proposal: testing: ability to tell if test is running in parallel

[jim-minter]

Proposal Details

Unit tests can indicate their ability to be run in parallel by making the opt-in call t.Parallel().

However sometimes a test might opt in but then mistakenly change some global state which could race with something else. In this case, the parallel test should actually be a serial one. In a large codebase it can be tricky to protect against this happening.

For example,

func TestShouldNotBeParallel(t *testing.T) {
    t.Parallel()
    // ...
    changeGlobalState(t) // a mistake, given above t.Parallel()
    // ...
}

func changeGlobalState(t *testing.T) {
    // proposal: could we check `t` here to determine that we're running in parallel when we shouldn't be, and error out?
    if t.IsRunningInParallel() {
        t.Fatal("changeGlobalState called from test with t.Parallel()")
    }

    // ...
}

The proposal here is a function like t.IsRunningInParallel() that could return true in the case that the current test is running in a parallel context. I don't think there's a way to check this currently?

[gabyhelp]

Related Issues and Documentation

• proposal: testing: add a `ParallelCount` func that reports the `go test -p` flag value #50719 (closed)
• proposal: testing: run tests in parallel by default #21214 (closed)
• proposal: add testing.T.RunParallel for simplifying data race detection in tests #21847 (closed)
• proposal: cmd/test: add flag to run tests in parallel by default #24335 (closed)
• testing: make -parallel=-1 force sequential runs

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[seankhliao]

This feels too niche, nobody is going to think to add something like this until they actually run into a problem, and at that point, this doesn't help you track down the issue.

[abhinav]

I've definitely felt a need for this when writing test utilities that modify global state (e.g. changing a global logger temporarily to capture output, or stubbling a global value). Any such utility will misbehave if called from parallel tests, and there's no way for the utility to prevent this misuse.
Right now, such misuse is caught by go test -race if the global state being manipulated happens to hit that. But if the API for changing the global state is safe for concurrent use (e.g. Zap's ReplaceGlobals), then you're out of luck and the tests that are misusing the non-parallel test helper will just have a hard-to-debug intermittent test failure.

As an stdlib data point in favor of exposing this information to external libraries:
testing.T.Setenv also does not want to be called from parallel tests, and it's able to guard against it by inspecting the internal state of testing.T.

Naming-wise: I suggest IsParallel since Parallel is already taken.

[seankhliao]

if you can modify a global value, then you can also hold a lock or some other token protecting the global value, which you can protect properly rather than doing as hoc checks of if a test is running in parallel

[Jorropo]

As a workaround you can do:

func TestShouldNotBeParallel(t *testing.T) {
    t.Parallel()
    // ...
    changeGlobalState(t) // a mistake, given above t.Parallel()
    // ...
}

var noParallel bool

func changeGlobalState(t *testing.T) {
    noParallel = !noParallel // trips the race detector

    // ...
}

[abhinav]

That's true, @seankhliao, except that's not always an option.
TB.Setenv and the incoming TB.Chdir are examples of this.
They both rely on being able to check whether the test is being run with t.Parallel to prevent misuse.
It's currently not possible for a helper that lives outside the standard library to implement similar functionality.

Edit:
That's a good workaround, @Jorropo.

[adonovan]

The workaround is a good one, but it still relies on the race detector. I don't think the request to publish t.CheckParallel is unreasonable as it would give earlier notification when a (known) concurrency-unsafe test helper is accidentally used by a parallel test.

[seankhliao]

I was thinking more along the lines of:

var globalResource sync.Mutex

func TestMyThing(t *testing.T) {
    // or Lock() if you want to just sequence the tests
    gotLock := globalResource.TryLock()
    if !gotLock {
        t.Fatal("...")
    }
    defer globalResource.Unlock()

    // ...
}

It's only wrong for tests that mutate the same set of resources to run in parallel, not for anything else to run in parallel.

[adonovan]

TryLock only probabilistically reports the problem, with low probability because the critical section is short. By contrast, T.checkParallel reliably reports the problem.

The analogy with t.Chdir is instructive: Chdir is a short-lived operation, but it has long-lasting global effects; the risk is not that two tests call Chdir at the same time, but that any test calls it while another test is running.

[seankhliao]

A successful TryLock ends with you holding the lock for the entire duration of the test, covering any change and restore operation you make on the global resource. Any other test using the same lock would fail if run in parallel.

[Jorropo]

This could happen:

• Test A start
  • Test A calls parallel
• Test B start
  • Test B calls parallel
  • Test B TryLock
  • Test B do it's thing
  • Test B release lock
• Test A TryLock
  • ...

Which should have been caught but was not.