test: fix child-process-fork-regr-gh-2847

[santigimeno]

The test would sometimes time out on some platforms. Take the initial
version of the test and instead of sending 100 handles, just send 2.
It still fails when run with the code before the change was introduced
and passes afterwards.

[santigimeno]

I've checked that it passes when run on 36b969f (where the test was initially introduced) and fails with the previous commit 6ac79bc in OS X and Debian Jessie 64.

[Trott]

This one has been failing a lot on pi2-raspbian-wheezy in CI a lot lately, so here's a stress test with this fix:

• https://ci.nodejs.org/job/node-stress-single-test/428/nodes=pi2-raspbian-wheezy/console

And here's a stress test using current master:

• https://ci.nodejs.org/job/node-stress-single-test/427/nodes=pi2-raspbian-wheezy/console

(CI is closed off from public view for another couple of days until the security release of node is out, but I'll summarize the results here.)

[Trott]

A change that just landed on master today was to mark this test as flaky in test/parallel/parallel.status so this PR should probably update that file too.

[Trott]

I stopped the master branch stress test because the test failed 5 times out of 9 runs. So there's the unsurprising confirmation that the test is highly flaky in current master.

With this fix in place, the stress test has run 250 times without a single failure. So there's confirmation that this fix improves the situation.

[Trott]

@santigimeno Once you push the change to the parallel.status file, I (or someone else) will run this through a full CI. (Nice work, as always!)

[santigimeno]

@Trott PR updated. Thanks!

[Trott]

CI: https://ci.nodejs.org/job/node-test-pull-request/1574/

[Trott]

LGTM, but given the extent of the rewrite, I'd like to get a second from at least one of:

• @indutny (who wrote the test)
• @mhdawson (who made the test AIX-friendly)
• @mscdex (who made the test not-flaky on Windows)

[mscdex]

Has this been stress tested on Windows? Also, I know this commit didn't change this part, but I'd be a bit concerned about the low timeout value (especially on pi1).

[Trott]

Stress tests on:

• pi1-raspbian-wheezy
• win-vs2013

[bnoordhuis]

I don't understand this comment. Where does the test call process.disconnect()?

[Trott]

@bnoordhuis The original version of that comment landed in 36b969f and was authored by @indutny. I think it might be referring to target.disconnect() being called in lib/internal/child_process.js in channel.onread() but I'm not actually sure.

[santigimeno]

Yes, what @Trott says. It was in the original version of the test. I can try to find out the reason

[mscdex]

It looks like the pi1 stress test ran 0 times?

[Trott]

@mscdex That's what happens when I tell the test runner to run the test 000 times when I mean to type 999. Guess I better try that again. Here's hoping no typos this time: https://ci.nodejs.org/job/node-stress-single-test/442/nodes=pi1-raspbian-wheezy/console

[mhdawson]

I'm ok with the change provided it still catches the original problem and it passes on AIX. CI run on AIX here: https://ci.nodejs.org/job/node-test-commit-aix/37/ (unfortunately is does take a while to run)

[mscdex]

LGTM

[Trott]

I confirmed that the test still fails (as expected) in Node.js v3.3.1 (with ES6-isms removed from common.js) and passes in Node.js v4.0.0. (This is a bit redundant because @santigimeno already confirmed that it fails with the fix commit backed out.)

LGTM but awaiting AIX results...

[mhdawson]

lgtm tests still passs on AIX.

[Trott]

Landed in 4e4b260

Feels good to get that one off the flaky list...Thanks, @santigimeno !