Open Stack

Yuriy Taraday <yorik.sar at gmail.com> wrote:

> On Fri Feb 20 2015 at 9:14:30 PM Joshua Harlow <harlowja at outlook.com> wrote:
> > This feels like something we could do in the service manager base class,
> > maybe by adding a "post fork" hook or something.
> 
> +1 to that.
> 
> I think it'd be nice to have the service __init__() maybe be something like:
> 
>   def __init__(self, threads=1000, prefork_callbacks=None,
>                postfork_callbacks=None):
>      self.postfork_callbacks = postfork_callbacks or []
>      self.prefork_callbacks = prefork_callbacks or []
>      # always ensure we are closing any left-open fds last...
>      self.prefork_callbacks.append(self._close_descriptors)
>      ...
> 
> (you must've meant postfork_callbacks.append)
> 
> Note that multiprocessing module already have `multiprocessing.util.register_after_fork` method that allows to register callback that will be called every time a Process object is run. If we remove explicit use of `os.fork` in oslo.service (replace it with Process class) we'll be able to specify any after-fork callbacks in libraries that they need. 
> For example, EngineFacade could register `pool.dispose()` callback there (it should have some proper finalization logic though).

+1 to use Process and the callback system for required initialization steps
and so forth, however I don’t know that an oslo lib should silently register
global events on the assumption of how its constructs are to be used. 

I think whatever Oslo library is responsible for initiating the Process/fork
should be where it ensures that resources from other Oslo libraries are set
up correctly. So oslo.service might register its own event handler with
oslo.db such that it gets notified of new database engines so that it can
associate a disposal with it; it would do something similar for
oslo.messaging and other systems that use file handles.   The end 
result might be that it uses register_after_fork(), but the point is that 
oslo.db.sqlalchemy.create_engine doesn’t do this; it lets oslo.service
apply a hook so that oslo.service can do it on behalf of oslo.db.

So, instead of oslo.service cutting through and closing out the file
descriptors from underneath other oslo libraries that opened them, we set up
communication channels between oslo libs that maintain a consistent layer of
abstraction, and instead of making all libraries responsible for the side
effects that might be introduced from other oslo libraries, we make the
side-effect-causing library the point at which those effects are
ameliorated as a service to other oslo libraries.   This allows us to keep
the knowledge of what it means to use “multiprocessing” in one
place, rather than spreading out its effects.