The original article was posted by Michael Veksler on Quora

A very well known fact is that code is written once, but it is read many times. This means that a good developer, in any language, writes understandable code. Writing understandable code is not always easy, and takes practice. The difficult part, is that you read what you have just written and it makes perfect sense to you, but a year later you curse the idiot who wrote that code, without realizing it was you.

The best way to learn how to write readable code, is to collaborate with others. Other people will spot badly written code, faster than the author. There are plenty of open source projects, which you can start working on and learn from more experienced programmers.

Readability is a tricky thing, and involves several aspects:

1. Never surprise the reader of your code, even if it will be you a year from now. For example, don’t call a function max() when sometimes it returns the minimum().
2. Be consistent, and use the same conventions throughout your code. Not only the same naming conventions, and the same indentation, but also the same semantics. If, for example, most of your functions return a negative value for failure and a positive for success, then avoid writing functions that return false on failure.
3. Write short functions, so that they fit your screen. I hate strict rules, since there are always exceptions, but from my experience you can almost always write functions short enough to fit your screen. Throughout my carrier I had only a few cases when writing short function was either impossible, or resulted in much worse code.
4. Use descriptive names, unless this is one of those standard names, such as i or it in a loop. Don’t make the name too long, on one hand, but don’t make it cryptic on the other.
5. Define function names by what they do, not by what they are used for or how they are implemented. If you name functions by what they do, then code will be much more readable, and much more reusable.
6. Avoid global state as much as you can. Global variables, and sometimes attributes in an object, are difficult to reason about. It is difficult to understand why such global state changes, when it does, and requires a lot of debugging.
7. As Donald Knuth wrote in one of his papers: “Early optimization is the root of all evil”. Meaning, write for readability first, optimize later.
8. The opposite of the previous rule: if you have an alternative which has similar readability, but lower complexity, use it. Also, if you have a polynomial alternative to your exponential algorithm (when N > 10), you should use that.

Use standard library whenever it makes your code shorter; don’t implement everything yourself. External libraries are more problematic, and are both good and bad. With external libraries, such as boost, you can save a lot of work. You should really learn boost, with the added benefit that the c++ standard gets more and more form boost. The negative with boost is that it changes over time, and code that works today may break tomorrow. Also, if you try to combine a third-party library, which uses a specific version of boost, it may break with your current version of boost. This does not happen often, but it may.

Don’t blindly use C++ standard library without understanding what it does - learn it. You look at std::vector::push_back() documentation at it tells you that its complexity is O(1), amortized. What does that mean? How does it work? What are benefits and what are the costs? Same with std::map, and with std::unordered_map. Knowing the difference between these two maps, you’d know when to use each one of them.

Never call new or delete directly, use std::make_unique and [cost c++]std::make_shared[/code] instead. Try to implement usique_ptr, shared_ptr, weak_ptr yourself, in order to understand what they actually do. People do dumb things with these types, since they don’t understand what these pointers are.

Every time you look at a new class or function, in boost or in std, ask yourself “why is it done this way and not another?”. It will help you understand trade-offs in software development, and will help you use the right tool for your job. Don’t be afraid to peek into the source of boost and the std, and try to understand how it works. It will not be easy, at first, but you will learn a lot.

Know what complexity is, and how to calculate it. Avoid exponential and cubic complexity, unless you know your N is very low, and will always stay low.

Learn data-structures and algorithms, and know them. Many people think that it is simply a wasted time, since all data-structures are implemented in standard libraries, but this is not as simple as that. By understanding data-structures, you’d find it easier to pick the right library. Also, believe it or now, after 25 years since I learned data-structures, I still use this knowledge. Half a year ago I had to implemented a hash table, since I needed fast serialization capability which the available libraries did not provide. Now I am writing some sort of interval-btree, since using std::map, for the same purpose, turned up to be very very slow, and the performance bottleneck of my code.

Notice that you can’t just find interval-btree on Wikipedia, or stack-overflow. The closest thing you can find is Interval tree, but it has some performance drawbacks. So how can you implement an interval-btree, unless you know what a btree is and what an interval-tree is? I strongly suggest, again, that you learn and remember data-structures.

These are the most important things, which will make you a better programmer. The other things will follow.

Studying a functional programming language is a good way to discover new approaches to problems and different ways of thinking. Although functional programming has much in common with logic and imperative programming, it uses unique abstractions and a different toolset for solving problems. Likewise, many current mainstream languages are beginning to pick up and integrate various techniques and features from functional programming.

Many authorities feel that Haskell is a great introductory language for learning functional programming. However, there are various other possibilities, including Scheme, F#, Scala, Clojure, Erlang and others.

Haskell is widely recognized as a beautiful, concise and high-performing programming language. It is statically typed and supports various cool features that augment language expressivity, including currying and pattern matching. In addition to monads, the language support a type-class system based on methods; this enables higher encapsulation and abstraction. Advanced Haskell will require learning about combinators, lambda calculus and category theory. Haskell allows programmers to create extremely elegant solutions.

Scheme is another good learning language -- it has an extensive history in academia and a vast body of instructional documents. Based on the oldest functional language -- Lisp -- Scheme is actually very small and elegant. Studying Scheme will allow the programmer to master iteration and recursion, lambda functions and first-class functions, closures, and bottom-up design.

Supported by Microsoft and growing in popularity, F# is a multi-paradigm, functional-first programming language that derives from ML and incorporates features from numerous languages, including OCaml, Scala, Haskell and Erlang. F# is described as a functional language that also supports object-oriented and imperative techniques. It is a .NET family member. F# allows the programmer to create succinct, type-safe, expressive and efficient solutions. It excels at parallel I/O and parallel CPU programming, data-oriented programming, and algorithmic development.

Scala is a general-purpose programming and scripting language that is both functional and object-oriented. It has strong static types and supports numerous functional language techniques such as pattern matching, lazy evaluation, currying, algebraic types, immutability and tail recursion. Scala -- from "scalable language" -- enables coders to write extremely concise source code. The code is compiled into Java bytecode and executes on the ubiquitous JVM (Java virtual machine).

Like Scala, Clojure also runs on the Java virtual machine. Because it is based on Lisp, it treats code like data and supports macros. Clojure's immutability features and time-progression constructs enable the creation of robust multithreaded programs.

Erlang is a highly concurrent language and runtime. Initially created by Ericsson to enable real-time, fault-tolerant, distributed applications, Erlang code can be altered without halting the system. The language has a functional subset with single assignment, dynamic typing, and eager evaluation. Erlang has powerful explicit support for concurrent processes.

Computer Programming as a Career?

What little habits make you a better software engineer?

Let’s face it, fad or not, companies are starting to ask themselves how they could possibly use machine learning and AI technologies in their organization. Many are being lured by the promise of profits by discovering winning patterns with algorithms that will enable solid predictions… The reality is that most technology and business professionals do not have sufficient understanding of how machine learning works and where it can be applied.  For a lot of firms, the focus still tends to be on small-scale changes instead of focusing on what really matters…tackling their approach to machine learning.

In the recent Wall Street Journal article, Machine Learning at Scale Remains Elusive for Many Firms, Steven Norton captures interesting comments from the industry’s data science experts. In the article, he quotes panelists from the MIT Digital Economy Conference in NYC, on businesses current practices with AI and machine learning. All agree on the fact that, for all the talk of Machine Learning and AI’s potential in the enterprise, many firms aren’t yet equipped to take advantage of it fully.

Panelist,  Michael Chui, partner at McKinsey Global Institute states that “If a company just mechanically says OK, I’ll automate this little activity here and this little activity there, rather than re-thinking the entire process and how it can be enabled by technology, they usually get very little value out of it. “Few companies have deployed these technologies in a core business process or at scale.”

Panelist, Hilary Mason, general manager at Cloudera Inc., had this to say, “With very few exceptions, every company we work with wants to start with a cost-savings application of automation.” “Most organizations are not set up to do this well.”