0004-dunder-runner.md 6.6 KB

  • Feature Name: Rename __salt__ to __runner__ in the Runner functions
  • Start Date: 2019-01-02
  • RFC PR:
  • Salt Issue:

Summary

As this 4 years old TODO states, we are indeed overloading __salt__ too much. In the recent releases, we have added the salt.cmd Runner function to be able to invoke Execution Functions on the Master side, however the usage is counter-intuitive and often misleading. Additionally, this prevents a drawbacks when it comes to whatever template rendering on the Master side.

Motivation

One of the main arguments is solving that 4-years old TODO mentioned in the Summary. I can potentially understand what were the reasons back then, when the __salt__ dunder has been added to be available on the Master side, however I believe these reasons are now obsolete: Salt has evolved and still has much to grow in terms of glueing different parts together; for a very long time it felt like the Execution Modules, Runners, et. al., are separate entities, completely decoupled from the rest. But more recently, the modern Salt, feels much more like a whole, and not separate conglomerates. Renaming the __salt__ object to, e.g., __runner__, and re-mapping __salt__ to the Execution Functions on the Minion side, would be a step forward in this direction.

Historically, from a Runner, we didn't have native support for accessing arbitrary Execution Functions. In the recent releases, we have a new Runner salt.cmd as a workaround, however the usage is not particularly straight forward, e.g., __salt__['salt.cmd']('system.get_system_date'). Renaming the existing __salt__ to __runner__ and re-mapping __salt__ to the Minion modules, the usage becomes: __salt__['system.get_system_date'] just like on the Minion side, by re-using the existing code, and executing it on the Master.

This equally has an impact on the template rendering pipeline, for anything that is rendered on the Master side (anything SLS rendered on the Master). A good example is Reactor SLS files: when I firstly used in a Reactor salt.<module>.<function> I wasn't sure whether I should call an Execution Function or a Runner Function. Few years later, I still need to double check that. For this reason, overloading __salt__ on the Master side becomes misleading, while having a separate object would make it self-explanatory and expand the capabilities.

Besides all of these, the biggest issue I hit was in the context of network automation: without diving into too many details beyond the current scope, I managed to have network devices managed from the Master only (i.e., without any Minions or Proxy Minions). The implementation is based on a simple Runner, it works very well, except when trying to invoke States. For more context, when working with network devices, we usually aren't able to install the regular Salt Minion on the devices we target, and typically we use the Proxy Minion which is installed on whatever server. In other words, the State system doesn't care where it's executed from (i.e., from what physical machine). For this reason, a "proxyless" implementation is literally blocked by a simple object naming.

Design

The basic implementation is as simple as renaming the __salt__ dunder to __runner__ into the Lazy Loader. To resolve also the second part of the problem, have __salt__ point to the Minion modules. In code, this change is as simple as:

 diff --git a/salt/loader.py b/salt/loader.py
 index e046b449a5..b4bd45472b 100644
 --- a/salt/loader.py
 +++ b/salt/loader.py
 @@ -915,7 +915,7 @@ def call(fun, **kwargs):
      return funcs[fun](*args)


 -def runner(opts, utils=None, context=None, whitelist=None):
 +def runner(opts, functions=None, utils=None, context=None, whitelist=None, proxy=None):
      '''
      Directly call a function inside a loader directory
      '''
 @@ -927,16 +927,18 @@ def runner(opts, utils=None, context=None, whitelist=None, proxy=None):
          _module_dirs(opts, 'runners', 'runner', ext_type_dirs='runner_dirs'),
          opts,
          tag='runners',
 -        pack={'__utils__': utils, '__context__': context},
 +        pack={'__utils__': utils, '__context__': context, '__proxy__': proxy},
          whitelist=whitelist,
      )
 +    if functions is None:
 +        functions = minion_mods(opts,
 +                                utils=utils,
 +                                context=context,
 +                                whitelist=whitelist,
 +                                proxy=proxy)
 -    ret.pack['__salt__'] = ret
 +    ret.pack['__salt__'] = functions
 +    ret.pack['__runner__'] = ret
      return ret

While testing, I've noticed that the time to load the Minion functions is quite considerable (a few seconds), so I thought about adding a new option: runner_load_minion_mods defaulting to False which can be used to request whether the Minion modules should always available in the Runners.

A more extensive change can be followed at: https://github.com/saltstack/salt/compare/develop...mirceaulinic:mircea/dunder-runner?expand=1#diff-f6777cb7fca1276c97cd4b4a6b9f085a which should cover most of the core changes required for this. Another step required together with the change linked above is replacing __salt__ with __runner__ in all the Runner modules, e.g., sed -i 's/__salt__/__runner__/g' salt/runners/.

Even though in code that's a very simple change, it is breaking backwards compatibility and the users would need to execute the above rename command in their Runner extension directory salt://_runners/.

To help with this transition, for a couple of major releases, we can pack both __salt__ and __runner__:

ret.pack['__salt__'] = ret
ret.pack['__runner__'] = ret

This wouldn't change anything, but allow the users to prepare in advance the handover to the new variable name. At the same time, we may add a warning to mention that they should no longer use __salt__ to invoke Runner functions from a Runner, and switch to using __runner__. I would be tempted to suggest adding these two changes (the duplicate packing and the warning) straight away into the develop branch to be included in the very next major release, Neon, and targeting the whole change for Magnesium.

Nevertheless, I am going to handle the documentation for these changes before and after the releases mentioned.

Drawbacks

The drawback I'm seeing here is the breaking change for the users, and I have elaborated under the Design paragraph on how to approach this.