Re: invalid tid errors in latest 7.3.4 stable.

I said:
> Okay, I'll work out some extension of the APIs to let us propagate the
> snapshot request down through SPI and into the Executor, rather than
> using a global variable for it. (Unless someone has a better idea...)

I've committed the attached patch into CVS HEAD. I am now wondering
whether to back-patch it to the 7.3 branch or not. It's a bit larger
than I would have liked, and really needs more testing before being
shoved into a stable branch.

The simplest test case I was able to generate for Wade's bug is this:

-----------
create table t1 (f1 int primary key);
create table t2 (f1 int references t1 on delete cascade);
create table t3 (f1 int);

create or replace function t2del() returns trigger as '
begin
update t3 set f1 = f1 + 1;
return old;
end' language plpgsql;

create trigger t2del before delete on t2 for each row
execute procedure t2del();

create or replace function t3upd() returns trigger as '
begin
perform count(*) from t3;
return new;
end' language plpgsql;

create trigger t3upd before update on t3 for each row
execute procedure t3upd();

insert into t1 values(1);
insert into t2 values(1);
insert into t3 values(1);

delete from t1;
-----------

Until this commit, CVS HEAD generated
ERROR: attempted to mark4update invisible tuple
CONTEXT: PL/pgSQL function "t2del" line 2 at SQL statement
7.3 branch generates a different spelling of the same error:
WARNING: Error occurred while executing PL/pgSQL function t2del
WARNING: line 2 at SQL statement
ERROR: heap_mark4update: (am)invalid tid

AFAICT you need a minimum of two levels of triggers invoked by an RI
trigger to make this happen, so it may be a corner case best left
unfixed in the 7.3 branch.

Opinions anyone?

regards, tom lane