> Until we fix that (maybe for 7.2, maybe not) your existing hack is
> probably pretty reasonable. You could save some cycles by avoiding
> conversion to text, though --- instead return an opaque datum that is
> pointer-to-tuple-slot and let the dblink_tok function extract fields
> from the tuple. Look at SQL function support and the FieldSelect
> expression node type for inspiration.
>
I changed the dblink() function to return a pointer instead of concatenated
text, and dblink_tok() to use the pointer. FWIW, a query on a small (85
tuples) remote (a second PC on a 100baseT subnet) table takes about 34
milliseconds (based on show_query_stats) versus about 4 milliseconds when
run locally. It actually takes a bit longer (~65 milliseconds) when run
against a second database on the same PC. The original text parsing version
was about 25% slower.
Although shifting from text parsing to pointer passing is more efficient, I
have one more question regarding this -- for now ;) -- is there any way to
check the pointer passed to dblink_tok() to be sure it came from dblink()?
Thanks,
-- Joe
