Friday, September 11, 2015

... about fear of pair programming

Image result for who's afraid of virginia woolf
Although pair programming has been around since 1999, I only heard about it 5 to 10 years later. Software quality was high on my list of values, but finishing a task within the estimated time was a lot higher. And because all our estimates were wrong (too low), I was too engaged in meeting the deadline. I was in a constant state of haste that transformed into a state of panic when the sprint drew near its end. 

Back then, I thought pair programming was wonderful but a bit of a waste of time because you'd put two developers on a task that could be done by one. I knew the arguments about the lower bug rate and better design when doing pair programming, but I didn't really believe they were sufficient to compensate for the wasted time. Finally, I was pretty sure I could not sell the idea to my project manager.


ffwd 10y

That was almost 10 years ago. I think differently now. Meeting initial estimates is not that important to me anymore. It's not at the top of my list anymore. Reducing software complexity is. Understanding the business domain too.

For me, pair programming has lots of advantages. It's an effective tool to get new team members up to speed. It's a way to avoid wasting too much time in dead end streets. It's a great way to learn from others. But most importantly: it's the best way to have a shared understanding of what the domain is, what the solution should look like and how it should be solved.


Out of fashion

Image result for ouderwets

I found that selling the idea of occasional or regular pair programming to project managers is a non-issue. I've yet to meet a manager who's against the idea of pair programming. They seem to understand the added advantages and trust their team to not needlessly waste time.

The problem is not with the management. 

It's with - some/most of - my team members. They don't really want to. Not all will admit it, but I can feel enthusiasm sink fast whenever I suggest to pair program in order to solve a certain task. It's like pair programming has gotten out-of-fashion (some argue that it never got into fashion in the first place). Perhaps it's a personal thing and they just don't want to pair with me. I don't know. 

When asking my colleagues I get reactions similar to the following: 
I don't like pair programming because it's too intense. I want to take time to understand the problem and look for solutions.  
I'm afraid we'll waste too much time and that won't look good on our next stand-up meeting.
I want to drive in dead-end streets, return on my steps and try different routes. I want to be able to make mistakes and bang my head against a wall.
I think it might make me look stupid when I don't understand what my co-programmer wants me to do.

I understand all of this. When you're new to a project, things might look overwhelming at first. Pair programming makes you vulnerable. You have to leave that safe shell, called your desk, and directly expose your thoughts and ideas to your partner. This can only happen when there's enough trust within the team and when every team member is willing to learn from others.

Together everyone achieves more?

Image result for a-team Development is all about team work. It's scrum teams, agile teams, team retrospectives, daily team stand-ups. It's team everywhere. Everyone's a team player, playing that team song. Everything is put in place to make that team work as efficient and effective as possible. 

However, the moment that daily team moment is over, we all go back and work in our own little individual world, developing our own little code in our own particular style having our own thoughts about what clean code should be. We think of all sorts of solutions and abstractions, protocols and schemes, but we neglect to communicate them to the rest of the team. Banging our head against the wall at every misunderstanding of other developer's intent. 

That's not a team. That's only a group.


Conway in the small 

For me, pair programming is the best way to keep your software project coherent. To keep it in a consistent state. To mix my ideas and style of programming with that of my team members. 

There's different ways of doing pair programming. Driver-conductor, tutor-pupil, mob programming, ... I don't really care about that. Any form will do for me.

When pair programming does not occur, the program inevitably breaks down in an aggregate of styles. Of pieces that I wrote and pieces that my colleague wrote but didn't reach any synergy. Of abstractions that my team member found useful but are not used any further because I have other abstractions in mind that conflict with hers. It's Conway's law in the small. When not pair programming, every team member creates her own little program and coherence is lost.

After 10 years, I don't have that sense of emergency anymore at the end of a sprint when I realize that my estimates were overoptimistic. Nowadays, I get that feeling of discomfort, when I realize that the intent of some piece of code was misunderstood by other team members. That the project is going in all directions at the same time. That complexity is needlessly added in the name of standardization or "because that's how we did it in another project"

I just hope I will be able to convince my colleagues about the value of pair programming and how much value it adds to any project. Perhaps we do need more trust first.

Greetings
Jan

Friday, September 4, 2015

... about growing dependencies in stateless services

Our architecture - sort of - always looks the same:
  1. A web request is handled by a spring controller. 
  2. The controller dispatches a command that is handled by a command handler. 
  3. The command handler retrieves some sort of aggregate by calling a repository, 
  4. calls an intention revealing method on the aggregate and 
  5. saves the result in the database.

Code is divided into command handlers, aggregates, value objects, repositories and domain services. Everything has its place. We put the command handlers in the application layer, aggregates in the domain layer and repositories in the infrastructure layer. It's a cookie cutter pattern. We can do this with our eyes closed.


Feauture creep

We grow our software by adding more and more features. Before implementing each new functionality we ask ourselves if we can put the logic into an existing class. If we can find a service or repository that already does similar things we put it there. If not, we create a new class. For example, if we need to retrieve information about a customer from the database and we already have a CustomerRepository, we add the new method there.  

Things start out really simple in the beginning but after each sprint iteration, more functionality is added. A service that starts with a dependency on one repository, ends up having tons of them. An extra repository here, an additional validation service there. As time goes by, complexity increases because the number of dependencies increases too.

Take a service that validates your holiday plans, for instance. It starts simple by receiving an unvalidated holiday and returning errors. It only has a dependency on a hotelRepository to look up the hotel where the customer will stay during his holiday:

Loading code....

But that's only the beginning. Within a few sprint iterations, our poor HolidayValidator also has a dependency on a car booking agency, a flight booking agency, a restaurant reservation service, a repository for looking up previous holidays and a customer repository:


Loading code....


Hidden complexity

The result still looks quiet elegant though. It's nicely split into small methods that try to do only 1 thing. You don't really notice the increased complexity. The public method signatures stay the same. The complexity however, appears in the constructor that the IOC calls in order to setup the service.  It remains unnoticed to the developer because everything's @Autowired@Resource-d or @Inject-ed.

The place where you do notice that things get out of hand, is in your test code. It becomes more complex because now, you need to create all those stubs or mocks the service depends on.

Loading code....

In our test we don't stub-out all methods of those dependencies. Only the ones that will be called from within our service. This means our test needs to know what methods our service will call. That's annoying and prone to false-negatives when the implementation changes.


What went wrong? 

The problem is not the growing number of dependencies. That's just a result of added functionality. It's not something you can avoid. The real problem is in the way we use dependencies:


1. the method signature doesn't reveal its dependencies:

When calling a method, you have no idea what dependencies are involved. You need to look at the implementation to know what's going on. The method validating the hotel only really needs the HotelRepository, not the other 5 dependencies. Some other method of the HotelValidator needs the other dependencies but this one only needs a HotelRepository. Yet you can't tell looking at the method signature. It just says ...
Optional<Error> validateHotel(Holiday holiday).
It would be nice if we could see what dependencies it needs.


2. services depend on more than just the functionality they need

Usually, a class with a dependency to a repository, is really only dependent on 1 or 2 methods of that repository. It doesn't care about the zillion other services that repository offers. 

The validateHotel-method validates the hotel of a holiday object. To do that it needs to find the hotel based on a HotelId. The HotelRepository is able to fulfill this need and that's the reason why it's added as a dependency. 

HotelRepository has tons of other methods. But, the validateHotel-method doesn't care about saving, updating or deleting hotels. It doesn't want to lookup hotels by address or name, it only wants a way to find a hotel based on an id. That's just a function from ...
HotelId -> Hotel

Let's try to fix things.

Instead of injecting entire containers of functionality in the form of a repository, we'll only inject the functionality we need to fulfill our task. 


Loading code....

Now, our validator as no dependency on HotelRepository but instead it depends on Function<HotelId, Hotel>. Like any DI-style application it gets injected and it doesn't care where it comes from. That's the job of the IOC container.

Now, the real dependencies of the validator are shown. They are as lean as they could be and only show what they really depend on instead of masking it with a dependency to an entire range of functionalities.


Getting functions

There's different ways to get a function from HotelId to Hotel. We can transform all HotelRepository's methods into static ones by adding its dependency - the dataSource - as an extra argument. This would turn ...


Hotel findBy(HotelId id)
... into ...
static Hotel findBy(HotelId id, DataSource dataSource)

We could then use partial application, to turn the method into a Function<HotelId, Hotel>

There's something else we could do, though. We can turn ...


Hotel findBy(HotelId id)
... into ...
static Function<HotelId, Hotel> findBy(DataSource dataSource)

This has the advantage of being very explicit about dependencies: methods return functions and dependencies are passed as arguments. This way, HotelRepository is turned into a collection of static methods that take dependencies as arguments and return functions that do the correct behavior:

Loading code....

When doing the same for our validator, it becomes something like this:

Loading code....

There's nothing object oriented about stateless services

Our repository and validator have stopped becoming stateless services. All that is left is a collection of functions. We can do the same for other services. In fact, we could do the same for all stateless services, repositories, validators, factories, ... Each would become loose groups of functions that are there, because we put them together. Because we think they belong together. Not due to some technical reason. We can put 2 functions together because they deal with the same thing like persisting a hotel. Because we can easily find them back later. Because they often need to be changed together. In fact, we have total freedom to choose where to put those function-returning-static-methods. 

But there's another advantage. Tests become easier. Because dependencies now become as lean as it can possibly get, it's much easier to stub or mock them. What used to be a complex test setup, now becomes as simple as it can get:

Loading code....

Removing stateless services in favor of static functions, doesn't mean we should not have any services defined as objects at all. Sometimes it does make sense to wrap functionality inside a stateless service. For example when the service should be called in a strict order following some sort of life-cycle. Or when the analogy of a repository as being a store for your aggregates seems important to you. Nevertheless, in a lot of other cases, services can easily be transformed into static functions without losing any of the useful abstractions.

Naked dependencies


Finally, when your dependencies only consist of functions, it allows you to use a different kind of reasoning when implementing a service. Because all of our stateless services have disappeared, we're free to come up with any dependency we need. Instead of thinking in terms of what services and objects are already available, you can now just think about what function you need to fulfill your task. Because dependencies have become completely naked, it's easier to reason about why a service has certain dependencies.


References

Some more stuff that I found interesting:






Greetings
Jan