Have you ever edited code directly on a production server? I’ll admit that I have, years ago, before I knew better. It’s easy and fast and gets emergency fixes out there as quickly as possible. You get to know what works and what doesn’t because you make tiny changes. If you’re working with a system that anyone cares about, it’s also dangerous and stupid.
I’ve also worked with development organizations that take the opposite extreme. Even the most trivial server updates needed to be scheduled weeks in advance. In order to push the updated code, you had to notify everyone, take all servers down in the middle of the night, run through a long series of manual steps, and then do a lot of manual testing. When things in production aren’t quite how they are in the test environment, this is followed by a hasty rollback or panic-induced middle-of-the-night attempts at bug fixing.
Both of these extreme approaches are (maybe) appropriate in some environments. But neither are appropriate for a new web startup that’s trying to move quickly and still provide a reliable trustworthy experience to their customers.
At the previous company I worked for, we often talked about IMVU-Style Continuous Deployment as our ideal process, but it was always something “in the future”. We were hesitant (some of us more than others) to do automatic deployment without at least a little manual intervention. We always wanted to have more test automation, or a smoother deployment system, or whatever.
Since it seemed to be hard (for me anyway) to move to an existing development organization to a continuous deployment system, I started to wonder what would happen if you do it that way from day one? I got a chance to answer that question when I co-founded a startup last year. One of the very first things I did, before we had anyone using the site, was to create an solid automated test & deployment system that was as fast and easy as possible without being dangerous and stupid.
Here’s the basic workflow that happens in our office multiple times every day.
Step 0. We make changes on our local dev envirnments, with a bias toward making the smallest possible change that adds value. That could be a bug fix, correcting a typo, a stylistic tweak, a stubbed-out new feature, whatever. Once I’m confident in my local (manual and automated) testing that the change is good (not perfect, not feature-complete, but just better), I push that to my github repository.
From there, the continuous integration server pulls down the new code and does the following:
Step 1. Does the code still compile. If not, the build fails and everything stops.
Step 2. The build agent runs the unit tests (where “unit tests” are defined as tests that run with no external dependencies, these take just a few seconds). For anything that does require external (generally slow) dependencies (network API, databases, filesystem, whatever) we use test doubles (fakes, mocks, stubs, whatever).
This first feedback loop is about catching and preventing errors in core business logic and is generally measured in seconds, not minutes.
Step 3. The build agent runs a set of tests that rebuild the database from a reference schema and exercises all of the repository layer code.
Step 4. The build agent runs another set of tests that test our dependencies on external APIs (twitter, geolocation services, etc.)
These two sets of tests run in a few minutes, so the feedback loop isn’t quite as tight, but it’s still pretty darn fast. Basically, they make sure that the assumptions that we make with our test doubles in our unit tests aren’t totally wrong.
I’ve written about these sorts of automated test distinction a couple of years ago, in a post about the Automated Testing Food Pyramid.
Step 5. Provided that the entire gauntlet of tests has passed so far, the code gets automatically deployed to a staging server.
Step 6. There’s an additional set of tests that run against the staging web server. These tests can find configuration problems and code that just does the wrong thing in a web context. These tests are pretty shallow. They hit all of the user-facing pages/JSON endpoints and fail if anything is totally broken.
Step 7. The build artifacts are copied from TeamCity to a new folder on our productionvserver, and then the web server is reconfigured to serve from that folder instead of the folder it had been serving from.
At this point, we’ve verified that the core business (game, in this case) is OK, verified that the persistence stack works as expected, that our integration with external APIs works as expected, and that the code doesn’t completely break in a web context. We’ve done a zero-downtime deploy to the production web server.
That’s cool, but we’re not quite done yet. There’s two more steps.
Step 8. Run a set of tests against the production web site to make sure that all of the pages that worked a few moments ago still work.
Step 9. Have external monitoring systems in place, so if your changes make things slow or unresponsive. You’ll know. We use pingdom.
Yikes! There’s a bunch of distinct steps here, and it seems really complicated (because it is). But it’s all totally automated. All I need to do ?
git push origin master
Because there’s zero deployment effort on my part, I do this all the time. I find it very energizing to know that I can just do stuff in minutes instead of hours or days or (heaven forbid) months.
If (when) something goes wrong, I’ll know immediately. If a bad bit of code manages to roll through the test gauntlet, I can roll back easily (just reconfiguring the web server to use the last known good set of code). I’ve only had to roll back a couple of times over the course of several months and 326 deployments.
Just like when the folks at IMVU wrote about this process, I’m sure that some people in the audience convinced that I’m a crazy person advocating a crazy way of working. Here are the objections I’ve heard before.
Yeah, but this doesn’t give your QA team any time to test the code before it goes out!
We’re a small startup. We don’t have a QA team. Problem solved.
Yeah, but isn’t that incredibly dangerous?
No. The safest change you can make to a stable production system is the smallest change possible. Also, we design the individual parts of the system to be as encapsulated as possible, so we don’t tend to have crazy side-effects that ripple through and create unintended bugs.
When we make a change or add a new feature, we can manually test the hell out of that one thing in isolation (before checking in) instead of feeling like we need to spend a lot of time and effort manually testing everything in order to ship anything.
Yeah, but what about schema changes?
For schema changes that are backwards compatible with the code that’s out there (e.g. new tables, whatever). We have a simple system that executes the appropriate DML on application startup.
For non-compatible schema changes and things like server migrations, we have to take down the site and do everything manually. Fortunately, we’ve only had to do that twice now.
Yeah, but you have to spend all of that time writing tests. What a waste!
The time we spend writing tests is like the time a surgeon spends washing their hands before they cut you open. It’s a proven way to prevent bugs that can kill you.
Also, we get that time back, and then some, by not having to spend nearly as much time with manual testing and manual deployments.
Yeah, but what about big new features you can’t implement in just one sitting?
Traditional software development models (both waterfall and agile methods like Scrum) are organized around the idea of “multiple features per release (or iteration)“. Continuous deployment is organized around the idea of “multiple releases (or iterations) per feature“. As a result, we end up pushing a lot of code for features that aren’t done yet. For the most part, these simply unavailable through the UI or only exposed to users who have particular half-built features associated with their accounts. I credit Flickr with this general approach.
Yeah, but that might work for a solo developer, but it can’t work for a team.
There are actually three developers on the team.
Yeah, but I’m sure this only works with very experienced developers
One of the guys on the team has only been programming for the last year or so and hasn’t ever worked on a web project before. Tight feedback loops help everyone.
Yeah, but what about code you want to share with other that you don’t want released?
We use github, so creating additional branches and sharing with them is trivial. We also have a dedicated “preview” branch that triggers a parallel test/deploy gauntlet that sends code to a staging server instead of the production servers.
Yeah, but this will never work at my organization because…
OK. That’s cool. Don’t try it if you feel that it won’t work for you. You’re probably right. You’re not going to hurt my feelings either way. I found something that’s working really well for me, and I want share my experience to show other people that it’s possible.
What this really means
Half of what makes this process work is that we’re honest with ourselves that we’re human and will make mistakes. If we have multiple tight feedback loops between when we’ve broken something and when we know we’ve broken it, it’s faster and easier and cheaper to fix those mistakes and prevent similar mistakes from happening again.
The other half is the idea that if you design, implement, test, and release exactly one thing at a time, you know with certainty which change introduced a problem instead of having to ask the question “which of the dozen or so changes that we rolled out this month/sprint/whatever are causing this problem”.
About the site
Victors United is an online turn-based strategic conquest game. You can play asynchronously or in real time. You can play against robots or humans. If you’re playing against humans, you can play against your friends or against strangers. Unlike some other popular web based social games that I don’t like to mention, this is a real competitive game where strategy and gameplay matter.
About the tech
The tech here is kind of beside the point. This general approach would work just as well with different technology stacks.
The test gauntlet is a series of distinct nUnit assemblies, executed by TeamCity when we push new code to GitHub. There’s a single custom PowerShell script that pulls down the build artifacts and tells IIS to change what directory it serves code from.