On Sat, Dec 14, 2013 at 11:47 PM, Andreas Karlsson <andreas@proxel.se> wrote:
On 12/14/2013 10:59 AM, Alexander Korotkov wrote:
This patch allows to use index for order-by if order-by clause and index has non-empty common prefix. So, index gives right ordering for first n order-by columns. In order to provide right order for rest m columns, sort node is inserted. This sort node sorts groups of tuples where values of first n order-by columns are equal.
I recently looked at the same problem. I see that you solved the rescanning problem by simply forcing the sort to be redone on ExecReScanSort if you have done a partial sort.
Naturally, I'm sure I solved it at all :) I just get version of patch working for very limited use-cases.
My idea for a solution was to modify tuplesort to allow storing the already sorted keys in either memtuples or the sort result file, but setting a field so it does not sort thee already sorted tuples again. This would allow the rescan to work as it used to, but I am unsure how clean or ugly this code would be. Was this something you considered?
I'm not sure. I believe that best answer depends on particular parameter: how much memory we've for sort, how expensive is underlying node and how it performs rescan, how big are groups in partial sort.