Performance on many RDBMs can be heavily dependent on row size. While a “purist” approach might say that the application should be completely independent of the underlying data structure, over many rows improvements like using a smaller integer could shave gigabytes off table size, which makes more of the table fit in memory, which drastically improves performance.

I would use a small integer like this for say, a primary key on a table that would always have <100 rows, especially when this is stored as a foreign key in a table I expect to grow very large. While the size is implementation defined, it’s safe to assume that it is greater than 127 at the very least.