Hi Scott, Just a couple of clarification points: (a) My common sense comment was referring to the fact that since pooling is not spec'd, any pooling semantic must be invisible to the app, otherwise portability will suffer. (b) As far as read only properties/refs, don't forget that an important use case in SCA is the ability to use Java components as SCA component implementations which may have been created before SCA existed. Especially in these cases, but not limited these cases, it is very likely that a Java component may want to modify itself. I agree it is unlikely for a truly stateless component, but I think it's heavy handed to have such a constraint. (c) Issue-65 in the OASIS spec http://www.osoa.org/jira/browse/JAVA-65 is going to have to face the primary question you are asking.
Dave ----- Original Message ----- From: Scott Kurz To: [email protected] Sent: Monday, October 13, 2008 5:17 PM Subject: Re: [jira] Created: (TUSCANY-2635) Pool stateless-scoped Java impl instances Putting aside conversations, let's discuss another example: property injection. Suppose the user's business logic modifies a stateless instance's injected property. @Property protected String currency; // configured via SCDL and injected by container myMethod() { .... currency = "blah"; } So Dave, I take it you're arguing that, if the container were to pool stateless instances, it would be responsible for ensuring that in later dispatches of this same instance, the 'currency' property would appear as if it were freshly-injected upon the business logic reading its value. This capability strikes me to be of so little value to the app programmer that I think the ideal situation would be for the spec to disallow this in the stateless programming model and to say that the injected data is read-only. I don't see how this is a common sense issue either. This is a new programming model, with a new choice to be made. Going back to EJB again, we see a model where one had to understand the difference between instance creation and going in and out of the ready pool between dispatches. Maybe you're saying this is obviously undesirably complex, but on the flip side here is a component model that was widely used which employed these concepts. So I don't see it as too complicated to say, "it's stateless, so you don't own the instance, so don't mess with the injected instance variables." Scott On Mon, Oct 13, 2008 at 4:01 PM, scabooz <[EMAIL PROTECTED]> wrote: Hi guys, I'm having a hard time following this latest branch in the discussion, but I think there is a basic principle that needs to be asserted, which is that instance pooling must be unobservable to the business logic. This is just common sense, and anything else will make the programming model more difficult to understand. More specifically, when the runtime hands an instance to the business logic, the state of that instance should be identical regardless of whether the container new-ed up the instance or obtained it from a pool. Someone is going to jump on the "where does it say that in the spec?" question. It's probably not explicit, but it should be, and there are open issues in the OASIS work to clarify and lock it down. All I can say to this question is that the above is the spec intention, and is a principle that will cause the least surprise and most success for users. Dave ----- Original Message ----- From: Scott Kurz To: [email protected] Sent: Friday, October 10, 2008 4:16 PM Subject: Re: [jira] Created: (TUSCANY-2635) Pool stateless-scoped Java impl instances Simon, Sorry my line of reasoning didn't come through so well. I was actually trying to argue the point that a new conversation should NOT be started just because a stateless instance is obtained from a pool. However, I was trying to address the opposite point of view (a new conversation should be started), which I think is where you were coming from. If the specs had said something like "at the beginning of a scope, references are injected", then we'd have an answer (opposite to my reading). If you take Java C&I lines 193-195 to mean the same, we'd also contradict my interpretation. In contrast, I'm taking the lack of clarity as a signal that this is up to the runtime implementation to work out So the net is (per my reading), that you can not assume a new conversation per stateless dispatch, if you're relying on reference injection to start the conversation....and you have to use programmatic means if you want this. Scott On Fri, Oct 10, 2008 at 12:18 PM, Simon Laws <[EMAIL PROTECTED]> wrote: On Fri, Oct 10, 2008 at 4:49 PM, Scott Kurz <[EMAIL PROTECTED]> wrote: OK, So On Fri, Oct 10, 2008 at 10:08 AM, Simon Laws <[EMAIL PROTECTED]> wrote: Hi Scott Can you say a little more about what the implications for pooling stateless instances would be, for example, 1st call - Create object @Reference injected @Init run 1st Operation execution ends @Destory called Reset any conversational/callback state help in the components references object put in pool 2nd call - Retrieve object from pool @Init called 2nd Operation execution ends @Destory called Reset any conversational/callback state help in the components references object put in pool Is that the sort of thing you were thinking of in order to remove the injection overhead? Simon, yes, unlike, say ejbCreate() for a SLSB, the @Init is supposed to get called every time for a stateless invocation So let's look at conversational state, which is the type of concern I was hoping to get to. I'm pretty ignorant of what this looks like at a low level. At a high level, though, a new conversation is supposed to begin every time a conversational-interfaced-reference is injected into the client component. So, my naive take on this would be that this isn't all that useful when the client component is stateless, since you don't know when the container is going to give you a new stateless instance (with an already-started-conversation) or an existing one. So if you need better control over starting a conversation from a stateless component, you need to get your conversational reference programmatically. I suppose you could read into the specs the idea that they intended the references to be injected at the start of the scope..... there is this (which I don't believe helps): Java C&I - 1.2.2.1 192 If @Reference marks a public or protected field, the SCA runtime is required to provide the appropriate 193 implementation of the service reference contract as specified by the field type. This must done by setting 194 the field on an implementation instance. When injection occurs is defined by the scope of the 195 implementation. I don't think it helps since I don't see that the specs define this with respect to the Java impl scopes. So... given that my model is, we don't really consider this is a smart use case and we leave the conversation dangling in such a case... do we still have anything to worry about for conversations? Scott Hi If you did read into the specs that injection was intended to happen at the start of each scope, i.e. in this case when an object is retrieved from the pool, wouldn't that kill the performance improvement you are looking to achieve? Not sure I understand the last sentence. So... given that my model is, we don't really consider this is a smart use case and we leave the conversation dangling in such a case... do we still have anything to worry about for conversations? If you don't clean up the conversation object then when you call through the reference of a previously pooled component instance you will likely be continuing a conversation that is already started. There is state referenced behind the reference proxy and also in the conversation manager about onging conversations. This state is removed when a component instance is removed and we'd have to work out how to do the same if we were to pool stateful instances. This is of course in the case where the business logic doesn't end the conversation properly itself. Simon
