Thanks for the descriptions Alex. I’m
not completely sold, but let me describe in another way what you are
suggesting. Essentially it is declaring that scopes have the following lifecycle
(where all but step 5 is what we do today):
1 - Scope initialization (variables, partnerLinks,
etc)
2 - Scope schedules event handlers
3 - Scope executes Activity
4 – Scope waits for event handlers
to complete
5 – Scope validates that message
exchanges have completed (no outstanding replies for locally declared
partnerLinks or messageExchange receives)
6 – Scope completes
7 – Scope source links execute…
We already do wait for event handlers to
finish before the scope completes so checking that partnerLinks are completed
prior to “completing” seems ok. I don’t think we have a case
where a fault handler actually waits or checks for anything prior to completing
but now we are saying it would execute step 5 (it’s actually this point
that I am not sure about). If we do that then like any other fault
generated by a fault handler it would be thrown to the parent scope.
- Chris
From: Alex Yiu [mailto:alex.yiu@oracle.com]
Sent: Tuesday, December 13, 2005
6:10 PM
To: Alex Yiu
Cc: Danny van der Rijn; chris.keller@active-endpoints.com; 'Yuzo Fujishima';
'ws bpel tc'; Alex Yiu
Subject: Re: [wsbpel] Issue 221 -
discussion carried over from Issue 123
Hi, all,
[discussion carried over from Issue 123]
Ok ... let me summarize again on where to throw missingReply fault (to make
sure everyone is on the same page)
I believe that there are 2 variant of suggestions on the table:
Variant [A]: (which I am prefering)
It has two checkpoints:
checkpoint [A1]: S2 (at the end of normal execution) => S2
checkpoint [A2]: S2 (at the end of fault handler) => S1
[** A typical fault propagation flow would be:
- At the end of normal execution, a missingReply is
thrown by S2 to S2.
- Then, a fault handler (including the default one)
will be triggered, if the missingReply fault is NOT handled (aka default),
the fault will be propagated to S1. ** ]
Variant [B]:
It has only one checkpoint.
checkpoint [B1]: S2 (at the end of normal execution or fault handler) => S1
(You guys can think of [B1] is the morphed combination of [A1]+[A2] plus
modification.)
Concerns I have on Variant [B] are:
- It seems to me that S2 does not receive any fault
signal in the case of at the end of normal execution. What is the state
of S2? Is S2 considered normally completed and up for
compensation? Or, is it faulted?
- If S2 is considered
normally completed, that is very strange. A scope with an open-unfinished
receive and a missing reply should not be considered completed by itself.
- If S2 is considered
faulted, a fault signal should be sent to S2 first (which may in turn
propagate to S1, as depicted above).
- Say, if we go for this variant [b]:
- <onEvent>
scenario: S2 belongs to an <onEvent> construct. Now, one of
<onEvent> S2 instances throws this missingReply fault directly to
S1. What can S1 do? Perform the compensation on S2? But, remember we can
only do group compensation of all S2 instances? Can someone justify why
previous successfully completed S2 need to be affected?
Also, you guys can also think of scopes as components in a business process.
Component S1 makes use of component S2. If component S2 forgot to send a reply
(or make any other system-level mistake e.g. forgot to close some system
resources), it should be S2 considered at fault or faulted. If S2 can clean up
its own mess, S2 should handle it first. If S2 cannot handle, then it will be
S1's turn to handle it.
I hope I point out concerns that I have on Variant [B] more precisely this
time.
I also hope Variant [A] seems a better solution to you guys now.
Thanks!
Regards,
Alex Yiu
Alex Yiu wrote:
Hi all,
I am OK with moving this part of discussion to Issue 221 also, because I also
want to pass Issue 123 resolution in a timely fashion and the decision on where
to throw missingReplying is taking some discussion time than expected.
Thanks!
Regards,
Alex Yiu
Danny van der Rijn wrote:
I agree with Yuzo and Chris. I still think,
though, that we should save this discussion for 221. 123 only deals with
messageExchange, while 221 holds for any of the 3 values.
Chris Keller wrote:
I support Yuzo’s position completely. A missing
reply at the completion of S2 should be thrown to S1. The fact that we
are saying that the missingReply is thrown is when the partnerLink or
messageExchange goes out of scope says it all. How does it go out of
scope, it goes out of scope because the scope completed. After completion
we can’t come back into ourselves to catch a fault. All that said
I’d rather not have this delay 123. I’m ok with leaving the
text as it stands in the 123 proposal and solving the question of which scope
catches the fault in issue 221.
- Chris
From: Alex Yiu [mailto:alex.yiu@oracle.com]
Sent: Monday, December 12, 2005
11:50 PM
To: Yuzo Fujishima
Cc: Danny van der Rijn; Chris
Keller; 'ws bpel tc'; Alex Yiu
Subject: Re: [wsbpel] Issue 123 -
Proposal for vote
Hi Yuzo,
Thanks for following this discussion thread. Really appreicate it ... :-)
[a]
The main motivations behind "S2-to-S2" for missingReply fault are:
==> Finer-grain reply / compensation
- If the fault is
thrown to S1 instead of S2, S1 cannot do anything to fix the missing reply
situation. Because, one of resources (partnerLink or messageExchange) used
in the open IMA is already out of scope (S2) and S1 has no access those
out-of-scope resources anymore.
- The missingReply
situation may be caused by the situation where some other faults happened
and the enclosed scope (S2) does not handle those fault properly. (Please
see more analysis of "programming-error-vs-runtime-error"
below.) ... In the <onEvent> or parallel <forEach>
situation, assuming there are 5 instances created and 4 of them got
executed successfully without any missingReply situation. But, one of them
ran into this missingReply situation. If we throw the fault to S1 and if
we want to every compensate the work of S2, all 5 of them need to be compensated.
We cannot compensate just the faulty instance by itself. On the other
hand, if we throw to S2, we can do a partial compensation in its fault
handler.
[b]
About "Reason 1" : "programming-error-vs-runtime-error"
consideration:
I agree with you this is more likely a programming error, not a true runtime
error. That is why Dieter introduced this exit-on-standard-fault facility to
allow people to stop process in case of this programming error.
With or without using this exit-on-standard-fault facility, as a complete
specification, we need to specify snapshot state of the process - i.e., what
scopes are faulted and not, when the "exit" takes place.
If you think of the <onEvent> example described above, S1 should not be
faulted, when the "exit" happens. Similarly, the previous 4 instance
of "S2" should not be faulted either. Only the true missingReply
scope should be faulted.
I also agree with you that people should not rely on catching this missingReply
fault to do any complicated business logic. At most, people should just log the
fault and attempt doing the partial compensation, as described above. (This
applies to most of standard fault catchers).
I am just promoting scope-modularity. Definitely, not promoting any weired
coding style. If you want, I think we can add a caution statement along the
intention of the passed resolution of Issue 192 (Dieter's exitOnStandardFault
proposal) to spec to state that: "If people do not set exitOnStandardFault
to true, it is a good practice to restrict logic of any standard fault fault
handler to some minimalistic repair work (e.g. error logging and partial
compensation)."
Actually, the same wisdom applies to most of conventional programming language.
E.g. in Java, if one got an IOException, SQLException, or Security permission
related Exception, there are only few things that a Java application programmer
can do / should do (similar to above).
[c]
About "Reason 2: When to judge if a reply is missing":
Yuzo wrote:
we should expect a fault handler to send a reply.
Yes, I totally agree with you here. :-)
Our intended behavior are actually very similar. :-) ... I guess the
related text in my previous email writeup is muddled. Sorry for the
not-organized enough text.
Let me add some clarification and refinement here:
==================================
There are two
checkpoints to decide whether a bpws:missingReply fault needs to be thrown in a
scope (S2):
- After the normal execution of the scope (S2) is finished
but before the scope is considered completed and the fault handlers of the
scope (S2) are uninstalled, if no fault has been thrown to the scope (S2)
during the normal execution.
- After the execution of a fault handler of the scope (S2) is
finished, if a fault has been thrown to scope (S2) during the normal
execution of the scope (S2), and that fault triggers the fault handler of
the scope (S2), and no fault has been thrown by execution of the fault
handler
The bpws:missingReply fault thrown by missingReply
checkpoint logic in
the case #1 will be caught by a fault handler at the scope (S2)
which matches bpws:missingReply fault, if there is one.
If the triggered fault handler (either
by bpws:missingReply fault or other kinds of fault) does not
generate proper replies to
any open IMAs, a missingReply fault MUST be thrown. This is the case #2 checkpoint
logic.
Important Notes:
- In case #2, there is NO an infinite loop
situation, because any faults triggered or thrown by a fault handler (e.g.
<rethrow> activity) will NOT be caught by any fault handler of the
same scope (S2) and they can only be caught by its ancestor scopes (e.g.
S1).
- If a fault is thrown either during the normal execution of
the scope (S2), checkpoint logic of case #1 will not be executed. That
implies, if such a fault happens and the scope (S2) does not have a
matching fault handler, this fault will be propagated to the parent scope
(S1) and it will overshadow overall missingReply logic.
- If a fault is thrown or rethrown during the execution of a
fault handler of the scope (S2), checkpoint logic of case #2 will not be
executed. That implies, if such a fault happens, this fault will
overshadow missingReply logic of case #2.
Here is a fragment of process definition that illustrates the relationship of
scope S1 and S2 described above:
------------------
<scope
name="S1">
...
<scope
name="S2">
...
<partnerLinks>
<partnerLink name="pl1" ... />
</partnerLinks>
...
<sequence>
<receive partnerLink="pl1" ... />
<!-- receive without reply -->
</sequence>
</scope>
...
</scope>
------------------
Note: The bpws:missingReply fault is thrown by scope S2 to scope S2 itself.
==================================
[I highlighted the clarification changes in GREEN.]
To summarsize, we agree more than disagree. :-) ....
Case #2 is where the fault is thrown by S2 to S1.
I am merely adding one more differentiated checkpoint (case #1) to allow the
fault thrown by S2 to S2 to achieve the semantics of scope-modularity and
finer-grain repair and compensation. (IMHO, that is very important for
<onEvent> cases).
I hope I have explain my intended behavior better this time. :-)
More thoughts on this topic ... ? :-)
Thanks!
Regards,
Alex Yiu
Yuzo Fujishima wrote:
Alex,
Alex Yiu wrote:
Hi Danny and Chris,
Please see inline.
...
*Q1*: /have the process' faultHandlers been
uninstalled?/
Answer: As I mentioned in the my previous email: (please see the example)
http://lists.oasis-open.org/archives/wsbpel/200512/msg00050.html
I prefer the direction of which: the fault is thrown by scope S2 to scope S2.
I am not a big fan of this "S2 to S2" semantics for
the following reasons:
Reason 1: Credibility of the use case.
I am not convinced that there are (many) cases where
scope S2 fails to reply without throwing a fault (except
missingReply). If S2 does, it should be a programming
error rather than a runtime error.
More likely are, in my opinion, the cases where
scope S2 fails to reply due to some fault.
For such cases,
S2 should define a fault handler to catch the fault and
send a reply there.
Not catching the cause fault and instead waiting for the
missingReply to occur doesn't seem to be a good programming
practice. The language should not promote that.
Reason 2: When to judge if a reply is missing.
A missingReply should be thrown only after the whole execution
of the scope S2, including the fault handlers, completed
without sending a reply. (As explained in the example for Reason 1,
we should expect a fault handler to send a reply.)
Then the natural destination of the missingReply fault should be
the parent of S2, i.e., S1.
Yuzo Fujishima
NEC Corporation