testing: add -shuffle and -shuffleseed to shuffle tests

[cristaloleg]

This is a revive of #10655

Motivation

Consider the following code & corresponding tests:

package pkg

// build and cache regexp, reuse between clients
var re *regexp.Regexp

type Client struct {
	// ...
}

func NewClient(pattern string) *Client {
	if re == nil {
		re = regexp.MustCompile(pattern)
	}
	return &Client{
		// ...
	}
}

func (c *Client) HasPattern(s string) error {
	if !re.MatchString(s) {
		return errors.New("meh, incorrect param")
	}
	return nil
}

package pkg

func TestFoo(t *testing.T) {
	s := NewClient("^foo.*")
	err := s.HasPattern("foo123")
	if err != nil {
		t.Errorf("expected to pass")
	}
}

func TestFooBar(t *testing.T) {
	s := NewClient("^foobar.*")
	err := s.HasPattern("foobar1123")
	if err != nil {
		t.Errorf("expected to pass")
	}
}

Those tests pass, everything looks fine, but they're order dependent. Running them in another order will fail.

To prevent such hidden and hard to debug mistakes we need to make the order of test random for each test build.

Current workarounds

1. Manual ordering of tests.
2. Boilerplate code to specify the order of tests.
3. For table-driven tests we can use a map instead of slice, ex:

func TestSomething(t *testing.T) {
	testCases := map[name]struct{
		a,b int64
		res int64
	}{
		“simple case”: {1, 2, 3},
		“less simple”: {3, 3, 23},
	}

	// due to behaviour of map test cases will be 
	// in a different order for each run
	// but this solution is limited and not suitable for test funcs
	// aka `func TestXXX(t *testing.T)`
	for name, tc := range testCases {
		t.Logf(“test: %s”, name)

		res := foo(tc.a, tc.b)
		if res != tc.res {
			t.Errorf(“want %v, got %v, res, tc.res)
		}
	}
}

Possible solution

We need to specify a test run to execute tests with a random/desired order:

1. -shuffle to run tests with a random order (used random seed may or should be printed to output depending or not on a -v flag).
2. -seed <int> to specify a user-defined seed to have an ability repeat a failed build.

Open questions

1. Should we randomize order of benchmarks and examples? (looks like not)
2. This makes test runs 'flaky' but this helps to reveal possible implementation caveats.
3. This might not happen for 1.12 'cause proposal is submitted too late.

PS. if/when it will be accepted - will be happy to work on it.

[mvdan]

Interesting problem - reminds me of go test -race a bit. I'm a bit worried that every repository will now be expected to run three test commands on its CI; go test; go test -race; go test -shuffle.

On the other hand, making go test randomize the order could make its output non-deterministic.

I wonder if this is something we could enable when one runs go test ./.... That is, since a package's test output and results are printed all at once, we could run the tests in random order and then sort the results before printing them. That would be ideal in my opinion, since every existing user would benefit and we wouldn't need an extra flag.

That suggestion wouldn't apply to go test or go test -v, since those do need to run the tests in order - the results are printed as they happen.

[cristaloleg]

That is, since a package's test output and results are printed all at once

Might be a problem with #24929 😉

and then sort the results before printing them

But how we can know an order that triggered a test run failure? Sort only success run?

[mvdan]

Might be a problem with #24929

Oh, I thought that proposal would affect go test -v, not go test -v ./.... If it affects both, then you're right.

But how we can know an order that triggered a test run failure? Sort only success run?

Potentially. Or one could take the usual approach like with any other test flake - run the tests over and over again with -count or stress until we find the failure again or we're convinced that it's fixed.

[rsc]

More command-line flags seems problematic.
What if instead we added a method on testing.M so that TestMain could call m.Shuffle(time.Now().UnixNano()) in tests that wanted to opt in to randomized execution?
Those tests could also define their own seed flags instead of having to add one to all tests.

[rsc]

@aclements and @dr2chase and I were talking about the problem of benchmarks being sensitive to earlier benchmarks, which this would partly address too.

What if testing.M exposed a slice of test identifiers (strings or some object with a Name method) and a slice of benchmark identifiers, and you could reorder or filter or duplicate those, by rewriting the slices, before calling m.Run? We could still give an m.Shuffle helper for that common operation.

[cristaloleg]

Quick draft of what you're saying. Did I get it right?

package testing

type M struct { ... }

type Tester interface {
	Name() string
}

func (m *M) Tests() []Tester { ... }
func (m *M) Benchmarks() []Tester { ... }

func (m *M) Shuffle(tests []Tester) []Tester { ... }

[rsc]

@cristaloleg No, it would have to be fields in M (or have setters too), so that they could be rewritten before calling m.Run.

[aclements]

@aclements and @dr2chase and I were talking about the problem of benchmarks being sensitive to earlier benchmarks, which this would partly address too.

For benchmarks, this is specifically in the context of iterated runs. Ideally, I'd want a way to take, say, N runs of a set of M benchmarks and shuffle the whole N*M sequence of runs. This would get much more robust benchmark results when benchmarks aren't totally isolated (which they never are, at least because of GC state).

For reproducibility of order in the benchmark case, you could imagine printing the sequence randomization seed as one of the key/value headers in the output.

[ianlancetaylor]

Retitled per suggestion above.

[andybons]

Seems like we'd like to do this, but we're blocked on an API definition. Is that correct?

[cristaloleg]

@andybons looks so