I think this sounds reasonable- do you foresee there being any additions here? I’m not sure there’d be any more call-points, so it seems like a good way to move forward by adding flexibility.
On Apr 18, 2014, at 2:15 PM, Mark Chu-Carroll <mchucarr...@apache.org> wrote: > I'm contemplating making a change to the current configuration plugins. I'd > like to get a couple of other eyes to glance over it - I'm unsure whether > this is sound, or just a reaction to an immediate problem that I > encountered working on logging. > > The way that configuration plugins work right now in the noun/verb > framework is super simple. A configuration plugin is an object with an > "execute" method. During command dispatch, after the command parser > processes the arguments and a command-appropriate context is created, the > framework iterates over the plugins, and calls the execute method of each > one, with the context object and parsed command line parameters. > > There are two problems that I've encountered working with these in practice: > > 1. Some plugins really want to be run before the command-line arguments > are processed, and some want to be run after. For example, for a > distributed logging system, we'd like to be able to have the logging system > optionally enabled before the command is dispatched, so that the dispatch > process can use logging. But others, for managing things like proxies, need > to be called after arguments are processed, because they rely on arguments. > 2. Some plugins need cleanups. For example, in logging again, for best > performance, we'd like to use multiprocessing to do asynchronous > communication with a logserver. But to do that, we need to do a join after > the command completes execution. In the current framework, there's no way > to do that. > > What I'm thinking about is going more aspect-oriented in the plugins. There > will be three points during the execution of a command where plugins can > get injected: > > 1. before any argument processing is performed. The only thing that the > plugin would be allowed to look at would be the raw, unprocessed > command-line parameters. It would return a (potentially) modified copy of > the parameters, which would be passed to other plugins, and finally to the > actual noun/verb for the command being executed. > 2. after arguments have been processed, but before the noun/verb is > executed. This is exactly what the current plugin "execute" method does. > 3. after the command has completed executing, but before the client > exits. > > The code signatures would be: > > def beforeDispatch(self, args): > """code to execute before dispatching to a noun/verb command. > args: a list of the command-line parameters used > to invoke aurora. > Returns: a possibly modified list of command-line arguments. If > this throws an exception, then the execution > of the command will be aborted. > """ > > def beforeCommand(self, context): > """Code to execute after command-line arguments have been > processed, but before executing the command. If the > code throws a CommandException, then the execution of > the command will be aborted. > context: the context object that will be used to invoke > the command. > > > def cleanupAfterCommand(self, context, returnval): > """Code to run after the execution of a noun/verb command has > completed, before the command-line application exits. > context: the context object that was used to execute the command. > returnval: the exit code returned by executing the command. > """ > > > The order of execution of plugins would remain undefined.
