I just found out that this is a feature introduced in PostgreSQL 10,
back in 2017.
Besides this being the standard way to define an “auto incremental
column” introduced in SQL:2003[0], called “identity columns”, in
PostgreSQL the new syntax has the following pros, according to [1]:
* No need to explicitly grant usage on the generated sequence.
* Can restart the sequence with only the name of the table and column;
no need to know the sequence’s name.
* An identity column has no default, and the sequence is better
“linked” to the table, therefore you can not drop the default value
but leave the sequence around, and, conversely, can not drop the
sequence if the column is still defined.
Due to this, PostgreSQL’s authors recommendation is to use identity
columns instead of serial, unless there is the need for compatibility
with PostgreSQL older than 10[2], which is not our case.
According to PostgreSQL’s documentation[3], the identity column can be
‘GENERATED BY DEFAULT’ or ‘GENERATED ALWAYS’. In the latter case, it is
not possible to give a user-specified value when inserting unless
specifying ‘OVERRIDING SYSTEM VALUE’. I think this would make harder to
write pgTAP tests, and the old behaviour of serial, which is equivalent
to ‘GENERATED BY DEFAULT’, did not bring me any trouble so far.
[0]: https://sigmodrecord.org/publications/sigmodRecord/0403/E.JimAndrew-standard.pdf
[1]: https://www.2ndquadrant.com/en/blog/postgresql-10-identity-columns/
[2]: https://wiki.postgresql.org/wiki/Don't_Do_This#Don.27t_use_serial
[3]: https://www.postgresql.org/docs/15/sql-createtable.html
jordi fita mas