Re: Gripes about walsender command processing

Michael Paquier <michael(at)paquier(dot)xyz> writes:
> On Mon, Sep 14, 2020 at 11:34:44PM -0400, Tom Lane wrote:
>> (I don't quite follow your comment about repslot drop, btw.)

> There is already a command tag equivalent to DROP_REPLICATION_SLOT:
> $ git grep REPLICATION -- */cmdtaglist.h
> src/include/tcop/cmdtaglist.h:PG_CMDTAG(CMDTAG_DROP_REPLICATION_SLOT,
> "DROP REPLICATION SLOT", false, false, false)

[ blink ... ] So why is it that DropReplicationSlot does

SetQueryCompletion(&qc, CMDTAG_DROP_REPLICATION_SLOT, 0);
EndCommand(&qc, DestRemote, false);

when the caller will immediately after that do

SetQueryCompletion(&qc, CMDTAG_SELECT, 0);
EndCommand(&qc, DestRemote, true);

StartLogicalReplication has a similar weirdness.
The more I look at this code, the more broken it seems.

Anyway, independently of whether walsender should be sending multiple
command-complete messages, I don't think I really approve of including
replication commands in the CommandTag enum. It still seems like a
type pun. However, it looks like we'd have to duplicate
SetQueryCompletion/EndCommand if we don't want to do that, so maybe
I'd better just hold my nose and do it that way.

regards, tom lane