Fix flaky parallel/test-child-process-exec-kill-throws

[gibfahn]

• Version: master
• Platform: Windows
• Subsystem: test

Creating this issue to document that we need to work out and fix the issue. Will mark as flaky in a separate PR. EDIT: #12054

See discussion in #11038.

I think there might be a genuine bug with kill() on Windows, so maybe marking it as flaky will keep it on the radar.

cc/ @nodejs/platform-windows

[Trott]

PR to fix coming in a few minutes.

The child process isn't failing because of exceeding maxBuffer. It's failing because it's trying to load internal/child_process but it is not being passed the expose-internals flag. Only the parent gets that flag.

Fixed by moving the require() for that module to where the child process won't execute it.

I'm running a stress test now to make sure these changes actually get rid of the flakiness. (UPDATE: Looks good. https://ci.nodejs.org/job/node-stress-single-test/1146/nodes=win-vs2015/console)

As for how this bug triggers flakiness on Windows (assuming the flakiness on Windows is related to it), I'm not 100% sure, but I do see this: On Windows, it's getting UV_ESRCH a lot, indicating the process is already exited by the time ChildProcess._handle.kill() is called. (This is why .killed is never set. It is only set if the return code is 0 but it's often getting UV_ESRCH on Windows.) Interestingly, @sam-github brought up ESRCH in a comments on the original PR. Maybe it would be clear and obvious to him what the problem is here? Seems like a race condition, of course.

[Trott]

Here it is: #12111

[gibfahn]

Cross-posting from #12111 (comment)

@Trott this seems like it could be related to #11052, also a child-process-exec test that has issues on windows.