Switching wings mid-flight

Posted by on September 4, 2012

How would you choose to release a completely redesigned version of your web application? Capistrano? Git hooks? Copying files over SFTP?

Now imagine your application serves over 21,000 active, multi-user accounts. It’s a very busy app, and forms the backbone of many of those users’ livelihoods. At that point, the question perhaps becomes less about “how”, and more about “when”. 1am? And how long is the downtime going to be? Even if it’s minutes, even if it’s seconds, what if something goes wrong?

We did exactly this recently. Our answer to the “when” was 10am on a busy Wednesday morning. Wednesday the 13th, just to be sure. Our answer to “how” was one line on a production console:

new_design.update_attributes(:status => 'on')

Feature switches rock

new_design is a feature switch. If you haven’t come across feature switches before, the basic idea is simple. You have a list somewhere of
“switches”, maybe with a “status” (though that could be implied simply by present/not present), and your application code looks something like
this:

if has_feature?(:my_cool_new_feature)
  do(:new_hotness)
else
  do(:old_and_busted)
end

where has_feature? checks the list. Congratulations: you are now using feature switches.

We first used basic one-shot feature switches when we rolled out our stock management system, and then extracted the logic into a little system we call Flip. It’s super simple, but the simplest tools are usually the most powerful. It turns out that feature switches have a load of benefits.

Launch

First, and most obviously, releasing a new feature no longer requires a code deployment. Well, it still requires a deployment, but the release
of the feature is no longer tied to the deployment of the code. You can deploy the code, check that the old code all still works as expected, then throw the switch at your leisure.

Feature.create(:name => 'new_feature', :status => 'on')

Boom: your new feature is now deployed.

You also know that your exit strategy is bulletproof, since your users have been using the fallback code paths since your deployment. Kill the switch if anything goes wrong, and it’s all good.

Feature.find_by_name('new_feature')
       .update_attributes(:status => 'off')

Phased Deployment

Pretty soon, you want to check the feature works in production before you throw the switch for reals. Yes, you’ve tested it in development, in
an integration environment, in staging. It’s been through peer review and QA and is perfect. But even so, it would be nice if you could throw
the switch in production just for you so you can take a peek behind the curtain and make sure it’s firing on all cylinders. So our
has_feature? helper to takes into account the current user’s company, and extended the Features to have many switches:

c = Company.last
f = Feature.find_by_name('new_feature')
f.update_attributes(:status => 'company')
f.feature_switches.create(:switchable => c)

So now, I can not only release my feature whenever I like, I can release it to my dev team ahead of a release, in production, and have them
hammer on it till it breaks. Which it doesn’t, of course. Winning!

Getting Involved

The downside of that, however, is that we all know how good dev teams are at testing applications in the real world. We test the code, but we
miss really obvious UI issues, or workflows that just don’t make sense. You know who doesn’t miss those things?

Users.

Feature switches turn out to be a great tool for releasing new work to beta testers. There’s no testing as thorough as living with new code while working with real, personal data.

I’m sold

So were we. In fact, we were so sold that we decided to use switches as the backbone of our redesign. We split the redesign project into
manageable chunks (generally one project per “section” of the app), and wrapped the work involved in each project in its own switch. This let us have a team that methodically worked its way through the app, section by section, without forking the codebase. When a section was done, we
pushed the code into production. Development could continue on the application, as the same underlying code was driving both switched and
non-switched views.

It also meant that, if something urgent landed, the team could be pulled off the redesign project at the end of a section without worrying about their work going stale. Once something was in production, it was done done.

By having the code in production sooner, we could enlist a team of intrepid beta testers (you know who you are) who could hammer on our work far earlier in the process. We also gained great feedback early enough to either do something about it before launch, or understand better where we needed to be headed further down the line.

Thanks to all of this, throwing that switch mid-week, mid-morning was a no-brainer. We knew everything worked (as well as anyone can ever know that) and we knew we could back out trivially if everything went wrong.

SHIP IT.

But what about…?

It can’t all be wine and roses though. There are some legitimate concerns about taking this sort of “all in” approach.

Performance

At launch, we have a codebase littered with has_feature? checks. Literally, every view in the app checks to see if the redesign is
switched on. Doesn’t this impact on performance?

It does, but minimally. Our Flip system stores all the switches in memcached, so certainly no end users will be
able to see any slowdown.

Testing

How do you go about testing this? We chose to brute-force it, by simply duplicating our view tests, running them once with the switch, once
without. We experimented with only testing the changes, but the problem with that approach is that you lose coverage.

One of the advantages we gained by approaching this project with feature switches was that we could continue to enhance the existing application while the redesign was underway: we didn’t have to feature-lock a section while it was being redesigned. That meant we needed reliable coverage of both “before” and “after” views, and we needed to know if work on one had been replicated on the other.

It feels like an ugly solution, but the reality was that we were really running two separate view-layers, and so doubling the test coverage makes sense. Besides, we knew we were going to tidy everything up.

Housekeeping

Speaking of which, we now have the task of going through and pulling out the switches (and forked tests) so that the app simply has one code path through the views. This is a bit of a pain, but we’ve decoupled this completely from the act of deploying code to our users. The extraneous code isn’t impacting on performance, and because each section is wrapped in its own separate feature switch, we can tackle the clean up in manageable chunks.

Conclusion

Feature switches played a key role in allowing us to both develop and deploy the most complex change to the app we’ve undertaken so far. Just
about everything we now deploy (small or large) gets wrapped in a switch, as the wins are just too good to ignore.

Hopefully, we’ll get our Flip module open-sourced in the future, but the truth is that the code inside it is simplicity itself. You don’t need some fancy gem: you just need a model, a helper, and a cache.

Have at it.