Re: row filtering for logical replication

On Wed, Jan 19, 2022 at 7:45 AM houzj(dot)fnst(at)fujitsu(dot)com
<houzj(dot)fnst(at)fujitsu(dot)com> wrote:
>
> On Tues, Jan 18, 2022 8:35 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> Attach the V67 patch set which address the above comments.
>

Some more comments and suggestions:
=================================
1. Can we do slot initialization in maybe_send_schema() instead of
introducing a new flag for it?

2.
+ * For updates if no old tuple, it means none of the replica identity
+ * columns changed and this would reduce to a simple update. We only need
+ * to evaluate the row filter for the new tuple.

Is it possible with the current version of the patch? I am asking
because, for updates, we now allow only RI columns in row filter, so
do we need to evaluate the row filter in this case? I think ideally,
we don't need to evaluate the row filter in this case as for updates
only replica identity columns are allowed but users can use constant
expressions in the row filter. So, we need to evaluate the row filter
in this case as well. Is my understanding correct?

3. + /* If no filter found, clean up the memory and return */
+ if (!has_filter)
+ {
+ if (entry->cache_expr_cxt != NULL)
+ MemoryContextDelete(entry->cache_expr_cxt);

I think this clean-up needs to be performed when we set
exprstate_valid to false. I have changed accordingly in the attached
patch.

Apart from the above, I have made quite a few changes in the code
comments in the attached top-up patch, kindly review those and merge
them into the main patch, if you are okay with it.

--
With Regards,
Amit Kapila.