Some responses inline
On 8/24/2010 8:03 AM, Peter Niblett wrote:
Hi Eric
I have taken another look at PC
#1 and
I think I understand your analysis. Underlying all this is the
question
of what encapsulation / assembly is for and what it means for
events.
My view is that its main purposes
are
to simplify the depiction of complex event processing
applications, and
to allow common subassemblies to be reused. The most common use
case I
see is where an assembler groups together a set of components,
with well
defined producers and / or consumers, that can be then treated
as a single
component by some higher level user. The examples in the WD and
your
"easy scenario" on page 1 are examples of this kind of use case.
The approach we have in the WD is well suited to this use case,
and
I think it is important to preserve the simplicity that it
provides for
this case. In this case the composite is a way of implementing a
component
What makes the Pathological Case
different
is that
a) you have events spontaneously
produced
by components in the contained assembly, and consumed by
components in
the assembly as well as being emitted by the assembly
b) the consumer in the contained
component
consumes events (on a single channel)
It's the combination of both
these points
that makes it difficult, and I am still left wondering how
important a
case it is.
<vdR>
IMO it's a pretty standard case. To repeat my use case from the
meeting yesterday, I'll make an analogy to slf4j. Consider some
subsystem that includes D as a log sink that writes to logfiles.
'A' represents any number of components that emit log events.
Compose that with B/E where B represents more emitters, and E
represents another sink, this time to a database. And compose again
with C/F where F filters some of the events and shows them on a user
console. The point of using events for logging is exactly that you
don't know who's producing the events, and furthermore, there are so
many ways and reasons to consume them that you don't want to be tied
down to using a service.
IMO the "fir tree" is exactly correct here: the A/D component needs
to say that A and D are coupled, but doesn't care to say on what
channel binding. B/E are coupled in a similar fashion, and
furthermore, B/E are coupled to the same channel binding. Etc.
</vdR>
Of all your options, the last one
is
the cleanest, but it does raise a question, and that is whether
the composite
implements a component, or a channel - or both. As you have
drawn
it, it looks like it is a combination of both, and I am worried
that complicates
things by muddling the concepts.
<vdR>
In the last two options, I think Eric is trying to show that the
*coupling* of the producer and consumer as part of the
componentType. In this last example that you're highlighting, since
the consumer and producer are shown separated, there's certainly the
possibility of ambiguity in what this might mean in the graphical
form. If you overlay the semantics that the runtime merely treats
them as referring to the same thing, then this is just
constraining/delegating to the runtime to connect producers and
consumers, in the same way that the runtime connects services and
references.
</vdR>
On the direct wiring point, I
agree
with your "shorthand" comment. I see them as useful shorthand
for simple non-shareable channels used within an assembly.. You
would
still use explicit channels for the more complex cases.. I am
not suggesting
we get rid of explicit channels.
Regards
Peter Niblett
IBM Senior Technical Staff Member
Member of the IBM Academy of Technology
+44 1962 815055
+44 7825 657662 (mobile)
From:
Eric Johnson
<>
To:
Anish Karmarkar
<>
Cc:
Date:
17/08/2010 19:40
Subject:
Re: [sca-assembly]
A look at use-cases for composition with eventing, alternate
approaches
to make them work better
Hi Anish,
Let me take a stab at answering some of this. This might be
repeating
some of what we discussed in the call today. Sorry about
that.
On 08/16/2010 11:52 PM, Anish Karmarkar wrote:
> Eric,
>
> Thanks for taking the time to do this. I have a few
questions/comments
> on the material (listed below). It would be a good idea
to have you
> present this on one of our calls. I should also note that
we should
> evaluate this from the POV of existing usecases that we
want to solve.
>
> 1) page 1, last two bullets: Isn't that the same? IOW, if
you promote
> channels won't they have to appear in the CT of the
composite that
> they occur in?
The difference here is perhaps just how you think about it.
Do you
primarily promote producers and consumers, and occasionally
tie them
together (and that shows in the CT), or do you indicate
channels in the
CT, but (we expect) mostly only use one half of the channel
from the
component?
>
> 2) page 2, diagram for PC#1, following current WD, but
avoiding global
> channels and bindings:
> why are there two channels in the outermost and middle
composite?
Not
> that this is wrong, but one channel would suffice and be
simpler.
Depends on how you think of compositions. If I build A &
D, and
know
that I want to send messages from A & D, somehow the
intent has to
be
carried to the developer/designer of the outermost composite
to know
that A should be linked back to D. You've hit the fundamental
question
- how do I enable composability, achieve a clear result, and
also convey
that intent?
As near as I can tell, from going through this exercise, if I
cannot
convey the intent somehow in the CT, then you cannot address
any of the
use-cases I drew up here, and still achieve a clear result, or
composability.
>
> Another alternative for the same is to not have any
channels in the
> innermost and middle composite, promote all the
consumer/producers
all
> the way to the outermost composite and have a single
channel that
> connects C, F and the promoted consumer/producer.
And you've conveyed the intent that that needs to happen how?
>
> 3) page 3, direct wires: I see this as problematic. It
does not scale
> well and has problems with dynamic systems. Consider
domain-level
> components that want to use eventing and everyone
connected to the
> same channel (as a producer and as a consumer). Any
single component
> being undeployed (or deployed) causes changes to all
other components.
I'm perfectly happy to treat the wires as a short-hand.
Unlike wires
from references, I don't see multiple wires from references
implying
multiple channels. That is, if component A wires to D, E, and
F,
that
could perfectly well be a single channel/Destination when
ultimately
deployed. If you go with the direct wires approach, you might
question
whether or not that is an appropriate short-hand, but it seems
possible,
at least.
>
> 4) page 4, PC#1: Option: Channels as components, coupled
> producers/consumers:
> I'm not sure I understand this option. More explanation
would help.
Let me try again. Drop the current notion of channel
altogether.
Instead, replace it, in your head, with the notion that *any*
component
type can couple a producer and consumer in a way that
indicates
"pass-through". That is, "all" (approximately)
events sent to the
consumer will end up at the coupled producer.
Or, if you want to think about it in reverse, "channels" are
just a
specialization of a component that follows specific
characteristics in
the formation of its component type. For that matter, since
it is
just
a specialization of a component type, there's no reason to
exclude all
the other attributes of a component type, including services,
references, and properties.
>
> 5) page 6, PC#2, per WD: avoiding global channels and
bindings:
> You don't need three channels in the outermost composite.
One channel
> would suffice.
As we discussed in the call today, if you don't have the
multiple
channels, what is implied, at least is that at least two
copies of each
message would be delivered to certain destination components.
Having
thought about it a while, at deployment time, even if all the
"channels"
have exactly the same characteristics, I don't know for
certain that a
runtime could infer that the multiple channels can be
collapsed into
one, and then only one copy of a message would be received. I
drew
it
that way to prevent the duplicate message scenario.
>
> 6) page 7, PC#2, Using channels as components, producer
and consumer
> tied:
> Not sure I understand this. Seems like direct wiring.
>
> 7) page 8, PC#3, per WD avoiding global channels and
bindings:
> You can simplify this by using only one channel in the
outer composite.
Again, you might, but that only makes sense if you've
communicated that
intent somehow. If that communication on intent appears in
the model,
presumably I'll find it in the component type? Otherwise, I
think
I
just have to "know" what to do. And that just circles back
to the point
of the issue - that isn't composability, that's just a
monolithic batch
of components/composites that happened to be built as many
pieces, when
either of the two versions on the first page would be simpler.
-Eric.
>
> -Anish
> --
>
>
> On 8/4/2010 10:29 PM, Eric Johnson wrote:
>> I've had an action item to pull together use cases
relevant
for
>> discussing the various options around how to think
about composing
with
>> eventing.
>>
>> Well, maybe it wasn't an official action item at
first, but I
took it as
>> a useful exercise to apply some rigor to what I
proposed, and
actually
>> put my proposal through its paces, along with a bunch
of other
options,
>> to see how they all fare.
>>
>> I definitely tried to do this fairly. it is possible
that I've
>> overlooked some aspect of the current WD, or that I
didn't implement
>> other options fairly, or that there are yet alternate
ways that
we could
>> try to tackle the problem.
>>
>> Which, by way of introduction, I mean to say, please
send back
comments,
>> arrows, darts, alternate use-cases/scenarios you'd
like to see
me pursue
>> in any of the different approaches. Or, please feel
free
to suggest
>> alternate approaches.
>>
>> Unfortunately, I'm on vacation next week, so I'll be
tardy in
catching
>> up with whatever discussion we might have, at least
any that happens
>> after Friday afternoon.
>>
>> Attached, please find both the PDF and ODG format.
>>
>> -Eric.
>>
>>
>>
>>
---------------------------------------------------------------------
>> To unsubscribe from this mail list, you must leave
the OASIS TC
that
>> generates this mail. Follow this link to all your
TCs in
OASIS at:
>> https://www.oasis-open.org/apps/org/workgroup/portal/my_workgroups.php
>
>
---------------------------------------------------------------------
> To unsubscribe from this mail list, you must leave the
OASIS TC that
> generates this mail. Follow this link to all your TCs in
OASIS
at:
> https://www.oasis-open.org/apps/org/workgroup/portal/my_workgroups.php
---------------------------------------------------------------------
To unsubscribe from this mail list, you must leave the OASIS
TC that
generates this mail. Follow this link to all your TCs in
OASIS at:
https://www.oasis-open.org/apps/org/workgroup/portal/my_workgroups.php
Unless stated otherwise
above:
IBM United Kingdom Limited - Registered in England and Wales
with number
741598.
Registered office: PO Box 41, North Harbour, Portsmouth,
Hampshire PO6
3AU
S/MIME Cryptographic Signature
