child_process: stdout/stderr flaky on CentOS 5

[Trott]

This is currently (but hopefully not for long--I have a PR to fix it) the contents of test/parallel/test-process-argv-0.js:

'use strict';
var util = require('util');
var path = require('path');
var assert = require('assert');
var spawn = require('child_process').spawn;
var common = require('../common');

console.error('argv=%j', process.argv);
console.error('exec=%j', process.execPath);

if (process.argv[2] !== 'child') {
  var child = spawn(process.execPath, [__filename, 'child'], {
    cwd: path.dirname(process.execPath)
  });

  var childArgv0 = '';
  var childErr = '';
  child.stdout.on('data', function(chunk) {
    childArgv0 += chunk;
  });
  child.stderr.on('data', function(chunk) {
    childErr += chunk;
  });
  child.on('exit', function() {
    console.error('CHILD: %s', childErr.trim().split('\n').join('\nCHILD: '));
    assert.equal(childArgv0, process.execPath);
  });
}
else {
  process.stdout.write(process.argv[0]);
}

From time to time on CentOS 5 only, stderr and stdout from the child process stop working after the first line is written to stderr.

See, for example this centos5-32 test result or this centos5-64 test result.

See further discussion at #2541.

CentOS Project will support CentOS 5 until March 31, 2017.

[mscdex]

Shouldn't the test be listening for the child's close event instead of its exit event to ensure that all output is received?

[Trott]

If I'm understanding you (@mscdex) correctly, you're saying that there's no guarantee that exit on the child process will fire after the data events on the stdio streams? And when it doesn't, this test blows up? That this issue is (was) a bug in the old test code and not in node?

And if so, do you feel confident about that and closing this issue is in order? Or would you prefer someone else weigh in?

[mscdex]

exit can definitely fire before all output from a child process is received. close was added a long time ago to combat this. I think the commit you pushed to fix this effectively does the same thing though, it's just that the assertion will happen a little bit later than the child process's close event.

So yes, my guess is that changing exit to close would have fixed the test and it wasn't a bug in node/iojs. However, this can be closed now since your fix will work too.

[Trott]

Awesome. Closing because this isn't a bug in node.