Why “variables” in CSS are harmful
– an essay

Adding any form of macros or additional scopes and indirections, including symbolic constants, is not just redundant, but changes CSS in ways that make it unsuitable for its intended audience. Given that there is currently no alternative to CSS, these things must not be added.


There are certain kinds of extensions to CSS that people have regularly asked for over its (now 14 years) history but that nevertheless don't belong in CSS. They are extensions to bring CSS closer to a programming language (as opposed to extensions to support more of typography). They would at first sight be useful for advanced users, but in reality they are not, and are in fact harmful for the “semantic Web,” i.e., for the idea that information should be re-usable.

The proposed extensions include such things as conditional expressions (if-then-else), macros, expressions over properties, delegation (i.e., define an element's style by reference to another element), mark-up injection, mark-up transformation, and symbolic constants (a limited form of macros). This essay only deals with the last of these, constants, but the arguments apply in large part to the other proposed extensions as well.

Why this essay now?

Constants have been regularly proposed and rejected over the long history of CSS and neither the content of the Web nor the software that processes it has changed (or is expected to change) significantly, so there is no reason why constants should be useful now when they weren't before. However, it has become necessary to repeat the arguments, because of three recent developments: WebKit (the core of Apple's Safari browser) briefly experimented with an implementation of constants, a proposal for constants has been published by two people who are members of the CSS working group of W3C, and there is an implementation of constants done the “right way,” i.e., modular and independent of CSS.

The existence of the WebKit implementation proves that it is possible to add macros to CSS. (Something nobody doubted.) However, some people believe that the existence of an implementation also proves that macros are useful.

The other implementation is written in PHP. It proves that it is not necessary to add constants to CSS. Just like the existence of the WebKit implementation cannot be taken as proof that constants in CSS are useful, so the PHP implementation cannot prove that either. But the PHP implementation has the benefit of letting authors determine the usefulness for themselves, without modifying CSS on the Web.

The proposal that was put forward by the two working group members is called “CSS variables” (where “variable” is used in the sense of something whose value you may not know at the time that you read it, rather than as something that can change its value over time, like a variable in a computer program). That's why the title of this essay uses “variables” although it talks about symbolic constants.

Implementation effort

We know that it is possible to implement symbolic constants for CSS (the WebKit experiment shows that) and also that it is not necessary (the PHP implementation shows that authors can use constants without them being defined by CSS). But how big is the implementation effort?

If we look at the latest proposal, it only defines global constants. There is no scoping. That means that an implementation needs a symbol table, but no stack. A stack would require a little bit more memory, but mostly it would make implementations more complex. (Although every programmer has, one hopes, learnt to program a symbol table with lexical scope during his training.) Constants in CSS are thus easier than, e.g., XML Namespaces, which are lexically scoped.

It is different for those CSS implementations that provide a CSS Object Model (an API for manipulating a style sheet in memory). Those implementations do need to keep track of scope in some way, because adding or removing a line of the style sheet can make a previously redundant definition become meaningful.

Even if the implementation cost of constants would be very small, it is still better to leave them out of CSS. That is a consideration for every addition to CSS: extending CSS makes implementing more difficult and programs bigger, which leads to fewer implementations and more bugs. That has to be balanced against the usefulness of the extension. In the case of, e.g., hyphenation, the balance is in favor of adding the extension to CSS: adding all breakpoints as soft hyphens with an external program is possible, but not practical. In the case of constants, the balance is in favor of leaving them out of CSS: not only is replacing constants with an external program practical, it is actually desirable, because it makes the style sheet shorter (and easier to read and re-use, as we'll argue below).

Effort of maintaining style sheets over time

When writing computer code, whether C programs or CSS, there are things that you need to keep in memory for some time and others that you can forget as soon as you reach the end of the line. Syntax questions, such as putting the right number of colons and semicolons, are of the latter type and that makes them easy to deal with. Predefined keywords, such as all properties and other keywords in CSS, or standard library functions in C, are also of that type: to a certain function belongs a certain name, which is always the same, independent of the code you are writing.

But variables and functions in C, or names of counters in CSS, require more effort. You need to remember that a certain non-standard functionality exists and what its name is. An experienced programmer is so used to that effort, that he hardly notices… until he has to revise a program that he wrote more than a month ago.

Of course, in C you need those variables and functions. You need to divide up a problem in smaller ones and solve them one by one. Otherwise you would have to hold even more in memory. And many programming environments help the programmer by showing the list of user-defined names in a window. The computer screen thus becomes an extension of the programmer's memory.

That last point is important: the screen is an extension of the author's memory. What is on screen doesn't have to be remembered (or at least uses less effort). That is, e.g., the principle of drop-down menus: you don't have to recall the name of a command, you just have to recognize it when you see it. Added to that is the effect of location: the (relative) location of a function in the 2D space of the screen is an additional key with which to retrieve it.

Programmers will recognize this memory effect also in another way: Dividing up a problem into smaller ones is only one reason for defining functions. Just as important is the fact that a function that fits on one screen is easier to write than one that needs scrolling.

Thus there is a cost to user-defined names: memory (when writing the code) and understanding (when reading). That cost is offset by the cost of the alternative: long functions. How is that for CSS?

On the one hand, typography doesn't lend itself very well to the divide-and-conquer approach. The @import rule of CSS can, e.g., separate the rules for colors from the rules for sizes and margins, but in reality they are hardly independent. On the other hand, CSS style sheets are short. They are not much bigger than one editor window. Very few people (only professional designers, it seems) write style sheets longer than a hundred lines. Compare that to C programs, which are typically tens of thousands of lines.

Style sheets thus don't need to be split up into named functions. Doing that would add a cost (remembering user-defined names) without a benefit (avoiding problems that are longer than one screenful).

The above is an argument against macros in general. It should be noted that the recent proposal for CSS variables isn't proposing ways to make style sheets shorter. It actually makes them longer. It replaces values by user-defined names as a way to make style sheets self-documenting and easier to treat with search-and-replace functions.

But self-documenting by means of an identifier isn't very effective, a comment would be clearer. And the search and replace is hardly an issue in a style sheet in which a value typically occurs less than a dozen times. What remains is the cost of remembering and understanding user-defined names.

Reusing style sheets

One of the best features of HTML and CSS is that they make it easy to learn from and re-use other people's efforts. CSS is fairly easy to learn to read, even if some of its effects can be quite subtle. When there is a page you like, you can look at its style sheet and see how it's done.

Style debugging tools, such as certain extensions for the Firefox and Opera browsers, make that even easier. You can see exactly which part of the style sheet affects a given part of a document.

But reusing other people's style sheets is more difficult if those style sheets contain user-defined names. Class names are an example. Their names may suggest why the author created them (assuming they are in a language you understand), but typically you will have to look at the document to see where they occur and why. Symbolic constants make that problem worse.

Somebody who reads a style sheet with symbolic constants has the same problem as the author reading his own style sheet after not looking at it for some time, only worse. The author may manage to remember the function of some name after a few hints, any other person will have to hypothesize and check the function fully from what is written in the style sheet itself.

For many people, style sheets with constants will thus simply not be usable. It is too difficult to look in two places at once, the place where a value is used and the place where it is defined, if you don't know why the rule is split in this way. Many people are confused by indirection anyway and adding an extra one, in addition to the element and class names, has the same effect as obfuscating the style sheet.

Indeed, it is quite likely that somebody who is trying to learn CSS will give up learning it when he sees that style sheets that occur on the Web don't actually look like the tutorials he started from. Difference in upper- and lowercase or in pretty-printing are hindrances to learning, too, but limited ones: you soon learn to ignore those differences. But symbolic constants are different in each style sheet and have to be interpreted and understood each time anew.

Of course, this is a conjecture. It is difficult to test and predict how many people will not manage to learn CSS because of macros. But from looking at people writing style sheets and other computer code and seeing the kinds of things that make them fail, I'm pretty sure that this effect is quite strong.

Learning CSS

There are many things in CSS that make CSS too difficult to learn for many people. The concept of putting the style in a style sheet with rules linked to types of elements rather than to actual text, is already too much for many Web authors, but the benefit of separating style and structure is such that, all things considered, CSS is still a good idea for the Web.

The value of the semantic Web isn't defined by how well structured the best documents are, but by how well structured the vast majority of documents is. There is no alternative to CSS when it comes to separating style and structure of HTML documents. XSL is only an alternative at the high end, for advanced users. Thus we need to take care that CSS remains usable for that majority of documents. At least until an alternative emerges.

People who understand CSS in principle may still have trouble understanding the indirection provided by the em unit, many have trouble with the advanced selectors in level 3, and many more won't understand, e.g., any of the properties that have to do with bi-directionality or vertical text. For each feature that is added to CSS there must be a careful balance (based on an informed guess, because these things are difficult to test) between the number of users that will be excluded by that feature and the number for whom it is essential that it be added. Vertical text properties are on the whole beneficial: there are people who need them and the people who don't understand them are unlikely to encounter a style sheet that contains them. Advanced selectors are more controversial. They enable more highly structured documents and enlarge the applicability of CSS beyond HTML, but they are quite likely to be found in style sheets on-line by people who don't understand them.

The balance for symbolic constants is against including them. They are too difficult to learn for many people and at the same time they are not essential for anybody. People who feel that they need them can easily use them without causing them to appear in style sheets that other people may encounter. Solutions like XSL or the PHP implementation mentioned earlier already exist, along with more generic solutions, such as SSI, JSP, or preprocessors such as CPP, m4 or sed.


The Web is modular and that has several benefits: almost every part is optional, can be replaced or updated without affecting the others, and can be reused for tasks that were invented after the module itself was invented. For example, the links in HTML aren't specific to HTML, but they are URLs and URLs have since found their way into many other Web technologies. Similarly, style is not encoded in HTML, but in CSS, and thus HTML and CSS can be developed on their own and each can be used without the other. Images (PNG, JPEG) are another example. HTML uses images, but doesn't define them. And therefore CSS can use the same images as HTML.

CSS itself has given rise to two modules: selectors and media queries were developed for use in CSS but very quickly it was realized that they can be used on their own. Media queries, for example, are also used in HTML and XML.

Symbolic constants is a concept that looks like a candidate for a module. SGML entities, CURIEs and XML Namespace prefixes are examples on the Web of the same concept. And off the Web there are many more. SSI, tag libraries in JSP and various template processors for HTML are examples that are used in close relationship with the Web. It's clearly a concept that doesn't apply only to CSS.

Treating symbolic constants as an independent module would make them available for use in other contexts than CSS, would make them available to precisely the people who need them without hindering other people, would allow them to be developed without impact on CSS, and allow them to be developed more easily into what they are sure to develop into anyway: full macros, able to replace anything, not just CSS values, and able to make files shorter instead of longer.

If developed, it should be a module that is used off the Web, though. Macros are a help for authors, not for the semantic Web. The Web's main formats, HTML, CSS, SVG, PNG, JPEG, etc., already have all the structure they need. Adding an extra, independent structure would just make them more difficult to process.

Bert Bos, W3C/ERCIM, <bert@w3.org>
created 30 June 2008, last modified $Date: 2024/03/22 17:28:51 $