Re: partitioning - changing a slot's descriptor is expensive

On 2018/06/27 14:55, Andres Freund wrote:
> On 2018-06-27 14:46:26 +0900, Amit Langote wrote:
>> There is however similar code that runs in non-error paths, such as in
>> ExecPrepareTupleRouting(), that is executed *if* the leaf partition and
>> the root parent have differing attribute numbers.  So, one might say that
>> that code's there to handle the special cases which may not arise much in
>> practice, but it may still be a good idea to make improvements if we can
>> do so.
> 
> Yea, I was referring to all do_convert_tuple() callers, and some of them
> are more hot-path than the specific one above. E.g. the
> ConvertPartitionTupleSlot() call in CopyFrom().

Just in case you haven't noticed, ConvertPartitionTupleSlot(), even if
it's called unconditionally from CopyFrom etc., calls do_convert_tuple
only if necessary.  Note the return statement at the top of its body.

HeapTuple
ConvertPartitionTupleSlot(TupleConversionMap *map,
                          HeapTuple tuple,
                          TupleTableSlot *new_slot,
                          TupleTableSlot **p_my_slot)
{
    if (!map)
        return tuple;

map is non-NULL only if a partition has different attribute numbers than
the root parent.

Thanks,
Amit