On Tue, Sep 1, 2015 at 6:13 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:
> My thought is that there's a fair amount of places where we do string
> comparison for not a great reason. Perhaps a better example is data that
> comes back from a trigger; AFTER/BEFORE, INSERT/UPDATE/..., which is more
> expensive to setup the variables for (strdup a fixed string, which means a
> palloc), and then comparisons are done as text varlena (iirc).
>
> Instead if this information came back as an ENUM the variable would be a
> simple int as would the comparison. We'd still have a raw string being
> parsed in the function body, but that would happen once during initial
> compilation and it would be replaced with an ENUM value.
>
> For RAISE, AFAIK we still end up converting the raw string into a varlena
> CONST, which means a palloc. If it was an ENUM it'd be converted to an int.
>
> If we're worried about the overhead of the enum machinery we could create a
> relcache for system enums, but I suspect that even without that it'd be a
> win over the string stuff. Especially since I bet most people run UTF8.
I agree with Pavel on this one: creating an extra type here is going
to cause more pain than it removes.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company