Select

You can find all the code for this chapter here

You have been asked to make a function called WebsiteRacer which takes two URLs and "races" them by hitting them with an HTTP GET and returning the URL which returned first. If none of them return within 10 seconds then it should return an error.

For this, we will be using:

  • net/http to make the HTTP calls.

  • net/http/httptest to help us test them.

  • goroutines.

  • select to synchronise processes.

Write the test first

Let's start with something naive to get us going.

func TestRacer(t *testing.T) {
	slowURL := "http://www.facebook.com"
	fastURL := "http://www.quii.dev"

	want := fastURL
	got := Racer(slowURL, fastURL)

	if got != want {
		t.Errorf("got %q, want %q", got, want)
	}
}

We know this isn't perfect and has problems, but it's a start. It's important not to get too hung-up on getting things perfect first time.

Try to run the test

./racer_test.go:14:9: undefined: Racer

Write the minimal amount of code for the test to run and check the failing test output

func Racer(a, b string) (winner string) {
	return
}

racer_test.go:25: got '', want 'http://www.quii.dev'

Write enough code to make it pass

func Racer(a, b string) (winner string) {
	startA := time.Now()
	http.Get(a)
	aDuration := time.Since(startA)

	startB := time.Now()
	http.Get(b)
	bDuration := time.Since(startB)

	if aDuration < bDuration {
		return a
	}

	return b
}

For each URL:

  1. We use time.Now() to record just before we try and get the URL.

  2. Then we use http.Get to try and perform an HTTP GET request against the URL. This function returns an http.Response and an error but so far we are not interested in these values.

  3. time.Since takes the start time and returns a time.Duration of the difference.

Once we have done this we simply compare the durations to see which is the quickest.

Problems

This may or may not make the test pass for you. The problem is we're reaching out to real websites to test our own logic.

Testing code that uses HTTP is so common that Go has tools in the standard library to help you test it.

In the mocking and dependency injection chapters, we covered how ideally we don't want to be relying on external services to test our code because they can be

  • Slow

  • Flaky

  • Can't test edge cases

In the standard library, there is a package called net/http/httptest which enables users to easily create a mock HTTP server.

Let's change our tests to use mocks so we have reliable servers to test against that we can control.

func TestRacer(t *testing.T) {

	slowServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		time.Sleep(20 * time.Millisecond)
		w.WriteHeader(http.StatusOK)
	}))

	fastServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		w.WriteHeader(http.StatusOK)
	}))

	slowURL := slowServer.URL
	fastURL := fastServer.URL

	want := fastURL
	got := Racer(slowURL, fastURL)

	if got != want {
		t.Errorf("got %q, want %q", got, want)
	}

	slowServer.Close()
	fastServer.Close()
}

The syntax may look a bit busy but just take your time.

httptest.NewServer takes an http.HandlerFunc which we are sending in via an anonymous function.

http.HandlerFunc is a type that looks like this: type HandlerFunc func(ResponseWriter, *Request).

All it's really saying is it needs a function that takes a ResponseWriter and a Request, which is not too surprising for an HTTP server.

It turns out there's really no extra magic here, this is also how you would write a real HTTP server in Go. The only difference is we are wrapping it in an httptest.NewServer which makes it easier to use with testing, as it finds an open port to listen on and then you can close it when you're done with your test.

Inside our two servers, we make the slow one have a short time.Sleep when we get a request to make it slower than the other one. Both servers then write an OK response with w.WriteHeader(http.StatusOK) back to the caller.

If you re-run the test it will definitely pass now and should be faster. Play with these sleeps to deliberately break the test.

Refactor

We have some duplication in both our production code and test code.

func Racer(a, b string) (winner string) {
	aDuration := measureResponseTime(a)
	bDuration := measureResponseTime(b)

	if aDuration < bDuration {
		return a
	}

	return b
}

func measureResponseTime(url string) time.Duration {
	start := time.Now()
	http.Get(url)
	return time.Since(start)
}

This DRY-ing up makes our Racer code a lot easier to read.

func TestRacer(t *testing.T) {

	slowServer := makeDelayedServer(20 * time.Millisecond)
	fastServer := makeDelayedServer(0 * time.Millisecond)

	defer slowServer.Close()
	defer fastServer.Close()

	slowURL := slowServer.URL
	fastURL := fastServer.URL

	want := fastURL
	got := Racer(slowURL, fastURL)

	if got != want {
		t.Errorf("got %q, want %q", got, want)
	}
}

func makeDelayedServer(delay time.Duration) *httptest.Server {
	return httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		time.Sleep(delay)
		w.WriteHeader(http.StatusOK)
	}))
}

We've refactored creating our fake servers into a function called makeDelayedServer to move some uninteresting code out of the test and reduce repetition.

defer

By prefixing a function call with defer it will now call that function at the end of the containing function.

Sometimes you will need to clean up resources, such as closing a file or in our case closing a server so that it does not continue to listen to a port.

You want this to execute at the end of the function, but keep the instruction near where you created the server for the benefit of future readers of the code.

Our refactoring is an improvement and is a reasonable solution given the Go features covered so far, but we can make the solution simpler.

Synchronising processes

  • Why are we testing the speeds of the websites one after another when Go is great at concurrency? We should be able to check both at the same time.

  • We don't really care about the exact response times of the requests, we just want to know which one comes back first.

To do this, we're going to introduce a new construct called select which helps us synchronise processes really easily and clearly.

func Racer(a, b string) (winner string) {
	select {
	case <-ping(a):
		return a
	case <-ping(b):
		return b
	}
}

func ping(url string) chan struct{} {
	ch := make(chan struct{})
	go func() {
		http.Get(url)
		close(ch)
	}()
	return ch
}

ping

We have defined a function ping which creates a chan struct{} and returns it.

In our case, we don't care what type is sent to the channel, we just want to signal we are done and closing the channel works perfectly!

Why struct{} and not another type like a bool? Well, a chan struct{} is the smallest data type available from a memory perspective so we get no allocation versus a bool. Since we are closing and not sending anything on the chan, why allocate anything?

Inside the same function, we start a goroutine which will send a signal into that channel once we have completed http.Get(url).

Always make channels

Notice how we have to use make when creating a channel; rather than say var ch chan struct{}. When you use var the variable will be initialised with the "zero" value of the type. So for string it is "", int it is 0, etc.

For channels the zero value is nil and if you try and send to it with <- it will block forever because you cannot send to nil channels

You can see this in action in The Go Playground

select

You'll recall from the concurrency chapter that you can wait for values to be sent to a channel with myVar := <-ch. This is a blocking call, as you're waiting for a value.

select allows you to wait on multiple channels. The first one to send a value "wins" and the code underneath the case is executed.

We use ping in our select to set up two channels, one for each of our URLs. Whichever one writes to its channel first will have its code executed in the select, which results in its URL being returned (and being the winner).

After these changes, the intent behind our code is very clear and the implementation is actually simpler.

Timeouts

Our final requirement was to return an error if Racer takes longer than 10 seconds.

Write the test first

func TestRacer(t *testing.T) {

	t.Run("compares speeds of servers, returning the url of the fastest one", func(t *testing.T) {
		slowServer := makeDelayedServer(20 * time.Millisecond)
		fastServer := makeDelayedServer(0 * time.Millisecond)

		defer slowServer.Close()
		defer fastServer.Close()

		slowURL := slowServer.URL
		fastURL := fastServer.URL

		want := fastURL
		got, _ := Racer(slowURL, fastURL)

		if err != nil {
			t.Fatalf("did not expect an error but got one %v", err)
		}

		if got != want {
			t.Errorf("got %q, want %q", got, want)
		}
	})

	t.Run("returns an error if a server doesn't respond within 10s", func(t *testing.T) {
		serverA := makeDelayedServer(11 * time.Second)
		serverB := makeDelayedServer(12 * time.Second)

		defer serverA.Close()
		defer serverB.Close()

		_, err := Racer(serverA.URL, serverB.URL)

		if err == nil {
			t.Error("expected an error but didn't get one")
		}
	})

We've made our test servers take longer than 10s to return to exercise this scenario and we are expecting Racer to return two values now, the winning URL (which we ignore in this test with _) and an error.

Note that we've also handled the error return in our original test, we're using _ for now to ensure the tests will run.

Try to run the test

./racer_test.go:37:10: assignment mismatch: 2 variables but Racer returns 1 value

Write the minimal amount of code for the test to run and check the failing test output

func Racer(a, b string) (winner string, error error) {
	select {
	case <-ping(a):
		return a, nil
	case <-ping(b):
		return b, nil
	}
}

Change the signature of Racer to return the winner and an error. Return nil for our happy cases.

The compiler will complain about your first test only looking for one value so change that line to got, _ := Racer(slowURL, fastURL), knowing that we should check we don't get an error in our happy scenario.

If you run it now after 11 seconds it will fail.

--- FAIL: TestRacer (12.00s)
    --- FAIL: TestRacer/returns_an_error_if_a_server_doesn't_respond_within_10s (12.00s)
        racer_test.go:40: expected an error but didn't get one

Write enough code to make it pass

func Racer(a, b string) (winner string, error error) {
	select {
	case <-ping(a):
		return a, nil
	case <-ping(b):
		return b, nil
	case <-time.After(10 * time.Second):
		return "", fmt.Errorf("timed out waiting for %s and %s", a, b)
	}
}

time.After is a very handy function when using select. Although it didn't happen in our case you can potentially write code that blocks forever if the channels you're listening on never return a value. time.After returns a chan (like ping) and will send a signal down it after the amount of time you define.

For us this is perfect; if a or b manage to return they win, but if we get to 10 seconds then our time.After will send a signal and we'll return an error.

Slow tests

The problem we have is that this test takes 10 seconds to run. For such a simple bit of logic, this doesn't feel great.

What we can do is make the timeout configurable. So in our test, we can have a very short timeout and then when the code is used in the real world it can be set to 10 seconds.

func Racer(a, b string, timeout time.Duration) (winner string, error error) {
	select {
	case <-ping(a):
		return a, nil
	case <-ping(b):
		return b, nil
	case <-time.After(timeout):
		return "", fmt.Errorf("timed out waiting for %s and %s", a, b)
	}
}

Our tests now won't compile because we're not supplying a timeout.

Before rushing in to add this default value to both our tests let's listen to them.

  • Do we care about the timeout in the "happy" test?

  • The requirements were explicit about the timeout.

Given this knowledge, let's do a little refactoring to be sympathetic to both our tests and the users of our code.

var tenSecondTimeout = 10 * time.Second

func Racer(a, b string) (winner string, error error) {
	return ConfigurableRacer(a, b, tenSecondTimeout)
}

func ConfigurableRacer(a, b string, timeout time.Duration) (winner string, error error) {
	select {
	case <-ping(a):
		return a, nil
	case <-ping(b):
		return b, nil
	case <-time.After(timeout):
		return "", fmt.Errorf("timed out waiting for %s and %s", a, b)
	}
}

Our users and our first test can use Racer (which uses ConfigurableRacer under the hood) and our sad path test can use ConfigurableRacer.

func TestRacer(t *testing.T) {

	t.Run("compares speeds of servers, returning the url of the fastest one", func(t *testing.T) {
		slowServer := makeDelayedServer(20 * time.Millisecond)
		fastServer := makeDelayedServer(0 * time.Millisecond)

		defer slowServer.Close()
		defer fastServer.Close()

		slowURL := slowServer.URL
		fastURL := fastServer.URL

		want := fastURL
		got, err := Racer(slowURL, fastURL)

		if err != nil {
			t.Fatalf("did not expect an error but got one %v", err)
		}

		if got != want {
			t.Errorf("got %q, want %q", got, want)
		}
	})

	t.Run("returns an error if a server doesn't respond within the specified time", func(t *testing.T) {
		server := makeDelayedServer(25 * time.Millisecond)

		defer server.Close()

		_, err := ConfigurableRacer(server.URL, server.URL, 20*time.Millisecond)

		if err == nil {
			t.Error("expected an error but didn't get one")
		}
	})
}

I added one final check on the first test to verify we don't get an error.

Wrapping up

select

  • Helps you wait on multiple channels.

  • Sometimes you'll want to include time.After in one of your cases to prevent your system blocking forever.

httptest

  • A convenient way of creating test servers so you can have reliable and controllable tests.

  • Uses the same interfaces as the "real" net/http servers which is consistent and less for you to learn.

PreviousConcurrencyNextReflection

Last updated