Heikki Linnakangas <hlinnakangas@vmware.com> writes:
> On 01/13/2015 02:08 PM, Alexey Bashtanov wrote:
>> By varying the parameters and reading source code I determined that
>> memory usage linearly depends on (plan nodes count)*(overall columns
>> count), thus it quadratically depends on number of tables unionized.
>>
>> To remove this excessive memory usage I propose
>> to run deparse_context_for_planstate+deparse_expression in a separate
>> memory context and free it after a plan node is generated.
> Hmm, something like the attached? Seems reasonable...
This looks pretty unsafe to me: it assumes, without much justification,
that there is no memory allocated during show_expression() that will be
needed later.
I suspect the O(N^2) consumption comes directly from some workspace
allocated for variable deparsing in ruleutils.c. A safer fix would
be to do retail pfrees on that.
regards, tom lane