testify
testify copied to clipboard
stretchr
Running Parallel Tests in Suite
When running multiple tests in a suite which have been set to suite.t().Parallel(), tests will pass even if they should fail.
That's true, but as it's currently structured it isn't possible to run suite tests in parallel. This is because before and after each suite test SetupTest and TearDownTest is run.
Sure
On Sunday, November 1, 2015, Ernesto Jiménez [email protected] wrote:
Related: #52 https://github.com/stretchr/testify/issues/52
— Reply to this email directly or view it on GitHub https://github.com/stretchr/testify/issues/187#issuecomment-152894963.
Reviving this issue. Presumably it should be possible to run suite tests in parallel.
The problem with parallelism here is that the suite members used in tests have a shared single instance for all. Tests running in parallel could potentially access these members at the same time, causing them to "collide" with each other via these members.
However, my suggestion is that testify will add a special Parallel() flag to the suite. This new flag will instantiate new instance members for every new Parallel() test, calling Setup/TearDownTest() as well.
This will of course require extra work from the test developer, to make sure Setup/TearDownTest() are suitable for being called with/without parallel (As some tests will be run in parallel, and some don't).
Thoughts?
I just whipped up a slightly different approach to running tests in parallel: https://gist.github.com/dansimau/128826e692d7834eb594bb7fd41d2926
Similar to #369, it introduces a RunParallel, except that it uses reflection to create a new instance of the suite for each test run.
Because a new suite is initialised before every test, it means tests can't share data on the suite struct. This is a change in behaviour, though I imagine you could change the patch to make a copy of the suite struct instead of initialising a new one (that way tests could still share data and setup/teardown code using SetupSuite).
I actually prefer the share-nothing approach though; it feels like a bad practice to modify data on the suite struct while tests are running, especially now in parallel. I can't see why tests should ever share data (outside what is set up in SetupTest); but maybe I'm missing a use case.
Hi! Any updates on this? Are you planning to do work on this feature request?
One solution:
package main_test
import (
"testing"
"time"
"github.com/stretchr/testify/assert"
"github.com/ysmood/got"
)
func Test(t *testing.T) {
got.Each(t, beforeEach)
}
func beforeEach(t *testing.T) Suite {
t.Parallel()
return Suite{assert.New(t)}
}
type Suite struct { // struct that holds subtests
*assert.Assertions
}
func (s Suite) A() { // test case A
time.Sleep(time.Second)
s.Equal(1, 1)
}
func (s Suite) B() { // test case B
time.Sleep(time.Second)
s.Equal(2, 1)
}
As you can see, test A is not affected:
$ go test
--- FAIL: Test (0.00s)
--- FAIL: Test/B (1.00s)
main_test.go:31:
Error Trace: main_test.go:31
value.go:475
value.go:336
each.go:42
value.go:564
asm_amd64.s:20
Error: Not equal:
expected: 2
actual : 1
Test: Test/B
FAIL
exit status 1
FAIL goplay/default 1.254s
Reviving this issue. Presumably it should be possible to run suite tests in parallel.
The problem with parallelism here is that the suite members used in tests have a shared single instance for all. Tests running in parallel could potentially access these members at the same time, causing them to "collide" with each other via these members.
However, my suggestion is that testify will add a special
Parallel()flag to the suite. This new flag will instantiate new instance members for every newParallel()test, callingSetup/TearDownTest()as well.This will of course require extra work from the test developer, to make sure
Setup/TearDownTest()are suitable for being called with/without parallel (As some tests will be run in parallel, and some don't).Thoughts?
@oryband can we implement it now?
func (s *mySuiteA) TestA2() { // test case A
t := s.T()
t.Parallel()
assert.Equal(t, 3, 1)
}
cache the *testing.T pointer can solve the problem.
This issue is caused by the following code snippet
test := testing.InternalTest{
Name: method.Name,
F: func(t *testing.T) {
// ...
defer func() {
// ...
suite.SetT(parentT)
// ...
}()
// ...
method.Func.Call([]reflect.Value{reflect.ValueOf(suite)})
}
}
... and when run test in parallel, defer function executed from TestA even if the body of TestB not be executed, T set to the parent T so the child test failed but reported as passed. Because the fail happened on the parent T instead of the child T.
func (s *mySuiteA) TestA2() { // test case A t := s.T() t.Parallel() assert.Equal(t, 3, 1) }cache the
*testing.Tpointer can solve the problem.This issue is caused by the following code snippet
test := testing.InternalTest{ Name: method.Name, F: func(t *testing.T) { // ... defer func() { // ... suite.SetT(parentT) // ... }() // ... method.Func.Call([]reflect.Value{reflect.ValueOf(suite)}) } }... and when run test in parallel, defer function executed from
TestAeven if the body ofTestBnot be executed,Tset to the parent T so the child test failed but reported as passed. Because the fail happened on the parent T instead of the child T.
Well, it is not a solution because as tests grow the first place cache may cache a wrong pointer also.
Then I'd suggest we pass the inner T if the test method meet the signature, i.e.,
func (s *mySuite) TestT(t *testing.T)
being able to force parallelism for suites would be helpful for us as well, we are looking to ditch ginkgo but want to make sure we can force random execution order to find flaky tests and keep our overall suite runtime low
~Implementing SetupTestSuite following way should work as expected:~
No. The same problem is described in the above thread. Missed.
// SetupTest will run before each test in the suite.
func (s YourSuite) SetupTest() {
s.T().Parallel()
}
Here is a solution we've used. Our use case is for e2e tests that we want to run in parallel. All of them are defined in the same suite.
type E2ETestSuite struct {
suite.Suite
ts map[string]*testing.T // Map of test names > *testing.T
}
func (suite *E2ETestSuite) BeforeTest(_, testName string) {
t := suite.T()
if suite.ts == nil {
suite.ts = make(map[string]*testing.T, 1)
}
suite.ts[testName] = t
t.Parallel()
}
// T() overrides suite.Suite.T() with a way to find the proper *testing.T
// for the current test.
// This relies on `BeforeTest` storing the *testing.T pointers in a map
// before marking them parallel.
// This is a huge hack to make parallel testing work until
// https://github.com/stretchr/testify/issues/187 is fixed.
// There is still a small race:
// 1. test 1 calls SetT()
// 2. test 1 calls BeforeTest() with its own T
// 3. test 1 is marked as parallel and starts executing
// 4. test 2 calls SetT()
// 5. test 1 completes and calls SetT() to reset to the parent T
// 6. test 2 calls BeforeTest() with its parent T instead of its own
// The time between 4. & 6. is extremely low, enough that this should be really rare on our e2e tests.
func (suite *E2ETestSuite) T() *testing.T {
// Try to find in the call stack a method name that is stored in `ts` (the test method).
for i := 1; ; i++ {
pc, _, _, ok := runtime.Caller(i)
if !ok {
break
}
// Example rawFuncName:
// github.com/foo/bar/tests/e2e.(*E2ETestSuite).MyTest
rawFuncName := runtime.FuncForPC(pc).Name()
splittedFuncName := strings.Split(rawFuncName, ".")
funcName := splittedFuncName[len(splittedFuncName)-1]
t := suite.ts[funcName]
if t != nil {
return t
}
}
// Fallback to the globally stored Suite.T()
return suite.Suite.T()
}
I've opened a PR to try to allow Test... methods to receive their own t as a parameter: https://github.com/stretchr/testify/pull/1322 that partially allows to run suite tests in parallel.
This is the approach suggested by @tisonkun in https://github.com/stretchr/testify/issues/187#issuecomment-869029721
#1109 is the only good fix for this, but is a breaking change. It might be that a testify/v2/suite is made one day to address this, but for all of testify v1; do not use parallel tests or subtests with suite.
