Font subsetting

A typeface file carries every character its designer drew: Latin, the accented extensions, Greek, Cyrillic, arrows, currency, a few hundred pieces of punctuation nobody has typed since 1994. An English-language interface uses maybe a hundred of them. Subsetting is the step that throws away the rest before the file ever reaches a browser, and it is usually the largest single saving available on a page that uses a web font, because it cuts the bytes rather than merely rescheduling them.

There are two shapes to it. Static subsetting builds one smaller file: you point a tool at the font and a character set, and it writes out a face containing only those glyphs, plus the layout tables that still apply to them. Dynamic subsetting splits the face into several files, one per script or block, and declares each with a unicode-range descriptor inside its own @font-face rule. The browser then reads the page, works out which ranges are actually needed, and fetches only those files. This is what the big hosted font services emit, and it is why a page of English text served a family with Cyrillic support never downloads the Cyrillic.

The failure mode is cutting too deep. A subset that covers your copy today does not cover the em dash an editor pastes in tomorrow, the curly apostrophe an author’s word processor inserted, or a customer’s name with a diacritic in it. When a character is missing from every declared range the browser goes looking elsewhere, and the reader gets either a fallback font mid-word, which is visible as a change of shape and colour, or tofu, which is a box where a letter should be. Both are worse than the kilobytes you saved. The usual insurance is to keep the Latin-1 range whole rather than trimming to observed characters, and to subset from real content rather than from a guess.

Two smaller notes worth carrying. Subsetting can break things that live in the font’s tables rather than in its outlines: kerning pairs, ligatures, and the OpenType features a stylesheet later asks for, so a tool that keeps the layout tables for the glyphs it retains is the one to use. And on a variable font the saving works the other way around too: dropping unneeded axes, or pinning one to a single instance, is a subset of the design space rather than of the character set, and it can be worth more than the characters.