On Sat, Jan 19, 2008 at 10:33:00PM -0500, Tom Lane wrote:
> The difference comes from the fact that the HashAggregate step --- which
> is being done on hashing columns (data, usuario) --- is effectively
> reading out in hash-value order for usuario, meaning that that is the
> order in which we make index insertions. 8.2 had an extremely chintzy
> hash function for integers --- basically just return the negative of the
> integer value --- while 8.3 takes it seriously and produces a nicely
> randomized hash value. This means that the usuario values are returned
> in a relatively well ordered fashion in 8.2 and a nearly totally random
> one in 8.3.
A random thought: Is there not some way of representing this difference
in the planner? We store the correlation for tables, could it not be
argued that the output of a hash plan is correlation zero, whereas an
indexscan may have higher correlation... A low correlation works
against plans returning lots of rows.
Have a nice day,
--
Martijn van Oosterhout <kleptog@svana.org> http://svana.org/kleptog/
> Those who make peaceful revolution impossible will make violent revolution inevitable.
> -- John F Kennedy