Re: row filtering for logical replication

On Tue, Dec 21, 2021 at 5:58 AM Euler Taveira <euler(at)eulerto(dot)com> wrote:
>
> I reviewed 0003. It uses TupleTableSlot instead of HeapTuple. I probably missed
> the explanation but it requires more changes (logicalrep_write_tuple and 3 new
> entries into RelationSyncEntry). I replaced this patch with a slightly
> different one (0005 in this patch set) that uses HeapTuple instead. I didn't
> only simple tests and it requires tests. I noticed that this patch does not
> include a test to cover the case where TOASTed values are not included in the
> new tuple. We should probably add one.

The reason I changed the code to use virtualtuple slots is to reduce
tuple deforming overhead.
Dilip raised this very valid comment in [1]:

On Tue, Sep 21, 2021 at 4:29 PM Dilip Kumar
<dilipbalaut(at)gmail(dot)com> wrote:
>
>In pgoutput_row_filter_update(), first, we are deforming the tuple in
>local datum, then modifying the tuple, and then reforming the tuple.
>I think we can surely do better here. Currently, you are reforming
>the tuple so that you can store it in the scan slot by calling
>ExecStoreHeapTuple which will be used for expression evaluation.
>Instead of that what you need to do is to deform the tuple using
>tts_values of the scan slot and later call ExecStoreVirtualTuple(), so
>advantages are 1) you don't need to reform the tuple 2) the expression
>evaluation machinery doesn't need to deform again for fetching the
>value of the attribute, instead it can directly get from the value
>from the virtual tuple.

Storing the old tuple/new tuple in a slot and re-using the slot avoids
the overhead of
continuous deforming of tuple at multiple levels in the code.

regards,
Ajin Cherian
[1] - https://www.postgresql.org/message-id/CAFiTN-vwBjy+eR+iodkO5UVN5cPv_xx1=s8ehzgCRJZA+AztAA@mail.gmail.com