On 08.02.2011 23:50, Branko Čibej wrote: > Well, here it is, I fixed the thinko in the actual_props query and got > all prop_tests to pass with this version. Did I say that the way > ACTUAL_NODE is separate from NODE makes these kinds of WC operations > (that merge results from both tables) quite complex and expensive?
By the way, there's a simpler query that'll return the same result, but it uses an outer join instead of an inner join and takes forever -- it'd be twice as slow as the plain-vanilla trunk is right now. I suspect that if instead we "materialized" that outer join by simply merging NODES and ACTUAL_NODE into one table, denoting the ACTUAL_ bit with a new (higher) op_depth value, then my patch could have dropped a join and a whole query, making the proplist about twice as fast as the patch currently makes it. I'm sure a number of other queries could be improved that way, too. It would mean dropping some NOT NULL constraints, and would make the NODES table a bit larger; but performance-wise, having the whole shebang in a single table and avoiding outer joins (either programmatic or in queries) would open the door for further performance improvements and code simplification. -- Brane
