test-cluster-shared-leak failure on windows

[Fishrock123]

not ok 28 test-cluster-shared-leak.js
# events.js:142
# throw er; // Unhandled 'error' event
# ^
# 
# Error: write EPIPE
# at exports._errnoException (util.js:873:11)
# at ChildProcess.target._send (internal/child_process.js:607:18)
# at ChildProcess.target.send (internal/child_process.js:508:19)
# at sendHelper (cluster.js:701:8)
# at send (cluster.js:505:5)
# at cluster.js:480:7
# at SharedHandle.add (cluster.js:83:3)
# at queryServer (cluster.js:472:12)
# at Worker.onmessage (cluster.js:428:7)
# at ChildProcess.<anonymous> (cluster.js:715:8)

https://ci.nodejs.org/job/node-test-binary-windows/91/RUN_SUBSET=3,VS_VERSION=vs2015,label=win2012r2/tapTestReport/test.tap-28/

(Run against #3954, unrelated.)

cc @Trott

[Trott]

Failed again here:

https://ci.nodejs.org/job/node-test-binary-windows/176/RUN_SUBSET=3,VS_VERSION=vs2015,label=win2012r2/console

Definitely one to add to the flaky tests list...

[Trott]

Stress test confirming flakiness: https://ci.nodejs.org/job/node-stress-single-test/133/nodes=win2012r2/console

[Trott]

The degree of flakiness apparently depends on which machine you end up with. (There are slow ones and there are fast ones.) Check out the massively different results of these identical stress test runs. The only difference is likely which Windows host the test ended up getting assigned to.

• https://ci.nodejs.org/job/node-stress-single-test/135/nodes=win2012r2/console
• https://ci.nodejs.org/job/node-stress-single-test/136/nodes=win2012r2/console

[Trott]

Pinging @nodejs/build because someone with more knowledge of the build infrastructure might be able to provide some useful insight on the previous comment.

The parameters for the two stress tests are identical and they checkout the same commit hash for testing, but they have nearly opposite results. The first has no failures (as of this writing--it's only about halfway done) after thousands of runs while the second has just 7 successes and hundreds of failures as of this writing.

Possibilities that leap to mind for me:

• The flakiness is related to memory, CPU, or other resource.
• The flakiness is related to some other configuration difference on either the host or the network where the host resides.

I suspect the correct fix is still to rewrite the test to accommodate what the issue is, but it seems like it would be useful to know what the issue is.

[Trott]

Possible fix: #4173