In my experience this problem tends to be caused by heavily mocking things out more so than the unit tests themselves. Mocking things out can be a useful tool with its own set of downsides, but should not be treated as a requirement for unit tests. Tight coupling in your codebase can also cause this, but in that case I would say the unit tests are highlighting a problem and not themselves a problem.

Perhaps you're talking about some other aspect of unit tests? If that's the case then I'd love to hear more.

Not mocking the database and other pipes is the single best improvement everyone can make on their test suites.

We also had a test suite that followed the same principle but started all the services together to reduce the mocked surface, it executed every hour and was both incredibly useful and reliable too.

The test pyramid is false wisdom.

I wanted to drop in and say we had a version of this discussion internally while I was putting this post together. Your observation about fixing a bunch of tests for a simple one line change is something I have seen as well. What we ultimately landed on is that, especially in our service-heavy environment (though not necessarily micro services), the cost of creating and maintaining integration testing infrastructure that is reliable, reasonably fast, and reflective of something prod shaped turns out to be even more expensive. Specifically, we looked at things like the costs of creating parallel auth infra, realistic test data, and the larger, more complex test harness setups and on balance it actually ends up being more expensive on a per-test basis. In fact, in some cases we see meaningful gaps in integration testing where teams have been scared off by the cost.

This isn't to say that unit tests, especially those with heavy mocking or other maintenance issues don't carry their own costs, they absolutely do! But, and I think importantly, the cost-per breakage is often lower as the fix is much more likely to be localized to the test case or a single class. Whereas problems in integration tests or E2E tests can start to approach debugging the prod system.

As with any "experiential opinion" like this, YMMV. I just set out to try to contribute something to the public discourse that's been reflective of our internal experience.

The testing-type divide feels similar to the schism around ORMs, where one camp (mine) find that ORMs end up costing far more than the value they bring, while the other claim they've never had such issues and they would never give up the productivity of their favorite ORM.

Both sides appear to be describing their experiences accurately, even though it feels like one side or the other should have to be definitively right.

It might be useful to provide a little more context for why I wanted to write this in the first place - Over the last 15 or so years we have been tremendously successful at getting folks to write tests. And like any system, once you remove a bottleneck or resource constraint in one place, you inevitably find one somewhere else. In our case we used to take running our tests for granted, but now the cost of doing so now has actual cost implications that we need to consider. I also observed some in internal discussions that had become a little to strident about the absolutes of one kind of test or another, and often in such a way that treated terms like "unit" or "integration" as a sort of universal categories, completely ignoring the broad, practical implications we have bound together into a few shorthand terms.

My goal when trying to develop this idea was to find a way to succinctly combine the important set of tradeoffs teams should consider when thinking, not about a single test, but their entire test suite. I wanted to create a meme (in the Dawkin's sense) that would sit in the background of an engineer's mind that helped them quickly evaluate their test suite's quality over time.

I don't find this to be the case if the unit tests are precise (which they should be).

That is, if you are writing non-flaky unit tests which do all the "right" unit-testy things (using fakes/dependency injecting well and so isolating and testing only the unit under test), you should end up with a set of tests that

- Fails only when you change the file/component the test relates to

- Isn't flaky (can be run ~10000 times without failing)

- Is quick (you can do the 10000 run loop above approximately interactively, in a few minutes, by running in parallel saturating a beefy workstation)

This compares to integration/e2e tests which inherently break due to other systems and unrelated assumptions changing (sometimes legitimate, sometimes not), and can have rates of flakyness of 1-10% due to the inherent nature of "real" systems failing to start occasionally and the inherently longer test-debug cycle that makes fixing issues more painful (root causing bug that causes a test to fail 1% of the time is much easier when the test takes .3 CPU-seconds than when it takes 30 or 300 CPU-seconds).

Very few tests I see are actually unit tests in the above sense, many people only write integration tests because the code under test is structured in inherently un- or difficult- to test ways.

IME, testable pretty much just means referentially transparent.

> What I would like to see instead …

If you hire me, I can address those needs.