Interesting. I globally ok with the patch.

Could be good to add a test in DefaultTest.py.
You could create a temporary file in with a dummy worker class in it, add this file directory into python loading path and try loading it. That should cover you new case in your patch.

I'm curious to know what's your use case? :)

I'm curious to know what's your use case? :)

I have a python library that has a pretty heavy initialization overhead, because it loads yaml configs and runs plugins to discover facts about nodes (such as the BMC address for a node). The (existing) program flow is library => clustershell => exec the library, potentially resulting in two initializations. Using a native worker allows a much more direct path.

I don't expect that my worker would be desired upstream, and distributing it with the library makes it easier to iterate features without having to worry about maintaining cross-version compatibility between the worker and the library.

Could be good to add a test in DefaultTest.py.

I'll admit I'm not familiar with the test suite. I'll give it a try and look into adding a test. Thanks

Instead of looking for a '.' to indicate that the module is 'fully qualified', should we instead just manually map 'short' known ClusterShell workers into their fully qualified module name and then attempt to import whatever is left?

    if workername in [ 'exec', 'pdsh', 'rsh', 'ssh', 'worker' ]:
        modname = "ClusterShell.Worker.%s" % workername.capitalize()
    else:
        modname = workername

If you don't like the idea of keeping a list of in-tree workers, we could detect valid workers at runtime instead (with slightly more overhead).

I was just trying to rationalize if it made sense to require out-of-tree workers to be in a subpackage, or if a top-level module could be a valid worker. I don't see any reason that it couldn't.

Thanks for this update. Thanks also for adding the tests.

I'm not fond of explicitly listing the in-core supported modules. The idea was that it could be possible, as an addons, to add new Workers just adding a new class in this directory (you could if you need to). So, they need to be able to work, just adding the file.

I'm thinking if we could not just first try to import like it is a ClusterShell modules, and if it fails, try importing using a global module name?

We probably also add somewhere in the doc that a fully qualified module name is also now supported.

Except for the python3 incompat, I'm OK with the patch.

Please update the commit message subject to have the same kind of pattern we use.
like:

Defaults: xxxx

Probably good to add a Signed-Off line too