ansible: env vars set in vars plugin not present in play for localhost

[manicminer]

This might sound strange, and may even be an unexpected behavior of Ansible to begin with.

I have a vars plugin which sets some environment variables in-process. When I use the stock linear strategy, these vars are visible for plays running on localhost. When I use Mitogen, the vars are not visible.

Reproducing

# vars_plugins/test.py

from __future__ import (absolute_import, division, print_function)
__metaclass__ = type

from ansible.plugins.vars import BaseVarsPlugin
import os

class VarsModule(BaseVarsPlugin):
    def __init__(self, *args):
        super(VarsModule, self).__init__(*args)
        os.environ['FOO'] = 'bar'

    def get_vars(self, loader, path, entities, cache=True):
        super(VarsModule, self).get_vars(loader, path, entities)
        return {}

---
# playbook.yml

- hosts: localhost
  strategy: linear
  connection: local
  tasks:
    - shell: "env | grep FOO"

- hosts: localhost
  connection: local
  tasks:
    - shell: "env | grep FOO"

I fully admit this is abuse of a vars plugin, but I use it to set context-specific AWS_* variables, so that the correct AWS account is chosen based on some runtime logic.

Look forward to any feedback on this scenario! Is this sane? Fixable in Mitogen or antithetical to Mitogen's design? Thanks!

[manicminer]

I just noticed that the above doesn't reproduce, but this does:

---

- hosts: localhost
  strategy: mitogen_linear
  connection: local
  tasks:
    - debug:
        msg: noop

- hosts: localhost
  strategy: linear
  connection: local
  tasks:
    - shell: "env | grep FOO"

- hosts: localhost
  strategy: mitogen_linear
  connection: local
  tasks:
    - shell: "env | grep FOO"

When I remove the first play, the env var is visible to the second shell task.

[dw]

Funsies! So I guess this is because the vars plugin executes in the top-level process, and in vanilla Ansible that process is repeatedly forked for every new task. Environment variable modifications will make it into the forked children, which local actions execute as a subprocess of.

With Mitogen the local context is started once, as a subprocess of a fork of the top-level process, and that fork happens long before any vars evaluation would have a chance to run.

This is reminiscent of another annoying bug I've only seen once -- #122

---

As for a fix, we must preserve vanilla's semantics everywhere so this should be made to work.

The environment variables will propagate to new WorkerProcesses running under Mitogen just like they did in vanilla, so there needs to be a mechanism to ship them to the long-lived local worker.

Give me a few hours to think about this one! There is an obvious "special case localhost" option, but I'd prefer to avoid that.

Thanks for reporting!

[manicminer]

Wow! Can confirm this fixes the underlying issue. Great stuff, thanks!