I don’t know what programming language you use, but my experience of recent coding in Java, Ruby, JavaScript, PHP, Python, C++, and Rust tells me that the principle, which I will try to convince you to adhere to — is universal for all languages. It’s about the naming of test files. It may look to you like a question of low importance, but let me try to demonstrate that it’s not. How do you name your files with test classes? How many of them do you create in the src/test/java directory? Where do you place a class that is used only in a test but is not a test by itself? To most of these questions, the answer most of you would give is “Whatever!” So let’s try to find a better answer.

It often happens, especially in Java, that a few places in the method are potential exception originators. Usually, we make a large method-size try block with a single catch at the bottom. We catch all the exceptions, usually even using grouping. This helps us minimize the noise, which is the exception catching. However, such large try blocks jeopardize maintainability: we are unable to provide proper error context inside catch blocks.

Sometimes we rethrow exceptions. In Java we do this more often than in other languages, because it has checked exceptions. Sometimes we must catch and rethrow a few exceptions that originated from different places in a method. Java 7 introduced grouping of different types of exceptions in a single catch block. But even without the grouping, it is possible to just catch IOException or even Exception and provide a single catch block for all types and all originators (methods that throw). Recently I realized that this is a bad practice. Here is why.

My first academic paper was accepted and published in 1998. My most recent one was rejected by SPLASH just a week ago. I’m writing two papers right now and co-authoring about ten others. So far, my results are very poor: way more rejections than I expected. The key lesson I’ve learned so far: the reason most papers get rejected is not because they are essentially wrong. Instead, reviewers reject them because they have no enthusiasm for decrypting a poorly structured text, even though it may contain potentially interesting thoughts. In this blog post I suggest how to structure a research paper. Take it with a grain of salt though.

A few months ago I made a small Java library, which is worth explaining since the design of its classes and interfaces is pretty unusual. It’s very much object-oriented for a pretty imperative task: building a pipeline of document transformations. The goal was to do this in a declarative and immutable way, and in Java. Well, as much as it’s possible.

I suggested this idea a few weeks ago on Twitter and got mostly negative reactions. That’s why I wrote this blog post, to elaborate on the subject in an attempt to convince you. Here is the rule I’m suggesting: always submit changes to the code separately from the changes to its unit tests. Simply put, in one pull request you modify the tests, maybe marking some of them as “disabled.” You merge this pull request and then make the second one, modifying the code and most probably removing the “disabled” markers from the tests. You don’t touch the body of the tests in the second pull request.

A few years ago I created xcop, a simple command line tool that can check the style of an XML file. It’s similar to Checkstyle (for Java) and Pep8 (for Python), but for XML. It’s pretty easy to use xcop: just run it with a few command line arguments and it returns the list of errors found in your XML file, if there are any. However, some of the arguments may be convenient to have as defaults and instead of passing them through the command line on every execution, we could store them in some configuration file. The question is: What is the best format for this file? YAML, JSON, or TOML? None of them! I suggest plain text.

Automated tests are the ones that are usually called unit tests or integration tests, or just any tests that are being executed automatically. That’s the difference between them and manual tests. What is the purpose of automated tests? First and foremost, they reduce the amount of routine work: we don’t need to remember how to test a module, the tests remember. We just click a button and a suite of tests, which may consist of hundreds or thousands, runs and reports errors, if any are found. Saving time is important, but it’s not the only and, if you ask me, not the most important purpose of automated tests. A more important one is their role as a safety net.

When I make a slide deck for a new presentation, invent a new domain name, think about a name for a new Java class, itemize bullet points in an academic paper, even write an email—I try to follow a simple principle which helps me make my content more solid. Well, at least I believe it does. Maybe it will help you as well. The principle is simple: at all costs, try to squeeze the content into one word, one sentence, one paragraph, or one page.

Reflective programming (or reflection) happens when your code changes itself on the fly. For example, a method of a class, when we call it, among other things adds a new method to the class (also known as monkey patching). Java, Python, PHP, JavaScript, you name it—they all have this “powerful” feature. What’s wrong with it? Well, it’s slow, dangerous, and hard to read and debug. But all that is nothing compared with the coupling it introduces to the code.

For each accepted explanation of a phenomenon, there may be an extremely large, perhaps even incomprehensible, number of possible and more complex alternatives. The principle of parsimony, also known as Occam’s razor, suggests we prefer the simplest one. For example, “I can’t open the door and can’t attend the meeting” is a description of a problem, which could be reduced to “I can’t open the door” without losing any information, which might be important for those who are waiting for me in the meeting room. I suggest applying the same principle to bug reports.

A few days ago, DeepMind (acquired by Google in 2014) released AlphaCode and self-published a paper explaining how their artificial intelligence (AI) can “understand” a programming contest task written in English and then write a Python, Java or C++ program, which would work in about 30% of cases. Earlier last year OpenAI ($1B-funded by Microsoft in 2019) released Codex and published a paper, claiming that their AI can also solve around 30% of the programming tasks it was tested with. Wired, the Financial Times, The Verge and many others have already announced the victory: AI will replace programmers and we are all going to lose our jobs.

A few months ago I got an opportunity to teach a single course for 3rd-year BSc students at Innopolis University (Russia). The title was “System Software Design.” The size of the group was about 150 people and the duration was 8 weeks. I was supposed to give them sixteen lectures, two lectures per week. And I was asked to examine their knowledge by the end of the course. Pretty much a normal job for a university teacher. And you know, in my opinion, I failed most parts of it. Here is what I learned.

Since the time of Kernighan and Ritchie we share binary code in libraries. You need to print some text with printf() in C++? You get libc library with 700+ other functions inside. You need to copy a Java stream? You get Apache Commons IO with copy() and 140+ other methods and classes. The same happens in all languages I’m aware of, like Ruby, Python, JavaScript, PHP: you need an object, or a class, or a function, or a method—you have to add the entire library to your build. Wouldn’t it be more elegant to deal with individual objects instead?

It’s a well-known problem nowadays: how can we measure the performance and productivity of individual contributors who do non-routine creative work? The best examples are research and development (R&D) teams, which usually consist of software engineers, designers, scientists, architects, quality experts, product managers, and so on. Such professionals deliver results that are hard to get down to simple numbers. Many authors argue that the very idea of measuring individual performance is toxic and may only lead to negative consequences. We tried to challenge this point of view and did an experiment in our team, which demonstrated that individual performance can indeed be measured, even if people’s work involves creativity, and results are hard to predict. We designed a system of Calibrated Achievement Points (CAPs), which are rewarded to those who deliver visible and tangible results of different kinds. This article explains how CAPs work and summarizes the results of the experiment.

Here is a very simple management framework, which we have used in our teams for the last two years. We came to it experimentally, trying to merge some Agile principles, PMBOK ideas, and common sense. Our experience so far is positive, even though the proposed rules of work are not really about project management, but more about keeping an eye on the situation and making sure it’s not falling apart. This is the best most modern teams can afford anyway.