Dean Thrasher

I’m surprised that in 2017 more developers and IT departments haven’t heard about blue-green deployment. But apparently that’s been a problem since at least 2010, when Martin Fowler noted that blue-green deployment hadn’t gotten the recognition it deserved. This is a DevOps pattern used widely for testing and deploying websites and webservices with minimal downtime in the cloud. Amazon Web Services now has a Code Deploy template for blue-green deployments. Microsoft lists the blue-green deployment technique first in its Azure Continuous Deployment whitepaper. It’s safe to say that blue-green deployments have graduated from an edge technique to current state-of-the-art.

The idea behind blue-green deployment strategy is simple. When you want to test a new build of your website, you create a copy of your “blue” production instance in a stand-by or “green” environment. Then you deploy the latest code to the new green environment and perform a battery of tests on it. Once the green environment passes your load, performance, security and other tests, your direct traffic from the live “blue” infrastructure to the “green” instead.

You can leave the original blue infrastructure in place temporarily, just in case you want to roll back to the previous build. You could also decommission the blue infrastructure to save operating costs. When it is time for the next release, you make a new blue environment, deploy the release to it, and perform the tests again. If the release passes, you cut over from the currently live green environment to blue.

If releases are frequent enough or if you are doing this with physical servers, you can skip the decommissioning step, and simply alternate releases between the two environments. As long as you make sure to “reset” the standby environment back to a known baseline before your deploy your latest release to it.

There are three huge advantages to performing releases this way:

1. Your standby environment is a copy of production and fully scaled. You can run full acceptance, performance and load tests on this environment and be guaranteed that it will work the same in production — because it will be production, as soon as you direct traffic to it.
2. Cutover to a new build is almost instantaneous, since all you need to do is redirect traffic. There should be no downtime.
3. Rolling back to the previous is almost instantaneous, since all you need to do is redirect traffic to the previous environment. Disaster recovery is easy and your disaster recovery process is tested with every release.

Of course, there are some challenges to this approach:

1. You do have to have a fully scaled standby environment, which will increase hardware and software costs. You can minimize these in a cloud setting, however, because that standby environment can be decommissioned when it isn’t needed.
2. You have to invest in DevOps. If you are building and tearing down environments on a regular basis, you’ll need automated scripts for provisioning infrastructure, deploying code, and making configuration changes.
3. Your platform should be as stateless as possible. When a cutover happens, a user’s next request will be handled by the new environment. Any state that must be maintained from one request to the next must be shared by both environments.
4. Handling schema updates for databases must be handled differently from regular code deployments.

Of the challenges, the most resistance to the idea usually comes with the hardware and software costs, usually because companies haven’t embraced consumption pricing yet. With physical hardware or perpetual software licensing, you do wind up paying double for your production setup. But for critical applications it’s probably worth the expense.

Schema updates remain a difficult engineering challenge too, but they always have been, regardless of the deployment model.

Here’s a few other links on the blue-green deployment pattern, if you’d like to learn more:

• https://www.trendmicro.com/azure/blue-green-deployments/
• https://lostechies.com/gabrielschenker/2016/05/23/blue-green-deployment/
• https://blog.anynines.com/blue-green-deployment/
• http://shawnmclean.com/blue-green-deployments-with-octopus-deploy-and-azure-websites-from-tfs/
• https://acaseyblog.wordpress.com/2015/10/31/blue-green-deployments-with-azure-web-apps/
• https://www.thoughtworks.com/insights/blog/implementing-blue-green-deployments-aws

 

Like many web developers, I learned the art of writing code by trial and error, rather than through formal instruction. I’d look at examples on the Internet, read forum and blog posts, and occasionally consult a book or two. This worked surprisingly well — I’ve made a career out of it, and gained a lot of knowhow through hard-won experience.

You can’t learn everything by trial and error, though. Sometimes the experiments take too long to run, or you can’t afford to have the results blow up in your face. Although I’d figured out basic software architecture principles by observing what worked about past projects and what didn’t, I wasn’t very systematic about it. But hey, I could scribble a boxes-and-arrows diagram on a whiteboard and talk folks through it, and that was enough for most web applications.

My current employer, EPAM, takes software architecture seriously. I found this out when I was invited to serve on an assessment panel for Solution Architects seeking promotion, and I developed a bad case of impostor syndrome. It turns out that there are names for the different kinds of diagrams I’d been doodling, and methods for determining which customer requirements were architecturally significant. You could subject these ideas to analysis and evaluate tradeoffs in a systematic way instead of, you know, by gut feel.

My spidey-sense for software project disasters is pretty well attuned by this stage in my career, so I think I was able to ask reasonable questions of our crop of upcoming architects. But I could tell that several of our candidates were winging it, and I wondered if it was obvious that I was groping my way along as well.

So I decided to do something about it. After a bit of Googling, I found out about TOGAF and the SEI, and the fact that there’s an international standard for software architecture, ISO/IEC 42010. Neat. And then I found the book Software Architecture in Practice.

I’ve spent the last few weeks reading it through, chapter by chapter. As textbooks go, it’s well-written. It’s also (as befits a book about architecture) very well-structured, so it’s something I plan on using as a handy reference when I want to know about tactics for improving performance, or about tradeoffs between usability and security.

It’s also interesting to see a bit of what software architecture looks like in other domains. As a web developer, I don’t often need to think about the constraints imposed on real-time or embedded systems. But I do think a lot about scalability, availability, and some of the other quality attributes described in the book. It’s nice to have these cataloged, with approaches for how to address these needs and what the implications are.

I agree with the authors that the term “quality attributes” is much better than the more commonly used “non-functional requirements.” Most people seem to assume that the non-functional requirements are non-interesting and non-important, when in fact they mark the difference between unicorn and fail whale.

Later chapters address documenting and evaluating software architecture, and software architecture’s relationship to modern software development methodologies, like the various flavors of Agile.

I was really glad to have found a book that captured so much of what I’d learned in the school of hard knocks, and that offered a few new ideas and tools for me to apply to future projects. Naturally, I loaned it out to one of my coworkers and immediately recommended it to another Solution Architect I’d been mentoring at EPAM.

“Oh yes,” he said, “I know that book. It’s on the curriculum for EPAM’s Software Architecture University. That’s a good one.”

On the curriculum…? I guess I still have a way to go before I can ditch that impostor syndrome.