For a toy project, I'm currently using a 64-bit integer of which the first bit is a sign bit (I'm using Postgres' bigint), the next 32-bits make a UNIX epoch timestamp in seconds, and the rest of the 31 bits make a random number. In this way, the ID is very easily sortable by time, although it's a bit of a gimmick.

The scheme allows for 2 billion possible records per second, although if you consider birthday collisions, it comes down to a lot less (but I could re-insert on duplicates). I could later also use some of the random bits to inscribe extra data into the ID.

The bits from the two halves are then interspersed to make sure the first half looks random, and transformed to a 11-character id, for example /^[0-9A-Za-z]{11}$/. I can reject some IDs beforehand if they match a list of bad strings. The result is receiving an id such as "Vxe61yU5Cci" or "WgOUBEO88gb".