OASIS Mailing List ArchivesView the OASIS mailing list archive below
or browse/search using MarkMail.

 


Help: OASIS Mailing Lists Help | MarkMail Help

wsrm message

[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]


Subject: RE: [wsrm] CF4 detailed proposal for V1.07


Title: RE: [wsrm] CF4 detailed proposal for V1.07

Doug:
I have no strong opinion either way - I'd go with what the TC decides.
Just can't convince myself that it deserve much of our sweat at this time...
I think we may need a use case to appreciate the value of
delivering as many messages as possible for a group that we know is broken.
(since with our "strong" definition of ordering, we decided anyway to stop deliver any
message beyond the faulty/expired one)

Jacques

-----Original Message-----
From: Doug Bunting [mailto:Doug.Bunting@sun.com]
Sent: Wednesday, August 04, 2004 11:34 PM
To: Jacques Durand
Cc: WSRM (E-mail)
Subject: Re: [wsrm] CF4 detailed proposal for V1.07


Jacques,

We have different perspectives on this issue and I would much appreciate
hearing from others.  I am certainly hoping for validation but mainly want
a shared (TC) decision on whether an ordered group should terminate as soon
as any message (in or out of order) faults or expires.

More on my perspective:

My concern is not really about expired[1] messages that fall after a 1
message gap though that was my example.  The issue is more about long gaps
which might be filling in without problem until the oddball expires.  That
expiration event is seemingly unrelated to the status of the messages
successfully flowing from Sending to Receiving RMP.  I agree the "group is
doomed"[2] but does that mean every in flight[3] message in the group must
immediately go up in flames?

Bursts of traffic due to RMP recovery will easily result in large gaps of
message order seen at the Receiving RMP.  My preference is to deliver "as
many messages as possible" in this situation, not based on requirements
from that old document of ours but on predictability and a need to limit
the magnitude of failure.

Perhaps I am overly focused on the non-deterministic nature of this
situation.  Perhaps I am thinking about an illegitimate optimization of our
protocol.  I am not convinced either is the case.

thanx,
        doug

[1] shorthand for "expires or faults", "expire or fault", "expired or
faulted", ...
[2] because every message after the one that expired will never be delivered
[3] not yet delivered and earlier in the sequence

On 04-Aug-04 22:26, Jacques Durand wrote:

> Doug:
>
> -----Original Message-----
> From: Doug Bunting [mailto:Doug.Bunting@sun.com]
> Sent: Tuesday, August 03, 2004 8:28 PM
> To: Jacques Durand
> Cc: WSRM (E-mail)
> Subject: Re: [wsrm] CF4 detailed proposal for V1.07
>
>
> Jacques,
>
> Modulo a touch of word smithing such as "different from" becoming "other
> than", this looks pretty good.  I agree that, under our specification's
> assumption that messages arrive intact, we have only one fault that should
> be considered transient.  I can incorporate this change without additional
> input.
>
> Except... I do wonder (now that you reminded me of Section 5.1.3.5) about
> closing an entire group when otherwise only messages after the one in
> question would automatically fail.
>
> For example,
> - I have message 2 sitting waiting for message 1 in an ordered group.
> - Message 2 expires while I am waiting.
> - At this point, should I reject message 1 even if it has not expired?  It
> is without question the case that messages 2, 3 and higher will never be
> delivered but...
>
> <JD> well, even if technically you could deliver msge 1 "in order" with
> all previous ones, msg 1 arrives after that 2 expired, so the group
> itself is doomed before
>
> you do so..., and its is expected that the group be terminated as soon
> as it fails
> (we do not have requirements to deliver "as many messages as possible"
> that are in order,
> in a failed group !) SO I would not bother...
>
>
> Similar things could occur if an out of order (later) message arrives that
> encounters a permanent failure.
>
> Should these cases become a bit more specific, allowing lower numbered
> messages to arrive and be processed as normal and aborting the group
> (closing it or whatever the correct words are) as soon as no gaps remain
> prior to the problematic one?  At worst (the Sending RMP does not try to
> fill unacknowledged holes in the sequence), this would mean the group does
> not close until one of the other termination conditions occurs.
>
> <JD> again, i see this as an optimization - and even maybe not a
> legitimate one:
> one could argue that as soon as one of its message has expired before
> being delivered,
> an ordered group is broken and should be terminated right away and no
> message be delivered
> from this time on (even if in order with previous ones). That's a gray
> area in our spec, but this interpretation is as good
>
> as yours... so I would just leave it as is for now.
>
>
>
>
>
> thanx,
>         doug
>
> On 03-Aug-04 20:04, Jacques Durand wrote:
>  > --------- CF4 on V1.07:
>  >
>  > Important note:
>  > The three faults codes for which we recommend to terminate resending in
>  > 3.2.1,
>  > are kind of arbitrary:
>  > there is no chance that resending the exact faulted message will change
>  > its status from fault to succeed, no matter what the fault is, except
>  > for the fault: MessageProcessingFailure
>  >
>  > In consequence, and in the light of latest discussions on CF4, my
>  > proposal is:
>  >
>  > --------------------------
>  > L580:
>  > Replace:
>  >  "A Sending RMP SHOULD NOT resend a message for which an RM-Reply with
>  > one of the
>  >  following Fault types has been received and MUST notify its Producer of
>  > a delivery failure instead: <bullet list>"
>  > with:
>  > "A Sending RMP SHALL NOT resend a message for which an RM-Reply with a
>  > Fault type other than
>  > MessageProcessingFailure has been received, and MUST notify its Producer
>  > of a delivery failure instead."
>  > (note that the bullet list disappears).
>  >
>  >
>  > -------------------------------------------------------
>  > Section 5.1.3.5 (termination by ordering failure), the Triggering event
>  > line (in both Sender and Receiver)
>  > should be modified as:
>  >
>  > replace on "Receiver side" part:
>  > "Triggering event: in an ordered group, a received message expires
>  > before delivery."
>  > with:
>  > "Triggering event: in an ordered group, a received message expires
>  > before delivery, or a received message is faulted with a fault code
>  > different from MessageProcessingFailure ."
>  >
>  >  replace on "Sender side" part:
>  > "Triggering event: in an ordered group, a non-acknowledged message
>  > expires."
>  > with:
>  > "Triggering event: in an ordered group, a non-acknowledged message
>  > expires, or a sent message is faulted
>  > with a fault code different from MessageProcessingFailure ."
>  >
>  > -------------------------------------------------------
>  > Section 3.2.3 (Ordered Message Delivery), after L632 (at the end of
>  > section) add:
>  >  "A Sending RMP and a Receiver will terminate the group as specified in
>  > 5.1.3.5 (Termination by Ordering Failure)
>  > when respectively receiving and publishing Faults other than
>  > MessageProcessingFailure."
>  >
>  > (note that the normative requirement for this (MUST) is in 5.1.3.5)
>  >
>  >
>  > Jacques
>  >
>



[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]