Horton

Thread

Commits

  1. Force NO SCROLL for plpgsql's implicit cursors.

  2. Avoid misbehavior when persisting a non-stable cursor.

  1. BUG #17050: cursor with for update + commit in loop

    The Post Office <noreply@postgresql.org> — 2021-06-08T07:49:25Z

    The following bug has been logged on the website:

Bug reference:      17050
Logged by:          Алексей Булгаков
Email address:      bulgakovalexey1980@gmail.com
PostgreSQL version: 12.7
Operating system:   Red Hat 4.4.7-23
Description:        

create table public.test_tuple_stream (
	   id serial,
	   nm text,
	   dt timestamptz,
	   num bigint
	   );

CREATE OR REPLACE PROCEDURE public.test_tuple_stream()
 LANGUAGE plpgsql
AS $procedure$
  declare
	  l_cur cursor for
	    select id
	    from public.test_tuple_stream
	    order by id
	    for update;
	begin
	  for rec in l_cur loop
	    update public.test_tuple_stream
	    set num = num + 1
	    where id = rec.id;
	   
	    commit;	   
	  end loop;
	 
	  commit;
	END;
$proc	  
	  
--	truncate table public.test_tuple_stream;  
	  
	insert into public.test_tuple_stream(nm, dt, num)
	values ('A', now(), 1);
	insert into public.test_tuple_stream(nm, dt, num)
	values ('B', now(), 1);
	insert into public.test_tuple_stream(nm, dt, num)
	values ('C', now(), 1);

  call public.test_tuple_stream()

	select *
	from public.test_tuple_stream
	order by id

If run procedure test_tuple_stream then in result updated 2 rows of 3.
Why?
if remove in procedure "for update" or "commit in loop" then updated 3 rows
of 3

  2. Re: BUG #17050: cursor with for update + commit in loop

    Tom Lane <tgl@sss.pgh.pa.us> — 2021-06-08T18:54:15Z

    PG Bug reporting form <noreply@postgresql.org> writes:
> If run procedure test_tuple_stream then in result updated 2 rows of 3.
> Why?

So what's happening here is that at the point of the first COMMIT,
we need to save the output of the cursor query, because we have to
release locks and so forth.  The expectation is that subsequent
fetches will read from a tuplestore rather than the live query.

To do that, PersistHoldablePortal supposes that it can rewind and
re-read the query's executor run.  But that only works if the
query is guaranteed stable, which anything involving FOR UPDATE
is not.  In the example at hand, when we re-read the row with ID 1,
nodeLockRows sees that its state is now TM_SelfModified thanks to the
UPDATE we just did.  This causes it to not return that row, so that
what ends up in the tuplestore is only the rows with IDs 2 and 3.
But since the cursor "knows" that one row has already been read out,
the next fetch will return the row with ID 3, and the procedure's
loop never visits the row with ID 2.

I imagine that similar misbehavior could be constructed around
queries containing volatile functions (e.g. random()), rather than
FOR UPDATE.

The only obvious way to fix this is to always save aside the output
of a cursor query in case we need to persist it later, so that
PersistHoldablePortal doesn't have to assume that rewinding is safe.
That would be pretty catastrophic for performance, though, so I doubt
anybody will be happy with that answer.

For cursors that aren't marked scrollable, we might be able to say
that we only save the *rest* of the query output, and then adjust
the cursor state appropriately for that choice.  Seems possibly
nontrivial though, and there's still the question of what to do
for scrollable ones.

Anyway, many thanks for the report!  But don't hold your breath
for a fix; this is going to take some thought.

			regards, tom lane

  3. Re: BUG #17050: cursor with for update + commit in loop

    Tom Lane <tgl@sss.pgh.pa.us> — 2021-06-08T20:01:46Z

    I wrote:
> The only obvious way to fix this is to always save aside the output
> of a cursor query in case we need to persist it later, so that
> PersistHoldablePortal doesn't have to assume that rewinding is safe.
> That would be pretty catastrophic for performance, though, so I doubt
> anybody will be happy with that answer.
> For cursors that aren't marked scrollable, we might be able to say
> that we only save the *rest* of the query output, and then adjust
> the cursor state appropriately for that choice.  Seems possibly
> nontrivial though, and there's still the question of what to do
> for scrollable ones.

Actually ... maybe it's not that bad.  The nonscrollable case seems
to be quite simple to fix, and as for the scrollable case, maybe
we can just say it's on the user's head that the query produce
stable results.  There's already a prohibition on using FOR UPDATE
with SCROLL, and the DECLARE CURSOR reference page has some warnings
about volatile queries with WITH HOLD, which is basically the same
case as we're worried about here.

I think the DECLARE CURSOR page needs some modernization to mention
that cursors in procedures are basically the same as WITH HOLD.
But as far as code changes go, the attached seems sufficient.

			regards, tom lane