On 2021/03/29 13:55, Michael Paquier wrote:
On Mon, Mar 29, 2021 at 10:53:14AM +0900, Fujii Masao wrote:
I understand the motivation of this. But the other DMLs like UPDATE also
do the same thing for foreign tables? That is, when those DML commands
are executed on foreign tables, their changes are WAL-logged in a publisher
side,
e.g., for logical replication? If not, it seems strange to allow only TRUNCATE
on foreign tables to be WAL-logged in a publisher side...
Executing DMLs on foreign tables does not generate any WAL AFAIK with
the backend core code, even with wal_level = logical, as the DML is
executed within the FDW callback (see just ExecUpdate() or
ExecInsert() in nodeModifyTable.c), and foreign tables don't have an
AM set as they have no physical storage. A FDW may decide to generate
some WAL records by itself though when doing the opeation, using the
generic WAL interface but that's rather limited.
Generating WAL for the truncation of foreign tables sounds also like a
strange concept to me. I think that you should just make the patch
work so as the truncation is passed down to the FDW that decides what
it needs to do with it, and do nothing more than that.
Agreed.
Regards,
--
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION
