Ask HN: Software Testing Philosophy? | Hacker News Reader

Ask HN: Software Testing Philosophy?

7 pointsdoitLP10 comments

What is your philosophy for deciding if your software is well-tested?

Beyond the amorphous "do test reports inspire developer confidence?" I realize I don't know how to think about different testing thresholds.

I am taking over engineering at a young startup. Their approach to testing has been some automated smoke tests, and "90%" code coverage for unit tests. 90% seems too high for the speed tradeoff we have to make, but where to draw the line? Why 90% and not 84%?

Should we just try to find something that works and let production bugs be our guide? I would also like to explore other methods of testing I've never used, e.g. crowd testing, but I'm unsure how to fit that into my mental model of knowing well-tested software when I see it.

Any insight or articles are welcome. Thanks!

View on HN

I've found 80% to be a good number for unit tests. BUT things like integration testing and end-to-end user story testing can really make all the difference, and complement a solid unit test suite. Also focussing on the numbers can leave you going down the wrong path and trying to shoehorn as much testing in as possible without giving any tangible benefit for the sake of a higher number.

As it's a young startup I would focus on writing tests for the business-critical use cases, which should allow a team the confidence to "move fast and break things" knowing comfortably that the day-to-day is still functioning as per usual.

Ultimately, I would say most code should be covered with unit tests, but functionality covered by automated integration/end-to-end testing. To me that would be 'well-tested'.

(and no flakiness!)

I broadly agree with this approach. Automated functional end-to-end tests may not pinpoint what line of code has an issue but they will identify there’s a problem somewhere which is better than not knowing. Your product code and tests can be setup to help isolate where the source of the issue may be. Have developers continue to write unit tests but if you’re big enough to it have one developer focused on writing end-to-end automated functional tests it will give everyone a sense of assurance at each release that nothing is fundamentally broken. If you’re not big enough, like the parent comment said, try to do the major ones eg. sign up flow, core functionality.

Someone else made the point about using data effectively and I agree with that too. I’ve been involved with automated testing and tooling for a few years now and solid test reporting and test case management is something I advocate. I’m the lead dev and founder at Tesults (http://www.tesults.com) - try it and push your results data to it and house your manual test cases. There’s a free tier and if you contact me I’m happy to expand it. Basically if you set up your CI system to run your tests regularly, or even if you schedule them to run regularly (E.g. nightly/daily) then having something like this boosts your ROI because you can track history easily, link bugs, and do things like get indications about which tests may be flaky.

We don’t know exactly how small your team is, it may be you have to rely on manual testing if there’s no one to spare but I’d try to move some of your end-to-end to be automated sooner rather later if you end up growing otherwise you’ll make testing a bottleneck to releasing, full regression testing of even a moderately complex application can end up taking days.

Why would I want to link tests to bugs? Or track the history of tests? Or assign someone to a failing test? I've found they either fail or they pass, and you find why they failed and you fix the failure.

Just genuinely curious about the purpose of this software, as it's quite expensive, and most CI tools provide enough information for history / scheduling etc.

If you don’t see it as something you need, you probably don’t ... at least not yet.

Ive been an Engineering Manager and Head of Quality Engineering and when you end up with hundreds or thousands of test cases, either in one product or service or across many, managing the output of testing becomes a challenge. There may be low priority regressions that you want to fix later and create bugs for and link to failing tests so you know you’re waiting on those fixes before you expect the tests to pass.

If there is a problem with a test itself then assigning it to a team member to fix is useful. Tracking history helps identify flaky test cases - they’re automatically highlighted as being flaky. Consolidating tests from several sources helps understand the overall health of a project.

If you’re maintaining a product or service that goes through a lot of change and there’s say 10 people or more on a team managing all of this becomes increasingly difficult. With respect to pricing it’s free to try and $25/month per target so not particularly expensive.

What I do is classify parts as either prototype or release.

Prototype is code that is built for the sake of proving something - it might be to gauge user behavior, or convince investors/customers you can do this cool stuff. It might be because it's hard to plan something and easier to hack it together.

Prototype code is built disposable. You can throw it away and replace it with an optimized thing later, and they're decoupled enough to be disposed. For pre-PMF startups, nearly 100% of code is prototype.

For prototypes, I write tests, but it's a manual checklist. It's also usually user behavior focused, such as "user tries to upload 15 images from third party photo gallery and gets an error limiting them to 5 images at a time".

There's some point where it's more efficient to write tests than keep pressing buttons.

Once you decide a feature is concrete (probably someone is paying for it), then you can refactor that and do some TDD on it.

Disclaimer: context is king, we don't know your business so you'll get an "average" answer.

Google and read about the Modern Testing Principles [1], it's a set of principles and a community around it (join the Slack channel through the same link) that tries to bring change to the " test reports inspire developer confidence" state of mind. There's not a lot of tangibles around it yet but it's a good start.

As for coverage percentage, there are multiple papers showing that (usually) anything above 80%-85% does not contribute to the quality of the product. Also remember that unit test are not great at finding "good" regression bugs, especially those subtle bugs related to timing, concurrency or external resources (e.g. your DB).

I can throw dome ideas but you'll need to be more specific for them to be useful-

- Use Data, from CI tests through just-before-big-release tests to production bugs having more Data and better ways to utilize it will help you find problems earlier and solve them faster.

- Don't be afraid of some level of manual testing, sometimes it's the most cost effective solution especially for start ups. Mobiles are the best example, it's very costly and difficult to build a stable test environment especially if you want to have good phone models coverage.

- Cover bottom up (aka "The Test Pyramid"), have more unit tests, a bit less integration tests (contract tests, sub sets of modules where the rest is mocked), have good but less API tests and finally have only a handful of E2E tests and even less UI tests.

- None of the above is set in stone though if it works for you, for example if your UI is simple and stable enough go ahead and use it as your main interface.

Finally remember that testing is hard and knowing that you have good results is even harder than for your actual product, so iterate and learn as you go.

[1] https://www.angryweasel.com/ABTesting/modern-testing-princip...

On the frontend part of things, I tend to test the user behavior and not functions (unless critical).

The coverage tends to revolve around ~75% but that's not a number I actively think about. Instead I think about it in terms of interactions, possible errors etc..

Meaning you focus more on integration/e2e testing over unit tests?

That jives with my own inclinations. I’m ok with much lower unit test coverage levels on the UI where I can use something like Cypress or Robot over just testing functions.

Exactly, pairing testing-library and jest make these kind of tests super easy. No need to spin Cypress

I find this question useful: What is your objective evidence that the software does what it's supposed to?

Note that there's no "percent coverage" whatsoever in that question. Instead, it's a question of what behaviors of your code you need to test.

Reply on news.ycombinator.com