I noticed that the system is really pretty shaky about how it chooses
the datatype-specific operators to implement sorting and grouping.
In the GROUP BY case, for example, the parser looks up an operator
named '<' for the column datatype, and then sometime later the executor
looks up an operator named '=' for that datatype, and we blithely assume
that these operators play together and have the expected semantics.
This seems dangerous in a world of user-definable operators. (I think
it's already broken by the standard datatype "tinterval", in fact,
because tinterval's "=" operator doesn't have the semantics of full
What I'm thinking of doing instead is always looking up the "=" operator
by name, and accepting this as actually being equality if it is marked
mergejoinable or hashjoinable or has eqsel() as its restriction
selectivity estimator (oprrest). If we are looking for a "<" operator
to implement sorting/grouping, then we require "=" to be mergejoinable,
and we use its lsortop operator (regardless of name).
The only standard datatypes for which this would change the behavior
are tinterval, path, lseg, and line --- none of which could be sorted/grouped
correctly with the available operators, anyhow. User-defined datatypes
would stop working as sort/group columns unless the author were careful
to mark the equality operator as mergejoinable, but that's a simple
addition to the operator definition.
regards, tom lane
pgsql-hackers by date
|Next:||From: Tom Lane||Date: 2002-11-29 17:59:38|
|Subject: Aren't lseg_eq and lseg_ne broken?|
|Previous:||From: Joe Conway||Date: 2002-11-29 17:14:07|
|Subject: Re: One SQL to access two databases.|