Обсуждение: Counterintuitive locking behavior

On Sunday, January 06, 2013 7:48 AM Chris Travers wrote:

> I recently discovered that subselects in update statements don't assume that the select is for update of the updating
table.


> For example, if I do this:


> CREATE TABLE foo (
>   test int primary key,
> );


> INSERT INTO foo VALUES (1);


> then in one session:


> BEGIN;
> UPDATE foo SET test = 2 WHERE test in (select test from foo where test = 1);


> and then in the other session


> BEGIN;
> UPDATE foo SET test = 3 WHERE test in (select test from foo where test = 1);

The behavior will be same even for UPDATE foo SET test = 3 WHERE test =1;



> Is there a reason why we don't do locking this way?  (i.e. where on UPDATE foo, all rows selected from foo during the

> update are locked unless the subselect specifically states otherwise.)

The reason for this behavior is if it locks all rows during select, then it can so happen that Update will actually not
happenon the row but it will be locked. 
For example

UPDATE foo SET test = 3 WHERE test in (select test from foo where test = 1) and FALSE;

Now in this case if it locks the rows during subselect, then the rows will be locked during whole transaction
irrespective of the fact that they will not be updated.

With Regards,
Amit Kapila.

On Sunday, January 06, 2013 11:10 AM Amit kapila wrote:
On Sunday, January 06, 2013 7:48 AM Chris Travers wrote:




>> Is there a reason why we don't do locking this way?  (i.e. where on UPDATE foo, all rows selected from foo during
the
>> update are locked unless the subselect specifically states otherwise.)

>The reason for this behavior is if it locks all rows during select, then it can so happen that Update will actually
nothappen on the row but it will be locked. 
> For example

> UPDATE foo SET test = 3 WHERE test in (select test from foo where test = 1) and FALSE;

> Now in this case if it locks the rows during subselect, then the rows will be locked during whole transaction
> irrespective of the fact that they will not be updated.

In the above example and FALSE, I mean to say any other subquery which will yield the overall condition to not return
anyrow. 
Similarly there can be many more scenarios where only half of the selected rows (by one of the conds. ) will be actual
candidatesof Update. 

With Regards,
Amit Kapila.

2013-01-06 03:18 keltezéssel, Chris Travers írta:
> Hi everyone;
>
> I recently discovered that subselects in update statements don't assume that the select
> is for update of the updating table.
>
> For example, if I do this:
>
> CREATE TABLE foo (
>    test int primary key,
> );
>
> INSERT INTO foo VALUES (1);
>
> then in one session:
>
> BEGIN;
> UPDATE foo SET test = 2 WHERE test in (select test from foo where test = 1);
>
> and then in the other session
>
> BEGIN;
> UPDATE foo SET test = 3 WHERE test in (select test from foo where test = 1);
>
> When I commit both transactions, the second one chronologically always takes precedence.
>  In other words, the locks takes effect after the subselect but before the rows are
> updated.  This strikes me as quite error prone and quite a bit more error prone than a
> rule which says that unless stated otherwise subselects of the updated table are to be
> selected for update.
>
> This may strike some as a "do what I mean" kind of feature, but the way I am looking at
> it is that a SQL statement is usually written as a declarative block, and an assumption
> that the SQL statement is to be evaluated atomically is a good one for predicability of
> software (in other words, locks apply to the whole statement).
>
> Is there a reason why we don't do locking this way?  (i.e. where on UPDATE foo, all rows
> selected from foo during the update are locked unless the subselect specifically states
> otherwise.)

What you are seeing is the MVCC behaviour of PostgreSQL
in READ COMMITTED mode. You can use REPEATABLE READ mode
in 9.2.x or SERIALIZABLE mode in earlier generations.
Please read the "Concurrency control" section of the documentation.
http://www.postgresql.org/docs/9.2/interactive/mvcc.html

With this mode:

session 1:

zozo=> create table foo (test int primary key);
NOTICE:  CREATE TABLE / PRIMARY KEY will create implicit index "foo_pkey" for table "foo"
CREATE TABLE
zozo=> insert into foo values (1);
INSERT 0 1
zozo=> begin isolation level repeatable read;
BEGIN
zozo=> update foo set test = 2 where test = (select test from foo where test = 1);
UPDATE 1

session 2:
zozo=> begin isolation level repeatable read;
BEGIN
zozo=> update foo set test = 2 where test = (select test from foo where test = 1);
(session 2 is waiting for the lock on the row at this point)

session 1:

zozo=> commit;
COMMIT

session 2 threw an error after session 1 committed:

ERROR:  could not serialize access due to concurrent update

But updating a row blindly is quite silly even in this small example,
you can use SELECT ... FOR UPDATE for explicit locking.

session 1:

zozo=> begin;
BEGIN
zozo=> select * from foo where test = 1 for update;
  test
------
     1
(1 row)

session 2:

zozo=> begin;
BEGIN
zozo=> select * from foo where test = 1 for update;
session 2 waits...

session 1:

zozo=> update foo set test = 2 where test = 1;
UPDATE 1
zozo=> commit;
COMMIT

session 2 now gives the result:

zozo=> select * from foo where test = 1 for update;
  test
------
(0 rows)

and your application can warn the user that the row is not there anymore,
so no point in the subsequent UPDATE. You should do a new query to find
the row you wanted.

Best regards,
Zoltán Böszörményi

--
----------------------------------
Zoltán Böszörményi
Cybertec Schönig & Schönig GmbH
Gröhrmühlgasse 26
A-2700 Wiener Neustadt, Austria
Web: http://www.postgresql-support.de
      http://www.postgresql.at/

Chris Travers wrote on 06.01.2013 03:18:
> i.e. where on UPDATE foo, all rows selected from foo during the
> update are locked unless the subselect specifically states
> otherwise.

That would strike *me* as wrong.

The UPDATE should only lock the rows that were actually updated, not those that were selected.

Thomas