Re: row filtering for logical replication

On Wed, Jan 26, 2022 at 8:37 AM houzj(dot)fnst(at)fujitsu(dot)com
<houzj(dot)fnst(at)fujitsu(dot)com> wrote:
>
> On Monday, January 24, 2022 4:38 PM Peter Smith <smithpb2250(at)gmail(dot)com> wrote:
> >
> >
> > 3. src/backend/utils/cache/relcache.c - RelationBuildPublicationDesc
> >
> > +RelationBuildPublicationDesc(Relation relation)
> > {
> > List *puboids;
> > ListCell *lc;
> > MemoryContext oldcxt;
> > Oid schemaid;
> > - PublicationActions *pubactions = palloc0(sizeof(PublicationActions));
> > + List *ancestors = NIL;
> > + Oid relid = RelationGetRelid(relation); AttrNumber invalid_rfcolnum =
> > + InvalidAttrNumber; PublicationDesc *pubdesc =
> > + palloc0(sizeof(PublicationDesc)); PublicationActions *pubactions =
> > + &pubdesc->pubactions;
> > +
> > + pubdesc->rf_valid_for_update = true;
> > + pubdesc->rf_valid_for_delete = true;
> >
> > IMO it wold be better to change the "sense" of those variables.
> > e.g.
> >
> > "rf_valid_for_update" --> "rf_invalid_for_update"
> > "rf_valid_for_delete" --> "rf_invalid_for_delete"
> >
> > That way they have the same 'sense' as the AttrNumbers so it all reads better to
> > me.
> >
> > Also, it means no special assignment is needed because the palloc0 will set
> > them correctly
>
> Think again, I am not sure it's better to have an invalid_... flag.
> It seems more natural to have a valid_... flag.
>

Can't we do without these valid_ flags? AFAICS, if we check for
"invalid_" attributes, it should serve our purpose because those can
have some attribute number only when the row filter contains some
column that is not part of RI. A few possible optimizations in
RelationBuildPublicationDesc:

a. It calls contain_invalid_rfcolumn with pubid and then does cache
lookup to again find a publication which its only caller has access
to, so can't we pass the same?
b. In RelationBuildPublicationDesc(), we call
GetRelationPublications() to get the list of publications and then
process those publications. I think if none of the publications has
row filter and the relation has replica identity then we don't need to
build the descriptor at all. If we do this optimization inside
RelationBuildPublicationDesc, we may want to rename function as
CheckAndBuildRelationPublicationDesc or something like that?

--
With Regards,
Amit Kapila.