Edwin,
Thanks very much for the
thoughtful reply. You've raised some very good points; the one I most
appreciate is the reminder that there are costs and benefits for all the
language features we create, or, I would add, exclude. Users will bear some of
the costs, after all.
My comments are in-line with
yours, below:
Edwin Khodabakchian wrote:
We should be driven by uses cases NOT analogies.
In this case case, the use case is about re-use. As pointed out, there are
multiple ways to address that need:
(1) Encapsulation of the logic to be re-used in a
seperate process
(2) Let the tooling address this need through
templates
(3) Create a new construct at the language level
(called subfunction)
It seems that the only benefits of (3) over (1), (2) or
a combination of (1) and (2) is that the use of subfunction signatures can
having to avoid creating a new message type and the assign activities needs
to initialize and read data from the message
types.
This is not an insignificant advantage. Crafting new
message types, and writing "gnarly" assign activities are not simple tasks,
and can be regarded as "overhead" getting in the way of the process author.
Put another way, our domain-specific language requires a lot of solution
domain elements; sub-functions could serve reduce the number of solution
domain elements the author need worry about.
[edwink] I think that good tools will hide this complexity. In
both cases, you need a mapping between the variables defined in the
process and the signature of the subfunction. This mapping can be either
defined in the assign activity or inlined in the subfunction. So the
benefit from the process author is not the mapping (because that
part is not going away) it is only the type definition (XML schema part)
and that can be absorbed by the tool.
There might be also an argument about (3) offering
better support for privacy and packaging and deployment.
But there are also drawbacks/costs with (3) that need
to be weighted:
(1) Complexity: to really deliver on
the benefit of sub-functions, we need to replicate a very large number of
concept and constructs already existing in the process and scope
elements.
I think your general point about increased complexity is true, but I'm
not sure about the need to replicate much of what already exists. Yaron's
proposal seemed a very straight forward extension of those concepts, not
duplicative. Perhaps you were referring to implementations, rather than the
language itself?
[edwink] I am only talking about the
complexity of the language, NOT the implementation.
(2) Lack of communication
channel: we need to create a communication mechanism between
the subfunction and the calling activity so that the subfunction can report
its status. This last need is much better addressed with option (1) because
already supported by scope event handlers and receive activities.
An interesting point. I regard the proposed sub-function invocation
mechanism as a simple extension of that used by existing activity types, such
as <invoke>. Variables are passed by reference, either as input, or
output. The chief difference is cardinality -- subfunctions allow multiple
input (and output?) variables. The general mechanism, from the user's
perspective, is the same.
I'm not sure what you mean by a subfunction
reporting its status. Are you envisioning some sort of asynchronous
subfunction invocation?
[edwink] The difference I am pointing to here is that the call to
the subfunction will be a blocking call and the subfunction has no way to
report its status to the caller (by status I mean: I am at
subfunction.activity1, subfunction.activityN, I am about to finish, etc...On
the other hand, it the re-used code fragment is instead modeled as
a process, it can notify the caller and the caller can handle
those notifications using eventHandlers, pick/onMessage or simply
receive.
This is an area where BPEL is very different from
traditional flow languages and this is why I am not sure that your "other
process language do this" argument is valid.
I don't think BPEL is so different from existing languages, just
simplified. The adoption of an abstract messaging model for work performers
and requesters does eliminate (or move elsewhere) some of the complexities
found in other languages. In some other languages it is theoretically possible
to model other processes as work performers/requesters, but this, as far as I
know, has never been carried to the extreme of requiring that it be done so.
Regardless, I don't think BPEL is so different that the need for for a
compositional mechanism supported by the language itself has been
eliminated.
[edwink] I do not agree with you on this. BPEL is
both richer and different. Richer because it has much better support
for exceptions, events and transactions. Different because it is about
the coordination of message exchange patterns (NOT work management).
(3) Lack of
consistency:
This is a software point but I find the passing
arguments to the subfunction not consistent with the rest of the spec: you
need to create a new subfunction interface definition language (not WSDL),
you need to learn that sometime you can do assign (when using invoke,
receive, pick) for data manipulation and sometime you can plug expressions
directly in the signature (this is at least what Yaron is
suggesting).
Right. There would be a new concept, the subfunction, which is invoked in
a syntactically different, seemingly more direct way than services. This
distinguishes subfunctions from (external) services, but at the cost of the
user knowing the difference. The actual "pass by reference" semantics would
remain the same, keeping things consistent at that level.
[edwink] This goes back to my point about complexity. I am not
sure of what you mean by pass by
reference.
I believe that the single most important success factor
for BPEL to succeed is to remain simple (which is not trivial given that we
inherit the complexity of XML Schema and some of the exotic variations of
WSDL). Given that there is a significant
complexity "cost" associated with adding this feature to the language,
I would like to recommend following Ugo's advice and finding use cases where
the problem can not be addressed using a combination of (1) and (2) and
therefore require us to go the extra mile of spec'ing
out (3).
Unfortunately we have already established that we will not simply reduce
BPEL to the simplest language possible. (I find this minimalist approach very
attractive, but this has been rejected by the TC. BPEL is a modelling language
as well as an executable one. This will increase language complexity /
clutter.)
I still find Ugo's suggestion (that you refer to) as being
decidedly biased, and it certainly misses the mark in this issue. As you
mentioned earlier in your note, the question is about reuse: what language
support for reuse should we have? This affects readability, composability, and
(less directly) the granularity of reusable "components" that are feasible.
None of these factors are examined when simply looking for use cases that
cannot be addressed.
Your point about cost/benefit
trade-off is well taken. Adding more domain-specific concepts to a language
makes it more complex to implement, while making it easier for end-users to
utilize. Conversely, reducing the number of domain-specific concepts makes it
harder for end-users to use the language, but makes implementations
easier to realize. There is a balance to be achieved, not by always favouring
minimal language features, but, as Satish reminded us, by using our judgement.
How are we to judge the cost/benefit of language-based support for
reuse?
[edwink] I am not sure that I am biased here. At the end of the
day, we should be able to translate every aspect of the language (and the
conceptual complexity it entails) to a requirement/use case/something the
developer wants to do with solutions based on BPEL. And then make trade
offs. The tradeoff I am suggesting we make here is to promote (1) and (2) as
the best practice for reusing process logic instead of introducing the notion
of subfunction.
Cheers,
Edwin
Best
regards,
-Ron