Re: Planner statistics, correlations

Tobias Brox wrote:
> [Heikki Linnakangas - Fri at 10:41:34AM +0000]
>> I thought about partitioning the table by state, putting rows with
>> state=4 into one partition, and all others to another partition.
>
> That sounds like a good idea - but wouldn't that be costly when changing state?

In PostgreSQL, UPDATE internally inserts a new row and marks the old one
as deleted, so there shouldn't be much of a performance difference.

I'm not very familiar with our partitioning support, so I'm not sure if
there's any problems with an update moving a row from one partition to
another. I think you'll have to create an INSTEAD OF UPDATE rule to do a
DELETE on one partition and an INSERT on the other partition. Depending
on your application, that might be a problem; UPDATE is different from
DELETE+INSERT from transaction isolation point of view.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com