As I mentioned in passing a day or two ago, I've figured out how to
support aggregates whose input is flagged DISTINCT without too much
pain. Basically it can all be done inside nodeAgg.c, once we teach
the parser to put the DISTINCT flag bit into Aggref querytree nodes.
(a) If DISTINCT is not specified for a particular aggregate, then
nodeAgg.c runs the aggregate's transition function(s) as each input
tuple is presented, same as now.
(b) If DISTINCT is specified, then nodeAgg.c evaluates the aggregate's
input expression at each input tuple, and passes the resulting datum
into a sort operation that it's started. (Now that tuplesort.c has
a fairly clean object-based interface, it will be easy to start up
a separate sort operation for each DISTINCT aggregate.)
(c) At the end of the input table (or row group), nodeAgg.c does this
for each DISTINCT aggregate: * finish the pending sort operation; * scan the sort output, drop adjacent duplicate
values(the code for this can be borrowed from nodeUnique), and run the aggregate's transition function(s) for each
remainingvalue.
Finally, the aggregate result values can be computed for all the
aggregates (both DISTINCT and regular), and then the output tuple
can be formed.
This is looking like a day's work at most, and considering how often
it gets asked for, I think it's well worth doing.
A limitation of this approach is that an explicit sort of the aggregate
input values will always be done, even when the input is or could be
delivered in the right order anyway. It is certainly *necessary* that
nodeAgg.c be able to do internal sorts on-the-fly, in order to cope with
multiple DISTINCT aggregates, egSELECT COUNT(DISTINCT foo), AVG(DISTINCT bar) FROM table;
since there is no way to scan the table in an order that's sorted for
both simultaneously. But in simpler cases it might be a win if the
optimizer generated a plan that delivered the data in the right order
and nodeAgg.c could be told to skip the internal sort for a DISTINCT
aggregate. I'm not going to worry about that now, but it's a possible
future improvement.
regards, tom lane
