see my comments below;
Hi all:
Comments below.
Eran Chinthaka wrote:On Apr 7, 2005 10:07 AM, Deepal Jayasinghe <[EMAIL PROTECTED]> wrote:At the axis2 face to face we did number of changes to axis2 architecture,
had to get rid of EngineRegistry and introduce new set of contexts (
www.apache.org/~deepal/axis2f2f/axiscontext.jpg )
considering all those I will have to do some changes to the deployment
module too. This mail describes purpose changes to that, so please read that
carefully and feels free to give your comments and any suggestions.
Pre-dispatch phase:
This is a system inbuilt phase and which is not going to specify in the
server.xml as available phase , and all the handlers in this phase run for
each service, i.e they act as global handlers. The main rationality behind
introducing phase like this is to provide a way to add global handlers by
modules if they want.
If I add something more to this : We wanted somethings to happen before any of the dispatch handlers run. For example, security, WSA, erc., The name is just another name for a phase, but unlike other phases this will be burnt in to the engine. All the other phases are mentioned in the server.xml, or somewhere else, but we thought its better even not to mention pre-dispatch phase in server.xml as if we do that, one can go and delete that entry. So we will just put a comment in the server.xml about this phase, but not an entry.
This pre-dispatch phase can be considered as another global scoped phase, but these should before any of the dispatch handlers.
Um - guys? I don't recall talking about a "pre-dispatch" phase at all! (?) Where did this come from?
IIRC, we just said that since you need to know the operation before engaging operation-specific handlers/modules (which might actually include security, etc), it's important to do this level of dispatch as early as possible. So if an HTTP handler figures out the operation based on the action parameter of the MIME type, then great! That just means the operation specific (and by definition the service-specific also) handlers can be merged onto the ExecutionChain by the OperationDesc (or whatever we're calling the WOM Operation thing) really early on.
We also discussed the idea that our "baked-in" phases might actually all be immutable - there would probably only be a handful of these, such as "decryption", "reliability", etc., but we'd fix the relative order of those phases and make them always available even if nothing is in them, regardless of configuration. I don't think we settled on that as a path forward, but it came up in the same discussion.
Using its module.xml find the module name and then retrieve corresponding
"Module Context" form EC. If this is the first time server gets started up,
EC gives an empty Module Context else it gives the previously saved (when
server shutdown) Module context corresponding to the module.
But it's up to the Module, not the engine, to figure out how to get the context, right? I.e. the Module would be the one doing any recovery from DB or disk and filling in the ModuleContext in the appropriate place.
[Deepal]
agreed , I have no objection regrading who is going to do the recovery process, but I think if it is going to do by module he can do it the way he wants.
While processing the module.xml if there are any handlers with the phase property as "Pre-dispatch" phase then those handlers will be added to the Pre-dispatch phase by DeploymentEngine.
Again, this doesn't match my recollection. I thought that we just had all the regular phases, like reliability, security, etc... and the Modules just got to deploy anywhere they wanted within those (or define their own, which I believe we did decide to move forward with).
Oh, wait - maybe you were thinking of a "pre-dispatch ExecutionChain"? In other words, there is a "global" set of Handlers which is going to run by default.
[Deepal]
Yup , that is what we are talking about , which is going to be run for all the services.
You know this chain before dispatch happens (it's the
handlers+modules in the engine.xml), and that set acts as a template for each individual MessageContext's "real" ExecutionChain. So that set would likely include, for instance, the WSA handler and the WSS handler - but they'd still be deployed in whatever phases they wanted on the chain... i.e. no actual phase called "Pre-dispatch" (unless you wanted to custom build one, that is :)).
Does this match others' understanding?
For example, the security module ot WSA module may add some handlers to pre-dispatch phase.
(see above)
Then using the above Module context and EC, it calls the init() method of
module implementation class , from that we give more flexibly for module
developers to change the order of handlers in Pre-dispatch phase , add new
handlers to that chain, if the module need to deploy a new service to do that and etc..
this init() will only be called once and for all, during the life time of the engine. Basically this will called during engine startup or the corresponding module deployment ONLY.
If one needs to do some other stuff per message basis within the module, use the engage(...) method.
I'm wondering if Modules might need to do things like send messages *after* all the other Modules are started, so that global policies get correctly followed even during recovery. If so, we might need some kind of work-queue that Modules could add themselves to during startup to get pinged again after everyone's ready (i.e. register for the "engine is up!" event...).
Next add that module to the EC. Apply the same procedure for all the module. Next check to see whether there are any module references in srever.xml if
so Gets the corresponding module (s) form EC and gets handlers of those
modules. And add those handlers to the Post-dispatch phase , (those handlers are not
going to be ordered at the deployment time) Next if there are any service have been deployed (drop in to /service
directory) then load them one by one Read it service.xml and using service name retrieve the corresponding
"Service Context" and while processing the xml file fill the Service
context. Next check to see whether there are any module references if so gets
handlers corresponds to those. Next using above handler come from
modules (module references) service specific handlers (defined in service.xml) all the handlers in the Post-dispatch phase
Creates three kind of flows (or three kind of execution chains) called
1. In-Flow
2. Out-Flow
3. Fault-Flow
(Let's call those flows as serviceChain and those are also called pre
compute handler chains)
This was pretty confusing to read... :)
Didn't we talk about needing in/out/infault/outfault chains (potentially at each level global/service/operation)? You might want to do something very different for outgoing faults than for incoming ones....
[Deepal]
Agreed, I remember that we discuss about in-fault-flow and out-fault-flow , so in the deployment time all the four kind of chains will be crated.
Next using above service context invokes init method of the service
implementation class. Finally it will be added the created service to the EC.
� At the run time for each message arrive, it invokes
engage(serviceChain , EngineContext) method of each and every module , then
the depending on the WS-police module will decide what to do to the
serviceChain , 99% of the time it will do nothing. But if we consider about
1% their module developer can use either Phase rule or any other mechanism
to change the handler chain.
I thought we decided it was engage(chain, context) where context is simply the generic Context interface, so that Modules can be reused easily at all levels (it might be OperationContext, ServiceContext, EngineContext....)?
Also, I don't think engage() is getting called as often as you think it does. That method gets called when a given Module learns that it should be applied to a given chain. So for the [engine/service/operation]-specific Modules (ones that are referenced in service.xml, say), that happens at init time, once. If Modules are *available* in the engine but not actually *engaged*, it's possible to engage them at runtime via some Handler which is processing something like policy metadata in an EPR.
But engage() shouldn't always be called on every module for every message - should it? (it's late here, so maybe I'm misremembering, but I didn't think we went there)
>Its like this. When engine asks for an execution chain for a service, or for an operation, the engine configuration will do the phase resolving as usual for all the handlers, irrespective of whether they are in a module or not. But even after that, the engine will call engage method of each every module connected to the current message.Most of the time, we guess this engage method will do nothing. But if one needs to do some handler manipulation even after phase resolving, here is the time.
Hm. See above. (?)
--G
Deepal
