NullAway VS infer

Some patterns arise from language design

* You can't express `T` where `null` is forbidden in the type system so you get NullPointerException everywhere and defensive null checks.

* You express a sum type as a product type because your language does not have sum types .

* Your language doesn't have first class multiple return values (or tuples) so you return extra parameters via out parameters or thread local variables such as `errno`.

* Your language doesn't have exceptions (or algebraic effects) and can't do IO so you have monad transformers.

* Your language doesn't have set-theoretic types so you need hacks like `thiserror` .

* Your language doesn't have stackful coroutines or can't infer async IO for you so you have `async/await` spam or callback hell or "mono's".

* Your language doesn't have exhaustive checks (or pattern matching) so you need a fallthrough case check on switch statements .

* Your language doesn't have algebraic effects, so you need to pass context everywhere.

I know someone will reply about Java's null annotation checking options, so here is one of them: https://github.com/uber/NullAway .

If you must use Java, use Uber's Nullaway which gives null safety via Errorprone.

Does anyone have experience using this at Meta who can compare to https://github.com/uber/NullAway ?

A special kind of validation is enforcing that record fields are not null. (Un)fortunately, records do not have any special behavior regarding nullability. You can use tools like NullAway or Error Prone to prevent null in your code in general, or you can add checks to your records:

What do you think of https://github.com/uber/NullAway

Take a look at NullAway, a plugin for Error Prone.

PMD, Spotbugs, Nullaway: Java linting/static analysis (https://pmd.github.io, https://spotbugs.github.io, https://github.com/uber/NullAway)

In personal projects, I've had good experiences using the error-prone compiler plugin with uber's nullaway.

I think most development occurs on problems that can't be formally modeled anyway. Most developers work on things like, "can you add this feature to the e-commerce site? And can the pop-up be blue?" which isn't really model-able.

But that's not to say that formal methods are useless! We can still prove some interesting aspects of programs -- for example, that every lock that gets acquired later gets released. I think tools like Infer[0] could become common in the coming years.

[0]: https://fbinfer.com/

Using infer, someone else exploited null-dereference checks to introduce simple affine types in C++. Cppcheck also checks for null-dereferences. Unfortunately, that approach means that borrow-counting references have a larger sizeof than non-borrow counting references, so optimizing the count away potentially changes the semantics of a program which introduces a whole new way of writing subtly wrong code.

Meta/Facebook are long time OCaml users, their logo is on the OCaml website. Their static analysis tool and its predecessor are both written in OCaml.

Then this idea that the US government will tell tech companies how to write secure software. Let's get this straight, the private sector, especially big tech is miles ahead of US government in this regard. Microsoft literally invented threat modelling and modern exploit mitigations. Facebook has the best appsec processes pretty much in the whole world, including their own cutting edge code analyzer. AWS uses formal verification everywhere. Meanwhile the US government itself runs mission-critical systems that's almost literally held together by bubble gum and toothpicks. Maybe they could dial down the arrogance a tad, get their own shit together, learn how this cyber stuff is actually done and only then try lecturing everyone else.

Efforts: Dependabot, CodeQL, Coverity, facebook's Infer tool, etc

I notice there isn't fbinfer. It's pretty cool, and is used for this library.

"Move fast, break stuff" is a great approach when you aren't pushing the broken bits to production. Fuck, even Facebook, the big "move fast, break stuff" company, uses tools to detect errors in its continuous integration toolchain. https://fbinfer.com/

TBH, there's a non-zero amount of non-"ivory tower" tools you may have used that are written in functional languages. Say, Pandoc or Shellcheck are written in Haskell; Infer and Flow are written in OCaml. RabbitMQ and Whatsapp are implemented in Erlang (FB Messenger was too, originally; they switched to the C++ servers later). Twitter backend is (or was, at least) written in Scala.