How I build software quickly

Basically there are things you can't avoid that are not necessarily fast (e.g. compilation, docker build, etc.) and things that you can actually control and optimize. Tests and integration tests are part of that. Learning how to write good effective tests that are quick to run is important. Because you might end up with hundreds of those and you'll be spending a lot of your career waiting for those to run. Over and over again.

Here's what I do:

- I run integration tests concurrently. My CPUs max out when I run my tests. My current build runs around 400 integration tests in about 35 seconds. Integration test means the tests are proper black box tests that hit a REST API with my server talking to a DB, Elasticsearch and Redis. Each test might require users/teams and some content set up. We're talking many thousands of API calls happening in about 35 seconds.

- There is no database cleanup in between tests. Database cleanup is slow. Each build starts with an ephemeral docker container. So it starts empty but by the time the build is over you have a pretty full database.

- To avoid test interaction, all data is randomized. I use a library that generates human readable names, email addresses, etc. Creating new users/teams is fast, recreating the database schema isn't. And because at any time there can be 10 separate tests running, you don't want this anyway. Some tests share the same read only test fixture and team. Recreating the same database content over and over again is stupid.

- A proper integration test is a scenario that is representative of what happens in your real system. It's not a unit test. So the more side effects, the better. Your goal is to find anything that might break when you put things together. Finding weird feature interactions, performance bottlenecks, and sources of flakiness is a goal here and not something you are trying to avoid. Real users don't use an empty system. And they won't have it exclusive to themselves either. So having dozens of tests running at the same time adds realism.

- Unit tests and integration tests have different goals. With integration tests you want to cover features, not code. Use unit tests for code coverage. The more features an integration test touches, the better. There is a combinatorial explosion of different combinations of inputs. It's mathematically impossible to test all of them with an integration test. So, instead of having more integration tests, write better scenarios for your tests. Add to them. Refine them with detail. Asserting stuff is cheap. Setting things up isn't. Make the most of what you setup.

- IMHO anything in between scenario tests and unit tests is a waste of time. I hate white box tests. Because they are expensive to run and write and yet not as valuable as a good blackbox integration test. Sometimes you have to. But these are low value, high maintenance, expensive to run tests. A proper unit tests is high value, low maintenance and very fast to run (it mocks/stubs everything it needs, there is no setup cost). A proper integration tests is high value, low maintenance, and slow to run. You justify the time investment with value. Low maintenance here means not a lot of code is needed to set things up.

- Your integration test becomes a load and stress test as well. Many teams don't bother with this. I run mine 20 times a day. Because it only takes less than a minute. Anything that increases that build time, gets identified and dealt with. My tests passing gives me a high degree of certainty that nothing important has broken.

- Most of the work creating a good test is setting up the given part of a BDD style test. Making that easy with some helper functions is key. Most of my tests require users, teams, etc. and some objects. So I have a function "createTeam" with some parameters that call all the APIs to get that done. This gets called hundreds of time in a build. It's a nice one liner that sets it up. Most of my tests read like this: create a team or teams, do some stuff, assert, do more stuff, assert, etc.

- Poll instead of sleeping. A lot of stuff happens asynchronously so there is a lot of test code that waits for shit to happen. I use kotest-assertions which has a nice "eventually" helper that takes a block and runs that until it stops throwing exceptions (or times out). It has configurable interval that it tries again that backs off with increasing sleep periods. Most things just take a second or two to happen.

- If your CPUs are not maxed out during the test, you need to be running more tests, not less. Server tests tend to be IO blocked, not CPU blocked. And your SSD is unlikely to be the bottleneck. We're talking network IO here. And it's all running on localhost. So, if your CPUs are idling, you can run more tests and can use more threads, co-routines, whatever.

- Get a decent laptop and pay for fast CI hardware. It's not worth waiting 10 minutes for something that could build in about a minute. That speedup is worth a lot. And it's less likely to break your flow state.

This stuff is a lot easier if you engineer and plan for it. Introducing concurrently running tests to a test suite that isn't ready for it can be hard. Engineering your tests to be able to support running concurrently results in better tests. So if you do this properly, you get better tests that run faster. Win win. I've been doing this for a while. I'm very picky about what is and isn't a good test. There are a lot of bad tests out there.

In my day job I work with Go and while it's fine, I end up writing 10x more code for simple API endpoints and as soon as you add query parameters for filtering, pagination, etc. etc. it gets even longer. Adding a permissions model on top does similar. Of course there's a big performance difference but largely the DB queries dominate performance, even in Python, at least for most of the things I do.

This doesn't mean writing tests for everything, and sometimes it means not writing tests at all, but it means that I do my best to make code "testable". It shouldn't take more time to do this, though: if you're making more classes to make it testable, you're already messing it up.

This also doesn't mean compromising in readability, but it does mean eschewing practices like "Clean Code". Functions end up being as large as they need to be. I find that a lot of people doing especially Ruby and Java tend to spend too much time here. IMO having lots of 5-line functions is totally unnecessary, so I just skip this step altogether.

It also doesn't mean compromising on abstractions. I don't even like the "rule of three" because it forces more work down the line. But since I prefer DEEP classes and SMALL interfaces, in the style of John Ousterhout, the code doesn't really take longer to write. It does require some thinking but it's nothing out of the ordinary at all. It's just things that people don't do out of inertia.

One thing I am a bit of hardliner about is scope. If the scope is too large, it's probably not prototype or MVP material, and I will fight to reduce it.

EDIT: kukkeliskuu said below "learn one tool well". This is also key. Don't go "against the grain" when writing prototypes or first passes. If you're fighting the framework, you're on the wrong path IME.