Re: BUG #17376: Adding unique column with a function() default results in "could not read block 0 in file" error

Japin Li <japinli(at)hotmail(dot)com> writes:
> On Sat, 22 Jan 2022 at 01:36, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> However, I fear we don't have adequate infrastructure to tell which
>> table accesses *are* part of the ALTER TABLE machinery and which aren't.

> Right.

>> Maybe it'd be sufficient to check for an active ALTER TABLE in the
>> parser, but I'm not sure.

> Do you mean check the table is accessed by ALTER TABLE when calling SPI to
> execute the function? How can we get the table that is be accessed by
> ALTER TABLE in parser?

The reason there's not a problem like this with respect to cross-session
accesses is that the exclusive lock acquired by ALTER TABLE prevents
other sessions from touching the table. Conversely, the reason we do
have a problem here is that such locks don't prevent accesses within
our own session. So a rough sketch for fixing this might be "at
the point where we acquire a lock on some table for a new statement,
also check to see if some internal-to-our-session operation has
declared a need for exclusive access". The problem is that we don't
have a well-defined line as to what's part of the ALTER and what
is not. For example, I recall that ALTER TABLE handles index rebuilding
by creating textual ALTER ADD INDEX commands, and parsing and executing
those. How can we distinguish that behavior from a SQL function within
an index or CHECK constraint trying to execute ALTER ADD INDEX?

regards, tom lane