satishgupta

Some teams build software with ML features. They may do some in-house research or reuse existing models/architectures. The ML feature can be a critical part of the product, or it can be a nice-to-have.

I am more interested in this. Understand from the perspective of building ML-based software applications/product-features (updated the question).

That is a good point.

I looked at GitHub: https://madnight.github.io/githut/#/pull_requests/2020/2

And that matches with survey. How that can be reconciled? What are the biases here?

First and foremost, I does not treat anything as gospel. TDD, like everythingthing else, might not make sense in all scenarios.

I often use TDD for microservices. It forces me to define REST service interface from the perspective of consumers, and only then jump into implementing it. So I write tests that call the service for various use cases, and asserts for returned status code, pay load, and any invariances. All tests fail initially.

It is indeed cumbersome, requires disipline, but it becomes easier with practice. I find, it particularly useful for for microservices. Earlier I would implement, then test, and while doing both will keep figuring a better API and keep changing. With TDD, I upfront get the clarity what functionality I want to implement, so counterintuitively TDD makes me faster.

I use it mostly wherever interfaces are involved. For example, if my service is using a DB, unless the use cases are too simple, I write tests for that API.

Parts that does not have (external) interface, and I want it to be fluid/evolve, I write tests only after implementing. But service level (TDD) tests provide some safety net.

I can see this approach can be extended at all levels, but I am not sure about costs vs. benefits tradeoffs. I think if I do TDD only for macro (process/service/subsystem interfaces kind of things) and some leaf library type modules, tradeoffs are favourable. And if you have TDD for those, the returns of doing it for rest diminish.

Disclaimer: I haven't done detailed study, I am telling my experience, so only one data point.

I use 3-pass learning strategy: Mark, Drill, and Sweep.

Mark: Quick overview of the landscape marking key ideas/sub-topics.

Drill: Map task at hand to a sub-topic, drill as deep as needed to finish that task.

Sweep: If needed, systematically study the subject.

It is rare that I get to do sweep. It is "eventually" achieved through a lot of need based drilling.

... hackery of epic scale.

Code can indeed be poetry :-)