Re: row filtering for logical replication

On Mon, Jan 24, 2022 at 5:09 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Mon, Jan 24, 2022 at 10:29 AM Greg Nancarrow <gregn4422(at)gmail(dot)com> wrote:
> >
> > On Mon, Jan 24, 2022 at 2:47 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > >
> > > > (3) pgoutput_row_filter_exec_expr
> > > > pgoutput_row_filter_exec_expr() returns false if "isnull" is true,
> > > > otherwise (if "isnull" is false) returns the value of "ret"
> > > > (true/false).
> > > > So the following elog needs to be changed (Peter Smith previously
> > > > pointed this out, but it didn't get completely changed):
> > > >
> > > > BEFORE:
> > > > + elog(DEBUG3, "row filter evaluates to %s (isnull: %s)",
> > > > + DatumGetBool(ret) ? "true" : "false",
> > > > + isnull ? "true" : "false");
> > > > AFTER:
> > > > + elog(DEBUG3, "row filter evaluates to %s (isnull: %s)",
> > > > + isnull ? "false" : DatumGetBool(ret) ? "true" : "false",
> > > > + isnull ? "true" : "false");
> > > >
> > >
> > > Do you see any problem with the current? I find the current one easy
> > > to understand.
> > >
> >
> > Yes, I see a problem.
> >
>
> I tried by inserting NULL value in a column having row filter and the
> result it shows is:
>
> LOG: row filter evaluates to false (isnull: true)
>
> This is what is expected.
>
> >
> > But regression tests fail when that code change is made (indicating
> > that there are cases when "isnull" is true but the function returns
> > true instead of false).
> >
>
> But that is not what I am seeing in Logs with a test case where the
> row filter column has NULL values. Could you please try that see what
> is printed in LOG?
>
> You can change the code to make the elevel as LOG to get the results
> easily. The test case I tried is as follows:
> Node-1:
> postgres=# create table t1(c1 int, c2 int);
> CREATE TABLE
> postgres=# create publication pub for table t1 WHERE (c1 > 10);
> CREATE PUBLICATION
>
> Node-2:
> postgres=# create table t1(c1 int, c2 int);
> CREATE TABLE
> postgres=# create subscription sub connection 'dbname=postgres' publication pub;
> NOTICE: created replication slot "sub" on publisher
> CREATE SUBSCRIPTION
>
> After this on publisher-node, I see the LOG as "LOG: row filter
> evaluates to false (isnull: true)". I have verified that in the code
> as well (in slot_deform_heap_tuple), we set the value as 0 for isnull
> which matches above observation.
>

There are obviously multiple code paths under which a column can end up as NULL.
Doing one NULL-column test case, and finding here that
"DatumGetBool(ret)" is "false" when "isnull" is true, doesn't prove it
will be like that for ALL possible cases.
As I pointed out, the function is meant to always return false when
"isnull" is true, so if the current logging code is correct (always
logging "DatumGetBool(ret)" as the function return value), then to
match the code to the current logging, we should be able to return
"DatumGetBool(ret)" if "isnull" is true, instead of returning false as
it currently does.
But as I said, when I try that then I get a test failure (make
check-world), proving that there is a case where "DatumGetBool(ret)"
is true when "isnull" is true, and thus showing that the current
logging is not correct because in that case the current log output
would show the return value is true, which won't match the actual
function return value of false.
(I also added some extra logging for this isnull==true test failure
case and found that ret==1)

Regards,
Greg Nancarrow
Fujitsu Australia