Re: [SQL] Are PL/pgSQL calls atomic?

Mark Wright asked:

>
> If I call a PL/pgSQL function, and another client makes the same call, does
> the second client have to wait for the first to complete execution before it
> begins? If not, is there some sort of mechanism I can use to prevent more
> than one call to a function from happening at the same time (something like
> a mutex in Win32).
>
> I need to select a row and then mark it as unavailable for other clients to
> use. The table looks like:
> create table xyz (id serial, status char default 'N', ...);
>
> My function finds a row by doing:
> select id into my_id_variable from xyz where id
> = (select min(id) from xyz where status = 'N');
>
> and then marks that row as unavailable by setting status:
> update xyz set status = 'Y' where id = my_id_variable;
>
> Obviously, if a second client calls this function before the UPDATE takes
> place, it will pick up the same row as the first client, causing rending of
> hair and gnashing of teeth.
>
> How do I prevent this?

It's not a PL/pgSQL specific problem I think. Doing this in
two psql sessions:

S1: BEGIN;
S1: SELECT id FROM xyz WHERE id =
(SELECT min(id) FROM xyz WHERE status = 'N')
FOR UPDATE OF xyz;

-- Session 1 returns one id, say 42

S2: BEGIN;
S2: SELECT id FROM xyz WHERE id =
(SELECT min(id) FROM xyz WHERE status = 'N')
FOR UPDATE OF xyz;

-- Session 2 blocks now, but...

S1: UPDATE xyz SET status = 'Y' WHERE ID = 42;
S1: COMMIT;

-- Session 2 continues and returns id 42

The reason why session 2 returns the same id is that the
qualifying subselect finding the lowest id where status = 'N'
didn't block. The outer SELECT FOR UPDATE did, but that
didn't check the state any more.

Unfortunately a

SELECT min(id) FROM xyz WHERE status = 'N' FOR UPDATE OF xyz;

doesn't work either because FOR UPDATE isn't allowed with
aggregates.

From an application you could use a whole table lock before
looking up the id. Here now is a PL/pgSQL problem. LOCK
TABLE is a utility statement and PL/pgSQL cannot execute
them. I think I have to work on that for v6.6.

What you could do (if the number of rows with status = 'N'
isn't high) is the following:

CREATE FUNCTION my_func .... AS '
DECLARE
xyz_rec RECORD;
BEGIN
FOR xyz_rec IN SELECT * FROM xyz WHERE status = ''N''
ORDER BY id FOR UPDATE OF xyz
LOOP
-- If more changes in xyz are to be made than just setting
-- status to Y, do them all in one UPDATE. The record is
-- locked now and the lock will release only when our entire
-- transaction commits or rolls back - not when we update it.

UPDATE xyz SET status = ''Y'' WHERE id = xyz_rec.id;
...

-- Now we return from inside the loop at the first
-- row processed. This ensures we will process one
-- row at max per call.
RETURN _whatever_my_func_returns_;
END LOOP;

-- If we reach here, we did not find any row (left) with
-- status N. Hmmm - is this an error or not?

RAISE ERROR ''no (more) xyz rows with status N found'';
END;' LANGUAGE 'plpgsql';

Why it's important that the number of rows having status =
'N' isn't very big is because at the FOR xyz_rec IN time all
of them are fetched into memory, even if the loop will break
at the first one got.

Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#======================================== jwieck(at)debis(dot)com (Jan Wieck) #