Re: Speeding up INSERTs and UPDATEs to partitioned tables

On 9 October 2018 at 15:49, Amit Langote <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
> On 2018/10/05 21:55, David Rowley wrote:
>> I'm not so sure we need to zero the partition_tuple_slots[] array at
>> all since we always set a value there is there's a corresponding map
>> stored. I considered pulling that out, but in the end, I didn't as I
>> saw some Asserts checking it's been properly set by checking the
>> element != NULL in nodeModifyTable.c and copy.c. Perhaps I should
>> have just gotten rid of those Asserts along with the palloc0 and
>> subsequent memset after the expansion of the array. I'm undecided.
>
> Maybe it's a good thing that it's doing the same thing as with the
> child_to_parent_maps array, which is to zero-init it when allocated.

Perhaps, but the maps do need to be zeroed, the partition_tuple_slots
array does not since we only access it when the parent to child map is
set.

In any case, since PARTITION_ROUTING_INITSIZE is just 8, it'll likely
not save much since it's really just saving a memset(..., 0, 64), for
single-row INSERTs on a 64-bit machine. So likely it won't save more
than a bunch of nanoseconds.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services