uima-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Danai Wiriyayanyongsuk" <dan...@gmail.com>
Subject Re: Cannot share Session object when using PEAR class loaders.
Date Mon, 01 Oct 2007 20:57:45 GMT
Hi Adam,

Thanks again for the clarification. I really appreciated it :)

Speaking of Case1 testing, did your two testing annotators get the same
Session object reference? Mine did. Actually, by not using PEAR class loader
in Case1, both of my annotators consistently got the same Session object
reference for all three AnalysisEngine instances in the AnalysisEnginePool
of a MultiprocessingAnalysisEngine_impl instance.

Anyway, given the limitation/behavior of PEAR isolated class loader, which
approaches/practices would you recommend to share an object, say a large
array of bytes, between each annotator that run in its own PEAR isolated
class loader?

Thanks,
Danai Wiriyayanyongsuk

On 10/1/07, Adam Lally <alally@alum.rpi.edu> wrote:
>
> Danai,
>
> On 9/28/07, Danai Wiriyayanyongsuk <danaiw@gmail.com> wrote:
> > Hi folks,
> >
> > I'm trying the PEAR class loaders in UIMA 2.2. I think I found an issue
> > that, if a PEAR description file is specified in the aggregate analysis
> > engine description, the session object returned from the method
> > org.apache.uima.analysis_component.JCasAnnotator_ImplBase.getContext
> ().getSession()
> > from the related annotator will be different from other annotators. In
> other
> > words, the annotators cannot share the same session object if the PEAR
> class
> > loader is used.
> >
>
> Actually the Session objects aren't really fully implemented - we
> should have said this in the javadocs, sorry about that.  Here's the
> idea, though:
>
> Even without PEARs, the Session objects of two different annotators
> are supposed to be different.  The Session object was intended as a
> way for multiple instances of the *same* annotator to share
> information (within a remote service deployment where multiple
> instances were created for performance reasons), not as a way to share
> information between different annotators.  Each annotator gets a
> different Session object so that there would not be conflicts with the
> key names chosen by the different annotators.
>
> The idea is that in a remote service deployment where there are
> multiple instances of an annotator and multiple clients, that the
> Session object is a place for an annotator to store information on
> behalf of a particular client.  So if client A makes a remote call and
> this is handled by annotator instance 1, and that instance puts
> information in the Session object, then if client A makes another call
> which is handled by annotator instance 2, that annotator instance
> could retrieve the information put there by annotator instance 1.
>
> That part works, but currently if another client B comes in, it will
> not get a distinct Session object.  That functionality has to be
> implemented in the service wrapper and this was never done.
>
> > Here is more detail on the situation.
> >
> > I have two annotators. Both of them sharing the same piece of
> information
> > via the Session object (Session session = getContext().getSession();). I
> > assume that both annotators will have access to the same session object.
> I
> > tried it and found the inconsistent behavior between the following
> cases:
> >
> > Case 1: The aggregate analysis engine description contains full
> descriptions
> > of both annotators  (no import tags).
> > Case 2: The aggregate analysis engine description contains import tags
> with
> > the PEAR description files specified (xxx_pear.xml) for both annotators.
> >
> > In case 1, for both annotators, the *same* session object are returned
> from
> > the method
> > org.apache.uima.analysis_component.JCasAnnotator_ImplBase.getContext
> ().getSession().
> > However in case 2, the *different* session objects are returned from the
> > method.
> >
>
> I'm surprised that Case 1 actually works.  I wrote a little test
> annotator:
>
>   public void process(JCas aJCas) throws AnalysisEngineProcessException {
>     String val = (String)getContext().getSession().get("foo");
>     System.out.println("foo = " + val);
>     getContext().getSession().put("foo", "bar");
>     val = (String)getContext().getSession().get("foo");
>     System.out.println("Set foo = " + val);
>   }
>
> and made an aggregate with two such annotators, and showed that the
> information put there by annotator 1 is not seen by annotator 2.
>
>
> You still have a point, though, about the PEAR components.  Normally
> the way for different annotators to share information is through a
> SharedResourceObject.  However this won't work for PEAR components
> because they don't share the same ResourceManager.  The way the PEAR
> component works, it is completely isolated from the rest of the
> aggregate AE (the main thing is that it can have its own class loader)
> and can't share information.  This ought to be better documented.
>
> Regards,
>   -Adam
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message