Hacker News new | past | comments | ask | show | jobs | submit login

I think it depends on the problem domain. In high-risk or regulated industries having requirements and specifications is a prerequisite to shipping. There are plenty of good sources for finding the standard expected formats out there. If you're an IEEE member you'll find the ISO requirements and specifications templates in the library.

Doing requirements gathering isn't an inherently bad process to adopt even in non-regulated settings. However I believe many developers will have a negative reaction due to prior experiences with, or having heard about, the waterfall method. The key to remember is that requirements don't, and shouldn't, have to come with an estimate. Great requirements demonstrate a thorough understanding of the problem. They're written from the perspective of the end-user.

Specifications are the dual of requirements and are what drive the implementation that solves the stated problem. Specifications is something I think we need to get better at. We some unit testing, some integration tests, and occasionally end-to-end user tests... some few to property tests; but rarely do we write formal, verifiable specifications in a modelling language that can be checked with a model checker. Rarer still do we write proofs as part of our specifications.

We often elide specifications or we do the minimum necessary to assure correctness. This is to the detriment of your team on a larger project. How do you know your designs don't contain flaws that could have been avoided? How much time are you spending before you realize that the locking mechanism you chose and shipped has a serious error in it that is causing your SLO's to slip?

For requirements I just use plain old Markdown with pandoc. For specifications I use a mixture of Markdown and TLA+. I use TLA+ for the hard parts of the system that we're unsure about and require correctness for. The rest of the specifications that aren't as interesting I simply use prose. It's a matter of taste but it does require an intuition for abstraction... when to do it and how to think about systems abstractly.

We could definitely use better tools for TLA+-style specifications, btw. Maye more libraries that can translate TLA+ state machines into the host language so that we can drive them from our integration tests, etc. Better IDE tooling. Better training.




Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: