Karl: thank you very much for the excellent summary. (which I must admit I would need a week to really digest.). One thing that I (and I think you also pointed out ) was that that in the case of AR not agreeing, there were not enough hints for the AI to initiate another request.. Any comments on this? #Also comments to show stopper #7 would be very much appreciated :-) Best Regards Toshi ----- Toshiyuki Nakata 中田 登志之 Executive Chief Engineer, Central Research Lab. NEC 1753, Shimonumabe, Nakahara-Ku, Kawasaki,Kanagawa 211-8666,Japan Tel +81-44-431-7653 (NEC Internal 22-60035) Fax +81-44-431-7609 (NEC Internal 22-60509)
-----Original Message----- From: Karl Czajkowski [mailto:karlcz@univa.com] Sent: Friday, August 24, 2007 6:55 PM To: Toshiyuki Nakata Cc: 'Dominic Battre'; 'GRAAP-WG' Subject: Re: [GRAAP-WG] Modification to the wiki Page on Renegotiating anestablished Agreement
Hi Dominic:
So, yes, there are problems but it looks like a lot of
On Aug 23, Toshiyuki Nakata modulated: people want to
have this extra commit message. Are there any written records on the pros and cons of having the 2PC?
I think this info. is very important. Unfortunately, I am not able to address this issue.
Karl or others, please respond..
I definitely don't want to repeat the debate which is in the GRAAP-WG email archives, but let me just summarize what I think were the final observations and conclusions.
WS-Agreement is aimed at a distributed, soft real-time (one might say isochronous) environment with limited trust between peers with bounded rationality. In other words, it is for establishing automated agreement between programmatic systems in an Internet-type environment where the agreement terms may involve wall-clock obligations which are at the root of the problem.
When you boil it all down, there is always going to be a synchronization hazard where due to communication delays or faults, or even intentional decision delays, one party does not know whether the other has accepted the agreement and therefore is at risk to perform according to the terms of agreement when not necessary, or ignore the terms of agreement when bound by them.
WS-Agreement formalizes this hazard to always place the burden on the agreement initiator who makes an offer and is bound by it if the responder accepts. Early drafts, derived from our SNAP work, had an extra signalling step to allow a sort of "pre-initiator" called an invitation. The group felt that this was too complicated to pursue, particularly because it could be approximated via advertisement.
An invitation or advertisement would be a non-binding expression of interest in receiving agreement offers, so that the typical client-server roles could be reversed somewhat. The only difference is that an invitation is seen as usually a 1:1 relationship while an advertisement is 1:many. However, due to the non-binding nature of these signals and the assumption of bounded rationality (we can ignore the possibility of bluffing and other psychological strategies), I am not sure these differences are significant. For privacy, one could imagine secure advertisement channels (with authorization) to further blur the boundaries.
Furthermore, 2PC can be synthesized with two WS-Agreement round-trips and an appropriate domain-specific agreement semantics. So, either advance reservation, combined with the right cost/penalty model, or the underlying invitation system can both look like 2PC in practice:
1.a. initiator offers advance reservation, OR 1.b. advertiser publishes interest
after this first stage, both parties are aware of interest and primed to negotiate...
2.a. responder accepts (or rejects) reservation offer, OR 2.b. initiator offers against advertisement (or ignores it)
after this stage, one party is obligated to satisfy and the other has a choice to make... (For the advance reservation, the initiator has a choice because the reservation includes a cost model with cheap cancellation under reasonable deadlines).
3.a. initiator offers claiming agreement (or cancels reservation), OR 3.b. responder accepts offer (or rejects it)
after this stage, both parties are obligated to satisfy...
4.a. responder accepts claiming agreement (or acknowledges cancellation), OR 4.b. initiator acknowledges receipt (e.g. transport-level ACK)
after this stage, both parties actually know what the other knows...
In case it isn't obvious, advertisement is done via WS-Agreement templates and advance reservation followed by claiming is anticipated in the usage scenario examples as well.
Best Regards Toshi ----- Toshiyuki Nakata 中田 登志之 Executive Chief Engineer, Central Research Lab. NEC 1753, Shimonumabe, Nakahara-Ku, Kawasaki,Kanagawa 211-8666,Japan Tel +81-44-431-7653 (NEC Internal 22-60035) Fax +81-44-431-7609 (NEC Internal 22-60509)
karl
p.s. I will add that the fundamental challenge I see for GRAAP-WG and WS-Agreement is that everyone approaches it with the same natural (but incorrect) intuitions about trust, signalling, and consensus. I think there is still a large social experiment to see who can deploy systems which are actually performing distributed agreement without a central authority... economics and our real-life world says it should work, most of the time at least. :-)
-- Karl Czajkowski Software Architect
Univa Corporation 1001 Warrenville Road, Suite 550 Lisle, IL 60532
karlcz@univa.com www.univa.com ________________________________________________________________ www.univa.com.
The Leaders of Open Source Cluster and Grid Software