Title: RE: [asap] questions/concerns on initial working draft doc (and scope)
Let me
try to refocus this discussion.
ASAP
does not specify how you implement the web service. If you wish to use
J2EE, you may use whatever technique. I thought that describing in terms
of stateful/stateless would help in understanding the concepts, and to
distinguish the various meanings of the word 'asynchronous'. I didn't mean
to imply that this was the actual implementation. ASAP does not proscribe
how this is implemented in the J2EE world, and so I don't want to discuss this
here.
I
still think "method" in the spec refers to an operation, because it consists of
a request message and a response message. The messages are described, in
places, as parameters. Please point out in the document where it means
something else than this. I propose that we change "method" to "operation"
in the document in order to conform with SOAP terminology.
Let me
ask you a question: do you believe that a special protocol is needed for an
operation that can not be monitored? For example, a service that you send
a message to, and a while later you get a response from. No possiblity to
change you mind. No possiblity to monitor status. Can't you do this
is basic SOAP protocol? Let's discuss if you think differently, but I am
going to assume that this is covered by basic SOAP.
The
"context data" is persistent data that exists in the service instance. It
can be sent as part of a reqeust (e.g. SetProperties), or as part of a response
(e.g. GetProperties). Furthermore, there are messages that contain the
schema of this context data as part of a response from the factory
object.
We
really should discuss usage scenarios. You touched upon the realestate
example, one of my favorites. You describe the service as being a
recurring query for houses that meet a specific criteria. The context data
includes the query, as well as the result set. (You might implement this any
number of ways, but for now lets imagine that we choose this one.) The
client would make a simple SOAP request to start this service, giving both the
initial query, as well as the address to send responses to. Everytime the
result set changes, a notification is sent to the client. It would be
really unpleasant if there was no way for the client to turn this off.
When you started the service, you had no idea how long you would be
searching. You in fact might put the serach end data as a context
variable. Later, you can change that end date by using "SetProperties"
thereby extending the duration of the search service, or ending it early.
You can also cancel the service by using the "ChangeState" to the closed state
(which is the terminal state).
The
client of this service need only know the name of the context variable that
holds the query (e.g. "query"), and the name of the context variable that holds
the results (e.g. "listings"). Probably, the service instance would not
allow the client to set the value of "listings". Attempts to set this
would result in an error. Some services can not allow changing of context
variables at any time, so attempts to set any of them would cause an
error. In this sense, the operation "SetProperties" is optional. It
is not really optional, but can be implemented to simply return an
error.
If you
don't have a client that can receive notifications, then you also can simply
pick up the latest value of "listings" at any point in time. In this case
you would not register an observer address. Also, you would never receive
a completed message.
Finally, others interested in the same listings (the same query) could
register to receive notifications of when the listings change. Note: you
are not registering to a specific query, because the query could be
change. Instead you are registering to a specific instance of the
service.
Now,
can we start over with the questions because I am having a hard time deraveling
them from the discussion below.
-Keith
-----Original Message-----
From: Michael Shenfield
[mailto:]
Sent: Tuesday, October 07, 2003 12:26
PM
To: Keith Swenson
Cc: ASAP (E-mail)
Subject:
RE: [asap] questions/concerns on initial working draft doc (and
scope)
Keith,
I went back to EJB spec to validate my point on stateless
(vs. stateful) session beans. The following is from an EJB 2.1 final draft
(section 1.2. What is New in EJB 2.1):
"We have
added support for web services. We have added the ability for a
stateless session bean to implement a web service endpoint. We have added support to
allow all enterprise beans to utilize external web
services."
In the
current J2EE/Java world only servlets and STATELESS session beans could
implement a Web Service endpoint. Neither can support the protocol described
in a current working draft (actually it's not true, one can come up with
a model when the state is shared by servlet instances using a ServletContext
object. It could be a sneaky way to implement monitoring but very
inefficient and non scalable
one)
Cheers,
MS
-----Original Message-----
From: Michael Shenfield
Sent: October 7, 2003 2:42 PM
To: Keith Swenson; ASAP
(E-mail)
Subject: RE: [asap] questions/concerns on initial working
draft doc (and scope)
Keith,
My
comments/answers are embedded in the orig. message. I suggested at the last
conference call that we define the scope of ASAP to clarify what this spec
is going to be about. It looks like we have different views on what it
should cover.
Thanks,
Michael
Shenfield
Advanced Research
Group
RIM
905-629-4RIM x4347
____________________
This
information is transmitted on a "WITHOUT PREJUDICE" basis. It is intended
only for the person or entity to which it is addressed and may contain
confidential and/or privileged material. Any review, retransmission,
dissemination or other use of, or taking of any action in reliance upon,
this information by persons or entities other than the intended recipient is
prohibited. If you received this in error, please contact the sender and
delete the material from any computer.
-----Original Message-----
From: Keith Swenson
[mailto:]
Sent: October 7, 2003 1:37
PM
To: Michael Shenfield; ASAP (E-mail)
Subject: RE:
[asap] questions/concerns on initial working draft doc (and
scope)
The whole point about ASAP is to set up a 'contract' to
have something done and to be able to check on its progress. [Michael
Shenfield] Re: "be able to check on its progress":
Then the name is misleading. The spec should either
be limited to dealing with async Web Services that support
monitoring (I see it as a small fraction) or to make monitoring
optional. From the wireless standpoint perspective monitoring is a waste
of valuable resources, for the "wired" client it could be useful but
requires much more complex implementation of a Web Service. As
already pointed out, the scalability will take a major hit as Web Service
instance cannot be stateless anymore.
Perhaps more
important is the ability to change your mind and cancel it. [Michael
Shenfield] Again, it's up to the group, but I don't think of it as a
solid requirement. These are all things that need a
"correlation" id of some sort. What ASAP oes it provide a simple way
to request something to be done and receive a correlation id that can be
used to check onthe progress.
There are many use cases that fit this model, but there
are also many that do not. A "stateless session bean" is a pattern
that ASAP is not designed to support. Instead, ASAP is designed for
the "stateful session bean" where you have a unique instance with an ID,
and subsequent requests go back to that same instance. This session
bean is endowed with a goal, and it is expected to last until that goal is
accomplished, and then a little while after that (see the timeout
setting). The goal could be "search four house listings for 3
months" or something abstract like that. [Michael
Shenfield] The general agreement in a Web Services J2EE world is
that Web Services should only be exposed by servlets or STATELESS session
beans - there're tons of references on the topic. I am not aware of
any implementations of Java Web Services running as stateful session beans
(reasons: scalability, thread safety, management/sharing issues,
etc.)
For many implementations of ASAP, the factory will be like
a stateless session bean in that it is a web service that simply "exists"
and the starting and stopping of it is invisible to the outside
world. The "service instances" then have a lifecycle that can be
externally monitored.[Michael Shenfield] I don't have any problems with a
COMesque "CreateInstanceRq", I've used it for years in COM, as well
as Home interfaces' "ejbCreate" in EJB. I just don't see Web Services
following the same pattern, being involved in many Web Services activities
I haven't seen many that mandate user managed instance creation. The other
reasons are as per previous comments.
The concept of "fire and forget" (uncorrelated
messages)[Michael Shenfield] "fire and forget" mode is a
user experience term. The message request and response(s) share same
correlation ID.
is probably not what ASAP is designed to support. This
would be a single call that later returns a result (or multiple
results). This seems like it can be handled sufficiently with
(reliable) asynchronous messaging. In this case you do not need a
correlation id, because there is no other communications with the
instance. As you point out, these do not need monitoring or other
support. Lets discuss this.
As for the concept of "multiple async responses", this is
quite definitely what ASAP is designed for. There is no reason that
ASAP is restricted to a single response at the end. Once the
correlation key is set up, the services on both sides can make further
requests to each other. The simplest way to implement this is every
new result causes a change notification event back to the observer.
The event might (or might not) carry the new data. If the event does
not carry the data, the observer can do a "GetProperties" in order to
fetch the changed data. The former implementation is most likely to
work with a generic ASAP implementation, but people are alowed to extend
the ASAP protocol to include special methods, e.g.
"NewSearchResult". In this case the ASAP sets up the basic instance
which represents the agreement between the client and realtor, and then
the services can inspect the interfaces supported, and if they both
understand this extended command, they use it.
Use of ASAP in the above example would allow the client to
modify the type of house they are looking for (after receiving a bunch of
uninteresting results) or terminate the search (in case they find their
dream house).[Michael Shenfield] see comment
#2
Terminology is always a big problem. We used to call
them "interfaces" and now we call the "port types". I think that our
usage of "method" should be "operation" now, right? [Michael
Shenfield] No, method in your definition means "message".
Operation is defined as a set of Requset, Response, and
Notification messages
If you see things
like this that obvisouly deviate from the SOAP norm, please bring these
up. Our goal is to be consistent. "Context data" & "result
data" on the other hand is a concept unique to ASAP for which there is no
other term (unless I have missed something). [Michael
Shenfield] Web Service operations are defined using Request and
Response (if any) messages. I think "Request data" and "Response data" if
you want to define parts (or group of parts) will be more
consistent.
These are parts of messages, but the key is that these structures
persist in the service instance, and can be queried and (in some cases
modified) by later operations.
BTW: We are designing Wf-XML as an extension of ASAP, and
in that the Factories themselves can be created and started. This
maps into the BPM ability to define and start new business
processes. ASAP is extended to add the new methods for the 'factory
lifecycle'.
-Keith Swenson
(408)
456-7667
-----Original Message-----
From:
Michael Shenfield [mailto:]
Sent: Tuesday, October 07, 2003 9:45 AM
To: ASAP (E-mail)
Subject: [asap]
questions/concerns on initial working draft doc (and
scope)
Hi all,
At the last conference
call we had a discussion on taking the initial working draft to the
general public. I suggested 'no' based on my initial impression from
scanning through the document 2 weeks ago. Finally, past weekend I was
able to review the document in detail. Here's the list of questions and/or
concerns that I would like to post here for your review:
- Factory "resource".
I am not
sure about the concept of "factory" and passing observerURI as a part of
CreateInstanceRq. Let's take Java Web Services as an example. When a Web
Service deployed as a servlet the hosting Web Server container plays the
role of the "factory" but this is "invisible" for the caller. The caller
posts to the WS endpoint URI an HTTP request that contains the SOAP
message to be invoked. Lifecycle management of the servlet (as a WS
instance) is not covered by invocation protocol and therefore "creation"
of the WS instance is non of the caller's concern. Similar model stands
for the Web Services deployed as stateless session beans, the EJB
container hides the creation process. For all async Web Services deployed
at RIM (to support application initiated push to a wireless device) we
have an "observer" URI (or rather requestor endpoint) passed as a part of
the Web Service message. The requestor URI could also be submitted at
registration (optionally with a correlationID), so that all following
correlated and uncorrelated notifications could be posted to the requestor
endpoint. This opens the stage for the next question:
- Multiple async responses for a single request or
uncorrelated notification messages.
One of the
possible deployments for async Web Services is when a single request
causes a number of responses. As an example, one of our demo Web Services
"Realtor" uses this model to send listings to a real estate agent
registering a new client. The model is obviously asynchronous as we don't
want our real estate clients wait (and pay for the airtime) till the
request is processed. The "client registration" message is a request that
carries the client search information, requestor endpoint URI and
generated correlator ID to be used for responses. The server is doing the
search in MLS and asynchronously sends found listings each in a separate
SOAP message (here we have multiple responses for a single req). Moreover,
when a new listing is posted to MLS the server is searching for interested
clients and sends the listing to their agents as a notification
(correlated notification model). Additionally, the Web Service is capable
of sending uncorrelated notifications. We implement the following model:
when real estate agent is logging in, the Realtor Web Service gets
endpoint URI of the service proxy (linked with the device) and when a new
information for the agent is available in the server (e.g. viewing
appointment), the Web Service sends uncorrelated notification to the
endpoint URI. In my opinion both models (multiple async responses and
uncorrelated notification) should be included in the scope of ASAP
- Monitoring support.
In the
document there is a lot of discussion on monitoring the execution process
and changing the state/properties during execution. I am not sure if this
should be mandatory for compliance, one can perfectly deploy the async Web
Service without support for monitoring or state change (see example
above). Moreover, as an implementer I am not clear on how to implement
this monitoring and state altering model in a scalable Web Service - in my
opinion it contradicts basic scalability considerations of J2EE or .NET
deployment. I think that ChangeState, GetProperties, SetProperties, and
StateChange msgs should be changed to optional. As for Subscribe,
Unsubscribe, and Completed, I am looking forward to the discussion on how
it should be defined in the ASAP context.
Introducing monitoring of the execution is identical to
polling for the completion/result and kills async efficiency. I understand
the async paradigm as the model where requestor "calls and forgets". The
implementation (and ASAP spec) should guarantee that the requestor gets
the results (or notified on completion) when ready. Am I missing something
here?
- Introduction Section. Language and terminology.
It looks like the verbage of section 1 intentionally
deviates from the standard Web Services terminology and definitions. Is
there a reason for this? Why should we call SOAP messages as "methods",
Request data as "Context data", Response Data as "Results data", etc.
Again, I might be mistaken here but my impression was that ASAP is
strictly about Web Services?
Sorry to send this 20 mins prior to the conference call,
but as I didn't see posted agenda we can as well discuss these
points.
Cheers,
Michael Shenfield
Advanced
Research Group
RIM
905-629-4RIM
x4347
____________________
This information is transmitted on a "WITHOUT PREJUDICE" basis. It
is intended only for the person or entity to which it is addressed and may
contain confidential and/or privileged material. Any review,
retransmission, dissemination or other use of, or taking of any action in
reliance upon, this information by persons or entities other than the
intended recipient is prohibited. If you received this in error, please
contact the sender and delete the material from any
computer.
To unsubscribe from this mailing list (and be removed from
the roster of the OASIS TC), go to http://www.oasis-open.org/apps/org/workgroup/asap/members/leave_workgroup.php.
