Flaky TestForDeps/loop-matrix test

[Ikke]

• Task version: 3.39.x
• Operating system: Alpine Linux
• Experiments enabled: None

I'm maintaining the Alpine Linux package for task. While upgrading to 3.39.1 and 3.39.2, the TestForDeps/loop-matrix failed occasionally with:

=== RUN   TestForDeps/loop-matrix
    task_test.go:2495: 
        	Error Trace:	github.com/go-task/task/v3/task_test.go:2495
        	Error:      	"linux/arm64\ndarwin/amd64\ndarwin/arm64\nwindows/amd64\nlinux/amd64windows/arm64\n\n" does not contain "linux/amd64\n"
        	Test:       	TestForDeps/loop-matrix

Retrying the build fixes it. This may indicate some kind of race condition. The failure did happen on somewhat slower arches (armv7, riscv64).

[pd93]

Strange, the test uses a SyncBuffer that should only be able to write one thing at a time, so it's odd that you're getting linux/amd64windows/arm64\n\n in your output instead of linux/amd64\nwindows/arm64\n. I suppose new newline is written separately somewhere. I'm guessing the quick fix is to remove the \n from the test, but this feels like sweeping the issue under the rug.

Do you know if this can happen in versions prior to 3.39.1 or was it introduced in that version?

[Ikke]

On 3.38.0, running test with -count 100:

    task_test.go:2399:
                Error Trace:    /home/buildozer/aports/community/go-task/src/task-3.38.0/task_test.go:2399
                Error:          "21\n\n3\n" does not contain "2\n"
                Test:           TestForDeps/loop-different-tasks
--- FAIL: TestForDeps (0.08s)

On 3.37.0 additionally:

=== RUN   TestForDeps/loop-explicit
    task_test.go:2326:
                Error Trace:    /home/buildozer/aports/community/go-task/src/task-3.37.0/task_test.go:2326
                Error:          "ba\n\nc\n" does not contain "b\n"
                Test:           TestForDeps/loop-explicit

Different subtests, but similar issues. So not introduced in 3.39

[pd93]

Thanks for checking. I will investigate

[pd93]

Just leaving this as a note to myself. This just happened in our CI https://github.com/go-task/task/actions/runs/10962583048/job/30444328303

[pbitty]

I have a clue as to why this might be happening. I added the following line inside SyncBuffer.Write() (inside the lock):

fmt.Fprintln(os.Stderr, "len:", len(p))

When I run the tests, I get this output:

go test -run TestForDeps/loop-explicit -count 1
len: 1
len: 1
len: 1
len: 1
len: 1
len: 1
PASS
ok  	github.com/go-task/task/v3	0.358s

This implies that each call to Write() contains a single character, so the lock won't prevent the task outputs from being interleaved, and explains why sometimes we get something like ba\n\nc\n.