You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
For some kinds of processing we don't need to submit complete pipelines to the cluster. Currently it's not possible unless we use a hack (create a pipeline with two jobs: the one you want to be executed and a dummy one).
We just need something like a JobManager class, like PipelineManager, but simpler.
The text was updated successfully, but these errors were encountered:
Very important. Then it would be very nice to create a simple decorator people could use to submit jobs, as many distributed processing libraries have: People could decorate functions they want to offload to a pipelinin queue for distributed processing. I have something similar in my Liveplots package: https://github.com/fccoelho/liveplots/blob/master/src/liveplots/xmlrpcserver.py
defenqueue(f):
"""Decorator that places the call on a queue"""defqueued(self,*args,**kw):
Q.put((f,(self,)+args))
queued.__doc__=f.__doc__queued.__name__=f.__name__returnqueued
where Q is a Queue object.
Also a customized map/reduce (this could be lifted from other python map/reduce libraries) framework would see great adoption.
The initial idea for this issue is just to implement something that I can submit a Job object instead of an entire Pipeline object (which is a group of Job objects). So, for the first implementation, the worker should exist on brokers to this approach work properly.
In a next release, maybe, we can think of remote code execution (we need to be more careful if implement this). I think it's needed but not for now -- @fccoelho, can you please create an issue describing this feature?
For some kinds of processing we don't need to submit complete pipelines to the cluster. Currently it's not possible unless we use a hack (create a pipeline with two jobs: the one you want to be executed and a dummy one).
We just need something like a
JobManager
class, likePipelineManager
, but simpler.The text was updated successfully, but these errors were encountered: