I was having a discussion the other day about the reasons for using SOA. If the liklihood of defects in a system are logarithmically proportional to the complexity of the system, I noted, then SOA is useful because you can create a collaboration of interacting systems, where each system is as simple as possible, and some logic moves to the collaboration or orchestration between them.
To which my friend replied: so if a team has 10 members, and one is not functional, the rest of the team can adapt, but if a team has 10 members, but communication is screwed up, then the team itself is dysfunctional. That's worse. So, can SOA create dysfunctional collaborations? Can we create a "team" of systems that hate each other?
What if one system is best served by mistakes that show up in another? Can that system engage in passive-agressive behavior with another system? What about codependency? Can two systems behave in a manner that is counterproductive to both, but makes both of them look effective from the outside?
Do our test plans need to start including common team dysfunctional behaviors as test scenarios?
I was reminded recently of the fact that long running transactions, especially those involving multiple databases, cannot be made to follow the ACID rules of database transactions. On its face, this is completely true. However, I'm thinking that there are mechanisms that could be used to allow the positive effects of ACID to remain, even when the actual implementation is not available in the automated manner we are used to.
As a refresher: A is atomicity (which means that the entire transaction has to occur or not occur... failure means to roll it back). C is consistency (if part of a transaction breaks a rule, then the entire transaction fails), I is isolation (two people performing actions on the data should not affect one another), and D is durability (committed transactions are not lost when power fails or other adverse events occur).
So if a long running transaction causes a change in Database D1, then is transmitted to a remote system, where the next day, it affects Database D2, (where it could fail), then we lose both Atomicity (because the transaction was committed to D1 even before it is known to be successful at D2) as well as Isolation (since a user could ask both databases for info in the mean time, and get two different answers.
However, the positive effects of ACID come when viewed from the viewpoint of the user. The user is not a concept. He or she is real. They have a goal and a purpose for using the database. If you can present ACID-like interactions to them, then these flaws can be minimized.
In order to do this, I'd suggest that a "system of record" is kept seperate from the systems interacting in the transaction. An interaction with the "system of record" would occur at the last step of the long running transaction. That interaction would only occur if all prior interactions were successful. All users who want the "correct" information would be encouraged to check there. This gives you a kind of atomicity, since a change would not occur in this system until all parts of the transction are complete.
Similarly to Atomicity, Isolation can be met from this location as well, since queries to this system would not return different results depending on the status of various transactions, until those transactions completed and updated the system.
So while long running transactions don't meet the ACID test, systems that support and defect long running transactions can be set up to provide the benefits of ACID transactions fairly readily.
I'm specifically looking for feedback on a workflow component I am working on.
We have implemented a workflow engine that maintains a distinction between ResponsibleFor and AssignmentTo. We have ambiguity around the combination of this concept with group assignment. I'd like to know what you think.
Here's the conundrum:
A workflow item has the alias of the person who is responsible for it in a particular stage of the workflow. Responsibility can span stages. Each stage is (ideally) aligned with a business task within an overall process.
Within a stage, we have steps. This "containment" allows us to model two of the three levels of abstraction within a single workflow model. Multiple people can be assigned to a different steps at the same time, and the workflow language has constructs for split and join to allow these activities to be coordinated. The interesting thing to note about this is that assignment does NOT span from one stage to another. When you leave a stage, all assignments are wiped clean, and the next stage gets an assignment to whomever is responsible for the item. This provides a way for the workflow modeler to use ad-hoc workflow, while constraining the users from messing up their own business process.
Now add the notion of "assign to group." Our customers have asked for a way to indicate that a group of people should be made responsible for a work item, and any one member of the group can "take responsibility" for the item. The taking of responsibility is not an event. It simply occurs. In our app, a person visits a web page, sees the items assigned to the group, selects the item he or she wants to own, and clicks "assign to me".
My questions, your thoughts:
In your organization, would it be more correct to say that a "group" is responsible for an activity, or that a person is responsible and that the person has delegated assignment to the group. In other words, if a group is responsible, who is accountable if the item isn't worked? If the item is ultimately escalated to someone, wasn't that person responsible all along?
If you go with the idea of group responsibility: does it span stages? If we accept that a group is responsible for a set of stages, then one person can take responsibility at one stage... does the responsibility revert to the group when the next stage begins? In effect, can the group be responsible but an individual be assigned?
Many different products claim to be effective for Enterprise Application Integration. There's about as many products as there are ways to integrate applications.
The first and most common approach to integration is data integration. I've seen integration from the standpoint of common data tables from multiple sources, mined data in common reports, and even "star data dispersion" where one "domain data" application acts as a source of data for a host of others. Domain data alignment is essential to integration, but it is only the first step.
After you have created a structure for sharing common data values across applications, and keeping them up to date, you need a way to share EVENTS. This is more important than sharing code or sharing services. You need to have a common understanding of what events are important to the business, and a common definition surrounding the conditions that define each event.
For example: let's say you have a social service case management system running in a state agency responsible for child welfare. A call comes in to a hotline, and a concerned citizen is calling to report that the child in the next apartment has been crying for hours. They suspect possible abuse. Your system needs to assign a task to someone to investigate. Do you create a case even though you don't know the name of the child? If there is no abuse, should it be a case or just an unattached incident report? If your system doesn't know the "magic" condition that means "a case is created" then you have no way to share the "NewCase" event.
The obvious question then is how and why to share this event. Do you have a reporting system that tracks the number of cases assigned to each worker? Do you have a financial system that tracks case costs? How about a system that instigates a long-running process to look up information on the suspect's address, to see if a known felon is known to frequent the location? Would you share the event with these systems? There could be a lot of value there.
In a loosely coupled system, you really shouldn't care what system subscribes to your events. You should send your event to a broker and let it decide who to send it to. More importantly, you need a way to insure that a system that is not online at the moment has a way to get the event when it returns to operation. Perhaps you are using a system of cooperating components to coordinate these messages... or perhaps you have a single point clustered broker. Either way, you are implementing a publish-subscribe messaging system. If it is truly loosely coupled, you have a message bus.
The point I want to make is: thes two integration mechanisms have Nothing to do with web services.
I can't count the number of times I've had to correct folks who talk about Integration only in terms of web services. Web Services are useful for a third kind of integration: the services bus. They are NOT the cornerstone of either data or event integration; mechanisms that are centered around data coordination and event brokering, respectively. While web services can be used as a communication end-point, especially for the messages bus concept, the fundamental technology is not web services or even SOAP. It's loosely coupled brokering.
Web services, and the new Indigo framework, are useful for creating a services bus. A services bus is a mechanism for registering, managing, and serving up a list of services. If a service has a way to advertise itself, and if an application has a way to find services that match its needs, then a services bus can connect the two, and allow an application (or a user) to consume a service without knowing who wrote it, in what language, or what server its running on.
This is very useful, and in some ways, fundamental to integration. However, it is only one aspect of integration. The services bus compliments the message bus and the shared data repository service. It does nothing to supplant it. On the contrary, I would posit that a services bus, without shared domain data or a mechanism for retrieving it, is fundamentally crippled.
So, the next time someone says "Use Web Services for Integration", think to yourself: "that's part of the story, but not all."
After all, a barber shop quartet sounds much better if all four vocalists are in the room.
I'm seeing the difference more clearly than before: how a team can use Feature Driven Development for development, and how that is different than using FDD for project management. How could I have missed this?
For Development, FDD means:
For Project Management, FDD means:
So... why did I detail all of this? Because I ran across a situation where a team was using FDD for dev, but PM wasn't using FDD for planning or management. It's a little like being hungry at a buffet table.
Enough for now. I hope this makes sense.