On 9/5/07, Mike Edwards <[EMAIL PROTECTED]> wrote: > > Ant, > > ant elder wrote: > > I'm looking at what we could do for TUSCANY-1559 which is about > unnecessary > > pass-by-value copies causing failures when arguments aren't > serializable. > > The problem is on line 260 of JavaComponentContextProvider where it > tries to > > determine if the pass-by-value invoker is required, but it doesn't take > into > > account whats at the other end of the invocation chain. > > > I'm a bit confused by this. These statements seem to imply that you > can't use the service/reference in the case where remoting is going on. > That will indeed be the case if you have non-serializable arguments. > > But why is the invocation chain of any relevance here? Or have I got > the wrong end of the stick. > > Any service or reference that is remotable must have an interface that > can be serialized onto a wire. SCA expresses this in terms of > mappability to WSDL. If it's remotable it must be expressible in WSDL > and serialization is part of this picture. > > > > How about changing this so its not the Java implementation types > > responsibility to handle pass-by-value and move it out into core, having > a > > separate RuntimeWireProcessor and Interceptor for pass-by-value support, > and > > don't do copies if the source and target interfaces use different > > databindings? > > > > ...ant > > > Different databindings surely is not part of the problem. Again, you > can map into and out of a "canonical" form (XML) if required - that is > where the WSDL mappability comes in. > > Where there IS an opportunity for avoiding serialization is for > remotable interfaces which are marked "AllowsPassByReference" - then, if > the target happens to be local you can use pass by ref and avoid any > serialization. Optimization of this kind will require some smarts in > the invocation chain - but it is all clearly marked in terms of the SCDL. > > But I fail to see how failures can occur for "unnecessary pass-by-value" > copies in any circumstance other than one that SHOULD be an error. > > > Or am I missing something here?
Maybe you are missing something but I'm not clear what you are suggesting or objecting to. Are you saying TUSCANY-1559 should be closed as working as designed? Or if not what are you objecting to in the above suggestion, or proposing as an alternative fix? If you look at the code what happens today in this scenario is the pass-by-value interceptor is trying to use Java serialization to copy the arguments and that fails as they are not serializable. We could try to fix the pass-by-value interceptor but the copy is unnecessary when its a Java component using a web service binding on the service or reference as the arguments will be transformed by the databinding framework into Axiom OMElements which is effectively the pass-by-value copy. ...ant
