Re: [wsbpel] invoke/receive-callback race conditions, even in the1.1 sample (p.22) - Assaf Arkin - org.oasis-open.lists.wsbpel

18 messages in org.oasis-open.lists.wsbpelRe: [wsbpel] invoke/receive-callback ...

From	Sent On	Attachments
Eckenfels. Bernd	Jul 4, 2003 8:42 am
Frank Leymann	Jul 5, 2003 9:25 am
Eckenfels. Bernd	Jul 6, 2003 8:14 am
Francisco Curbera	Jul 6, 2003 5:50 pm
Eckenfels. Bernd	Jul 7, 2003 12:41 am
Francisco Curbera	Jul 7, 2003 5:47 am
Monica J. Martin	Jul 7, 2003 6:05 am
Eckenfels. Bernd	Jul 7, 2003 7:11 am
Francisco Curbera	Jul 7, 2003 7:32 am
Yaron Y. Goland	Jul 7, 2003 11:00 am
Eckenfels. Bernd	Jul 7, 2003 11:22 am
Yaron Y. Goland	Jul 9, 2003 10:33 am
Assaf Arkin	Jul 9, 2003 10:52 am
Eckenfels. Bernd	Jul 10, 2003 2:46 am
Maciej Szefler	Jul 10, 2003 12:20 pm
David RR Webber - XML ebusiness	Jul 10, 2003 12:45 pm
Rajesh Pradhan	Jul 11, 2003 5:20 am
David RR Webber - XML ebusiness	Jul 11, 2003 12:12 pm

Subject:	Re: [wsbpel] invoke/receive-callback race conditions, even in the1.1 sample (p.22)
From:	Assaf Arkin (ark...@intalio.com)
Date:	Jul 9, 2003 10:52:50 am
List:	org.oasis-open.lists.wsbpel

Yaron Y. Goland wrote:

If all the messages are going to the same partner then the only race condition is that partner foo has returned the sendShippingPrice response and so now moves on to make the sendInvoice request. But if the process hasn't processed the response to sendShippingPrice before foo sends sendInvoice then foo will get an error message.

While the process performs these activities strictly in serial order, there is nothing here to preclude the partner from a) potentially doing them in parallel or b) treating them as separate activities (e.g. two different processes). In the later case the partner may end up sending them in any order.

I actually find this a very natural way to define a process in which the understanding of the modeler is: 1. I need response to sendShippingPrice so I can do something with it 2. I need input for sendInvoice so I can do something with it 3. Exact order in which messages are received is not material

From the perspective of the modeler this is very natural. I may call two people (or even the same one) and ask them to fax me some information and handle that information in some pre-defined order. If one is faster to response, or the same person has some information available and needs time to retrieve other information, I may end up with the faxes being received in order in which they were produced, different from order in which I expect to handle them.

Instead of calling one of them and saying "I'm not ready yet, please send later", I would probably put one of them in my inbox until I can attend to it.

Rather than write <receive> activities based on the timing in which messages may be recieved (low-level) I would expect people to model processes using <receive> activities based on the timing in which messages are to be processed (high-level). So I find that the expectation to have some form of queuing mechanism there to be consistent with the natural business-oriented modeling of the process.

arkin

There are three ways to solve that problem that I can think off: #1 - Put in a queue that will hold messages for a predetermined period to see if an operation that can accept them is activated and if not time the message out and return an error. #2 - Immediately return an error stating something like "That operation is not available yet, please try again later." #3 - Use a signaling system that lets the process send a signal to foo confirming that the response to sendShippingPrice has been successfully processed and the process is now ready to continue to the next step.

#1 is a combination of a configuration problem (which should probably be handled in the WSDL) and an error specification problem which I discussed in a previous letter and said BPEL probably needs to solve.

#2 is just the error specification problem I have already discussed.

#3 requires the introduction of a reliable messaging protocol with signaling capacity which is outside the scope of BPEL but could certainly be built underneath BPEL.

What other solution would you suggest?

Yaron

-----Original Message----- From: Eckenfels. Bernd [mailto:B.Ec...@seeburger.de] Sent: Monday, July 07, 2003 11:37 AM To: wsb...@lists.oasis-open.org Subject: RE: [wsbpel] invoke/receive-callback race conditions, even in the 1.1 sample (p.22)

Hello Yaron,

for the service provider, who gets an WS invocation, there is no way to know, when he is allowed to answer to that request. In best case he will answer immediatelly after the SOAP call is completed, worst case he wont answer at all.

Asuming the best case: It is not an error for a web service to answer fast. If the web service happens to answer faster than the requestor's engine enabled the next activity, then of course the requested service will get an soap fault like "no correlation, your callback is unexpected".

This will most likely confuse the responding service provider, and even more, it will break the process, since the answer is lost. The requestor's receive for the callback will time out, since it was already answered before it was enabled.

This is called a race condition, and quite common in programming. For example in java it is a deadly sin to use notify() to wake up an thread, without also setting a boolean flag, to avoid the thread to fall asleep at all. Same situation. In java you have to use synchronized, in BPEL there is no formal way, thats a lack in the spec. There is no atomic operation (specified) which allowd to invoke/receive.

Greetings Bernd

-----Original Message----- From: Yaron Y. Goland [mailto:ygol...@bea.com] Sent: Monday, July 07, 2003 8:17 PM To: wsb...@lists.oasis-open.org Subject: RE: [wsbpel] invoke/receive-callback race conditions, even in the 1.1 sample (p.22)

I can't help but wonder if this isn't an issue in error handling. The issue described below seems to specify a case in which someone called an interface before they were supposed to. I can't help but wonder if the correct behavior wouldn't be to return a generic SOAP error saying 'Sorry, that operation isn't currently available.' HTTP had standard errors for this sort of thing.

Yaron

-----Original Message----- From: Francisco Curbera [mailto:curb...@us.ibm.com] Sent: Monday, July 07, 2003 7:49 AM To: Eckenfels. Bernd Cc: wsb...@lists.oasis-open.org Subject: RE: [wsbpel] invoke/receive-callback race conditions, even in the 1.1 sample (p.22)

Bernd,

I agree - a crisp statement of your concern will help deal with it.

Paco

"Eckenfels.

Bernd" To: <wsb...@lists.oasis-open.org>

<B.Eckenfels@seeb cc:

urger.de> Subject: RE: [wsbpel] invoke/receive-callback race conditions, even in the 1.1 sample (p.22)

07/07/2003 10:25

Hello Francisco,

every abstract programming language has a well defined set of semantics, especially in the presence of concurrency. For example, if you

look at the

java memeory model, there is the requirement, that you cannot (and therefore must not) asume, that concurrent access to a freshly modified object works. (Interesed parties might look up the Double

Checked Locking

pattern for such an miss asumption).

The same is true for a sequence. If we want to allow BPEL scripts to run in the way the samples shows, then our engines must execute steps

in infinite

fast time. This is not a useful requirement for runtime engines and will result in engines who do ugly hacks to work around the problem (i.e. waiting while accepting not-found correlation events) or ignore the problem, which will result in BPEL scripts which suceed on one

engine, and

fail on the other.

I will follow Minicas advice and raise a formal issue here.

Greetings Bernd

-----Original Message----- From: Francisco Curbera [mailto:curb...@us.ibm.com] Sent: Monday, July 07, 2003 3:04 PM To: Eckenfels. Bernd Cc: wsb...@lists.oasis-open.org Subject: RE: [wsbpel] invoke/receive-callback race conditions, even in the 1.1 sample (p.22)

Bernd,

The meaning of the example you mention is that messages will be sent and consumed in a particular order. I think it is clear that the

specification

must assume that "the runtime engine behaves correct in this

situation"; I

don't know what else could be assumed.

The point you are making seems related to the set of assumptions we are (implicitly) making about the capabilities of the runtime environment in which BPEL-e will execute. Any language makes this kind of assumptions, more or less explicitly; languages become higher level (and more useful) often by pushing additional capabilities into the runtime. I

think we have

to assume that the runtime will be able to cope with things as simple as an invoke-receive sequence; otherwise we would make modeling the simplest message exchange protocol an exercise in concurrent programming

- I don't

think we want to go there, but that is just my humble opinion.

Regards,

Paco

This is different from writing a program that will

"Eckenfels.

Bernd" To: <wsb...@lists.oasis-open.org>

<B.Eckenfels@seeb cc:

urger.de> Subject: RE: [wsbpel] invoke/receive-callback race conditions, even in the 1.1 sample (p.22)

07/07/2003 03:55

Hello Francico,

But, again, it is a statement about the implementation environment, not about the correctness or

validity of the

process.

If common business process modelling asumes, that the runtime engine behaves correct in this situation (and I dont see a clean way for the runtime engine to do that), then the process is valid _only_ if the spec allows this asumption. I dont see any implementation constraints in the spec, and therefore the sample is highly unreliable.

Note: possible solutions for fixing this problem in the engine includes: reordering of activation (receive before invoke), pre-traversal

of soon to

be enabled activities to make correlation entries before invoke is commited, delay negative acceptance of incoming web service

requests with

unmatched correlation, till engine signals "stable" state (all activites activated), ...

All of them are hard work and not clean, since dealing with timeouts and making the system less perfomrant. Therefore personally I would

vote for a

syntactic regulation, not an implementation constraint.

Greetings Bernd

"Eckenfels.

Bernd" To: <wsb...@lists.oasis-open.org>

<B.Eckenfels@seeb cc:

urger.de> Subject: [wsbpel] invoke/receive-callback race conditions, even in the 1.1 sample (p.22)

07/04/2003 11:56

Hello,

in the 1.1 spec, there is a sample on p.22 featuring call back ports for the result of price calculation. This sample suffers from a

possible race

condition:

In this sample, requesting a invoice by Purchase Order, the

second invoke

adding shipping infos, and the reply waiting for the calculated invoice. So depending on the runtime engine, the receive activity might not get activated before the response is already received. Besides the

fact, that

this sample is not dealing with correlations, it might be rewritten as:

But this still is open for a race condition (as the activation

sequence of

a flow is undefined).

So, we beed to address two issues here:

a) is it required for the enactment service to support receives

which are

depending (by link or sequence) on the completion of invokes.

If yes, how

deep this dependency should be allowed. If it is not required for the enactment, then the sample is illegally constructed.

b) if we do not allow the sequencing and declare the sample as broken, then an alternative notation with a flow may be used, but also requires some sematic definitions:

'flows are activated top down, this meens if a receive activity

in a flow

is before the invoke (with no links), then engines have to

gurantee, that

the corellation entry is made before the service is invoked.'

'flows are activated, so that all ready receicves are invoked

before ready

invokes.'

We have the problem, that we can't use links to express a relationship like "if the receive has established the orrelation entry, then invoke". The other problem is, that my proposed regulation for flow does not

help with

receives, which are not directly nested below the flow. For example a

does not reliable work with the "first all receive" nor with

the "top down

activation" definition.

I started this discussion, to ensure my understanding of the

situation is

correct. If you agree, I will raise a formal issue.

Note: this is also reláted to issue 31, because generating a UUID is a requirement for the b) solution.

Greetings Bernd

	Start a set with this search
	Include this search in one of my sets
	Exclude this search from one of my sets