Wide Awake Developers

Inviting Disaster

| Comments

I’m reading a fabulous book called "Inviting Disaster", by James R. Chiles. He discusses hundreds of engineering and mechanical disasters. Most of them caused serious loss of life.

There are several common themes:

1. Enormously complex systems that react in sometimes unpredictable ways

2. Inadequate testing, training, or preparedness for failures – particularly for multiple concurrent failures

3. A chain of events leading to the "system fracture". Usually exacerbated by human error

4. Politics or budget pressure causing otherwise responsible people to rush things out. This often involves whitewashing or pooh-poohing legitimate criticism and concern from experts involved.

The parallels to some projects I’ve worked on are kind of eerie. Particularly when he’s talking about things like the DC-10 and the Hubble Space Telescope. In both of those cases, warning signs were visible during the construction and early testing, but because each of the people involved had tunnel vision limited to that person’s silo, the clues got missed.

The scary part is that there is no solution here. Sometimes, you can’t even place the blame very squarely. When half-a-dozen people were involved with unloading and handling of oxygen-generating cylinders on a ValuJet flight, no single individual really did something wrong (or contrary to procedure, anyway). Still, the net effect of their actions cost the lives of every single person on that flight.

It’s grim stuff, but it ought to be required reading. If you ever leave your house again, you’ll be much better prepared for building and operating complex systems.

Technorati Tags: operations, systems


New Interview Question

| Comments

So many frameworks… so much alphabet soup on the resumes.

Anyone that ever reads The Server Side or Monster.com knows exactly which boxes to hit when they’re writing a resume. The recruiters telegraph their needs a mile away. (Usually because they couldn’t care less about the differences or similarities between Struts, JSF, WebWork, etc.) As long as the candidate knows how to spell Spring and Hibernate, they’ll get submitted to the "preferred vendor" system.

Being one of those candidates is tough, but that’s not the part I’m concerned about now. I’m interested in weeding out the know-nothings, the poseurs, and the fast talkers.

When I’m interviewing somebody, my main criterion is this: would I want to work on a two-person project with this candidate? My secondary criterion is "Would I feel comfortable leaving this person along at a client site? Will they deliver value to the client? Will they look like an idiot, and by extension, make me look like an idiot?"

My friend Dion Stewart had a great idea for a weed-out question. No matter what frameworks the candidate shows on the resume, ask them what they disliked the most about the framework. (I have my top three list for each framework I’ve worked in… except NeXT’s Enterprise Objects Framework. But that’s another story.)

If they can’t answer at all, then they haven’t actually worked with the framework. They’re just playing buzzword bingo.

If they answer, but it sounds like bullshit, then odds are they’re bullshitting you.

If they have never thought about it, haven’t formed an opinion, or say "it’s all good", then they lack passion about what they do.

A candidate that is driven, that cares about the quality-without-a-name should be able to go on a rant about something in each framework they’ve actually worked with. In fact, you’ve really hit the jackpot if your candidate <i>can</i> go on a rant, but does it in a professional, reasoned way. I love to see a candidate that can show some fire without seeming like a loon. That’s when I can see how they’ll react when the client makes a decision the candidate considers boneheaded. (I’ve seen some spectacular pyrotechnics from consultants that forgot whose money they’re spending. But that’s another story.)

Technorati Tags: resume, jobs

JAI 1.1.3 in Beta

| Comments

I’ve been using JAI 1.1.2 for the past year. It’s an incredibly powerful tool, though I will confess that the API is more than a bit quirky.

Early this year, Sun made JAI an open-source project available at java.net. That project has been working on the 1.1.3 release for most of the year. It’s now in beta, with a few enhancements and a lot of bug fixes.

The most significant enhancement is that JAI can now be used with Java WebStart. Previously it had to be installed as a JRE extension.

Also, one of the big bugs is fixed. Issue #13 is fixed in the beta. It could cause the JPEG codec to use excessive amounts of memory when decoding large untiled images. (Which we do in our app a lot!)

Technorati Tags: java

Ruby Expressiveness and Repeating Yourself

| Comments

Just this week, I was reminded again of how Java forces you to repeat yourself. I had an object that contains a sequence of "things to be processed". The sequence has to be traversed twice, once before an extended process runs and once afterwards.

The usual Java idiom looks like this:

public void preProcess(ActionContext context) {
  for (Iterator iter = updates.iterator(); iter.hasNext(); ) {
    TwoPhaseUpdate update = (TwoPhaseUpdate) iter.next();

public void postProcess(ActionContext context) {
  for (Iterator iter = updates.iterator(); iter.hasNext(); ) {
    TwoPhaseUpdate update = (TwoPhaseUpdate) iter.next();

Notice that there are only two symbols different between these two methods, out of 20 semantically significant symbols. According to the Pragmatic Programmers, even iterating over the collection counts as a kind of repetition (and therefore a violation of DRY - don’t repeat yourself.)

The Ruby equivalent would be something like:

def preProcess(context)
   updates.each { |u| u.preProcess(context) }

def postProcess(context)
   updates.each { |u| u.postProcess(context) }

Now, there are two differening symbols out of 10 (20% variance instead of 10%). There’s been no loss of expressiveness, in fact, the main intention of the code is clearer in the Ruby version than in the Java version.

Can we make the variance higher? Perhaps.

def preProcess(context)
   each_update(:preProcess, context)

def postProcess(context)
   each_update(:postProcess, context)

def each_update(method, context)
   updates.each { |u| u.send(method, context) }

Now the two primary methods have 2 symbols out of 7 different or nearly 28%. The expressiveness is damaged a little bit by the dynamic dispatch via "send". It would be unthinkable to use reflection in Java to make the code clearer. (Anyone who’s worked with reflection knows what I mean.) Here, it’s not unthinkable, but it might just not help clarity.

Technorati Tags: java, beyondjava, ruby

MySQL 5.0 Stored Procedures

| Comments

The MySQL 5.0 release is finally adding stored procedures, triggers, and views. This is a welcome addition. With the strong storage management features, clustering, and replication from the 4.x releases, MySQL now has all the capabilities of an "enterprise" database. (Of course, the lack of these features didn’t stop thousands of users from deploying earlier versions in enterprises, even for "mission-critical" applications.)*

Here’s a fairly trivial example:

create procedure count_table_rows ()  reads sql data begin
      select table_name, table_rows from information_schema.tables;

* Somtime, I have to post about the perversions of language perpetrated by people in business. "Mission-critical" means "without this, the mission will fail." What percentage of applications labelled as mission-critical would actually cause the company to fail? Most of the time, the "mission-critical" label really just means "this application’s sponsor has large political clout".

Technorati Tags: mysql

The Dumbest Thing I’ve Seen Today

| Comments

I generally like Swing, but I just found something in the Metal L&F for JSlider that strikes me as a big WTF. The BasicSliderUI allows you to click in the "track" of the slider to scroll by a block. That’s either 10% of the span of the slider, or a minimum of 1 unit. The MetalSliderUI overrides that sensible behavior with a method that scrolls by just one unit. Period.

Here’s a quick fix:

JSlider slider = new JSlider(); 
slider.setUI(new MetalSliderUI() {
  protected void scrollDueToClickInTrack(int dir) {

Technorati Tags: java, swing

Programmer Productivity Measurements Don’t Work.

| Comments

Programmer productivity measurements don’t work.

The most common metric was discredited decades ago, but continues to be used: KLOC. Only slightly better is function points. At least it’s tied to some deliverable value. Still, the best function point is the one you don’t have to develop. Likewise, the best line of code is the one you don’t need to write. In fact, sometimes my most productive days are the ones in which I delete the most code. Why are these metrics so misleading?

Because they are counting inventory as an asset. Lines of code are inventory. Function points are inventory. Any metric that only measures the rate of inventory production is fatally flawed. We need metrics that measure throughput instead.

Technorati Tags: lean, agile

More Beanshell Goodness

| Comments

Thanks to the clean layered architecture in our application, we’ve got a very clear interface between the user interface (just Swing widgets) and the "UI Model". In the canonical MVC mode, our UI Model is part controller and part model. It isn’t the domain model, however. It’s a model of the user interface. It has concepts like "form" and "command". A "form" is mainly a collection of property objects that are named and typed. The UI interacts with the rest of the application by binding to the properties.

The upshot is that anything the UI can do by setting and getting properties (including executing commands via CommandProperty objects) can be done through test fixtures or automated interfaces. Enter beanshell.

After integrating beanshell, all of our forms and properties were immediately available. Today, I worked with one of my teammates to build a beanshell script to drive through the application. It creates a customer and goes through the entire workflow. Run the script a million times or so, and you’ve got a great pile of test data. Schema changes? Domain model changes? No problem. Just re-run the script (and wait an hour or so) and you’ve got updated test data.

Technorati Tags: java

Smalltalk Style Prototyping for Java?

| Comments

I’ve been eyeing Beanshell for some time now. It’s a very straightforward scripting language for Java. Its syntax is about what you would expect if I said, "Java with optional types and no need to pre-declare variables." So, a Java programmer probably needs all of about thirty seconds to understand the language.

What I didn’t expect was how quickly I could integrate it into my applications. Here’s an example. I’ve got a Swing desktop application for which I wanted to add a small command shell pane. I spent about an hour working with Swing’s JTextArea and rolling my own parser. It was late at night and I was short on bright ideas. Finally, about the time I realized I was going to need variables and flow control, I pulled the emergency stop cord and backed up.

After downloading the full beanshell JAR file (about 280K) and adding it to my build path, all I had to do was this:

JConsole console = new JConsole();   
frame.getContentPane().add(new JScrollPane(console), BorderLayout.SOUTH);  
Interpreter interpreter = new Interpreter(console);   
new Thread(interpreter).start(); 

Those four lines of code are literally all that is needed to get a beanshell prompt running inside of your application. From the prompt, you have access to every class and method within your application. If you’ve got some kind of object registry or namespace, or any kind of Singletons, then you’ll also have access to real object instances.

Beanshell has a built-in command called "desktop()". The one line will launch a Smalltalk-style IDE, with class browser and interpreter window. This desktop is still part of your application’s JVM. It lacks most of the power of Smalltalk’s library, which evolved together with the workspace to support highly dynamic programming. Nevertheless, the beanshell desktop retains the immediacy of working in Smalltalk.


  • Beanshell
  • Squeak - a Free, modern Smalltalk
  • Java Object Inspector - A simple Swing-based inspector, can be launched on any Java object from a Beanshell prompt. A great complement to Beanshell

Technorati Tags: agile, java

One of the Most Fun Features of My Current Project

| Comments

One of the most fun features of my current project is our "extreme feedback monitor". We’re using CruiseControl to build our entire codebase, including unit tests, acceptance tests, and quality metrics, every five minutes. To make a broken build painfully obvious, we’ve got a stoplight hanging on one wall of the room. (I may post some pictures later, if there’s interest.)

Kyle Larson found the stoplight itself in a gift shop (Spencer’s, maybe, I can’t remember… Kyle, help me out here). It had just one plug but you could push each light as a separate switch.

Well, it looks pretty dumb to walk over and push on the red light to show a broken build. It’s not pragmatic and it’s not automated. So, Kyle rewired it with two additional cords, so each lamp has its own plug.

I plugged each lamp into an X10 lamp module so each color could be turned on and off individually. I hooked a "FireCracker" wireless transmitter up to the serial port on the build box. With one switched receiver and two lamp modules, we were ready to go.

CruiseControl supports a publisher that is supposed to integrate directly with X10 devices over the serial port. Unfortunately, the installation and setup for Java programs to work with X10 devices on Linux is… problematic. First off, the JavaComm API appears to be totally stagnant. It does not support Linux at all, so you have to install the Solaris SPARC version, but supply an open-source Linux implementation of the API (www.rxtx.org), replacing a .properties file. Then you have to make sure that the user running your build loop is a member of the "tty" group. Then just cross your fingers.

I got all of the above to work from my Java test apps, but the X10 publisher built into CC still couldn’t open the serial port.

I finally gave up on the built-in publisher. I used wget, BottleRocket, and a shell script to check the build status web page every 30 seconds and change the lights accordingly.

Now, within a minute of a broken build, we can all see it. When the light is green, the build is clean.

If the red light means "broken build", and the green light means "good build", you might wonder what we use yellow for.

Yellow means that someone is in the process of synchronizing and committing code. Along with the FireCracker module, we also got a remote control. That normally sits in the middle of the tables in the lab. Whenever a pair needs to check in code, they grab the remote (i.e., take the semaphore) and turn on the yellow light. As an added "feature", the wireless switched receiver is the only module that makes an audible "click" when it switches. We use that one to control the yellow lamp, so we also have an auditory cue when a pair starts their commit dance.

After committing, the pair turns off the yellow light and replaces the remote, thus putting the semaphore and allowing the next pair to commit. In the event of multiple blocked pairs, FIFO behavior is not guaranteed. Semaphore holders have been known to be susceptible to flattery and bribery.

Technorati Tags: agile, automation, CruiseControl, pragmatic