I've dealt with variable-length headers on (binary) things before, and it creates annoyances. It screws up the memory alignment, for instance.

I have two minor, but useful, uses:

One is that the prefix is probably unique for small sets. So it's both easy to read and parse and then use just a prefix for doing something like stopping a docker container.

Second is that a short prefix on a larger set of data is probably not unique but at a rate I can predict. I can take a random, and repeatable, sample of data by selecting everything with the same prefix. This gives me a fast way of taking a random sample from a database (or file or anything else).

Good question. There are instances where it allows you to write cleaner (and faster) code if you can identify the parts without necessarily having to parse the metadata.