[dspillett]

There is a fourth reason: momentum.

We are just (finally) retiring a product with a large less-organised-than-we-care-to-about codebase that has been around in various states longer than our younger staff have been alive, that stored UUIDs wrong due to early misunderstanding by a long-gone expert. Several times during its life I made the case for converting it's use of UUIDS in strings to proper UUID types, even demonstrating significant performance improvements under load in a PoC, and storage savings. Everyone agreed it should be done that way, but there was considerably more work involved than just switching types (changing procs or type conversion could kill good query plans, dealing with the fact that at some point in history a smart person had used other strings as special values in one or two places, finding ad-hoc query preparation everywhere and making sure that declared the right type, ..., and of course the dreaded full regression test) so it never got done.

Especially in the last decade and a bit, because it would be retired in a year or two (ahem) so the time & effort would be better put to use elsewhere.