I’ve known about the Emacs autoload feature for a long time, but I never bothered to read about it. I decided to fix that and finally read the relevant part of the Emacs Lisp manual.
It turns out that autoloads are a pretty simple concept. They basically serve as a way of deferring loading a package until it is really needed. If you have a lot of packages, loading them on Emacs startup may take considerable time and memory. (Of course, nowadays we usually don’t care too much for the latter, since computers now have insane amounts of RAM. My first computer had 64 kilobytes of RAM, and my first PC had 4 megabytes, and GNU Emacs was written in times when 4 megabytes was a lot!) Autoloads are a way to tell Emacs when the package should be loaded by defining some functions as “entry points” to the package. (Actually, it is a bit more sophisticated than that, but that’s the main idea.)
If you write a package, you should mark the entry point(s) with the “autoload comment”. That means putting the string ;;;###autoload on the line before the relevant defun (or perhaps other form defining something, like defvar or cl-defun). When the package is installed, Emacs scans it for these comments and generates a special file containing information about these functions, variables etc. During startup, Emacs loads that special file (which is much, much smaller than the main package file or files, and hence faster to load). This means that all autoloaded functions become actually callable – but calling any of them causes Emacs to load the package in question first. This means that autoloaded functions are loaded “on demand” – when they are first called.
That is (of course!) not the whole story. For example, you can make Emacs put basically arbitrary code into the autoloads file by putting it after the autoload comment but on the same line. This way, that piece of code is not evaluated when loading the package (since it is commented out), but on Emacs startup (because it gets copied to the autoloads file). This may be useful if the package uses non-standard ways to define things.
I intend to put this knowledge into practice soon – I’d like to take some of my Elisp code and submit it as a package to Melpa, both to make it more discoverable and hence useful to other people and also to learn the process. Of course, when I do it, I will write about it here!