Debugging is the cornerstone of being a programmer. The first meaning of the verb to debug is to remove errors, but the meaning that really matters is to see into the execution of a program by examining it. A programmer that cannot debug effectively is blind.

Idealists that think design, or analysis, or complexity theory, or whatnot, are more fundamental are not working programmers. The working programmer does not live in an ideal world. Even if you are perfect, your are surrounded by and must interact with code written by major software companies, organizations like GNU, and your colleagues. Most of this code is imperfect and imperfectly documented. Without the ability to gain visibility into the execution of this code the slightest bump will throw you permanently. Often this visibility can only be gained by experimentation, that is, debugging.

Debugging is about the running of programs, not programs themselves. If you buy something from a major software company, you usually don't get to see the program. But there will still arise places where the code does not conform to the documentation (crashing your entire machine is a common and spectacular example), or where the documentation is mute. More commonly, you create an error, examine the code you wrote and have no clue how the error can be occurring. Inevitably, this means some assumption you are making is not quite correct, or some condition arises that you did not anticipate. Sometimes the magic trick of staring into the source code works. When it doesn't, you must debug.

To get visibility into the execution of a program you must be able to execute the code and observe something about it. Sometimes this is visible, like what is being displayed on a screen, or the delay between two events. In many other cases, it involves things that are not meant to be visible, like the state of some variables inside the code, which lines of code are actually being executed, or whether certain assertions hold across a complicated data structure. These hidden things must be revealed.

The common ways of looking into the ‘innards’ of an executing program can be categorized as:

Debugging tools are wonderful when they are stable and available, but the printlining and logging are even more important. Debugging tools often lag behind language development, so at any point in time they may not be available. In addition, because the debugging tool may subtly change the way the program executes it may not always be practical. Finally, there are some kinds of debugging, such as checking an assertion against a large data structure, that require writing code and changing the execution of the program. It is good to know how to use debugging tools when they are stable, but it is critical to be able to employ the other two methods.

Some beginners fear debugging when it requires modifying code. This is understandable---it is a little like exploratory surgery. But you have to learn to poke at the code and make it jump; you have to learn to experiment on it, and understand that nothing that you temporarily do to it will make it worse. If you feel this fear, seek out a mentor---we lose a lot of good programmers at the delicate onset of their learning to this fear.

Debugging is fun, because it begins with a mystery. You think it should do something, but instead it does something else. It is not always quite so simple---any examples I can give will be contrived compared to what sometimes happens in practice. Debugging requires creativity and ingenuity. If there is a single key to debugging is to use the divide and conquer technique on the mystery.

Suppose, for example, you created a program that should do ten things in a sequence. When you run it, it crashes. Since you didn't program it to crash, you now have a mystery. When out look at the output, you see that the first seven things in the sequence were run successfully. The last three are not visible from the output, so now your mystery is smaller: ‘It crashed on thing #8, #9, or #10.’

Can you design an experiment to see which thing it crashed on? Sure. You can use a debugger or we can add printline statements (or the equivalent in whatever language you are working in) after #8 and #9. When we run it again, our mystery will be smaller, such as ‘It crashed on thing #9.’ I find that bearing in mind exactly what the mystery is at any point in time helps keep one focused. When several people are working together under pressure on a problem it is easy to forget what the most important mystery is.

The key to divide and conquer as a debugging technique is the same as it is for algorithm design: as long as you do a good job splitting the mystery in the middle, you won't have to split it too many times, and you will be debugging quickly. But what is the middle of a mystery? There is where true creativity and experience comes in.

To a true beginner, the space of all possible errors looks like every line in the source code. You don't have the vision you will later develop to see the other dimensions of the program, such as the space of executed lines, the data structure, the memory management, the interaction with foreign code, the code that is risky, and the code that is simple. For the experience programmer, these other dimensions form an imperfect but very useful mental model of all the things that can go wrong. Having that mental model is what helps one find the middle of the mystery effectively.

Once you have evenly subdivided the space of all that can go wrong, you must try to decide in which space the error lies. In the simple case where the mystery is: ‘Which single unknown line makes my program crash?’, you can ask yourself: ‘Is the unknown line executed before or after this line that I judge to be executed in the about the middle of the running program?’ Usually you will not be so lucky as to know that the error exists in a single line, or even a single block. Often the mystery will be more like: ‘Either there is a pointer in that graph that points to the wrong node, or my algorithm that adds up the variables in that graph doesn't work.’ In that case you may have to write a small program to check that the pointers in the graph are all correct in order to decide which part of the subdivided mystery can be eliminated.

I've intentionally separated the act of examining a program's execution from the act of fixing an error. But of course, debugging does also mean removing the bug. Ideally you will have perfect understanding of the code and will reach an ‘A-Ha!’ moment where you perfectly see the error and how to fix it. But since your program will often use insufficiently documented systems into which you have no visibility, this is not always possible. In other cases the code is so complicated that your understanding cannot be perfect.

In fixing a bug, you want to make the smallest change that fixes the bug. You may see other things that need improvement; but don't fix those at the same time. Attempt to employ the scientific method of changing one thing and only one thing at a time. The best process for this is to be able to easily reproduce the bug, then put your fix in place, and then rerun the program and observe that the bug no longer exists. Of course, sometimes more than one line must be changed, but you should still conceptually apply a single atomic change to fix the bug.

Sometimes, there are really several bugs that look like one. It is up to you to define the bugs and fix them one at a time. Sometimes it is unclear what the program should do or what the original author intended. In this case, you must exercise your experience and judgment and assign your own meaning to the code. Decide what it should do, and comment it or clarify it in some way and then make the code conform to your meaning. This is an intermediate or advanced skill that is sometimes harder than writing the original function in the first place, but the real world is often messy. You may have to fix a system you cannot rewrite.

Logging is the practice of writing a system so that it produces a sequence of informative records, called a log. Printlining is just producing a simple, usually temporary, log. Absolute beginners must understand and use logs because their knowledge of the programming is limited; system architects must understand and use logs because of the complexity of the system. The amount of information that is provided by the log should be configurable, ideally while the program is running. In general, logs offer three basic advantages:

The amount to output into the log is always a compromise between information and brevity. Too much information makes the log expensive and produces scroll blindness, making it hard to find the information you need. Too little information and it may not contain what you need. For this reason, making what is output configurable is very useful. Typically, each record in the log will identify its position in the source code, the thread that executed it if applicable, the precise time of execution, and, commonly, an additional useful piece of information, such as the value of some variable, the amount of free memory, the number of data objects, etc. These log statements are sprinkled throughout the source code but are particularly at major functionality points and around risky code. Each statement can be assigned a level and will only output a record if the system is currently configured to output that level. You should design the log statements to address problems that you anticipate. Anticipate the need to measure performance.

If you have a permanent log, printlining can now be done in terms of the log records, and some of the debugging statements will probably be permanently added to the logging system.

Learning to understand the performance of a running system is unavoidable for the same reason that learning debugging is. Even if you understand perfectly and precisely the cost of the code you write, your code will make calls into other software systems that you have little control over or visibility into. However, in practice performance problems are a little different and a little easier than debugging in general.

Suppose that you or your customers consider a system or a subsystem to be too slow. Before you try to make it faster, you must build a mental model of why it is slow. To do this you can use a profiling tool or a good log to figure out where the time or other resources are really being spent. There is a famous dictum that 90% of the time will be spent in 10% of the code. I would add to that the importance of input/output expense (I/O) to performance issues. Often most of the time is spent in I/O in one way or another. Finding the expensive I/O and the expensive 10% of the code is a good first step to building your mental model.

There are many dimensions to the performance of a computer system, and many resources consumed. The first resource to measure is wall--clock time, the total time that passes for the computation. Logging wall-clock time is particularly valuable because it can inform about unpredictable circumstance that arise in situations where other profiling is impractical. However, this may not always represent the whole picture. Sometimes something that takes a little longer but doesn't burn up so many processor seconds will be much better in computing environment you actually have to deal with. Similarly, memory, network bandwidth, database or other server accesses may, in the end, be far more expensive than processor seconds.

Contention for shared resources that are synchronized can cause deadlock and starvation. Deadlock is the inability to proceed because of improper synchronization or resource demands. Starvation is the failure to schedule a component properly. If it can be at all anticipated, it is best to have a way of measuring this contention from the start of your project. Even if this contention does not occur, it is very helpful to be able to assert that with confidence.

Most software projects can be made 10 to 100 times faster than they are at the time that they are first released with relatively little effort. Under time-to-market pressure, it is both wise and effective to choose a solution that gets the job done simply and quickly, but less efficiently than some other solution. However, performance is a part of usability, and often it must eventually be considered more carefully.

The key to improving the performance of a very complicated system is to analyze it well enough to find the bottlenecks, or places where most of the resources are consumed. There is not much sense in optimizing a function that accounts for only 1% of the computation time. As a rule of thumb you should think carefully before doing anything unless you think it is going to make the system or a significant part of it at least twice as fast. There is usually a way to do this. Consider the test and quality assurance effort that your change will require. Each change brings a test burden with it, so it is much better to have a few big changes.

After you've made a two-fold improvement in something, you need to at least rethink and perhaps reanalyze to discover the next-most-expensive bottleneck in the system, and attack that to get another two-fold improvement.

Often, the bottlenecks in performance will be an example of counting cows by counting legs and dividing by four, instead of counting heads. For example, I've made errors such as failing to provide a relational database system with a proper index on a column I look up a lot, which probably made it at least 20 times slower. Other examples include doing unnecessary I/O in inner loops, leaving in debugging statements that are no longer needed, unnecessary memory allocation, and, in particular, inexpert use of libraries and other subsystems that are often poorly documented with respect to performance. This kind of improvement is sometimes called low-hanging fruit, meaning that it can be easily picked to provide some benefit.

What do you do when you start to run out of low-hanging fruit? Well, you can reach higher, or chop the tree down. You can continue making small improvements or you can seriously redesign a system or a subsystem. (This is a great opportunity to use your skills as a good programmer, not only in the new design but also in convincing your boss that this is a good idea.) However, before you argue for the redesign of a subsystem, you should ask yourself whether or not your proposal will make it five to ten time better.

Sometimes you'll encounter loops, or recursive functions, that take a long time to execute and are bottlenecks in your product. Before you try to make the loop a little faster, spend a few minutes considering if there is a way to remove it entirely. Would a different algorithm do? Could you compute that while computing something else? If you can't find away around it, then you can optimize the loop. This is simple; move stuff out. In the end, this will require not only ingenuity but also an understanding of the expense of each kind of statement and expression. Here are some suggestions:

The cost of each of these operations depends on your specific system. On some systems compilers and hardware do these things for you. Clear, efficient code is better than code that requires an understanding of a particular platform.

For a lot of problems, processors are fast compared to the cost of communicating with a hardware device. This cost is usually abbreviated I/O, and can include network cost, disk I/O, database queries, file I/O, and other use of some hardware not very close to the processor. Therefore building a fast system is often more a question of improving I/O than improving the code in some tight loop, or even improving an algorithm.

There are two very fundamental techniques to improving I/O: caching and representation. Caching is avoiding I/O (generally avoiding the reading of some abstract value) by storing a copy of that value locally so no I/O is performed to get the value. The first key to caching is to make it crystal clear which data is the master and which are copies. There is only one master---period. Caching brings with it the danger that the copy is sometimes can't reflect changes to the master instantaneously.

Representation is the approach of making I/O cheaper by representing data more efficiently. This is often in tension with other demands, like human readability and portability.

Representations can often be improved by a factor of two or three from their first implementation. Techniques for doing this include using a binary representation instead of one that is human readable, transmitting a dictionary of symbols along with the data so that long symbols don't have to be encoded, and, at the extreme, things like Huffman encoding.

A third technique that is sometimes possible is to improve the locality of reference by pushing the computation closer to the data. For instance, if you are reading some data from a database and computing something simple from it, such as a summation, try to get the database server to do it for you. This is highly dependent on the kind of system you're working with, but you should explore it.

Memory is a precious resource that you can't afford to run out of. You can ignore it for a while but eventually you will have to decide how to manage memory.

Space that needs to persist beyond the scope of a single subroutine is often called heap allocated. A chunk of memory is useless, hence garbage, when nothing refers to it. Depending on the system you use, you may have to explicitly deallocate memory yourself when it is about to become garbage. More often you may be able to use a system that provides a garbage collector. A garbage collector notices garbage and frees its space without any action required by the programmer. Garbage collection is wonderful: it lessens errors and increases code brevity and concision cheaply. Use it when you can.

But even with garbage collection, you can fill up all memory with garbage. A classic mistake is to use a hash table as a cache and forget to remove the references in the hash table. Since the reference remains, the referent is noncollectable but useless. This is called a memory leak. You should look for and fix memory leaks early. If you have long running systems memory may never be exhausted in testing but will be exhausted by the user.

The creation of new objects is moderately expensive on any system. Memory allocated directly in the local variables of a subroutine, however, is usually cheap because the policy for freeing it can be very simple. You should avoid unnecessary object creation.

An important case occurs when you can define an upper bound on the number of objects you will need at one time. If these objects all take up the same amount of memory, you may be able to allocate a single block of memory, or a buffer, to hold them all. The objects you need can be allocated and released inside this buffer in a set rotation pattern, so it is sometimes called a ring buffer. This is usually faster than heap allocation.

Sometimes you have to explicitly free allocated space so it can be reallocated rather than rely on garbage collection. Then you must apply careful intelligence to each chunk of allocated memory and design a way for it to be deallocated at the appropriate time. The method may differ for each kind of object you create. You must make sure that every execution of a memory allocating operation is matched by a memory deallocating operation eventually. This is so difficult that programmers often simply implement a rudimentary form or garbage collection, such as reference counting, to do this for them.

The intermittent bug is a cousin of the 50-foot-invisible-scorpion-from-outer-space kind of bug. This nightmare occurs so rarely that it is hard to observe, yet often enough that it can't be ignored. You can't debug it because you can't find it.

Although after eight hours you will start to doubt it, the intermittent bug has to obey the same laws of logic everything else does. What makes it hard is that it occurs only under unknown conditions. Try to record the circumstances under which the bug does occur, so that you can guess at what the variability really is. The condition may be related to data values, such as ‘This only happens when we enter Wyoming as a value.’ If that is not the source of variability, the next suspect should be improperly synchronized concurrency.

Try, try, try to reproduce the bug in a controlled way. If you can't reproduce it, set a trap for it by building a logging system, a special one if you have to, that can log what you guess you need when it really does occur. Resign yourself to that if the bug only occurs in production and not at your whim, this is may be a long process. The hints that you get from the log may not provide the solution but may give you enough information to improve the logging. The improved logging system may take a long time to be put into production. Then, you have to wait for the bug to reoccur to get more information. This cycle can go on for some time.

The stupidest intermittent bug I ever created was in a multi-threaded implementation of a functional programming language for a class project. I had very carefully insured correct concurrent evaluation of the functional program, good utilization of all the CPUs available (eight, in this case). I simply forgot to synchronize the garbage collector. The system could run a long time, often finishing whatever task I began, before anything noticeable went wrong. I'm ashamed to admit I had begun to question the hardware before my mistake dawned on me.

At work we recently had an intermittent bug that took us several weeks to find. We have multi-threaded application servers in Java™ behind Apache™ web servers. To maintain fast page turns, we do all I/O in small set of four separate threads that are different than the page-turning threads. Every once in a while these would apparently get ‘stuck’ and cease doing anything useful, so far as our logging allowed us to tell, for hours. Since we had four threads, this was not in itself a giant problem---unless all four got stuck. Then the queues emptied by these threads would quickly fill up all available memory and crash our server. It took us about a week to figure this much out, and we still didn't know what caused it, when it would happen, or even what the threads where doing when they got ‘stuck’.

This illustrates some risk associated with third-party software. We were using a licensed piece of code that removed HTML tags from text. Due to its place of origin we affectionately referred to this as ‘the French stripper.’ Although we had the source code (thank goodness!) we had not studied it carefully until by turning up the logging on our servers we finally realized that the email threads were getting stuck in the French stripper.

The stripper performed well except on some long and unusual kinds of texts. On these texts, the code was quadratic or worse. This means that the processing time was proportional to the square of the length of the text. Had these texts occurred commonly, we would have found the bug right away. If they had never occurred at all, we would never have had a problem. As it happens, it took us weeks to finally understand and resolve the problem.

To learn how to design software, study the action of a mentor by being physically present when they are designing. Then study well-written pieces of software. After that, you can read some books on the latest design techniques.

Then you must do it yourself. Start with a small project. When you are finally done, consider how the design failed or succeeded and how you diverged from your original conception. They move on to larger projects, hopefully in conjunction with other people. Design is a matter of judgment that takes years to acquire. A smart programmer can learn the basics adequately in two months and can improve from there.

It is natural and helpful to develop your own style, but remember that design is an art, not a science. People who write books on the subject have a vested interest in making it seem scientific. Don't become dogmatic about particular design styles.

The late, great Edsger Dijkstra has eloquently explained that Computer Science is not an experimental science[ExpCS] and doesn't depend on electronic computers. As he puts it referring to the 1960s[Knife],

Programming ought not to be an experimental science, but most working programmers do not have the luxury of engaging in what Dijkstra means by computing science. We must work in the realm of experimentation, just as some, but not all, physicists do. If thirty years from now programming can be performed without experimentation, it will be a great accomplishment of Computer Science.

The kinds of experiments you will have to perform include:

I don't think in this essay I can explain the design of experiments; you will have to study and practice. However, I can offer two bits of advice.

First, try to be very clear about your hypothesis, or the assertion that you are trying to test. It also helps to write the hypothesis down, especially if you find yourself confused or are working with others.

You will often find yourself having to design a series of experiments, each of which is based on the knowledge gained from the last experiment. Therefore, you should design your experiments to provide the most information possible. Unfortunately, this is in tension with keeping each experiment simple---you will have to develop this judgment through experience.