You give me an example of a test that would realistically be written in a dynamic language then. I've been out of dynamic languages for so long that I've forgotten how to produce working code in them (seriously, I tried to write some Django because I used to be quite good at it and it seemed like a good fit for a specific problem, and I just wrote code that had stupid bugs everywhere).

Here's a test I wrote recently. The code in question is part of a nodejs service that talks to a database and produces json objects to be used by a web client application.

There is a definition for the schema to a fruit table.

    Fruit: {
        list_id: 32
    }

The list_id indicates that this isn't a real table, but is actually stored in the list_options table. So my code has to generate a different query then for a normal table.

The actual options are populated from a list:

    const LIST_OPTIONS = [
        [32, "Blueberry", 1],
        [32, "Apple", 2],
        [32, "Orange", 45],
        [45, "Nope", 1]
    ] 

These get inserted into the database. Note that the first column is the list_id from the schema.

And for the actual test:

This actual sends the request to be processed.

       let response = await processRequest({
           type: 'QUERY',
           table: 'Fruit',
       })

Then I assert that the respone is exactly what I thought it would be:

      deepStrictEqual(response, {
           status: 'OK',
           records: [{
                id: "Blueberry",
                name: "Blueberry",
                position: 1
           }, {
               id: "Apple",
               name: "Apple",
               position: 2
           }, {
               id: "Orange",
               name: "Orange",
               position: 45
           }]

I would have written pretty much the exact same test if it were a statically typed language.

Ok, good example on some levels, though it's doing a lot of interfacing with untyped parts. In a real-world scenario I would be looking to extend the type safety by using thrift or similar rather than json, and ideally a strongly typed datastore (for a static list like this, probably plain old code rather than a database).

That aside, what pieces could actually go wrong in a statically typed language? You can't typo field names or the like. You could typo the names, but I don't believe typing them into the computer twice adds value - better to check them once in code review. People aren't going to be editing this code in a way that could lead to getting the names wrong.

You couldn't go down the wrong code path and end up with the "normal table" query, because having list_id instead of something else is part of the type. You couldn't pass some other number as the list id, because list ID is a typed thing that is different from all other IDs, and in any case where would an outside number come from? And you're definitely going to return a response that's formatted with a record list, because that's part of the type of the method. (You might need to test the JSON <-> String piece, but that's generic, so you only need an O(1) test for that, you don't have to test every JSON endpoint. In practice you're likely using an existing well-tested library. Similar reasoning applies to the database querying part). Assuming you have warn-for-unused-parameters enabled, you couldn't forget to apply the criterion and just return everything in the list.

So yeah, assuming this was for part of an existing system where I'd already tested generic facilities like database access and JSON serialization, I wouldn't feel the need to write this test at all.

> So yeah, assuming this was for part of an existing system where I'd already tested generic facilities like database access and JSON serialization, I wouldn't feel the need to write this test at all.

But this is the generic facility for database access. This is just an additional special case it has to handle. Does that change your perspective on whether you would write a test for it?

> That aside, what pieces could actually go wrong in a statically typed language?

Well, I could type the SQL wrong, but let's assume we are working in a sufficiently awesome statically typed language that would catch that kind of error.

I could filter on the wrong column, with the wrong operator, with the wrong value, on the wrong table. I could decode the name column using the wrong encoding. I could return multiple rows per option. I could do a huge number of other things that I couldn't think of. These may not be really likely to be mistakes I'd make (this is a pretty simple task), but several refactorings down the line, I can easily see an issue being introduced.

At the end of the day, even in a statically typed language where much of the functionality would be verified by the compiler, I'd still like to have a test on this.

> Ok, good example on some levels, though it's doing a lot of interfacing with untyped parts

Hmm... the fact that it deals with the boundary of the system and the fact that it deals only with converting data from one format to another might render it a poor sample (i.e. there's not really much interesting logic). I can come up with another if you are interested.

> But this is the generic facility for database access. This is just an additional special case it has to handle.

Presumably it would be using a lower layer - or more likely an existing library - for doing the actual query though?

> I could filter on the wrong column,

Not if you've got a typed representation of which column's which

> with the wrong operator,

Not if equality is the only operator available. This is one of my biggest reasons for using wrapper types around IDs - you shouldn't treat IDs as integers, because the idea of adding or multiplying or <=ing an ID makes no sense.

> with the wrong value,

How? You have to use the passed ID for something, and it can't do anything else.

> on the wrong table.

So make the table part of the type. A list ID is a different thing from a user ID, they shouldn't be compatible.

> I could decode the name column using the wrong encoding.

Encoding should absolutely be handled with types.

> I could return multiple rows per option.

Hmm, maybe. But I can't see how you'd do that by accident.

> At the end of the day, even in a statically typed language where much of the functionality would be verified by the compiler, I'd still like to have a test on this.

I started out thinking the same thing (I moved from Java/Python to Scala). I've gradually reduced the amount of tests I write over four years as I've found them to not add value. I was as surprised as anyone, and even privately entertained the thought that maybe I was some kind of super-programmer who was smarter than all those people who needed tests, but that month or two of trying to do stuff in Django made it very clear that it wasn't me, it was the language.

> Hmm... the fact that it deals with the boundary of the system and the fact that it deals only with converting data from one format to another might render it a poor sample (i.e. there's not really much interesting logic). I can come up with another if you are interested.

Yeah, happy to if you're interested. It's a useful exercise in thinking it through for myself as much as anything.

> I started out thinking the same thing (I moved from Java/Python to Scala).

Ah, Scala. I could be prepared to believe that this is true for Scala (and similarly advanced typed systems). It is when people claim they write fewer tests in java than in python that I found it really dubious.

> Yeah, happy to if you're interested. It's a useful exercise in thinking it through for myself as much as anything.

Ok here's another test:

These are action objects, they represent a sequence of user actions on a JobExtra.

        const actions = [
            JobExtra.actions.load(undefined),
            JobExtra.actions.field('changeAmount', Money.actions.set(new Decimal("12.45"))),
            JobExtra.actions.field('certifiedForemanRelated', Checkbox.actions.set(true)),
            JobExtra.actions.field('appliedPercentage', Percentage.actions.set(new Decimal("4.5")))
        ]

I construct the final state by reducing across all those actions

        const state = _.reduce(actions, (state, action) => JobExtra.reducer(state, action), undefined)

I verify that the job extra is considered valid afterwards.

        strictEqual(true, isJobExtraValid(state))

To be considered valid:

1. The changeAmount must be set to non-zero 2. The certifiedFormanRelated must be false or the appliedPercentage must be true

But, if the user modifies the appliedPercentage, the certifiedFormanRelated should be automatically set to true.

So this, and a number of related tests verify that under different sequences of actions I end up the result I expected.

Is there a way to push this into the type system?

Yep. Ensuring something is valid is a great fit for use of types - you make sure that the type is inherently valid and it's impossible to construct in an invalid state.

For 1., you could have a non-zero number type, or you could settle for a check in the JobExtra constructor. For 2., it sounds like this is a coproduct situation - a JobExtra is either certifiedFormanRelated or has an appliedPercentage. Scala doesn't quite have native support for coproducts the way it does for tuples, but you could represent it with a sealed class hierarchy:

    sealed trait ForemanOrPercentage
    case class ForemanRelated() extends ForemanOrPercentage
    case class AppliedPercentage(value: Decimal) extends ForemanOrPercentage

or being a bit lazier you could just use an Option[Decimal] and say that if it's Some it's the applied percentage and if it's None then it's certified foreman related.

In Java there isn't a "sealed" keyword, but you can emulate it with the visitor pattern - you'd define a ForemanOrPercentageVisitor and only ever operate on ForemanOrPercentages via the visitor. That way if someone creates their own subtype of ForemanOrPercentage it still has to behave like one or the other, because they still have to implement visit().

I don't think that works for me.

Since the user input modifies this object, the user can put the object into an invalid state. In fact, when the object is first created it is in an invalid state. So I can't simply define my types so that invalid state is impossible.

If your object has distinct possible states, maybe it should be two different objects. I don't think of user input as modifying my object (after all, this kind of object should be an immutable value) - rather I think of mapping my object to a set of form fields and mapping a set of form fields into a new object. (This goes well with having a component hierarchy model - just as my object is a value made of simpler values, its corresponding editor form is made of simpler editor forms). So I'd only create the (domain-level) object when the user's provided the right input to create the object. If you need a model of the user input in code, that's fine, but make that its own distinct type - then you have a single point at which to do validation, and the types make it very clear whether a given value is a business object (in which case it is necessarily valid) or unvalidated user input.

I see now.

That's very interesting, and shows how types can go much further then what we see in languages like Java. I'll have to try out one of these languages some day and see how well it works in practice.

Even in Java it's possible (but it's not a great environment to learn these techniques in - they usually end up being very verbose when expressed in Java) .

I don't know. The pure fact I can't be sure I won't have NullPtrExceptions in java code would make it pretty hard to accept not writing those tests.

If you always declare object fields as final then you're forced to initialize them in the constructor, so there's no risk of null there. Reading a field in the constructor should be a red flag, as should literal null in your code. Local variables should of course be final too. At that point you only need to check for null around library calls. I'm not sure how much a test actually helps with that - if you don't think of the possibility that a library call might return null then you wouldn't write a test for it either, if you do think of it then you wrap the call in a null test in the same line (don't try to get clever and store possibly-null values in variables).

> if you don't think of the possibility that a library call might return null then you wouldn't write a test for it either

No, I don't think so. I'll write a test that uses the library. If the library returns null when I don't expect it, it'll go into my code an cause an exception to tbe thrown. I'm not deliberately testing for nulls, I'm testing generally that my code does what I think it does.

> No, I don't think so. I'll write a test that uses the library. If the library returns null when I don't expect it, it'll go into my code an cause an exception to tbe thrown. I'm not deliberately testing for nulls, I'm testing generally that my code does what I think it does.

The kind of library function that returns null doesn't usually return it on the "happy path", it usually returns null for some weird condition you haven't thought of. I mean, if it's a function that returns null often or that you'd expect to sometimes return null, you'll wrap up the return value in an option straight away even without having written a test. The kind of library function that causes you to get exceptions in production is the kind that returns null in some weird corner case - but unless you think of the weird corner case, you wouldn't hit it in a test either.

If I was using a library with some complex stateful interface I might write tests for the interaction with that (I'd probably put a facade on it, keeping it separate from the business logic, and the facade would have unit tests) - but I'd try to avoid using such a library. And I'll always have a few high-level end-to-end tests (not really "unit" tests) as a basic "does it work at all", because that kind of mistake is easy to make. But the bread-and-butter unit tests that I wrote when I worked in Python just don't seem valuable or necessary.

I think you are assuming saner code then I'm used to.

Honestly, I can't say how it would work if I was using a java code base that consistently applied these kinds of ideas. You may be right, but at this point I have hard time believing.

I've worked on some bad codebases. You can convert a little bit at a time, treating the remainder of the codebase as a large library with a poorly factored interface. You absolutely will need unit test coverage at the boundary, which will probably exist for several years.