Simon Riggs <simon@2ndQuadrant.com> wrote:
> On 9 June 2013 12:58, Craig Ringer <craig@2ndquadrant.com> wrote:
>> We don't currently have OLD and NEW relations so we're free to
>> define how this works pretty freely.
> I think the best way, if we did do this, would be to have a
> number of different relations defined:
>
> OLD
> NEW
> INSERTED
> DELETED
> all of which would be defined same as main table
>
> and also one called
> UPDATED
> which would have two row vars called OLD and NEW
> so you would access it like e.g. IF UPDATED.OLD.id = 7
Well, there is the SQL standard, which has a couple paragraphs on
the topic which we might want to heed. For a delete there is just
an old table; for an insert just a new one. For an update you have
both, with the same cardinality. The rows in the old and new
tables have a correspondence, but that is only visible to FOR EACH
ROW triggers. For something like RI, why would you need to
establish correspondence? A row with the referenced key either
exists after the statement completes, or it doesn't -- why would we
care whether it is an updated version of the same row?
Syntax for how to refer to the these is defined by the standard.
As usual, I don't object to adding capabilities as long as the
standard syntax is also supported with standard semantics.
--
Kevin Grittner
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company