Re: [HACKERS] How embarrassing: optimization of a one-shot query doesn't work

Am 01.04.2008 um 13:14 schrieb Dave Cramer:
>
> On 1-Apr-08, at 6:25 AM, Michael Paesold wrote:
>
>>
>> Am 01.04.2008 um 01:26 schrieb Tom Lane:
>>> While testing the changes I was making to Pavel's EXECUTE USING
>>> patch
>>> to ensure that parameter values were being provided to the planner,
>>> it became painfully obvious that the planner wasn't actually *doing*
>>> anything with them. For example
>>>
>>> execute 'select count(*) from foo where x like $1' into c using $1;
>>>
>>> wouldn't generate an indexscan when $1 was of the form 'prefix%'.
>> ...
>>> The implication of this is that 8.3 is significantly worse than 8.2
>>> in optimizing unnamed statements in the extended-Query protocol;
>>> a feature that JDBC, at least, relies on.
>>>
>>> The fix is simple: add PlannerInfo to eval_const_expressions's
>>> parameter list, as was done for estimate_expression_value. I am
>>> slightly hesitant to do this in a stable branch, since it would
>>> break
>>> any third-party code that might be calling that function. I doubt
>>> there
>>> is currently any production-grade code doing so, but if anyone out
>>> there
>>> is actively using those planner hooks we put into 8.3, it's
>>> conceivable
>>> this would affect them.
>>>
>>> Still, the performance regression here is bad enough that I think
>>> there
>>> is little choice. Comments/objections?
>>
>> Yeah, please fix this performance regression in the 8.3 branch.
>> This would affect most of the JDBC applications out there, I think.
>>
> Was the driver ever changed to take advantage of the above strategy?

IIRC, it is used in most cases with the v3 protocol, as long as you
don't set a prepare-threshold.

Best Regards
Michael Paesold