The solution is to be religious about testing and writing tests. If you've done it well, refactors are pretty safe and your objects are pretty well understood.
Unfortunately, most people got a taste of static typing with Java and the initial language and library design were done in such a way that types were well, if not entirely useless, then certainly under-used. But if you look at many libraries for languages such as Rust, Scala, OCaml, Haskell, F#, ... (I haven't looked at Java in a while, but I don't hold high hopes on this specific front) you'll find many examples in which the API and the type system cooperate to guarantee that high-level protocols are enforced.
That being said, I absolutely agree that strong static typing does not mean that you don't need tests. As everything, if you want to be safe, you need a defense in depth, with good API design and many test layers.
The solution is to become a software atheist and just break things down when they no longer fit in your head.
Interestingly, this seems to be what microservices originally were about until they got twisted by hype.
It is also what OOP was about, until it got twisted by another religion.
But to counter myself, the perceived lack of developer speed and flexiblity of a strongly typed language (with Rust, I always get the feeling that I'm fighting with the compiler!) is also solved in the long run with practice and tooling.
Sure, test for sane failure modes in line with the documented contract, and that may include the occasional test that is de facto a type test, but often testing for types, especially in languages with poor type systems, but also in Ruby where we have alternatives, ends up with tests for classes which is frequently the wrong thing.
E.g. in Ruby never, ever check for somearg.kind_of?(IO) if all you ever do is somearg.read - if you absolutely must typecheck somearg, the Ruby way is to check for presence of "#read", e.g. somearg.respond_to?(:read), or try and fail responsibly (and often just allowing the NoMethodError to bubble up is the right way to fail).
Also think it's just fine for people to add and ship Sorbet type declarations for gems etc. to then signal those contracts so people can verify them if they choose. There's no reason not to offer that when it's reasonable to do so.
In Python or JavaScript/TypeScript, though, my experience is that failing to validate types at the borders (i.e. every single function/method/constructor/generator/... exposed in your API) is pretty much guaranteed to end up, months later, with developers attempting to sherlock out surprising breakages in production from logs that make no sense and traces that do not show anything remotely close to the actual culprit.
I have the scars to show it :) Of course, YMMV.
Hence don't check for an IO object when what you care about is the presence of a "#read" method.
Or don't check if something is an Array if what matters is that it supports "#map". Instead, either somearg.respond_to?(:map), or if it needs more than map, somearg.respond_to?(Enumerable) (this seems broad, but supporting Enumerable only requires implementing "#each" and including the module, so a caller "worst case" can reopen a class or wrap their object), or call one of Array(somearg) (tries "#to_ary" then "#to_a" then falls back to returning [somearg]) or Array.try_convert(somearg) (tries "#to_ary" then falls back on nil).
Consistently picking the most generic applicable options when faced with choices like that still enforces the contract on the boundary, but also tends to lead to code with much less ceremony. E.g. far fewer [something]Adapter classes, or glue code to convert data before calling methods that are being overly prescriptive.
But for most statically typed languages, when people talk about static typing, they still talk about a class and a type as interchangeable (there are exceptions, and it's getting better, and with increased type inference coupled with increased support for type annotations and analysis of dynamic languages, I expect there to be an increasing convergence, though).
Note that I'm writing "types" and not "classes". Interfaces do just as well, if not better!
For what it's worth, in the static realm, OCaml introduce features that let code check for presence of a method (instead of implementing an interface) ~25 years ago. Unfortunately, this proved rather unwieldy and, to the best of my knowledge, nobody uses that feature.
When there is a meaningful name to give, in Ruby you might define a module, and include it and use that as an interface:
module HelloWorld
def hello
raise "Not implemented"
end
end
class Foo
include HelloWorld
end
Foo.new.hello # Raises an exception
Foo.new.kind_of?(HelloWorld) # Returns true
There are plenty of contract-checking frameworks for Ruby [1] but they see little use, because most of our contracts tend to be extremely simple. Along the line of "implements method X" or "includes module Y".
Sorbet w/inference tools and optionally storing the types in a separate files (it supports inline too) might slowly change that (and effectively let you use modules as "proper" interfaces), since the biggest opposition to types in Ruby tends to be visual noise and forcing refactoring, and if it can be turned on/off in your IDE and regenerated, a lot of the objections fall away. It's not that we (well many of us) don't want the extra type checks, but that we don't want to pay the cost of making the code less readable.
[1] Here's one: http://egonschiele.github.io/contracts.ruby/ and of course there's Sorbet: https://sorbet.org/blog/2020/07/30/ruby-3-rbs-sorbet
Do you have examples? That's not my experience.
Of course, we may simply be working on very different types of code :)
But e.g consider basically almost any case with abstract member functions / methods. While that happens in Ruby (or the equivalent: Leaving a method which raises an exception unless overridden) it's far less common.
Typically you'll instead implement defaults in a module if there is lots of functionality that a client might want for itself). These are generally fine to name, and type check by the module name.
But when that is not* the case (any class consisting mostly of abstract methods), a Ruby implementation would be more likely to instead expect the presence of a single method and embed the default functionality in the calling class. That's the kind of scenarios where you end up with the kind of long, contrived class names people tend to mock Java for in particular.
But going further, a Ruby implementation will instead often take an argument that is expected to support "to_proc" or a block argument, so that you can provide a lambda/closure to do what you'd otherwise put in an "adapter" or "shim" class.
Conversely, any time where you demand (whether in a static language by type annotiations or in a dynamic one by checking for the class) an object of a very specific class and only call one or a few methods on it, you're sidestepping this problem by avoiding the creation of a more precise interface that'd clutter up the code. You don't create and name interfaces that encompass the specific operations needed by each method because it'd be a total mess.
A few common examples for which I rely upon static typing: - I have a proof that operation X is complete, because I have a protocol to follow; - this string is a Matrix room id (and not a Matrix user id or a Matrix server id, also strings); - this time is in milliseconds.
Sounds like very different use cases :)