Hi Danny and Najeeb,
[I]
I partially disagree. BPEL 2.0 Section 9.2
says:
Good catch. Invoke will be an exceptional case. I agree with you.
I do agree, though that those faults are only
thrown internal to the
BPEL process, and are not externally visible.
It's good that we are on the same page.
[II]
... or the instance matching the correlation
*has yet to be started.*
Ok with adding your addtional statement.
"... or the process instance that potentially matches the correlation
has yet to be started."
I will consider this as a friendly amendment.
[III]
This is the statement that I just don't buy as
defining this issue
worthy of standardizing. I'd like to have some more discussion about
when/how the BPEL infra can make such a determination. I posit that
it's enough of an edge case that it would be unwarranted to standardize
what to do in this case. To restate myself, this standard is about the
combination of SCA and BPEL. If we're positing behavior that is
outside the scope of the BPEL standard (IMO *way* outside), there is no
place in the BPEL C&I spec for how such a situation should be
handled.
To say it yet another way, you're attempting to create a standard way
of dealing with an event that many compliant (BPEL and SCA BPEL
C&I) implementations will never see. You're trying to create a
standard fault that we can't create a compliance test for.
Actually, you read my mind here. :-)
At the very beginning, when Najeeb wanted to open this issue, I have
quite a bit of hesitation. I'd rather not opening this can of worm.
But, it seems to me that I failed to convince Najeeb. The bar of
opening a new issue is intentionally low. So, I did not want to oppose
opening this issue that much.
I agree with you partially here. Without referring to a concrete
protocol (e.g. some version of WS-BA), the scope of the condition for
the infrastructure to raise this fault is:
(a) very small: I am referring to the "will never be matched"
condition and "a request-response
operation" bullet
(b) very abstract: I am referring to "one-way operation" and "this error situation MAY
be notified" bullet.
That's why I tried to convince Najeeb not to open this issue. Because,
what the SCA-BPEL spec can do here is very limited and the proposal
value giving people a sense of direction.
At the same time, once the issue is open, I want to make sure the spec
describes some correct and desirable behavior (allowing implementation
to pick a right protocol), even if the behavior description is
abstract.
How are you proposing this would occur?
Timeouts are specified on a
client. Providers don't send timeout faults. This fault is, in
essence, a statement of "I bet you're tired of waiting. Instead of
waiting for YOU to give up, I'm giving up on your behalf." While
possibly more informative, it's bizarre, and, I would say, incorrect.
Copied from the proposal:
If the message endpoint address of a
business
process leverages or participates a (transport or above transport
level) protocol that has a timeout or expiration duration value
specified,
That is yet another abstract part of the proposal. First of all, there
is a "IF" at the beginning - i.e. IF a protocol is used and IF a
protocol has a timeout/expiration feature. If such a protocol is used,
the initiator of the protocol (transport or above transport level) will
typically specify the timeout. (An example is Expiration feature in
WS-Coordination).
[IV]
Well, OK, I can buy what you said there. That
an active, waiting, IMA
can be uniquely identified by a correlation set. But I'm not sure what
that buys you in this conversation. All you may know is that no IMA is
*currently active* for a particular incoming message. I know you're
not suggesting that that's enough to trigger a dead-letter situation.
What more information would you need? Is that information anywhere in
either standard?
Actually, I still fail to understand the point you are trying to
express about whether BPEL CS identifies a process instance. IMHO, that
question has no direct impact on this issue and its proposal. :-)
To me, the core of a dead-letter message is about whether an incoming
message can be matched and dispatched to a process instance. BPEL CS is
just a part of the formula to match a message to a process instance.
:-)
[V]
Right now, I can see there are 2 options:
(1) close the issue with no change
(2) close the issue with the supplied proposal (maybe with some minor
friendly amendment, including dropping the protocol timeout bullets)
And, I am happy with both options, even I spent quite some time in
drafting the proposal.
Thanks!
Regards,
Alex Yiu
Danny van der Rijn wrote:
47574538.8090809@tibco.com" type="cite">
I blame my email client ;-)
Danny van der Rijn wrote:
4757448E.2090504@tibco.com" type="cite">
Najeeb - here's my argument. I'd like to talk with you about it before
I send it out. Perhaps my having put it in writing will help the
conversation between you and me? If we don't talk about it before I
leave today, I'm going to send it out, so it's on record for tomorrow's
call.
Danny
I still think that this proposal is badly flawed. Comments below.
Alex Yiu wrote:
Hi all,
Here is an amended proposal (take 2) for this Issue 3.
Let me repeat a few points here:
- "bpel:correlationViolation" is for BPEL's CorrelationSet
lifecycle violation - not CS mismatch. I am not sure we want to
overload it. All existing "bpel:*" fault thrown are for internally
consumption only. Not directly visible through BPEL partnerLink.
I partially disagree. BPEL 2.0 Section 9.2 says:
When a bpel:correlationViolation is thrown by an <invoke> activity because of a violation on the response
of a request/response operation, the response MUST be received before
the bpel:correlationViolation is
thrown. In all other cases of bpel:correlationViolation, the message that causes the fault MUST NOT be
sent or received.
I do agree, though that those faults are only thrown internal to the
BPEL process, and are not externally visible.
- We need to address the difference in one-way or
request-response MEP.
The proposal has two parts: background (non-normative) text and
normative text.
My proposal is to add both parts into the spec text:
The background text will be added as a non-normative appendix titled:
"Background about Dead Letter Messages in BPEL: (Non-normative)"
------------------------------------
[background text begins here ...]
When an inbound message comes into the SCA and BPEL infrastructure,
such a message is normally consumed by a matching inbound message
activity (IMA)(e.g. a <receive> activity). However, due to
process model error or runtime message data error, there is no matching
IMA at all or a matching IMA is not enabled within the expected time
limit of the (system/business level) protocol between the message
sender and receiver. This kind of messages, which do not have a
matching IMA, are termed as "dead message messages"
Examples of process model error are:
- matching IMAs are skipped by faults
- matching IMAs blocked by other activities within a sequence
or
an
impossible-to-fulfill control link transition condition.
- IMAs cannot receive message due to incorrect usage of
message
correlation mechanism, including BPEL correlation set and SCA
conversational interface
Examples of runtime message data error are similar to above, as the
above error are not inside the process definition itself but caused by
incorrect data values.
There might not be a universal way to determine a
message is truly a "dead letter message" without any additional
protocol between message senders and receivers. Consider the following
example, an message is dispatched to a BPEL process instance by SCA
conversational mechanism. At the moment when the message is matched
with the BPEL process instance, there might be no <receive>
activity enabled for the matching partnerLink and operation at all, or
there is a <receive> activity enabled for the matching
partnerLink and operation but with a mismatched correlation set. Some
users might think this is certainly a dead letter message situation.
However, a matching IMA may be enabled minutes, hours, or days later,
as the matching IMA might be blocked in the process model.
... or the instance matching the correlation *has yet to be started.*
On the other hand, there might be some cases that the BPEL
infrastructure can determine there will never be a
matching IMA enable in future. And, some advanced features in BPEL
infrastructure (e.g. process instance repair or process definition
repair) might make the detection of "dead letter message" cases more
difficult. However, with some additional system-level protocol
coordination between the message sender and receiver, it might make
detection easier.
This is the statement that I just don't buy as defining this issue
worthy of standardizing. I'd like to have some more discussion about
when/how the BPEL infra can make such a determination. I posit that
it's enough of an edge case that it would be unwarranted to standardize
what to do in this case. To restate myself, this standard is about the
combination of SCA and BPEL. If we're positing behavior that is
outside the scope of the BPEL standard (IMO *way* outside), there is no
place in the BPEL C&I spec for how such a situation should be
handled.
To say it yet another way, you're attempting to create a standard way
of dealing with an event that many compliant (BPEL and SCA BPEL
C&I) implementations will never see. You're trying to create a
standard fault that we can't create a compliance test for.
------------------------------------
------------------------------------
[normative text begins here ...]
If the SCA or BPEL infrastructure is able to determine that a message,
that has been sent to an endpoint address of a business process, will never be matched with a
corresponding inbound message activity (IMA) (i.e.
receive, onMessage or onEvent), then:
- If
the message is sent through a request-response operation,
"sca:DeadLetterMessageError"
fault SHOULD be replied to the
message sender
- If
the message is sent through a one-way operation and additional
system-level
protocol is used between the message sender and receiver, this error situation MAY be notified to the
message sender,
according
to the protocol used.
If the message endpoint address of a
business
process leverages or participates a (transport or above transport
level) protocol that has a timeout or expiration duration value
specified, and if no IMA can be matched with the inbound message within
the timeout / expiration duration, then:
- If the message is sent through a
request-response operation, "sca:DeadLetterMessageError"
fault MAY be replied to the message sender
How are you proposing this would occur? Timeouts are specified on a
client. Providers don't send timeout faults. This fault is, in
essence, a statement of "I bet you're tired of waiting. Instead of
waiting for YOU to give up, I'm giving up on your behalf." While
possibly more informative, it's bizarre, and, I would say, incorrect.
- If
the message is sent through a one-way operation and additional
system-level
protocol is used between the message sender and receiver, this error
situation
MAY be notified to the message sender,
according
to the protocol used.
Again, why are we telling people how to deal with such a situation
that's so far outside of the standard realm?
For example, if a
message is sent through a request-response operation and HTTP is used as a
transport protocol with the timeout duration set as 60 seconds, and if
SCA/BPEL infrastructure can determine ino IMA can be matched within 60
seconds, a SCA+BPEL infrastructure might reply "sca:DeadLetterMessageError"
fault as the response to pre-empt the transport level timeout error.
Again, the 60 seconds is a value known to the client implementation.
How does the SCA/BPEL infrastructure get notified of what the timeout
is? Even if it *could* know, it's going to have to determine some
amount of time *before* the 60 seconds to send its fault, so it can be
received before the client times out. What if the IMA can be matched
in that window?
------------------------------------
Besides using the "never" wording suggested by Michael Rowley, there
are some minor fine tuning of wordings in the first half of the
normative text.
The second half of the normative text is newly added. The logic and
style are very similar to the first half. It is explicitly targeting
the "protocol time out" situation that Danny mentioned in the last
email. If people do NOT want the spec to deal with "protocol time out"
situation explicitly, I am OK to remove it.
Thanks!
Regards,
Alex Yiu
|