The Simplest Thing

I recently took part in a code dojo, using the scoring for bowling kata, where we were interested in intention revealing and expressive code. The rules set up were typical, TDD, rotating pairs, no verbal explanation to new pairs, no talking to your current pair. Same old rules, same old results, no surprises. It seems to me that if you want different code to come out of the activity you may need to frame the activity differently.

The same old results of TDD as practiced in the typical code dojo is not a good thing. Write a test and implement a solution to it as quickly as possible. Satisfies each test with the least number of characters possible and in the shortest time possible.

Short and quick lead to producing the easiest solution to come to mind, or perhaps simply the first solution that comes to mind, and basically the most primitive solution. By primitive I mean that the solution is implemented as directly as possible in the primitives of the language.

But it's TDD, the goal is to write the simplest solution. Not to be simple minded about your solution. This depends on defining simple of course, but in my experience that's a measure that should be applied to the result code, not the process that generates it. TDD is not about doing the simplest work that results in something that satisfies the test, it's about doing the hard work of delivering something which is simple and satisfies the test.

Simple becomes a set of principles, some with potential metric support, some more like rules of thumb. Simple implementations for me are those which can't be used wrongly. Simple implementations make explicit the important things, not merely the static things that are easy in whatever language you're using but also the processes and dynamic properties of the system even if you have to twist the language to make them explicit. Simple often means hiding and wrapping primitives so that they don't interact and that the myriad things they might do but shouldn't are cut off.

Simple very often means qualities that are not captured in tests, simple is the judgement of a professional saying that the passing test is just the beginning, diligent professional conduct requires me to attend to these other values.

What I found was that we could have written better code faster in one of two ways, (the code I wanted had a layer abstracting frames, and chaining them together as the game plays). If I'd thought long and hard I could have presented the tests in an order that made the abstraction an immediate and obvious benefit. Or, trusting our professional intuitions and being principled, we could have implemented the highest interface in terms that corresponded to the description of the problem, that DDD approach would have automatically led to a frames based abstraction because that's how the domain experts talk about the problem. Or I could have trusted my little rules of thumb: wrap your primitives, abstract away the language. Perhaps if I'd been able to talk to my partner we'd've been able to establish that quality foundation, but once people started rotating the pressure was on to just follow suit on poor beginnings.

I don't think I've ever written code where an implementation that jumped from the highest level interface straight into code using raw language primitives was the best result. When working with code like that, grief and suffering are typical.

TDD is a process that helps me write good professional code by freeing me up from tracking whether my code is functionally correct, instead I can give more energy to all the other professional values that need to be embodied in my code. But this is my heresy, TDD says nothing about the quality of the implementation, but a poor quality implementation is a terrible cost a company.

The problem with that common type of TDD kata was that it punishes us for attending to code quality and architectural concerns (remember those other things in XP: coding standards and the system metaphor, they acknowledged that TDD needs balance).

So how do you get good, expressive, professional, intention revealing code? No guarantees, but you need at least to be able to identify good from bad, you need the boldness to say no to the bad, you need to be allowed to implement better than the poorest conceivable solution.

TDD is a game, it rewards getting stuff done quickly and crudely, it also rewards getting things that function. People who want the process to deliver quality are sadly disappointed, people who like to tick things off and generally score points can use it to churn out piles of muck. People who feel that quality matters often end up ignoring it when forced to practice with rules mavens and game players. What's really need is counter balancing games that rewards the other qualities of good code, qualities that have great business value but may not be easily measured like maintainability or job satisfaction. Such games may look like leader boards showing the latest model code for the team to strive to excede, perhaps as an ongoing vote or outcome from retrospectives. It may be a checklist of points that need to be judged as part of code review. TDD doesn't stand alone.

TDD: Making Bricks or Building Houses

TDD has shown itself to be amazingly productive practice. With it's tight feedback cycle it's great for learning. Having very specific disciplines and methods mean some excellent tooling has been developed to support it. It's a natural fit for the small stories and task we favour for iterative development. It's strengths are also it's weakness, specifically, it drives an intense focus on the smallest units of development. And as time goes by it's becoming coming coupled ever more tightly to other practices that also focus down at small units of development.

That plays to a very human bias, substituting a hard question with an easier one. We puzzle of over a tricky question: are we making good software? TDD leads us to answer yes by substituting different questions: Have we written good units? Or, have we worked hard with sophisticated tools?

Good units do not make a good program any more than good bricks make a good wall or good musical instruments make a good orchestra.

The general tide in TDD is to keep looking down at small units. This is reinforced by mocking libraries and test frameworks and a culture that favours unit tests over test that aggregate units, (no one questions having a unit test but it is frowned on to have higher levels tests without unit tests, and slow running aggregate tests may well get axed in favour of a faster running test suite).

TDD teaching methods in some quarters have had a similar bias, discouraging conversation and discouraging looking further ahead than the next test. Further reinforcement comes from common patterns of pairing, ping-ponging for example. Likewise practices that focus on stories with little context, either chunked into iterations or striving to have just a few upcoming and potentially unrelated stories in a stream. Add to that developers who take pride in starting what has been presented as the next story without further consideration.

All these things make it easier to focus on the finest grain of problem solving. But practiced without balances they take focus away from the system as a whole being delivered, hiding synergies and commonalities and large scale structure. Indeed it can cause great frustration when someone practiced with the tools and techniques of TDD plows on into chaos dragging in their wake a partner who wanted more time to reflect or consider their work in a context broader than the current test or story.

In all this the mainstream of TDD risks being a "greedy reductionist". Greedy reductionists are a parody in science that says everything can be described in terms of the lowest level realties, basic particles etc, and that higher order abstractions are needless. Dennett gives an example that illustrates the problem: imagine a calculator that returns 3 when you press 1 + 1, everything else it does just fine; you can describe the operation of this calculator in terms of electrons and semiconductors and completely fail to explain it's most interesting and distinctive behaviour. TDD when focused on the smallest granularity risks the same mistake.

Focusing on small problems is called narrow framing and it can a powerful tool. In many situations the human mind instinctively, perhaps unavoidably, resorts to narrow framing. But narrowly framing problems can be deeply misleading about what makes a good decision. Our tendencies to over spend to guarantee a win or to avoid loss can be amplified in damaging ways if we don't take the big picture into account. Treating the parts of decomposed problems in isolation can exaggerate biases: TDD works against balancing our decisions in the context of the greater whole.

TDD helps us build good bricks, but it doesn't really say much about building good houses. Of course it's hard to build a good house from bad bricks, just there's a lot more that we fail to attend to if we only do TDD.

When the world was introduced to TDD it was in the context of system and practices that balanced those weaknesses. I personally have become very wary of programming with anyone people who don't actively practice techniques to balance the focus on the fine grain that is inherent in TDD.

There are many possible practices that can provide the balance; XP as a system has several built in:

The embedded customer and the system metaphor both help focus us on the whole system and it's values as a product, (sadly, they've often been the things left behind in emerging hybrid and customised development methodologies).

Code standards can be used to improve the focus on the whole and on code that works together as a coherent whole. One style of code standard identifies those areas of the code that embody the values at stake, (not just classes functions or files, but whole modules packages or pages that work as cohesive systems). Another style of coding standard uses guiding aphorisms, for example: business logic and events should be represented explicitly as objects; fan out should be limited; always wrap third party code; always wrap primitives and native library code. That last is can be argued but ties into the DDD principle of ubiquitous language: write layers and abstractions so that you are implementing in terms that reflect your users language. The quest for an over arching language rooted in the users expression of the problem domain, and having that language pervade the implementation can be a powerful tool for creating a coherent whole.

Perhaps one of TDDs greatest strengths is that it's easier to do than the counterbalancing practices and we naturally give energy to the things we do well or that are at least easy to do. Real growth, however, requires giving attention to what we do poorly. So the next time you go to do some deliberate practice or try to formalise your team's process, have the boldness to ask yourself what you need to do to improve or to cover for your weaknesses rather than just shining a light on your strengths.

Valid States Only, Please!

I've been thinking about why debugging can be hard and some higher level design heuristics that can help. One thing that really gets me down is invalid states. This problem has a slightly different face in functional programming versus object oriented programming. Sometimes it's not clear cut that a state is invalid, but rather we may have interim states that have no real business meaning.

Lets start with the classic list processing style of functional programming. It's pretty common to take a list of elements conforming to some design, they represent things in the domain in some given state, and the initial states are normally pretty sensible. Then the list is processed via a series of list processing stages, maps and filters, reductions, and so forth. The problem is that these primitives don't distinguish convenient intermediate steps that are just implementation details from stages that are really significant and meaningful states from the domain perspective, valid business states.

The solution is chunking together list processing primitives to make higher order list processing functions take lists of valid items and produce similarly valid results, hiding the intermediate workings, (valid here means representing a state that is meaningful in the business domain).

Failing to break the list at the important points leaves any future developer puzzling over the meaning and correctness of the intermediate lists, particularly if they want to change the process by perhaps adding additional intermediate states or shifting process so that subsequent states are different, or if they are wondering why an object arrives at a particularly stage of processing in a given state.

An analogous situation plagues object oriented programs. A method transforms an object, it does so via a series of intermediate steps. If the method makes those changes on the object directly then the object goes through a series of intermediate states. This can cause all sorts of confusion if you're trying to modify a transformation, it leaves the developer puzzling over what precise intermediate state the object might be in if they what to make change at some given point, effectively the developer has to run the whole program in their head to understand the context of any given sub part of the process. I'll just mention parallel programming, no need to say more I hope.

The solution here is to avoid changing parts of the object, instead, as much as possible calculate the new state in local variables and method parameters, then change the object in a single step. Two phases, one of data crouching and then another of state changing.

Too many months or maybe even years of my working life have been spent trying to understand the state of programs that allow wide access to intermediate states and don't distinguish intermediate states from significant stable states with real meaning in the business domain.

So, some rules of thumb:

• Don't go tweaking state ad hoc, (getters and setters are for tools not programmers), state changes should correspond to meaningful business events.
• Aim to hide, even from other methods, the intermediate states that come about as you implement a method.
• Don't let raw list processing primitives become exposed in your business abstraction layer.
• Chunk list processing primitives together into higher order operations which take you from one business state to another.
• Know what constitutes a valid state, and don't let your objects ever be in invalid or intermediate states.

Getting Stuff Done

I'm planning another blog to track my other interests: songwriting and music in general, and perhaps gigs and other social stuff. Anyway, I've been reading The Frustrated Songwriters Handbook and re-reading Writing Better Lyrics and getting some great ideas and general motivation from both. They are very different books in almost every imaginable way except one.

Both books deal with the problem of slow startup time for creative activity.

The central activity frustrated songwriters are encouraged to try is the 20 song game: set aside a day to write 20 songs and then show them to other songwriters who have done the same. Yes, 20, in one day - then do it again. I'm planning my first effort soon, maybe this weekend, if not then in the week after Christmas.

One of the very first exercise in Pat's book is called object writing set aside 10 minutes each day to write concretely about senses and experiences surrounding some particular object. Stop at 10 minutes. Even if you're bursting with ideas, stop at 10 minutes. After just a few days you start to learn to skip the pretence and centring and all your little rituals, all you can do is as he says dive deep for the good ideas fast.

Both these books are addressing problems of being blocked, of procrastination, of inefficient work practices, and such, that plague songwriters. It's particularly a problem for songwriter because it's normally something you need to fit around the rest of your life. Obviously as an amateur, but even for professional performing musicians distractions like preparing for the next gig, or taking students, and so forth, get in the way of writing music.

These songwriting exercises are not directly about producing songs, they're training techniques that help you hit the ground running when you write.

It seems to me that a very similar problem effects many professional programmers. There's a lot of code that I want to write, more than just what I must write in the office. I want to polish up portfolios pieces, I want to explore new ideas and learn new languages, I want to demonstrate all manner things, and I never seem to have the time, family and friends need my attention, household chores must be done, and so on and so forth.

So, as programmers, are there similar exercises that can make small chunks of time more effective for practicing what we do?

Here's a challenge, the equivalent of object writing, half-hour programming sessions. Make yourself a promise that for some reasonable period of time, (a week, a month, ...), you will work on a project for 30 minutes each day. At the end of 30 minutes, no matter where you're at, stop. If you're in the flow, stop, and let that remind you to get rolling even faster tomorrow. Maintain the commitment to 30 minutes every day for your chosen period, don't work longer today and then give up tomorrow. Outside of that 30 minutes you might need to fix up your tools and environment to make sure you can get the most out of that 30 minutes.

Here's another, the equivalent of the 20 song game. Write 6 programs in one day. They can be anything you like. It would be more interesting if they were different types of programs, maybe even different languages. But the should actual work as deployed programs. For example, write: a game, some data visualisation, a to-do list, a network monitor, a shopping cart, and a web photo light box, all in just one day. Can you wake with no idea what you're going to write, and just make a bunch of programs. Can you do it again next week? What bits of your environment will you need to polish and tune to achieve that 6 program target, what will you need to practice so that you don't waste time next time? Unless you think they are exceptionally cool, they need only be shown to other programmers who've taken the same challenge.

Good luck, here's hoping for a more creative and constructive future.

Red and Blue Architects

I've been thinking about what I expect from a software architect in the highly skilled egalitarian teams doing iterative development that I'm part of, where we're all able to make major changes in response to changing requirements and discoveries in a changing environment.

The traditional software architect's role has been to present a right way of doing something in advanced of stubbing your toes on a problem. That presumes we know enough about our product and it's environment in advanced to make those decisions. It also tends to under-rate the capabilities of other developers to solve problems. Let's call this lot red architects. That was actually true, and good architects in that style were valuable, when I worked in the hardware business: the experts really did have deep knowledge that wasn't available without experience of the product line and general business domain, and the operating constraints of the hardware really were pretty clear in advanced. I haven't worked in that world for many years.

Now I'm working in startup land, it's much faster and looser. The problems of bad architecture still plague us but architecture delivered by red architects just hasn't really worked. We need a different sort of architect, I'll call these new folk blue architects.

Blue architects don't tell me what to do. But they insist that a certain class of question gets asked and answered. Questions that couple the business needs to the hardest technical problems so that they can't be ignored and everyone sees the value in solving them. These are the guys who ask how we're going to handle cache regeneration when we go through database failover, or how we're going to separate our background report from our web front end when the data volume makes the background processes so costly they'll impact the servers responsiveness. They ask what we're going to do when we no longer have a window for downtime because we've become globally successful. Questions that steer us away from the superficial and primitive toward effective and satisfying solutions.

Blue architects guide the whole team to produce better implementations by asking penetrating questions.

Of course everyone with some technical nouse can ask blue architecture questions. But I expect the guy who claims to have architecture specialisation to be better at it, to ask deeper questions, and to frame those questions in ways that lodge in the team consciousness. An architect fails if they can't communicate their insight; I expect anyone claiming specialisation to actually deliver.

In the last few years of work the red architects haven't had much success. On the other hand I've seen individuals deliver blue architecture questions to great effect. Let's have more blue architects.

Object orientation is more than a language primitive

Distinguishing object oriented languages from non-object oriented languages is like talking about free trade and regulated trade. All trade is regulated in both extreme and subtle ways; it's never free (think about slave trade, child labour, pollution control, safety, truth in advertising, and so forth). So called object orient languages provide primitives generally for just one object oriented abstraction; so for any given object oriented language there are many object oriented abstractions and features that it doesn't implement. All object oriented languages are just some person's or group's preferred abstraction (or merely easily implemented abstraction), not at all a universal truth.

I have a distrust of coupling my domain abstraction to language features. A lot of what a like about using functional style in imperative languages is that it helps me to recognise and explicitly represent some things that I might overlook if I just stayed in the mindset of the imperative language I'm working with. Object orientation abstraction is one thing regularly overlooked and left as a non choice. We generally fail to ask the question: what is the right way to use object orientation in this specific application?

It's important to me to recognise what different object oriented abstractions provide: different ways of expressing type relationships, different ways of sharing common code, different ways of protecting access to code and data, different ways to communicate between objects, different ways to get new objects, and so forth. With that understanding I want to build the most appropriate object oriented abstraction to get the benefits most relevant to my particular problem.

One tool to help reach this end is to sneer a little bit at the object oriented primitives in whatever language I'm using. By not letting myself lazily treat that set of primitives as automatically the right way to do object orientation I make myself ask valuable questions: what is the right way to partition code in this application, what is the right way of expressing type in application, what states or life cycle will my object pass through, and much more.

This has recently been brought home to me in discussions about programming Lisp. Programmers that I respect say they have no idea where to start writing a program in Lisp. I think they expect to have the language tell them to organise their code. But really in Lisp, beyond the beginners stage, you're job is to decide how to organise your code, how to represent object orientation if that's part of your application, or how to model data flow, or logical structure, how to represent what ever is important in your application.

In the JavaScript world there are a number of systems that offer support for building object hierarchies, providing abstractions that are generally more like Java's object oriented abstraction. This is a great thing if that object oriented model is the right abstraction for your client side programming. Sadly I think it's been done because people find it difficult to see object orientation as a choice of application specific abstraction rather than a constant of the tools they are using, and aim to make all their tools homogenous with regard to object oriented abstraction.

Oddly, the best object oriented code I've ever dealt with has been in C and Lisp, probably because the developers made choices about organising their code, and didn't treat the primitives of the language as the end of the story. Higher level abstractions of data flow, and of aggregating data and function, and object life cycle, were clearly chosen and expressed in code. Avoiding accidental design driven by language primitives.

Saying or acting as if there can be only one object oriented abstraction and one set of perfect primitives implementing it is frankly and obviously silly. There are many object orient abstractions out there being used productively every day. Good programmers should be able to choose the abstraction that's right for the application and implement it in the language that they are using for the application.

Dynamic dispatch is not enough to replace the Visitor Pattern

I've been bugged a little by how poorly understood the Visitor pattern is by people claiming it's not needed in their particular language because of some funky feature of the language. In the Clojure world dynamic dispatch or multi-methods are the supposed giant killer, (and I'm an Lisp fan of old, and Clojure is fast growing on me, this will be a rant about the fans who rail against an enemy and get it wrong).

The Visitor pattern has two parts, an Iterator and Dispatcher.

The mistake is that because Clojure has dynamic dispatch then it must be the dispatcher that gets replaced. In fact the dynamic dispatch really can help with the iterator, allowing different things in the structure to be treated differently. But as the example below shows it's not sufficient for some even more arbitrary dispatch types, not without adding extra information to parameters to facilitate the dispatch, which is an unnecessary added complexity; unnecessary added complexity being my basic definition of bad design.

My favourite example is a tree traversal engine, for something like an AST, with data like this: (def tree '(:a (:b :c :d) :e))

Now I want to do arbitrary things to this tree, for example, render it as JSON data:

{"fork": ":a",
 "left": {
   "fork": ":b",
   "left": ":c",
   "right": ":d"},
 "right": ":e"}

And also as an fully bracketed infix language: ((:c :b :d) :a :e)

There is absolutely nothing, but nothing, neither in the data nor the traversal that can be used to distinguish between rendering as JSON data or rendering for an infix language. This not a candidate for any sort of algorithmic dynamic dispatch based on any part of the data: the distinction is arbitrary based on functions chosen without regard to the data.

It's quite fair to note that there are two different types of nodes, branching and terminal, that need to be traversed in different ways, which can be usefully used for dynamic dispatch, but in that's dynamic dispatch in the traversal engine not the final functional/Visitor dispatch.

First define a generic traversal:

(declare ^:dynamic prefix 
         ^:dynamic infix 
         ^:dynamic postfix 
         ^:dynamic terminal)

(defmulti traverse #(if (and (seq? %) (= 3 (count %))) :branching :leaf))

(defmethod traverse :branching [[fork left right]] 
  (do
    (prefix fork)
    (traverse left)
    (infix fork)
    (traverse right)
    (postfix fork)))

(defmethod traverse :leaf [leaf]
  (terminal leaf))

Nice application of multi-methods and dynamic dispatch. Note that in the most recent versions of Clojure (1.3.0) dispatch defaults to a slightly faster common case of non-dynamic dispatch, so you have to declare the dynamic binding type. If you know you need dynamic dispatch you need to tinker with meta-data, and you can do that to other people's code from outside: (alter-meta! #'their-symbol assoc :dynamic true)

For ease sake I'm also going to provide a default set of behaviour, doing nothing. Notice the matching dynamic binding type specified as meta-data:

(defn ^:dynamic prefix   [fork] nil)
(defn ^:dynamic infix    [fork] nil)
(defn ^:dynamic postfix  [fork] nil)
(defn ^:dynamic terminal [leaf] nil)

Now to get it to do something real, rendering in a fully bracketed in infix style. Notice that binding doesn't need explicit dynamic meta-data, because dynamic is the heart of binding:

(binding [prefix   (fn [_] (print "("))
          infix    (fn [fork] (print " " fork " "))
          postfix  (fn [_] (print ")"))
          terminal (fn [leaf] (print leaf))]
  (traverse '(:a (:b :c :d) :e)))

How about printing the symbols space separated in reverse polish notation, as if for a stack language like PostScript or Forth. Notice how I only need to provide two of the four possible bindings thanks to the default handlers.

(binding [postfix  #(print % " ")
          terminal #(print % " ")]
  (traverse '(:a (:b :c :d) :e)))

The fully bracketed rendering illustrates one other very naive mistake with many peoples understanding of the Visitor pattern. Some nodes are the data to more than one function call during the traversal, the distinction being dependent on the traversal pattern and the nodes location in the structure. The final dispatch to the is not determined by the data of the node in isolation, the individual node data being passed to the function is insufficient to choose the function.

There are some useful applications of multi-methods in the final dispatch, for instance to distinguish between different types of terminals as in this example:

(defmulti funky-terminal #(if (number? %) :number (if (string? %) :string :other)))
(defmethod funky-terminal :other [terminal] (println "other: " terminal))
(defmethod funky-terminal :number [terminal] (println "number: " terminal))
(defmethod funky-terminal :string [terminal] (println "string: " terminal))
(binding [terminal funky-terminal]
  (traverse '(:a (:b 42 :x) "one")))

But that dispatch isn't the whole of the Visitor pattern, and frankly to hold it up as the Visitor pattern only demonstrates that you haven't grokked the complexity of dispatch from a complex data structure.

Consider another typical use of the Visitor pattern: traversal of graphs. Let's propose a traversal of a graph that is is guaranteed to reach all nodes and to traverse all edges while following edges through the graph. There are many graphs where this would require visiting some nodes and some edges multiple times. But there are many applications that would only want to give a special meaning to the first traversal of an edge or the first reaching of a node. There may even need to indicate the traversal has restarted discontinuously in a new part graph in those cases were there is not edge leading back from a node or where the graph is made of disconnected sub-graphs. Nothing in the node itself can tell you that this is the first or a subsequent reaching or traversal, much less that the traversal has had to become discontinuous with regard to following edges, (maybe you could infer that from reaching a node without passing an edge, maybe you could track nodes and edges, but why do all that work when a good visitor will tell you directly by calling a suitably named functions; if we're going to enjoy functional programming, lets not be ashamed to use functions in simple direct ways). All these different function calls represent different targets of dynamic binding in the model I've presented her. Of course you could use multi-methods to achieve dynamic dispatch if there are different types of nodes for example. But it would still be the Visitor pattern if there was only one type of node and no reason for dynamic dispatch.

Dynamic dispatch is a powerful feature in Clojure, but trying to justify it as a giant killer to the Visitor pattern shows a lack of understanding of Visitor pattern, and weakens the argument. Present the languages strengths, not your own weaknesses.