Re: row filtering for logical replication

On Thu, Sep 9, 2021 at 11:43 AM Peter Smith <smithpb2250(at)gmail(dot)com> wrote:
>
> PSA my new incremental patch (v28-0002) that introduces row filter
> validation for the publish mode "delete". The validation requires that
> any columns referred to in the filter expression must also be part of
> REPLICA IDENTITY or PK.
>
> [This v28-0001 is identical to the most recently posted rebased base
> patch. It is included again here only so the cfbot will be happy]
>
> ~~
>
> A requirement for some filter validation like this has been mentioned
> several times in this thread [1][2][3][4][5].
>
> I also added some test code for various kinds of replica identity.
>
> A couple of existing tests had to be modified so they could continue
> to work (e.g. changed publish = "insert" or REPLICA IDENTITY FULL)
>
> Feedback is welcome.
>
> ~~
>
> NOTE: This validation currently only checks when the filters are first
> created. Probably there are many other scenarios that need to be
> properly handled. What to do if something which impacts the existing
> filter is changed?
>
> e.g.
> - what if the user changes the publish parameter using ALTER
> PUBLICATION set (publish="delete") etc?
> - what if the user changes the replication identity?
> - what if the user changes the filter using ALTER PUBLICATION in a way
> that is no longer compatible with the necessary cols?
> - what if the user changes the table (e.g. removes a column referred
> to by a filter)?
> - what if the user changes a referred column name?
> - more...
>
> (None of those are addressed yet - thoughts?)
>

I think we need to remove the filter or the table from publication in
such cases. Now, one can think of just removing the condition related
to the column being removed/changed in some way but I think that won't
be appropriate because it would change the meaning of the filter. We
are discussing similar stuff in the column filter thread and we might
want to do the same for row filters as well. I would prefer to remove
the table in both cases as Rahila has proposed in the column filter
patch.

--
With Regards,
Amit Kapila.