How do you plan an application's architecture before writing any code?

One thing I struggle with is planning an application's architecture before writing any code.

I don't mean gathering requirements to narrow in on what the application needs to do, but rather effectively thinking about a good way to lay out the overall class, data and flow structures, and iterating those thoughts so that I have a credible plan of action in mind before even opening the IDE. At the moment it is all to easy to just open the IDE, create a blank project, start writing bits and bobs and let the design 'grow out' from there.

I gather UML is one way to do this but I have no experience with it so it seems kind of nebulous.

How do you plan an application's architecture before writing any code? If UML is the way to go, can you recommend a concise and practical introduction for a developer of smallish applications?

I appreciate your input.

I consider the following:

1. what the system is supposed to do, that is, what is the problem that the system is trying to solve
2. who is the customer and what are their wishes
3. what the system has to integrate with
4. are there any legacy aspects that need to be considered
5. what are the user interractions
6. etc...

Then I start looking at the system as a black box and:

1. what are the interactions that need to happen with that black box
2. what are the behaviours that need to happen inside the black box, i.e. what needs to happen to those interactions for the black box to exhibit the desired behaviour at a higher level, e.g. receive and process incoming messages from a reservation system, update a database etc.

Then this will start to give you a view of the system that consists of various internal black boxes, each of which can be broken down further in the same manner.

UML is very good to represent such behaviour. You can describe most systems just using two of the many components of UML, namely:

• class diagrams, and
• sequence diagrams.

You may need activity diagrams as well if there is any parallelism in the behaviour that needs to be described.

A good resource for learning UML is Martin Fowler's excellent book "UML Distilled" (Amazon link - sanitised for the script kiddie link nazis out there (-: ). This book gives you a quick look at the essential parts of each of the components of UML.

Oh. What I've described is pretty much Ivar Jacobson's approach. Jacobson is one of the Three Amigos of OO. In fact UML was initially developed by the other two persons that form the Three Amigos, Grady Booch and Jim Rumbaugh

I really find that a first-off of writing on paper or whiteboard is really crucial. Then move to UML if you want, but nothing beats the flexibility of just drawing it by hand at first.

You should definitely take a look at Steve McConnell's Code Complete- and especially at his giveaway chapter on "Design in Construction"

You can download it from his website:

http://cc2e.com/File.ashx?cid=336

If you're developing for .NET, Microsoft have just published (as a free e-book!) the Application Architecture Guide 2.0b1. It provides loads of really good information about planning your architecture before writing any code.

If you were desperate I expect you could use large chunks of it for non-.NET-based architectures.

I'll preface this by saying that I do mostly web development where much of the architecture is already decided in advance (WebForms, now MVC) and most of my projects are reasonably small, one-person efforts that take less than a year. I also know going in that I'll have an ORM and DAL to handle my business object and data interaction, respectively. Recently, I've switched to using LINQ for this, so much of the "design" becomes database design and mapping via the DBML designer.

Typically, I work in a TDD (test driven development) manner. I don't spend a lot of time up front working on architectural or design details. I do gather the overall interaction of the user with the application via stories. I use the stories to work out the interaction design and discover the major components of the application. I do a lot of whiteboarding during this process with the customer -- sometimes capturing details with a digital camera if they seem important enough to keep in diagram form. Mainly my stories get captured in story form in a wiki. Eventually, the stories get organized into releases and iterations.

By this time I usually have a pretty good idea of the architecture. If it's complicated or there are unusual bits -- things that differ from my normal practices -- or I'm working with someone else (not typical), I'll diagram things (again on a whiteboard). The same is true of complicated interactions -- I may design the page layout and flow on a whiteboard, keeping it (or capturing via camera) until I'm done with that section. Once I have a general idea of where I'm going and what needs to be done first, I'll start writing tests for the first stories. Usually, this goes like: "Okay, to do that I'll need these classes. I'll start with this one and it needs to do this." Then I start merrily TDDing along and the architecture/design grows from the needs of the application.

Periodically, I'll find myself wanting to write some bits of code over again or think "this really smells" and I'll refactor my design to remove duplication or replace the smelly bits with something more elegant. Mostly, I'm concerned with getting the functionality down while following good design principles. I find that using known patterns and paying attention to good principles as you go along works out pretty well.

http://dn.codegear.com/article/31863

I use UML, and find that guide pretty useful and easy to read. Let me know if you need something different.

UML is a notation. It is a way of recording your design, but not (in my opinion) of doing a design. If you need to write things down, I would recommend UML, though, not because it's the "best" but because it is a standard which others probably already know how to read, and it beats inventing your own "standard".

I think the best introduction to UML is still UML Distilled, by Martin Fowler, because it's concise, gives pratical guidance on where to use it, and makes it clear you don't have to buy into the whole UML/RUP story for it to be useful

Doing design is hard.It can't really be captured in one StackOverflow answer. Unfortunately, my design skills, such as they are, have evolved over the years and so I don't have one source I can refer you to.

However, one model I have found useful is robustness analysis (google for it, but there's an intro here). If you have your use-cases for what the system should do, a domain model of what things are involved, then I've found robustness analysis a useful tool in connecting the two and working out what the key components of the system need to be.

But the best advice is read widely, think hard, and practice. It's not a purely teachable skill, you've got to actually do it.

I'm not smart enough to plan ahead more than a little. When I do plan ahead, my plans always come out wrong, but now I've spend n days on bad plans. My limit seems to be about 15 minutes on the whiteboard.

Basically, I do as little work as I can to find out whether I'm headed in the right direction.

I look at my design for critical questions: when A does B to C, will it be fast enough for D? If not, we need a different design. Each of these questions can be answer with a spike. If the spikes look good, then we have the design and it's time to expand on it.

I code in the direction of getting some real customer value as soon as possible, so a customer can tell me where I should be going.

Because I always get things wrong, I rely on refactoring to help me get them right. Refactoring is risky, so I have to write unit tests as I go. Writing unit tests after the fact is hard because of coupling, so I write my tests first. Staying disciplined about this stuff is hard, and a different brain sees things differently, so I like to have a buddy coding with me. My coding buddy has a nose, so I shower regularly.

Let's call it "Extreme Programming".

"White boards, sketches and Post-it notes are excellent design tools. Complicated modeling tools have a tendency to be more distracting than illuminating." From Practices of an Agile Developer by Venkat Subramaniam and Andy Hunt.

I'm not convinced anything can be planned in advance before implementation. I've got 10 years experience, but that's only been at 4 companies (including 2 sites at one company, that were almost polar opposites), and almost all of my experience has been in terms of watching colossal cluster********s occur. I'm starting to think that stuff like refactoring is really the best way to do things, but at the same time I realize that my experience is limited, and I might just be reacting to what I've seen. What I'd really like to know is how to gain the best experience so I'm able to arrive at proper conclusions, but it seems like there's no shortcut and it just involves a lot of time seeing people doing things wrong :(. I'd really like to give a go at working at a company where people do things right (as evidenced by successful product deployments), to know whether I'm a just a contrarian bastard, or if I'm really as smart as I think I am.

I beg to differ: UML can be used for application architecture, but is more often used for technical architecture (frameworks, class or sequence diagrams, ...), because this is where those diagrams can most easily been kept in sync with the development.

Application Architecture occurs when you take some functional specifications (which describe the nature and flows of operations without making any assumptions about a future implementation), and you transform them into technical specifications.

Those specifications represent the applications you need for implementing some business and functional needs.

So if you need to process several large financial portfolios (functional specification), you may determine that you need to divide that large specification into:

• a dispatcher to assign those heavy calculations to different servers
• a launcher to make sure all calculation servers are up and running before starting to process those portfolios.
• a GUI to be able to show what is going on.
• a "common" component to develop the specific portfolio algorithms, independently of the rest of the application architecture, in order to facilitate unit testing, but also some functional and regression testing.

So basically, to think about application architecture is to decide what "group of files" you need to develop in a coherent way (you can not develop in the same group of files a launcher, a GUI, a dispatcher, ...: they would not be able to evolve at the same pace)

When an application architecture is well defined, each of its components is usually a good candidate for a configuration component, that is a group of file which can be versionned as a all into a VCS (Version Control System), meaning all its files will be labeled together every time you need to record a snapshot of that application (again, it would be hard to label all your system, each of its application can not be in a stable state at the same time)

I have been doing architecture for a while. I use BPML to first refine the business process and then use UML to capture various details! Third step generally is ERD! By the time you are done with BPML and UML your ERD will be fairly stable! No plan is perfect and no abstraction is going to be 100%. Plan on refactoring, goal is to minimize refactoring as much as possible!

I try to break my thinking down into two areas: a representation of the things I'm trying to manipulate, and what I intend to do with them.

When I'm trying to model the stuff I'm trying to manipulate, I come up with a series of discrete item definitions- an ecommerce site will have a SKU, a product, a customer, and so forth. I'll also have some non-material things that I'm working with- an order, or a category. Once I have all of the "nouns" in the system, I'll make a domain model that shows how these objects are related to each other- an order has a customer and multiple SKUs, many skus are grouped into a product, and so on.

These domain models can be represented as UML domain models, class diagrams, and SQL ERD's.

Once I have the nouns of the system figured out, I move on to the verbs- for instance, the operations that each of these items go through to commit an order. These usually map pretty well to use cases from my functional requirements- the easiest way to express these that I've found is UML sequence, activity, or collaboration diagrams or swimlane diagrams.

It's important to think of this as an iterative process; I'll do a little corner of the domain, and then work on the actions, and then go back. Ideally I'll have time to write code to try stuff out as I'm going along- you never want the design to get too far ahead of the application. This process is usually terrible if you think that you are building the complete and final architecture for everything; really, all you're trying to do is establish the basic foundations that the team will be sharing in common as they move through development. You're mostly creating a shared vocabulary for team members to use as they describe the system, not laying down the law for how it's gotta be done.

I find myself having trouble fully thinking a system out before coding it. It's just too easy to only bring a cursory glance to some components which you only later realize are much more complicated than you thought they were.

One solution is to just try really hard. Write UML everywhere. Go through every class. Think how it will interact with your other classes. This is difficult to do.

What I like doing is to make a general overview at first. I don't like UML, but I do like drawing diagrams which get the point across. Then I begin to implement it. Even while I'm just writing out the class structure with empty methods, I often see things that I missed earlier, so then I update my design. As I'm coding, I'll realize I need to do something differently, so I'll update my design. It's an iterative process. The concept of "design everything first, and then implement it all" is known as the waterfall model, and I think others have shown it's a bad way of doing software.

Try Archimate.