The medium obscures the message: why programming languages matter

Recently, I’ve been getting into confrontations in my workplace regarding our use of PHP. Even some who agree with me that PHP is a poor language still say that choice of language is mostly irrelevant to the bottom line.

I disagree. I have reason to believe that PHP, or any programming language that is less readable, predictable, and consistent than viable alternatives, is harmful in use and observably so. While I do not have figures that indicate this1, I will argue my point using basic analogies, commonly-accepted hacker wisdom, and examples that should appeal intuitively to technical readers.

We’re engineers

Often, after complaining of a surprising or erroneous behavior of PHP2 in our development chat, I am engaged by someone who takes issue with my criticism of PHP and our use of it.

Typically, the crux of their response is this: my criticisms of PHP are academic, aesthetic judgements that belong on the sidelines; at the end of the day, we are engineers that “solve problems” and so the choice of language for problem solution is outside of our concern so long as the language gets the job done.

I sympathize partially with this response. We are engineers; our daily activities are concerned with solving problems. However, much as the material that a civil engineer chooses for the construction of a bridge largely determines the stability of that bridge, so too does the choice of a medium, a programming language, determine the robustness of a solution that we code.

A serious engineer surely would not use clay to build a suspension bridge (though it may be possible), just as a professional web developer would not advise using TeX to build a high-traffic web application. These are extreme examples, but they both allude to a relative ordering in the quality of a solution in terms of its medium, ceteris paribus.

What is quality?

In order to have an intelligent discussion about efficacy, we must first define it. What does it mean for a piece of software to be quality? Specifically, what does it mean for a body of code to be quality?

Quality is highly contextual. A bridge of good quality could be described as a physical construction that provides a stable, human-navigable connection between two pieces of land while requiring as little maintenance as possible. Quality is comprised of a large number of independent metrics in most cases, but by Pareto’s principle, it’s likely that just a few of those dimensions will give us a good indication of total quality. The description that I have provided of a good bridge adequately, but perhaps not completely, represents the criteria used to determine whether a bridge is good or not.

I will characterize code of a good quality as being a mechanism which allows a user to supply input, efficiently effecting a desired result which varies on that input, and preferably in a way that requires the user to learn as little additional context as possible. Implicit in this definition of quality, and perhaps key to it, is that the software will perform as free of error as possible.

Stated simply: good code is, among other things, nearly bug-free. For simplicity, I will not extend this definition of quality to include extensibility, which, while important, is beside my point. I contend that choice of language is especially relevant to the number of bugs; that issue will be my primary concern.

Comparing mediums

So, I have demonstrated that the choice of a tool or material, i.e. the medium, plays a definite role in determining the quality of a solution. I have also given my definition of what quality is. Now, the relevant question becomes: does a use of PHP affect the quality in a significantly negative way, ceteris paribus, relative to the use of other languages?

I will argue that while the difference between PHP and Groovy or Python isn’t as drastic as the difference between clay and steel for use in the making of a bridge, it is significant. I base my reasoning on three metrics: predictability, readability, and compactness.

Predictability

I will define predictability as how easy it is for a human to reason about what a piece of code will do before running that code, or statically. Predictability is important when concerned with writing bug-free code because it allows the programmer to not only be clear about his intentions to other programmers, but it allows him to produce code that carries out his intentions with as little effort as possible.

The more clearly the intentions of the programmer are represented in code, the easier it will be for the author and other programmers to verify correctness.

What do you expect the following code example to result in?

<?php
php> $a = null;
php> $a->hotbaz = 1;

Conceptually, null (in most programming languages, and in PHP) is the single, canonical representation of nothing. It is the absence of value. From the PHP manual:

The special NULL value represents a variable with no value. NULL is the only possible value of type NULL.

So, what would you expect of the above code? Considering that null is nothing, most sensible programmers would expect an error to be thrown upon dereferencing nothing. Many sensible API designers would happily return null for functions in which the attempted retrieval of, say, an object fails.

Unfortunately, that isn’t what happens. Instead, PHP accepts the above as a valid operation. Here’s how $a looks afterwards:

<?php
php> = $a
<object #2 of type stdClass> {
  hotbaz => 1,
}

A stdClass object is automatically instantiated, from nothing, with no warning. A bug results when a user employing the fictional API mentioned above continues to use this object, expecting that an Exception should have been thrown and caught.

In other languages, namely Python and Groovy, this class of bugs is not possible:

>>> a = None 
>>> a.hotbaz = 1
Traceback (most recent call last): File "<stdin>", line 1, in <module>
AttributeError: 'NoneType' object has no attribute 'hotbaz' 
groovy:000> a = null
===> null
groovy:000> a.hotbaz = 1
ERROR java.lang.NullPointerException:
Cannot set property 'hotbaz' on null object
        at groovysh_evaluate.run (groovysh_evaluate:2)
        ...

Those two languages maintain a reverence for the definition of None and null, and thus behave in a way that a programmer capable of deductive reasoning would consider predictable.

The treatment of such a fundamental language construct should be indicative of other properties of a language. If a language behaves unexpectedly with fundamental properties such as the null value, how can we rely on it to do more complex operations in a predictable way?

Let’s consider another essential language property: function arity.

<?php
php> function baz($a) {
 ...   return $a + 1;
 ... }

php> = baz(1, 2, 3);
2

The PHP interpreter doesn’t complain when a function is passed more arguments than what it was defined to handle; this makes the execution of a program more difficult to reason about and therefore makes unexpected behavior harder to track down. What if an argument is left out of a function definition after refactoring? PHP will not inform the programmer of this, giving him less information to combat errors.

This sort of lenience is not permitted in other languages, where predictability as a tool for quality is recognized3:

>>> def foo(a):
...   return a + 1
... 
>>> foo(1, 2, 3)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: foo() takes exactly 1 argument (3 given)
groovy:000> def baz(a) {
groovy:001>   a + 1
groovy:002> }
===> true
groovy:000> baz(1, 2, 3)
ERROR groovy.lang.MissingMethodException:

No signature of method: groovysh_evaluate.baz() is applicable for argument
types: (java.lang.Integer, java.lang.Integer, java.lang.Integer) values: [1, 2, 3]
Possible solutions: baz(java.lang.Object), wait(), run(), run(), any(), grep()
        at groovysh_evaluate.run (groovysh_evaluate:2)
        ...

Readability

In a 1946 essay entitled Politics and the English Language, George Orwell argues that stark, simple English must be used (and demanded) by readers, else lies and atrocities could be masked by overly complex presentation.

The same applies to programs. The logical meat-and-potatoes of a program should be as plain as possible to readily reveal potential errors. Syntax should be simple and noise-free so that larger matters of correctness can be more easily considered.

When each variable in your language is prefixed by $, the signal-to-noise ratio is diluted. When object dereferences take two characters instead of one, the eye has more boilerplate to process. When the most commonly used data structure in your language is wrapped in array() instead of [], an individual critiquing for correctness (or simply attempting to understand) has more work to do than is necessary. Finding hidden or subtle faults will take more time and diligence.

Readability, or the level of ease and speed with which a competent practitioner can fully understand a piece of code, has an obvious effect on how amenable a program is to housing errors.

An example

Consider the following example of PHP4:

<?php
$x = 1;
$nums = array(10, 20, 30, 40);
$res = 0;

foreach ($nums as $n)
  if ($n > 15)
    $res -= $n*2 + $x;

Consider the same functionality in Groovy:

def x = 1
def nums = [10, 20, 30, 40]
def res = nums.findAll { it > 15 } 
    .collect { it * 2 + x } 
    .inject(0) {accum, val -> accum - val}

Consider the same functionality in Python, phrased even more succinctly by mrjbq7:

>>> def foo(x, nums):
...     return -sum(n*2+x for n in nums if n > 15)
... 

>>> foo(1, [10,20,30,40])
-183

The intent in the latter two examples is much clearer. We can attempt to emulate this succinct style in PHP, though it becomes an unreadable mess:

<?php
$x = 1;
$nums = array(10, 20, 30, 40);

$res = array_reduce(
        array_map(
            function($v) use ($x) { return $v * 2 + $x; },
            array_filter(
                $nums,
                function($v) { return $v > 15; })
            ),
        function ($v, $w) { return -$v - $w; }
);

When a language requires more effort to read, either because it pollutes code with a lot of noise
($, ->, array(), et al), or lacks succinct notation for common operations like collection –iteration,
–creation, and –manipulation, it is more difficult to judge the correctness of the algorithms it expresses.

Because of this, a more readable language will exhibit fewer bugs. By our definitions, this leads to an increase in the quality of code and a decrease in the amount of time that it takes to ship software with the same functionality.

Compactness

In The Art of Unix Programming, Eric S. Raymond defines compactness as a property indicating how well “a design can fit inside a human being’s head.” Compactness is a primary determinant in how predictable and readable a programming language is. As such, it is relevant when considering how effective a language is at producing code with less errors.

Stated differently, a language is compact if it is intuitive to use after an initial understanding of the language’s paradigms.

Regularity in paradigm

A confusion of design makes a language anti-compact. When object-orientation is half-baked into a standard library that is mostly procedural, unintuitive behavior and code that is difficult to read follows. Systems that are built on languages like this exhibit the same problem: see the confusion that results from the presence of stdClasss, structured arrays, and full-blown objects in Drupal, a CMS framework written in PHP.

Nodes, one of the most common datatypes in Drupal, are represented with stdClasss when, by nature of having behaviors and required attributes, they should be distinct classes.

<?php
php> $node = new stdClass();
php> $node->title = 'foobar';
php> $node->type = 'article';
php> node_save($node);

php> $node2 = new stdClass();
php> $node->title = 'heybaz';
php> node_save($node2);

Uncaught exception: EntityMalformedException: Missing bundle property on entity
of type node.

Persisting the node to a datastore can be attempted without the node having required attributes.

The association of functions with a stdClass instead of methods with a proper object means that there is no single point of truth for how Drupal expects a node to behave, thereby increasing complexity, confusion, and the likelihood of misuse. Instead of going to one source of documentation for how to work with a node, we’re subject to many sources. This is more time-consuming than it needs to be.

Consistency in the data model of PHP would encourage consistency in Drupal. This is how the above should look:

<?php
php> $node = new Node('article');
php> $node->title = 'foobar';
php> $node->save();

php> // better yet
php> $node = new ArticleNode($title);
...

Regularity in notation

In Python and Groovy, collections behave consistently.

Python 2.7.2 (default, Dec  5 2011, 20:11:10)
...
>>> for i in "abc":
...   print i,
... 
a b c
>>> for i in ['a', 'b', 'c']:
...   print i,
... 
a b c
>>> for i in {'a':1, 'b':1, 'c':1}:
...   print i,
... 
a c b

>>> [i for i in "abc"]
['a', 'b', 'c']

>>> [i for i in [1, 2, 3]]
[1, 2, 3]

>>> {"abc"[x]:x for x in range(3)}
{'a': 0, 'c': 2, 'b': 1}

Specifically, collections are objects that all implement a specific interface which allows them to be used by the same control structures. They are polymorphic to a common notation. This affords the Python or Groovy programmer regularity; he has to think less to accomplish what he wants. A reader doesn’t have to work as hard because the notation is consistent. This medium makes the message clearer.

PHP does not exhibit this kind of compactness.

<?php
php> foreach (array(1, 2, 3) as $x) {
 ...   echo $x;
 ... }
123

php> foreach ("abc" as $i) {
 ...   echo $i;
 ... }

Warning: Invalid argument supplied for foreach() in
/usr/local/lib/python2.7/site-packages/phpsh/phpsh.php(593) 
...

php> foreach(str_split('abc') as $x) {
 ...   echo $x;
 ... }
abc

This divides how a user operates with separate data structures that are, in many ways, analogous. If we were reading Chinese, this is like having to parse a technical manual written in both Mandarin and Cantonese. An editor having to check that text’s grammar and content now has twice as much work.

PHP also doesn’t respect referential transparency5, a large factor in compactness and predictability.

<?php
php> "abc"[0]
Multiline input has no syntactic completion:

Parse error: syntax error, unexpected '[' in Command line code on line 1

php> $a = "abc"
php> = $a[0]
"a"

php> $arr = array(1, 2, 3);
php> function get_arr() {
 ...   global $arr;
 ...   return $arr;
 ... }

php> = get_arr()[0]
Multiline input has no syntactic completion:

Parse error: syntax error, unexpected '[' in Command line code on line 1

php> = (get_arr())[0]
Multiline input has no syntactic completion:

Parse error: syntax error, unexpected '[' in Command line code on line 1

php> = $arr[0]
1

This kind of behavior stifles regularity, and therefore compactness. It places a needless burden on the programmer and his co-workers.

In short, anti-compactness wastes time.

The medium constrains the message

Mathematics is, by many accounts, the purest distillation of reason. It also features a language that is predictable, readable, and compact. The language of mathematics is so rich in these three metrics that the overwhelming majority of time is spent not on the notation itself, but understanding the ideas represented by the notation.

Looking back at early mathematical notation and its evolution, the language clearly wasn’t always concise as it is today.

Ancient Chinese notation for an arithmetic expression, from the article above.

Had we continued using the same cumbersome notation seen above for ideas as basic as addition and multiplication, how could we expect to develop and express complex ideas like differential calculus? How could we expect mathematicians to critique one another’s work in a timely way?

It’s no coincidence that mathematical notation is minimal and concise. It is a medium that makes its content as obvious as possible, which has enabled criticism and development to flourish.

The same evolution is true of programming languages. Using programming languages that make concise abstraction difficult results in code that is more prone to error because it is more difficult to develop and read. This is wasted time. This is wasted money6.

For this reason, the choice of language is relevant to the daily business and long-term success of a software company. Subsequently, it is the duty of a software engineer to make judgements about his tools and advocate those which allow him to operate most effectively.


  1. Studies correlating bug count with language use are apparently not of interest to academia.

  2. To name a few: injection of inner functions into the global symbol table (violation of lexical scoping), PHP’s unwillingness to allow array dereferences on functions that return arrays (foo()[0], or even (foo())[0]), PHP’s throwing of a fatal error on the second encounter of an inner function definition, PHP’s automatic stdClass construction from a null variable ($a = null; $a->foo = 'abc';), PHP’s inability to declare class-level attributes that are not strings or integers.

  3. Worth noting is that both Python and Groovy offer constructs for variadic arguments.

  4. This example was taken and modified from this article: http://justafewlines.com/2009/10/whats-wrong-with-php-closures/ and was kindly cleaned up by JS235 from reddit.

  5. Referential transparency on Wikipedia: http://en.wikipedia.org/wiki/Referential_transparency

  6. See Paul Graham’s experiences with Viaweb: http://www.paulgraham.com/avg.html