Error types

Last updated 2 months ago

You can find all the code here

Creating your own types for errors can be an elegant way of tidying up your code, making your code easier to use and test.

Pedro on the Gopher Slack asks

If I’m creating an error like fmt.Errorf("%s must be foo, got %s", bar, baz), is there a way to test equality without comparing the string value?

Let's make up a function to help explore this idea.

// DumbGetter will get the string body of url if it gets a 200
func DumbGetter(url string) (string, error) {
res, err := http.Get(url)
if err != nil {
return "", fmt.Errorf("problem fetching from %s, %v", url, err)
}
if res.StatusCode != http.StatusOK {
return "", fmt.Errorf("did not get 200 from %s, got %d", url, res.StatusCode)
}
defer res.Body.Close()
body, _ := ioutil.ReadAll(res.Body) // ignoring err for brevity
return string(body), nil
}

It's not uncommon to write a function that might fail for different reasons and we want to make sure we handle each scenario correctly.

As Pedro says, we could write a test for the status error like so.

t.Run("when you dont get a 200 you get a status error", func(t *testing.T) {
svr := httptest.NewServer(http.HandlerFunc(func(res http.ResponseWriter, req *http.Request) {
res.WriteHeader(http.StatusTeapot)
}))
defer svr.Close()
_, err := DumbGetter(svr.URL)
if err == nil {
t.Fatal("expected an error")
}
want := fmt.Sprintf("did not get 200 from %s, got %d", svr.URL, http.StatusTeapot)
got := err.Error()
if got != want {
t.Errorf(`got "%v", want "%v"`, got, want)
}
})

This test creates a server which always returns StatusTeapot and then we use its URL as the argument to DumbGetter so we can see it handles non 200 responses correctly.

Problems with this way of testing

This book tries to emphasise listen to your tests and this test doesn't feel good:

  • We're constructing the same string as production code does to test it

  • It's annoying to read and write

  • Is the exact error message string what we're actually concerned with ?

What does this tell us? The ergonomics of our test would be reflected on another bit of code trying to use our code.

How does a user of our code react to the specific kind of errors we return? The best they can do is look at the error string which is extremely error prone and horrible to write.

What we should do

With TDD we have the benefit of getting into the mindset of:

How would I want to use this code?

What we could do for DumbGetter is provide a way for users to use the type system to understand what kind of error has happened.

What if DumbGetter could return us something like

type BadStatusError struct {
URL string
Status int
}

Rather than a magical string, we have actual data to work with.

Let's change our existing test to reflect this need

t.Run("when you dont get a 200 you get a status error", func(t *testing.T) {
svr := httptest.NewServer(http.HandlerFunc(func(res http.ResponseWriter, req *http.Request) {
res.WriteHeader(http.StatusTeapot)
}))
defer svr.Close()
_, err := DumbGetter(svr.URL)
if err == nil {
t.Fatal("expected an error")
}
got, isStatusErr := err.(BadStatusError)
if !isStatusErr {
t.Fatalf("was not a BadStatusError, got %T", err)
}
want := BadStatusError{URL:svr.URL, Status:http.StatusTeapot}
if got != want {
t.Errorf("got %v, want %v", got, want)
}
})

We'll have to make BadStatusError implement the error interface.

func (b BadStatusError) Error() string {
return fmt.Sprintf("did not get 200 from %s, got %d", b.URL, b.Status)
}

What does the test do?

Instead of checking the exact string of the error, we are doing a type assertion on the error to see if it is a BadStatusError. This reflects our desire for the kind of error clearer. Assuming the assertion passes we can then check the properties of the error are correct.

When we run the test, it tells us we didn't return the right kind of error

--- FAIL: TestDumbGetter (0.00s)
--- FAIL: TestDumbGetter/when_you_dont_get_a_200_you_get_a_status_error (0.00s)
error-types_test.go:56: was not a BadStatusError, got *errors.errorString

Let's fix DumbGetter by updating our error handling code to use our type

if res.StatusCode != http.StatusOK {
return "", BadStatusError{URL: url, Status: res.StatusCode}
}

This change has had some real positive effects

  • Our DumbGetter function has become simper, it's no longer concerned with the intricacies of an error string, it just creates a BadStatusError.

  • Our tests now reflect (and document) what a user of our code could do if they decided they wanted to do some more sophisticated error handling than just logging. Just do a type assertion and then you get easy access to the properties of the error.

  • It is still "just" an error, so if they choose to they can pass it up the call stack or log it like any other error.

Wrapping up

If you find yourself testing for multiple error conditions dont fall in to the trap of comparing the error messages.

This leads to flaky and difficult to read/write tests and it reflects the difficulties the users of your code will have if they also need to start doing things differently depending on the kind of errors that have occurred.

Always make sure your tests reflect how you'd like to use your code, so in this respect consider creating error types to encapsulate your kinds of errors. This makes handling different kinds of errors easier for users of your code and also makes writing your error handling code simpler and easier to read.