Interesting.
I want to make sure I am
still following the main thread. Are we saying that:
- We want to have reusable
BPEL
- Since the unit of
reusability needs to also follow conventions of compensation,
reusability is along the lines of “scope”
- We want to be able to
deploy these reusable scopes anywhere (no restrictions?)
- “invoking” these scopes
should be as easy as with regular scopes (no complex mapping of state,
like with assign)
Close?
++harvey
-----Original
Message-----
From: Ron Ten-Hove
[mailto:Ronald.Ten-Hove@Sun.COM]
Sent: Thursday,
October 30, 2003
11:54 AM
To: ygoland@bea.com
Cc:
wsbpel@lists.oasis-open.org
Subject: Re: [wsbpel]
Issue - 2-
requirements for a sub function solution
Yaron,
The first time I read the BPEL4WS 1.0 spec, one of my
strongest reactions was "where are the subprocesses?" Over a year
later, one revision of the spec, and countless re-readings, and my
initial
reaction is unchanged.
Modelling subprocesses as separate web services is an
unsatisfactory substitute. This creates unreadable processes, with, as
you
observed today, really messy <assign> activities used to pack and
unpack
requests and responses. Not only are the assignments ugly, cluttering
the
process, but the dedicated BPEL reader will be faced with
1.
Discovering that an <invoke> isn't
"real" (involving a proper partner) -- there is no clear way to
distinguish such service invocations from subprocesses, except by
tricks like
naming conventions.
2.
Decoding the initial <assign> to
figure out
what the input parameters are for the subprocess.
3.
Decoding the trailing <assign> to
figure out
what the output parameters are for the subprocess.
This
doesn't sound like
the best of modelling approaches, and certainly makes run-time
implementations
more difficult. Can we do better?
Assuming that we can agree that sub-functions, as you have
defined them, can be used as a simple form of sub-process, then we may
be able
to converge on a satisfactory solution.
The rest of my comments are in-line:
Yaron Goland wrote:
Here is my view on what the requirements are for a solution to Issue 2.
Terminology:
Host - This is the process that called the sub-function.
Requirements:
Integration into host fault/compensation model - The sub-function MUST be
able to throw faults that will propagate to the host. The host MUST be able
to call compensation handlers defined in the sub-function.
I
would add that
sub-functions must integrate into the host's serializable scopes, when
shared
variables are passed by reference to the sub-function.
Functional Programming - One of the key concepts of functional programming
is the idea of isolation. Functions in functional programming systems are
expected to be able to run in their own context and to only have contact
with the host context through explicit value passing. The sub-function
solution MUST enable for isolation.
Human Usable - It MUST be reasonably easy for a programmer to read/write
sub-function definitions and calls without requiring the aid of a tool.
I
believe we established
this as a requirement for BPEL very early on, in the debate over
whether <sequence>
was necessary.
By Reference - Messages seem only to get larger with multi-megabyte XML
messages now commonly seen. Therefore when passing data into a sub-function
it is critical that it be possible to hand over the data by reference so
that the data does not have to be copied on its way in or out.
While
passing by
reference has good, pragmatic foundations in managing the impact of
those
ever-growing XML payloads, I believe another line of argument would
support the
need for clear, controllable parameter passing to the subfunction, such
that
the process author can clearly control and understand how data are
passed to
and from the sub-function.
Use in Expressions - It MUST be possible to call a sub-function anywhere one
could call a XPATH or other type of general/query/date expression and have
the sub-function return the desired value.
That
is an interesting
idea. Such sub-functions would need to have restricted "signatures"
to map to XPath functions; also, isn't XPath supposed to always be
side-effect
free? That would imply that all input variables of the sub-function
must be
read-only; a single output variable would be nominated as the function
result.
Import - A mechanism is needed to allow sub-functions to be defined in stand
alone files which can then be referenced by multiple BPEL processes.
Agreed.
-Ron