January 10 2020, 7:08PM
Staging environments are a distraction. Massive hours have been poured into making them coherent with production all to little effect. When staging environments become unbearable developers start resorting to alternative environments that can be spun up at will or are equally long-lived as staging environments are. These production clones feel safe to developers and product managers because they they aren't shown to customers. One uncomfortable fact that developers and product managers struggle with is that production doesn't mean seen.
I've written about this before but received some confused responses and I think I've realised why people feel uncomfortable about this concept. Feature flag aren't just for testing shades of blue. Feature flag services at as curtains over new functionality until you are ready for the big reveal. There are third party services out there but you can write your own feature flagging system to hide away details although there are some limitations with doing it yourself. I have long been a proponent of small pull requests; small changes give large boosts of energy, helping progress and leading to the eventual discovery that you've built a mountain when it has felt like a short jog. Developers who feel safe pushing changes start pushing a lot of changes, hence, I think having a great feature flagging system is pivotal to making the small pull requests approach feasible in a team.
It's understandable why there is a reluctance to have one environment. There is a natural pain associated with pushing bad code and frantically trying to fix it. Tucking things into the shadows means you are growing and building while no one else really notices. Then, one day they come around and notice the flourishing garden and sculptures you've built that they couldn't see before.
A basic feature flagging system is a key-value store for named tags, the key, and booleans, the value. Non-existent tags are always false to avoid strange behaviour. The steady state of the system is when all flags are off. Flags should persist across the whole of the architecture to reduce mismatch and bloat, which means a flag should be visible to all parts of the larger system or product. Flags should be persisted to long term to be robust in the case of failures.
Fancier feature flagging systems support things like traffic routing and mutual exclusion. As noted before, a user may be randomly assigned to a split in an A/B test and that particular flag they were assigned may be incompatible with other flags. This isn't needed out of the box unless your platform is already messy or you are suffering from too much load.
Buggy code or migrations can poison a production database. If you are not already taking regular snapshots of your database, fix that! Trying to prescribe solutions for various use cases could easily fill other articles, but I will say that despite it seeming scary that you are mixing feature-flagged code and steady-state code that both touch the same shared state, with some forethought it is far easier to curate one pool of data. If you can get back to a good known state, you can work towards a granularity of restoration that suits your products needs. A fantastic book on operations around databases that goes in much greater depth is Database Reliability Engineering.
Tying this all together, you should treat features as immutable migrations. An immutable migration is one that doesn't happen in place, such that if I have state A and want to be in state B, I first create state B and transition over. A mutable migration is the one most people are familiar with, changing a piece of pre-existing code, testing it locally and in a staging environment or similar, perhaps even prod, and hoping for the best. Another way to put it, there is a time where state A and B exist at the same time with immutable migrations, but with mutable migrations the migration happens as a single commit.
With immutable migrations and feature flags, you can push the code progressively to prod and test at-will. I'm a big fan of pushing code to production that doesn't isn't being used yet. Doing this a lot and using namespacing on a lexical level, e.g. with function or module names. When I wrote about this before I mentioned the idea of breathing room and immutable migrations give you just that. In fact, they ought to be your default form of migration unless you are certain doing an in-place change is going to make things quicker and be relatively pain-free.
Less is more; have one environment you put all your effort and love into. Hide things from your customers until you are confident of your release. Make immutable migrations the default instead of risky (albeit faster) in-place migrations. Put some thought in what you need to do to protect your shared state. With this your deployments will get more fearless and frequent as well as your changes smaller and easier to reason about. Production doesn't mean seen.