they're null. If the columns aren't there, or if they're there but they're null, then it's going to read from the old column, but we're going to try and write to both columns. In this way the application is going to lazily migrate the data from the old column to the new column. If we need to roll back our application, we can do an application roll back because the old address column contains the updated data, so we haven't lost any data in the course of doing that. If we then need to redeploy the new version, what we can do is we can just set the new columns to null, and have the new version of the app lazily migrate the data again.
In this way, we've decoupled the deployment of the database changes from the deployment of the application. They're independent of each other, and I can roll back and forth the app, and roll back and forth the database completely independently of each other.
However, this simplicity in terms of deployment, again as a cost in terms of the application, we have to add that extra code to look and see if the columns are there, and to write to both. That's the extra complexity you've incurred as a result of trading off for ease of deployment.
It's important to note that these kinds of patterns, there are tradeoffs in order to implement them. What you're doing is making it much cheaper and lower risk to do those deployments incrementally in pieces, and decoupled the deployment of the application from the other things that the application depends on. You can do that long before the deployment, and crucially during normal business hours. Adding new objects that the application doesn't care about, whether it's the static content, new versions of upstream services, database changes, they can be done during normal business hours because they do not impact the deployment of the application. So reducing risk and at the same time making sure that we can do most of our deployment work during normal business hours.
In the next section we'll talk about how to apply patterns to enable us to deploy the the application itself in ways that reduce the risk of deployment.