How will you make it if you never even try?

February 27, 2010

The CQRS Light Bulb Moment

Filed under: ASP.NET MVC, C#, CQRS, Domain Driven Design (DDD), nhibernate, nServiceBus, OLAP — charlieflowers @ 11:17 pm

As I recently blogged, the project I’m on has recently decided to move to CQRS (Command Query Responsibility Segregation). We’re going to use nServiceBus as a message bus that let’s us separate our system into 2 “channels”: The “Read side” and the “Write side” (aka, the “Query side” and the “Command Side”).

This decision has been the result of several “Light Bulb Moments”, in which various ones of us had a flash of insight that helped us see how an architecture that at first sounded wierd and unorthodox would actually solve a number of problems and help us tremendously.

I’ve decided to share here one of those Light Bulb Moments in raw form. Here’s the text of an email I sent to two other architects on our team (over the weekend, from my own account … we talk about this stuff all the time because we love it). It expresses well many of the reasons we made the move (although I understand more about CQRS at this point and would tweak a few details). (Note: Names changed to protect the guilty).

The Email…


I’m seeing the opportunity to do something truly awesome here. It is based on the CQS reading I’ve been doing while thinking about what our “dto’s” or “commands” or etc. should look like.

I have created, worked with, and seen first hand the power of an OLAP database for read operations. It really is unbelievable in terms of the freedom it gives someone looking at the data. And it lets reads be very fast. But a lot of projects I’ve been on have said, “Let’s build the transactional system first. It is so obviously core to our business that we need it, and we need it yesterday. Once we get that done, we can think about maybe doing OLAP.”

But the way people are approaching CQS as a architectural concept these days, we have the opportunity to do both at the same time. It should help us get to the finish line faster, with screaming fast software and high scalability.

And it’s not that big of a change from what we’re doing now. It boils down to this:

1. We make the “flat view models” you guys are working on. They are designed to serve the view that they populate, and nothing else.

2. We express our edits to the domain in terms of “Commands”. These are merely Declarative … you look at one and it intuitively makes sense.

3. Our Domain Objects accept those Commands and process them. Our Domain Objects apply rules to decide whether or not a Command is valid. The Domain Objects have complete authority over accepting or rejecting an Edit Command.

4. Once the Edit Command is accepted by the Domain Objects, it is “applied”.

Now, right now, you’re both saying, “No shit, that’s what I said on Friday.” Yes, but let’s take stock of where this puts us, and see what else it allows us to do.

5. Since those “flat view models” don’t enforce any important business rules, they don’t have to come from our Domain Objects. (They can STILL come from NHibernate if that’s important or helpful, but they don’t have to come from our Domain Objects). Remember, our Domain Objects are in charge of *writing* all updates. Therefore, the written data can include calculated fields and anything else necessary to ensure that what comes back in on the read side is valid and complete. Complete domain integrity is maintained by the Domain Objects, so Reading is simplified. Needing a bunch of business logic on read has some challenges to it, plus I don’t think we have very many (maybe not any) kinds of calculated fields that would really require a full domain object.

6. We *are* still talking about NHibernate pulling the data that ultimately goes into our View Models. So there are probably some “DTO’s” that are *also* mapped to the same NHibernate tables that our Domain Objects write to. But those DTO’s can be “screen-shaped” (more accurately, “task-shaped”, since we want to include web services and other users of our system besides just the web-based human interface).

7. Now, the domain no longer needs many (possibly any) getters or setters.

8. Every single Edit Command can cause 2 things to happen: 1) our normalized OLTP database can get updated by our Domain Objects with the new data. 2) The very same Edit Command can get queued somewhere else to cause an update to our OLAP database for read access. We can essentially get an OLAP database that doesn’t need ETL … it gets updated from our Edit Commands and only lags a few seconds behind our OLTP database.

9. The Edit Commands also make it easy for us to have *MANY* copies of the readable OLAP database. We can update 3 databases as easily as one. Now we can load balance between them, and they’re equivalent.

10. We don’t actually need the fix Billy added to submit disabled controls. After all, we *know* those values didn’t change. Why should we need them on a Post? Our Edit Commands can be as sparse as what the user actually changed. (This is a minor thing, but still worth mentioning).

11. Here’s one of the main benefits of the whole thing: Once we get to this point, when we make a new screen, we make a DTO for it and a View Model for it. *Both* are custom designed to fit the screen itself. They will be coupled to the purposes of that screen because they need to be … this is good coupling. However, that screen will not exert *any pressure whatsoever* on our Domain Model. Our Domain Model will simply be exactly what it needs to be to express the logic of the domain. Think about how much easier things will be for us than they are right now. Now, we have to have a Domain Object Graph, and a parallel DTO object graph, and (soon) a View Model that gets mapped from the DTO object graph. Keeping the parallel Domain Object and DTO in sync has proven to be something we invest a lot of time in. They were drifting apart before I added the ApplyEdits() stuff. I then added the Interfaces that sometimes have 4 or 5 generic types riding along. Sam went further with it and has cases with 8 or 12 generic types, including multiple levels of nesting. We’re working too hard here. *IF you LOVE THE SMELL OF DELETED BITS IN THE MORNING”*, then you are going to get enjoyment out of moving to an approach like this.

Normally, doing something to fix problems you are having requires some extra work you hadn’t anticipated, and is a bit of a setback, though necessary. In this case, the fix for some of the problems we’re running into actually opens up whole new vistas of possibility, and these opportunities basically come for free after applying the correct fix.


February 26, 2010

Using CQRS!

Filed under: C# — charlieflowers @ 2:10 am

Man, I’m tired. I’m hoping the little cup of coffee I just had will give me one more short burst of energy.

Why am I so tired? Because the project I’m working on is so freakin’ awesome that I’m working night and day on it.

Really, I’m so excited about the architecture we’re using and the technology stack that I think this may be the most fun I’ve had (at work) in 8 years.

So what is all this great stuff, you ask. Well, glad you asked. We’re moving to an architecture called “Command / Query Responsibility Segregation” (CQRS). In an oversimplified nutshell, the idea is this: You break your system into 2 “channels”, one for Writes (aka, “Commands”) and one for Reads (aka, “Queries”). You have Domain Objects (and we’re following DDD), but they are Write-Only. That will sound very strange if you’ve never heard of CQRS. But it does a ton of good things for you.

So if your Domain Objects are write-only, then how do you populate a screen with existing data? Well, you use the Query “channel” for that.

When a write occurs, your Domain Objects do it. They apply all of the intelligent behavior and business logic that you’ve so carefully built into them. They won’t let the write occur if anything is wrong with it.

If the Domain Objects do decide to let the write occur, then they will also fire an event saying that something has happened (as in, “A payment has been taken” or “A new account has been created”).

These events are listened to by many possible “subscribers”, one of which is the “Query channel”. It then records or updates projections of the relevant state.

This means that you can represent the data that is your “true state” in many different ways (aka, various read-only projections). For example, you might write some data into a OLAP star schema. You might also make a separate projection that is tailored to “Customer Edit” screen. And some of that same data might be pushed into a projection tailored for your “Nightly Financial Report”. Since these representations are read-only, they can be denormalized and tailored to the needs of the task that is doing the reading.

Then, when someone comes to the “Customer Edit” screen, you do not use the Domain Object to populate that screen. Instead, you read from the Query channel, from the particular read-only projection that was written there for the “Customer Edit” screen.

You see, in a sense this data still comes from the Domain Objects … just not directly. The Domain Objects still control every bit of the data used by the system … but you don’t “call them” to get the info. They put the info somewhere, and you get it from there.

I’m not going to go into too much more detail right now. But in a nutshell, here are some of the main benefits:

1. Your Domain Model is free to emerge and evolve into the right way to express your complicated domain logic. It no longer is encumbered with the additional job of providing thousands of little pieces of data on demand. No other parts of your system even need dependencies on your Domain Objects, and they certainly do not place any “demands” on the shape of those Domain Objects. (It is this benefit, moreso than scalability, that led me in particular to CQRS).

2. The ability to produce many different representations of the state on the “Read channel” is surprisingly powerful. By that I mean that, even though it sounds powerful, it turns out to be surprisingly more powerful than it sounds at first. Your Domain Objects (and some other “helpers” on that end of the system) can custom-tailor specialized “summaries” and representations of the data … one each for the various “Read tasks” that you have. This means that all your “Read operations” get much simpler … there is very little or no “transformation” work to do, because it was already done for you when the events were fired from the “Command channel”. Granted, this is merely moving work around, because you used to do that “transformation” work during the Read, whereas now you do it as an indirect consequence of the Write. But there are significant advantages of doing that work “back there” as a consequence of the Write … a lot of that transformation work is quasi-business logic, and it is very nice and clean to do it back there. And it’s wonderful to not have to do it when rendering an Edit screen or showing a report.

3. Scalability. Duh. This is the most highly touted benefit and it is a big one. Now you can optimize your Write channel for writes, and your Read channel for Reads. You can have many instances of your Read database, which you load balance between. Most systems have many times more Reads than Writes … so now your “Write channel” won’t be burdened with serving all those Reads.

4. Real-time OLAP. (This sort of fits under # 2, but I feel it deserves its own bullet point). Star schemas are so fantastic for presenting information about what has happened in your business. So many businesses that would benefit greatly from them don’t have them. It’s often approached as a nice-to-have after the OLTP system itself is working. But CQRS let’s you have a real-time OLAP schema, merely because those events from the Command channel can be captured in an out-of-band manner and recorded into a star schema. I say “real-time” because, even though there will be a lag time in seconds between the Command channel write and the OLAP update, “seconds” is still very much real-time to the business world. Plus, you don’t have to justify the OLAP schema as a separate project. It can be the foundation of your Read channel, and therefore it is necessary (and powerful) for the system you are building right now. But you can also do bits and pieces, rather than the whole Data Warehouse enchilada.

So I’m fired up. A co-worked told me the other day that he couldn’t sleep the other night because he was so excited about it. I’m fired up, but I have been able to sleep at least 🙂

If you want to learn more about CQRS, here are some great links:

February 11, 2010

Important new Term: “Prefactoring”

Filed under: C# — charlieflowers @ 5:37 am

“Prefactoring” is like “refactoring”, but much more proactive. It is when you tell a co-worker, “Hey, you know that code you were going to write this afternoon? Yeah, well … uh … why don’t you go ahead and let me write that instead? Why don’t you take some time to run a few errands, or get some fresh air?”

(Update: Or, perhaps, it is when a co-worker does that to you. I know it can happen to me. After all, I always presume myself “ignorant until proven guilty.”)

Blog at