Bug Hunting Adventures #4: Casino Royale

On August 18, 1913, something very strange happened at Monaco’s Monte Carlo casino: for the twentieth time in a row, the ball had fallen into a black pocket. More and more people gathered at the table and started betting like crazy for red — they were convinced that the time for red was more than overdue. Well, most of them lost a lot of money because it took another seven spins of the wheel until red finally arrived.

Today, I present a little Groovy program that I wrote in order to get insight into Roulette probability, to save us from what is now known as the Monte Carlo Fallacy.

When you call ‘runSimulation’ you can specify how many number of times you want to spin the wheel. With every spin, the method ‘spinWheel’ returns either 0 (for black), 1 (for red), or 2 (for Zero).

‘runSimulation’ keeps track of the length of a certain color series in three associative arrays (i. e. maps), one for every color (including the color Zero). The key into these maps is the length of the series and the associated value is the count of how often this series length was encountered during the experiment.

To make accessing these color length maps generic, their references are stored in a plain list (‘seriesMaps’). By using the color value returned from ‘spinWheel’ as an index, one can easily obtain the length map for a particular color. This list of length maps is returned at the end of the simulation. As an example, after 100 spins, the list might look like this:

[ [2:6, 3:3, 1:14, 5:1], [3:5, 1:8, 2:9, 4:1, 5:1, 7:1], [1:3] ]

In this simulation, for black, a series of length 2 was encountered 6 times, and a length of 5 one time. For red, there was one series length of 7, and our virtual ball landed three times on Zero, but there was never a series of Zeros (i. e. longer than 1). (Aside: in one experiment I did with the bug-fixed version, I spun the wheel more than 100 million times; I got a maximum series of 30 for black and — believe it or not — I once got a series of four times Zero in a row.)

But the program, as it is presented to you, contains a bug. It just doesn’t work as it is supposed to. Can you spot it? (Note: the bug has nothing to do with any Groovy idiosyncrasies.)


Bug Hunting Adventures #3: Silent Threads

‘select’ isn’t broken
— The Pragmatic Programmers

Many moons ago, when I tried to familiarize myself with POSIX threads, I wrote a simple test program that was based on a textbook example.

My program sported two threads, one printing ‘+’ characters, the other one printing ‘-‘ characters. Everything worked as expected: a mixed stream of ‘+’ and ‘-‘ characters was emitted to stdout.

But everything happened so fast! Literally thousands of characters were outputted at the blink of an eye, so I added a little extra code that made the threads sleep for a specified amount of time before printing the next character.

Alas, when I set the delay (SLEEP_SECS) to 1 second (or in fact any value different to zero) nothing was printed at all! It looked like the threads got locked up completely. I came up with the weirdest theories about what had happened, including a bug in the pthreads library and the implementation of ‘sleep’.

It wasn’t until the next morning that I realized my mistake. Once I again, I had blamed it on the good ones, when the real problem was blind stupidity.

What was my mistake?


Growing a Solid Software Company — Update

Yesterday, on October 21, all Lufhansa pilots went on strike in Germany. Nevertheless, Lufhansa managed to conduct half of their flights. They achieved this miracle by using a two-fold strategy: subcontractors and — lo and behold — their own managers, who, according to Lufthansa, work in most cases as part-time pilots, anyway.

Pilots that manage, managers that fly. Exactly my words, friends, exactly my words…

Let’s sing a song together, shall we?

I see trees of green, red roses, too,
I see them bloom, for me and you
And I think to myself
What a wonderful world.

Growing a Solid Software Company

Isn’t it a shame that so many software development managers don’t code anymore?

Since I am in a malicious mood today, I claim that they didn’t even write much code when they were still developers. But does it always have to be like this?

After years of deep contemplation I’ve come up with this rule:

Managers, regardless of their position in a hierarchy, should spend at least one-third of their time doing the work of their immediate subordinates.

Not only would this mean that every manager is productive and actively contributes to a project; it would also mean that managers stay current from a technological point of view. Especially the latter would ensure that their strategic decisions are based on much firmer ground.

Wouldn’t performance appraisals (and promotions) suddenly become more objective and fair? Wouldn’t it be much easier for managers to hire new people since they would know — from first-hand experience — what exactly to look for?

If this recursive rule were applied, even the software manager’s boss would write small parts of the software himself. His boss, in turn, would probably not code that much but maybe do some code-reviews or check the nightly build for compiler warnings.

Wouldn’t the code quality be much higher if developers knew that someone way up the corporate ladder scrutinizes their work and gave feedback? Wouldn’t everyone feel much better because they knew that their bosses really cared for what they do?

This is a rule for building up a hierarchy of software craftsmen, a rule that yields what I call a “Solid Software Company”: A company were everyone is a developer (at least to some extent), where everyone understands software’s true nature and developer’s needs.

Imagine you could travel back in time, to the early days of a once hip, now bureaucratic, politics-laden, inefficient, dreadful-to-work-for monster of a software company. You will arrive at a point at which they suddenly start to promote or hire people to be “just managers”.

So, viewed from a different, more negative angle, my rule can be rephrased like this:

The long and slow demise of a young, aspiring software company begins when its software development managers cease to write code.

Bug Hunting Adventures #2: Monitoring Temperature Sensors

Imagine a distributed control system that relies on correct and timely temperature measurements. Various temperature controllers distribute their temperature readings periodically over a bus for other controllers to consume.

To ensure that the temperature sensors work as expected, a ‘TemperatureMonitor’ was implemented. It listens for messages from the temperature controllers and checks whether they arrive in time (at the very latest every 20 ms) and whether their values are within the valid range.

This checking of temperature messages upon arrival obviously doesn’t catch cases where temperature controllers don’t send messages at all (or with a delay that would cause the timer to overflow multiple times) so there is an additional cyclic task that caters for situations like these. This cyclic task is executed every 100 ms and it additionally takes care of the so-called ‘idle period’.

During the ‘idle period’, which is 500 ms, ‘TemperatureMonitor’ is lenient and doesn’t report problems (if any) to allow for an undisturbed start-up of the whole system.

Once the idle period is over and ‘TemperatureMonitor’ detects abnormal conditions it notifies the global ‘ErrorManager’, which will decide how to handle problems based on its current error handling policy (eg. just log the error, reset or disable a temperature controller).

Since the cyclic task and the message handler (the one that is invoked upon the reception of a temperature sensor message) may run in parallel, access to shared data is protected by a simple (but sufficient) synchronization scheme based on enabling/disabling interrupts.

Time measurement is implemented based on a wrap-around 32-bit tick counter that counts raw clock cycles. These clock cycles are converted to a more convenient unit (ie. milliseconds).


Bug Hunting Adventures #1: Logging Binary Data

Normally, loggers are used to track down bugs, but today, I present a Logger class method that contains a bug itself — isn’t it ironic?

Logger::logbuf() takes ‘len’ bytes of binary data from memory pointed to by ‘buf’ and converts it into printable, zero-terminated hex strings ala “AA 01 B3 C4…”.

For efficiency and readability, the hex string is broken up into smaller parts. Every hex string part is fed to the existing Logger::log() method which is capable of outputting arbitrary, zero-terminated strings (a new-line character is appended automatically).

Happy bug hunting! I will post the solution in two weeks time.


Bug Hunting Adventures

I’ve always loved to find bugs in code through mental debugging — especially other people’s code. Besides being fun, bug hunting improves product quality as well as one’s programming and problem-solving skills.

One of the books I enjoyed most is “Find the Bug” by Adam Barr. It contains many examples of buggy code in various programming languages, ranging from assembly language to Python. I heartily recommend it to everyone who shares my passion.

What a pity that there aren’t more books like this on the market!

To relieve our misery, I’ve decided to setup a new series: at irregular intervals I will post a bug-afflicted piece of code and challenge you to spot the problem. I don’t want to limit this sport to particular programming languages or defect categories; some bugs will be straightforward while others may be intricate, potentially only showing under certain, favorable circumstances.

Anyway, I hope you enjoy this new series as much as I do — stay tuned, the hunting season is open!

The God’s Equation for Perfect Highballs

“Mix a stranger a good drink and no longer will he be a stranger”
— Yours truly

I’ve never been a fan of pure mathematics. Even though I must admit that pure maths bears a lot of elegance, it’s real-world applicability that makes mathematical discoveries shine.

Boy oh boy, how time flies! It’s summertime again, at least in the northern hemisphere, on this planet, and what a great time it is for drinking plenty of highballs. For those of you who don’t know: highballs are simple long drinks that are made up of two ingredients: liquor (eg. gin, whisk(e)y, rum) and a non-alcoholic beverage (eg. soda, juice, cola).

Even though there are usually only two ingredients it is not that easy to make a perfect highball. It all depends on the right mixture. Some like their highballs stronger, others lighter. So how do you determine the perfect ratio? The task is complicated by the fact that the strength of the liquor used might vary (for instance, rums come with alcoholic strength levels reaching from 35% to 80%).

So I sat down for a while and did a little math — to solve a real-world problem once and for all. What I’ve come up with is this:

Vn / Vl = R = (Sl / Sh) - 1

In this equation, Vn is the volume (amount) of non-alcoholic beverage, Vl is the volume (amount) of liquor, R is the mix ratio, Sl is the alcoholic strength of the liquor used and Sh is the desired target strength of the resulting highball. Isn’t this wonderful?

As an example, I happen to like my highballs with a strength of 10% and most liquor in Europe is sold with a strength of 40%. What’s the recipe for a perfect Gin and Tonic? Answer: 40% divided by 10% is four, minus one is three. Therefore, you would need three parts of tonic water for one part of Gin to make me smile.

If you prefer a 5% Gin and Tonic (are you a wimp, by any chance?), note that you will need seven parts of tonic water, not six, as a naive would-be drinks mixer would expect.

There you have it — the formula for perfect, repeatable quality drinks. People in the same spirit (pun intended!) — let’s raise our glasses to this discovery!


I know, I know. Some of you smug, mathematically inclined folks with an IQ of 180+ would probably have known this right away. I don’t mind if you point out that my pathetic “discovery” is obviously just a special case of something more profound. Congratulations! I will drink a very special toast to you tonight.

And — just in case you haven’t done the maths yourself in the meantime, here is how I derived it:

Strength of liquor Sl is volume of alcohol in liquor Val divided by the total volume of liquor Vl:

Val / Vl = Sl         Eq (1)
Val = Sl * Vl         Eq (2)

Total volume of highball Vh is volume of liquor Vl + volume of non-alcoholic beverage Vn:

Vh = Vl + Vn          Eq (3)

Strength of highball Sh is volume of alcohol from liquor Val divided by the total volume of highball:

Sh = Val / Vh         Eq (4)

Inserting (2) in (4):

Sh = (Sl * Vl) / Vh   Eq (5)

Inserting (3) in (5):

Sh = (Sl * Vl) / (Vl + Vn)
Sh * (Vl + Vn) = Sl * Vl
Sh * Vl + Sh * Vn = Sl * Vl
Vl * Sh - Vl * Sl = - Vn * Sh
Vl * (Sh - Sl) = -Vn * Sh
Vl / Vn = - Sh / (Sh - Sl)
Vn / Vl = (Sl - Sh) / Sh
Vn / Vl = (Sl / Sh) - 1


A Tale of Two Qualities

Green and rotten. Isolated odject. Element of design.

“Here then, as I lay down the pen and proceed to seal up my confession, I bring the life of that unhappy Henry Jekyll to an end.”

— Robert Louis Stevenson, The Strange Case of Dr. Jekyll and Mr. Hyde

The old saying “What gets measured gets done” makes immediate sense: before you can improve anything, you first have to gain insight into it. If you act, you want to see the consequences, as soon as possible, compare them to desired results and adapt accordingly. Further, measurements not only give you a picture of the past and present — through extrapolation you get an impression of the future as well.

In modern software development, we constantly measure the quality of our product right from the beginning. We check for build breakers and failing test cases, code coverage, memory consumption and execution times on a check-in basis, as part of our continuous integration process. We always know the quality of our product — there won’t be any big surprises at major milestones or the end of the project. Running a software project like this removes the chance factor — this is quite the opposite of what happens when you follow a waterfall model with its dreaded big-bang integration phases.

But let’s face it: we mainly focus on controlling external quality; that is, everything that is visible to the customer.

The level of external quality is usually (relatively) easy to determine since external requirements are specified such that they are unambiguously verifiable (at least they should be). Thus, developers and testers can implement automated tests that will reveal any deviations. Because of the fact that external quality is directly visible to customers and directly influences whether they buy (or return) a product or not, it is not difficult to get proper funding for people and tools. In this sense, external quality really lives on the sunny side of software life.

Sadly, external quality’s brother, internal quality, lives an unhappy life in the dark: internal quality requirements are usually not nailed down precisely, in fact, they are often fuzzy and neglected. Sure, there are coding guidelines that mandate a certain indentation style and whether to use tabs or spaces (among other things) but are there actually checks against violations? And what about compiler warnings, bad coding practices, size of classes/methods, cyclomatic complexity, coupling of classes, missing API documentation and so on? And who is willing to invest in work the customer doesn’t see, anyway?

In most cases the answer is a blatant ‘No’. Neither are internal quality requirements stated in a precise and verifiable way, nor are they given high priority and almost never are they automatically checked as part of the continuous integration process. Since internal quality is not observable by the user, you can get away without paying attention. But then you have to pay for something else: an ever-increasing amount of technical debt and the compound interest that ensues.

Why does this all matter? In order to compete, a software product undergoes a multitude of major and minor modifications over many years, often carried out by developers other than the original authors. Once the first release is shipped, maintainability becomes a crucial factor and internal quality determines whether future changes this will be easy (cheap), hard (expensive), or outright impossible. Don’t underestimate the impact of even the smallest things; even a single broken window may be responsible for decay and increase of crime level in town districts.

What we need is a shift in mindset: first, internal quality requirements must get the same priority as external quality requirements and second, internal quality is perpetually tracked (and acted upon) as part of the continuous integration process.

It is of utmost importance to keep track of internal quality right from the beginning, when the mental distance is low and issues can be corrected with the least amount of work. Further, the whole team immediately benefits from the improved quality and the learning effect ensures that the amount of effort that needs to be spent goes rapidly down. Postponing internal quality work to late phases of the project is a costly and frustrating experience.

For some internal quality metrics it is necessary to build special-purpose analysis tools that cannot be bought off the shelf, but dynamic/scripting languages and heuristic regex-based parsing can go a long way. But even if you need more precision there are often free/open-source tools and frameworks at your disposal: I know of one C++ project that implemented several clang plugins to ensure that (among other things) identifier names where in line with the coding conventions.

Every developer has different standards as to what good-enough internal quality is. To make matters worse, day-to-day routine and especially schedule pressure will lure developers into accepting the status quo of their code once testing has proved that external quality criteria are met. But the converse is also true: without indisputable measurements, developers might spend more time than needed, endlessly polishing their beloved code.

We should give both sides of the quality coin equal consideration. While external quality is the necessary basis for the short-term success of a product, it is internal quality that ultimately determines the long-term success of a software company. Hence, we should remove the chance factor of internal quality, too.

Introduction to PC-Lint

I’ve uploaded my presentation “An Introduction to PC-Lint” to Slideshare. Share and Enjoy!