On 2014-05-11 12:47:21 -0400, Tom Lane wrote:
> Another idea is that the main reason we do things like this is the
> assumption that for UPDATE, ModifyTable receives complete new rows
> that only need to be pushed back into the table (and hence have
> to already match the rowtype of the specific child table). What if
> we got rid of that and had the incoming tuples just have the target
> row identifier (tableoid+TID) and the values for the updated columns?
> ModifyTable then would have to visit the old row (something it must
> do anyway, NB), pull out the values for the not-to-be-updated columns,
> form the final tuple and store it. It could implement this separately
> for each child table, with a different mapping of which columns receive
> the updates. This eliminates the whole multiple-plan-tree business
> at a stroke ... and TBH, it's not immediately obvious that this would
> not be as efficient or more so than the way we do UPDATEs today, even
> in the single-target-table case. Pumping all those not-updated column
> values through the plan tree isn't free. The more I think about it,
> the more promising this sounds --- though I confess to being badly
> undercaffeinated at the moment, so maybe there's some fatal problem
> I'm missing.
Yes, that sounds like a rather good plan. There's probably some fun
keeping the executor state straight when switching more frequently than
now and we'd probably need some (implicitly?) added type coercions? I
also agree, while there probably are some cases where'd be slower, that the
majority of cases will be faster.
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services