As a UX developer.. have you drawn a picture of it? From: [email protected] [mailto:[email protected]] On Behalf Of Scott Barnes Sent: Thursday, 7 July 2011 2:07 PM To: ozDotNet Subject: MSMQ Distributed Saga across multiple machines?
I've got a bit of an architectural question (its .NET related). Using a Starbucks concept, say you have the following services. CashierService, BaristaService, CustomerService and AuditService. Typical workflow is that the Cashier, Barista and Customer all exist on their own instances along with the guy monitoring their behaviors (lets say Starbucks time in motion study is under way and audit guy is looking for a better way to optimize everyone's work). Scenario ~~~~~~~~~~~~~~ CustomerA walks into the store, orders their beverage of choice. Cashier takes the order down and responds after x cycles with a price. Barista hears the order (without asking Customer or Cashier) and begins making the beverage under the likelihood of the transaction failing being quite low. Customer then pays Cashier independent of Barista completion state and goes and waits for the drink. Barista hands the drink to the customer and the customer triggers a "complete" status by saying thank you to the Barista / cashier "thanks guys, cya next time" (loud voice). Given the entire workflow is done via Saga, i'm wondering how this could play out given each actor i guess represents a server/device? My thinking so far is the following: ~~~~~~~~~~~~~~ Customer Sends Message to Cashier. Cashier then Echos the Message to Barista and Audit (because it knows about these two given they belong to the Starbucks context/domain). Barista sends acknowledgement to both Audit and Cashier that the NewOrder was of a known type and begins the CoffeeMakingWorkflow. Now, the transaction is governed by an AuditService who's job is to verify the Saga as being complete through-out stages, meaning the Cashier transacts with the customer and every time a new update occurs it notifies Audit. Barista then does the same with both Cashier and Customer but also notifies Audit when it's completed its conditions of success (in this case, coffee was made and handed to customer). Audit then does a quick "Ok Money taken, Order handed off.. yes, we're done" and gives it an immediate final ApprovedForFinalDelivery stamp. Customer then does one last check "ie does the coffee exist in their hands" and then tells both cashier and/or barista that they are satisfied with the transaction. These two then both immediately notify the audit of a customer confirmation and the audit basically takes the first confirmation and ignores the rest (first to tell him wins a gold star sort of thing). Using MSMQ/ActiveMQ/ZeroMQ etc. Basically each time a node connects with one another it sends the address of where they exist in the gird (msmq://machinex/nodeX). Each node keeps a local copy of the URI and puts it into its ping/pong heartbeat check queue (who's sole job is to maintain a list of live connections for other services/workflows to use). Each machine has what i'd call a "gossip" service whereby its entire existence is to re-echo inbound messages through the URI queue (kind of a distributed multi-cast scenario - or it could very well be that :D). As it echo's it adds its URI to the "ToldMessage" list (ie the packet gets handed around the grid with basically a list of URI's that have already been told the message - based off its CorrelationId). Each node does this until they have heard the message in which case they ignore all future messages being sent of them (in case you have a situation where two nodes assume "BlahMachine" hadn't heard the message and so they in turn tell them (perks of fire and forget?) Question: ~~~~~~~~~~~~~~~~ Given I'm clearly a just a UX Developer, does the above hold water? or am I just smoking way to many Pixels lately to fully grasp the awesomeness of publish/subscription servicing across private/public cloud(s)? --- Regards, Scott Barnes http://www.riagenic.com
