On 02/02/2014 08:54 PM, Craig Ringer wrote:
> On 02/03/2014 09:09 AM, Craig Ringer wrote:
>
>> At a guess, we're looking at a case where a new child context is created
>> at every call, so every MemoryContextResetChildren call has to deal with
>> more child contexts.
> That would be "yes". After a short run, I see 32849 lines like:
>
> json_array_elements temporary cxt: 8192 total in 1 blocks; 8160 free (0
> chunks); 32 used
>
> under the context:
>
> PortalMemory: 8192 total in 1 blocks
> PortalHeapMemory: 7168 total in 3 blocks
> ExecutorState: 65600 total in 4 blocks
> ExprContext: 8192 total in 1 blocks
> json_array_elements temporary cxt: 8192 total in 1 blocks;
> 8160 free (0 chunks); 32 used
>
>
> The attached patch deletes the context after use, bringing performance
> back into line. It should be backpatched to 9.3.
[...]
> + MemoryContextDelete(state->tmp_cxt);
> + state->tmp_cxt = NULL;
> +
> PG_RETURN_NULL();
Hmm. I guess I was assuming that the tmp_cxt would be cleaned up at the
end of the function since it's a child of the CurrentMemoryContext. But
if I got that wrong I'm happy to fix it. I don't see the point in
setting state->tmp_cxt to NULL since it's about to go out of scope anyway.
cheers
andrew