This approach is fraught with complexity and the ability to code
mistakes, and is why, IMO, IBM explicitly did not implement it in WSFL,
nor in BPEL. For instance, if there is a data dependency between the
to-be-joined activity and its predecessors, direct and/or indirect?
Should it then be legal to do such a "short-circuit" evaluation? |
Prasad Yendluri wrote:
Issue 6 is speaking to a process terminating without waiting for the
completion of all "unnecessary" concurrent activities.
The issue I am raising is with a specific activity continuing after the
joinCondition is met even when all source activities for the links
coming into have not completed. Specifically when an activity needs
only one of its
source activities to be completed as in the case of an implicit-join,
it seems it is unnecessary (with serious performance implications) to
require that all source activities must be complete prior to evaluating
the join condition. If there are cases that in fact desire to wait for
all source activities to complete, I think
we should make that something one needs to explicitly capture in the
join condition definition.
I think the following text in the section 12.5.1 second paragraph
capture the spirit of what I am calling for:
"If the explicit join
condition is missing, the implicit condition requires the status of at least one incoming link to be positive".
The solution to this could be as simple as adding an attribute that
qualifies joinCondition. Should the be joinCondition evaluated only after all source activities are complete
(default?) or each time the status of an incoming link changes (a la
@suppressJoinFailure). Enabling this would be very beneficial IMO.
Dieter Koenig1 wrote:
IMO, the text in 12.5.1:
"If an activity that is ready to start in this sense has incoming links,
then it does not start until the status of all its incoming links has been
determined and the (implicit or explicit) join condition associated with
the activity has been evaluated."
DOES constitute a control dependency.
Continuing when the final outcome of the join is determined but before all
source activities have completed, is IMO something we do not have so far
and addressed in issue 6.
>From: Prasad Yendluri <email@example.com> >To: firstname.lastname@example.org
>Subject: Re: [wsbpel] Incosistency in Link Join condition Evaluation specification
No where in the spec we say control dependency exists regardless of
joinCondition. So, we have scope for interpretation and so, things look
fine to you when they did not to me. We need to minimally correct that.
Let us look at what it really means for the control dependency to exist
regardless of the joinCondition. So I want to wait for all source
activities to complete (successfully or not) and I want to proceed only if
the join-condition evaluates to true when it is implicit join or OR join on
all links. That is fine but this is modeling a very special case IMO,
generally we don't wait for things to complete, the success or failure
status of which we don't care about. Even if one cared one should perhaps
capture that via an explicit AND joinCondition rather than rely on an
implicit syntactic sugar.
Where as more simpler (perhaps common) case would be to wait on any one (or
n < total) of the source activities to complete successfully (like a
barrier in threads). How can one model that?
It seems we have an issue here.
Danny van der Rijn wrote:
Things look fine to me. For (2) and (3), you have to realize that a
control dependency exists regardless of the joinCondiiton. If you
don't want the control dependency, don't draw a link. (4) is
referring to an explicit piece of code (the Initial Example) in which
your stipulation ("...is only one of the incoming links...") is not
correct, since there is only one link target.
Prasad Yendluri wrote:
We have the following text dispersed in sections 12.5.1 and
12.5.2 that results in an inconsistent specification of
evaluation of Link Join conditions. The spec states:
(1) Section 2.5.1: Every activity that is the target of a link
has an implicit or explicit “join condition” associated with
it. If the explicit join condition is missing, the implicit
condition requires the status of "at least one" incoming link
to be positive (see below for an explanation of link status).
So, the implicit join is an "OR" and the join condition
evaluates to true as soon as the status of one of the incoming
links to the activity is positive.
(2) Section 2.5.1: If an activity that is ready to start in
this sense has incoming links, then it does not start until the
status of "all" its incoming links has been determined and the
(implicit or explicit) join condition associated with the
activity has been evaluated.
If it is an implicit join why should the activity not start as
soon as one (or more) of its incoming links goes to a positive?
Why should it wait until the status of all incoming activities
has been determined?
(3) Section 2.5.1: For each activity B that has a
synchronization dependency on A, check whether:
o B is ready to start (except for its dependency on
incoming links) in the sense described above.
o The status of all incoming links for B has been
Again why the requirement to wait until the status of all
incoming activities is determined?
(4) Section 12.5.2: If one of these invocations were to fault,
the status of the outgoing link from the invocation would be
negative, and the (implicit) join condition at the target of
the link would be false, but the resulting bpws:joinFailure
would be implicitly suppressed and the target activity would be
silently skipped within the sequence instead of causing the
If the link that was set to false (due to the failure) is only
one of the incoming links to an "implicit join", shouldn't the
join condition evaluate to false only when all other links to
the join also evaluate to false.
It seems either the parts related implicit join in (2),(3),(4)
above or incorrect or (1) itself was not correct. I believe 1
Either way it seems we need to fix this. Comments?