Re: [AXIS2] [Sandesha2] Saving the message context

[Aleksander Slominski]

Matthew Lovett wrote: Hi all, That approach will probably need to be revisited once we get the security stuff running. If the inbound security handler authenticates the user, and puts the user identity onto the current thread (or into message context) then simply serialising the message would lose this identity by the time we re-inject the message. Equally, running through the security handler a second time (with the original message payload) won't work, as the timestamps on the message might make the re-injection fail. I think this case helps explain why we need a proper architecture for serialising context, so that the owners of state can make their own decisions about what is important. I also agree 100% that we should serialise the bare minimum, and that Java serialisation isn't the way to go either! hi, one way to get around this problem and allow for easy load balancing is to allow serialize context into XML - this is really handy when you need to move processing to another node [1] or just restart processing in case of RM. best, alek [1] http://www.extreme.indiana.edu/labpubs.html#lifang:den:gw2005 Liang Fang, Aleksander Slominski, and Dennis Gannon. Web Services Security and Load Balancing in Grid Environment. "Chamikara Jayalath" <[EMAIL PROTECTED]> wrote on 01/08/2006 03:34:58: Hi Jaliya, Yes. Thats similar to what we are doing currently. We save the message context in the Sandesha Global In Handler and re-inject it into the handler chain after a failure. Chamikara On 8/1/06, Jaliya Ekanayake <[EMAIL PROTECTED]> wrote: Hi Chamikara, Don't know whether it is an efficient way; how about this - we can save the SOAP message after security handler using a custom handler that will only be deployed in the persistent mode. -Jaliya ----- Original Message ----- From: Chamikara Jayalath To: Jaliya Ekanayake Cc: sandesha-dev@ws.apache.org Sent: Monday, July 31, 2006 1:54 PM Subject: Re: [AXIS2] [Sandesha2] Saving the message context Hi Jaliya, Well, not exactly. In Sandesha2 scenario we process the message in several transactions. Processing of a message within a handler will be done in one transaction while the invocation will be done in another transaction. So we cannot simply abandon the message. We have to reinject it into our system (thats what we do). But if we serialize the message in the very begining of the handler chain we can asume that the context would not have been changed and saving the SOAP envelope would be enough. But this is not always a practicle solution since handlers like security will sometimes have to be present before RM. Chamikara On 7/31/06, Jaliya Ekanayake <[EMAIL PROTECTED]> wrote: Hi Chamikara, What I am suggesting is this. If we get the QoS information stored properly that will enable us to build a definite state after a crash. e.g. We don't need transport info because RM will handle it by way of re-transmissions. ServiceContext and ServiceGroupeContext; IMHO WS-Transaction should handle this. So if we keep the states of QoS layer then we can avoid this heavy serialization. Any thoughts? -Jaliya ----- Original Message ----- From: Chamikara Jayalath To: Jaliya Ekanayake Cc: sandesha-dev@ws.apache.org ; axis-dev@ws.apache.org Sent: Sunday, July 30, 2006 10:58 PM Subject: Re: [AXIS2] [Sandesha2] Saving the message context Hi Jaliya, Thats good news. But only the properties will not be anough. There are other things like the state of the options object, transports and callback object that also have to be serialized. There are also references to the other contexts (serviceContext, serviceGroupContext) from the Message Context and we will not want to loose these connections when the Message Context is deserialized. However if it can be declared that the referenced contexts will not be serialized when serializing one context, that paves the way for a solution. But not sure weather this is valid for all cases.(Still have to think more on reconstructing the context hierarchy) Chamikara On 7/31/06, Jaliya Ekanayake <[EMAIL PROTECTED] > wrote: Hi All, As far as I remember we spent some time during the design of axis2 to solve this problem. The final conclusion we made was to do our own serialization by serializing only the properties (serializable objects) in the property bag not the entire message context which has pointers to other contexts. Thanks, -Jaliya ----- Original Message ----- From: Chamikara Jayalath To: [EMAIL PROTECTED] Cc: sandesha-dev@ws.apache.org Sent: Sunday, July 30, 2006 8:40 PM Subject: Re: [AXIS2] [Sandesha2] Saving the message context Hi Bill, I agree that doing serialization within context objects is the best approach in a design perspective. The approach I followed was only possible due to MessageContext already having made its useful state public. I also originally tried to follow Externalizable approach and introduced externalizable methods to all the contexts (they hv now been removed due to not having any usages). The main problem I had in this approach was having to serialize the whole context hierarchy. Every message context has a pointer to the configurationContext so to be general (not to be specific to our scenario) in the serialization method we would have to serialize this object as well. Since this has pointers to all other contexts they will be serialied too. What I am saying is that when adding the externalizable method to the axis2 codebase we would have to serialize the configContext and other contexts as well (because some people may actually want to serialize the whole context hierarchy). But in our case it seemed like this would be a burden. Every deserialized message context with come up with its own context hierarchy and maching between two deserialized Message contexts will be extremely difficult. If you find a solution to this problem I agree that your and Anns approach is the best way to go and I would like to use that in my code :-) Chamikara On 7/29/06, Bill Nagy < [EMAIL PROTECTED] > wrote: On Fri, 2006-07-28 at 23:46 -0400, Rajith Attapattu wrote: Anne, Again I will advice again serializing the contexts using any form of serialization. This will not scale at all in a production environment. Hi Rajith, Could you please explain this last comment? Again this approach will be error prone and as chamikara mentioned there will be too many information saved in the database. I don't understand why you and Chamikara keep saying that there will be too much information serialized. You have the option of taking complete control of the serialization, thereby writing/reading only the information that you want and in the form that you want it to be in. I don't believe that Ann is arguing for simply using the default serialization, only about who should be in control of making the decisions as to what should be saved. I am looking at clustering certain information within the ctx heirachy for high availability and I would only do with the bare minimum. In my opinion the performance overhead of serializing and deserializing (and validations to avoid erros) is a lot more than saving the required info in a bean like what chamikara does for Sandesha and then reconstructing it. Having the objects persist their own state is far less error prone than having a third-party piece of code do the persistence. For one, anytime someone changes or adds a piece of information that needs to be saved in order to correctly restore the state, they have to remember to modify the external code. It's generally hard enough to remember to modify code embedded in the class itself, much less having to remember to modify a completely separate piece of code. Secondly, you require the objects that need to be saved to expose methods, to return the portions that you want to have serialized, that you may not have to expose otherwise. In effect, the approach that you've chosen has abandoned encapsulation and created fragile dependencies -- this is bad design. -Bill Regards, Rajith On 7/28/06, Chamikara Jayalath < [EMAIL PROTECTED] > wrote: Hi Ann, Yes. We had introduced Externalizable implementaitons for all of the Context hierarchy objects sometime back. But this approach seemed to be saving too much information on the database. For example at some point there may be following context objects available in a running axis2 instance. 1. configuration context object 2. service group context objects 3 service context objects 4. Operation context objects 5. A lot of message context objects If we try serializing starting from a message context, since we have to serialize every incoming message context all these objects will be serialized every time (recall that the message context hs a link to the configuration context which has links to all other context objects). Think how deficult the reconstruction would be. Every deserialized message context will come up with its own hierarchy of context objects which may not map with the context objects reconstructed by deserializing others message contexts. Thats why I went for this approach of saving only the relavent information. It seemed to be much cleaner and it was working :-) You had mentioned about serializing the AxisOperaion object. All the 'Axis*' objects in Axis2 (AxisConfiguration, AxisServiceGroupt etc.) contains deployment time information. So we can safely ignore them in the serializing process. Chamikara On 7/29/06, Ann Robinson < [EMAIL PROTECTED]> wrote: Hi Chamikara, Thanks for the information. Did you consider using java.io.Externalizable for the AXIS2 message context-related classes? (Having the work done by the AXIS2 objects would have simplified the actions that Sandesha needed to take in order to save the message context, so I am curious about any issues that were encountered. In the MessageStoreBean, how much of the various objects do you store as Strings? For example, the AxisOperation object contains several lists and the executionChain object contains a list of handlers and phases. Ann WebSphere Development, Web Services Engine IBM 11501 Burnet Rd IZip 9035G021 Austin, TX 78758 (512)838-9438 TL 678-9438 Inactive hide details for Chamikara"Chamikara Jayalath" < [EMAIL PROTECTED] > "Chamikara Jayalath" < [EMAIL PROTECTED] > 07/28/2006 07:23 AM Please respond to axis-dev@ws.apache.org To Ann Robinson/Austin/[EMAIL PROTECTED] cc axis-dev@ws.apache.org , sandesha-dev@ws.apache.org Subject Re: [AXIS2] [Sandesha2] Saving the message context Hi Ann, I did some work on serializing message contexts and reconstructing them. This was done as a part of the Sandesha2 Persistent Storage Manager implementation. Unfortunately could not commit the code into Apache due to a license issue (it was dependent on Hibernate). But will try to host it somewhere else soon. The approach i took was extracting the relevant information from the message context, and saving them in a java bean. Later this bean was used to recostruct the message context. The format of this bean was as follows. public class MessageStoreBean { private String SOAPEnvelopeString; private String storedKey; private int SOAPVersion = 0; private String transportOut; private String axisServiceGroup; private String axisService; private String axisOperation; private String axisOperationMEP; private String toURL; private String transportTo; private int flow; private String executionChainString; private String messageReceiverString; private boolean serverSide; private String inMessageStoreKey; private String messageID; private String persistentPropertyString; private String callbackClassName; private String action; } As you can see the aim was to avoid Java serialization. One defect here is SOAP envelope being saved as a string, which may not be possible in the RM +MTOM scenario. I guess we can support that when a better serialization mechanism gets available for SOAP envelopes. Chamikara On 7/27/06, Ann Robinson < [EMAIL PROTECTED] > wrote: Hi all, I have posted this note to both the AXIS2 and SANDESHA developer discussion lists, so I apologize in advance to those folks who get multiple copies of this note. I am investigating how to save and restore the message context in AXIS2. This is functionality that would be used by other quality-of-service layers, for example, by a WS-ReliableMessaging implementation - particularly one that is composed with WS-Security, to save the message in persistent storage and later resume the message processing. The AXIS2 message context is very complex (it includes references to several complicated objects) and does not lend itself to the default java serialization mechanism (java.io.Serializable). In order to save the message context, the possible solutions include the following: (A) Internal Message Context option Do a customized serialization using java.io.Externalizable in the complex objects and use the default serialization mechanism (java.io.Serializable) in the simple objects. - - This keeps the knowledge of the object's internals in the object and keeps the responsibility in the object for persisting and resurrecting its own state. - - This lets an object have a plugpoint where needed to manage "user" data. This would apply to the situation where an object maintains a set of properties or attributes that are supplied by users of the object. The plugpoint would define an interface so that the users of the object could save their properties/attributes appropriately. (B) External Layer option Put in get/set methods in all of the objects related to the message context in order to allow another layer or quality of service (QoS) to extract sufficient information from the message context in order to save and resurrect the information. - - The simplest form of this technique is saving just the message (and the message attachments). However, this means that any processing on the message has to be re-done from the beginning. - - If there is a requirement to maintain the security context with the message, then the security layer would need to provide additional interfaces to allow that message's security context to be acquired by that other layer. (C) Core Plugpoint option Have a plugpoint in the AXIS2 core that would provide an interface to capture essential message context data for saving and restoring it. - - This solution would be a compromise between (A) and (B) - - This requires knowledge of message context object-related internals inside of the plugpoint implementation, which is not good object oriented design Any other suggestions or comments? I understand that there has been a previous attempt to do this in AXIS2 based on Sandesha requirements and that this attempt did not work. I was wondering if anyone remembers what problems were encountered and what issues ultimately blocked that solution? Thanks, Ann WebSphere Development, Web Services Engine IBM 11501 Burnet Rd IZip 9035G021 Austin, TX 78758 (512)838-9438 TL 678-9438 --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] -- The best way to predict the future is to invent it - Alan Kay