views:

102

answers:

8

hi,

I have a method SaveApp() which will deactivate the existing records and insert a new one.

void SaveApp(int appID)
{
   begin transaction;
   update;
   insert;
   commit transaction;
}

Let's say in database table SalesApp, I have 2 records with appID equal to 123;

  1. record 1, appID 123, inactive
  2. record 2, appID 123, active

If I call this method SaveApp() in two threads at same time, the first transaction (let's call it T1) will update the existing two record while the second transaction (let's call it T2) wait.

after T1 finishes, there will be three records in this table now. however, somehow T2 is not aware of the newly inserted record, the update query in T2 only update the previous two records, and insert the forth one.

after these two method call, in database, we will now have 4 records, the 3rd and 4th one both are active which is wrong.

  1. record 1, appID 123, inactive
  2. record 2, appID 123, inactive
  3. record 3, appID 123, active
  4. record 4, appID 123, active

Do you know any solution can solve this problem? I have tried using isolation level serializable which does not work.

Thanks!

A: 

Can you push the updates onto a queue (AQ maybe) so that they're executed sequentially?

Another option might be to lock the records in question (SELECT FOR UPDATE NOWAIT or SELECT FOR UPDATE WAIT)

cagcowboy
A: 

Seems like it isn't really an Oracle issue, it's a concurrency issue in your application. Not sure what language this is; if it's Java can you just synchronise the method?

Alex Poole
It is .NET app, we have 8 web servers, I do not think synchronise will work.
Ding
A: 

@Alex is correct, it's not really an Oracle issue, it's an application issue.

Perhaps something like this might work for you:

Put your Oracle transaction in a stored procedure, and perform it this way:

BEGIN
  LOOP
    BEGIN
      SELECT * 
        FROM SaleApp
       WHERE appID = 123
         AND status = 'ACTIVE'
         FOR UPDATE NOWAIT;
      EXIT;
    EXCEPTION
      WHEN OTHERS THEN
        IF SQLCODE = -54 THEN
          NULL;
        ELSE
          RAISE error
        END IF;
    END IF;
  END LOOP;
  UPDATE ....
  INSERT ....
  COMMIT;
END;

The idea here is that the first transaction to grab and lock the current active record gets to complete. Any other transactions that try to lock that record will fail on the SELECT FOR UPDATE NOWAIT, and loop until they succeed.

Depending on how long it takes to perform a typical transaction, you may want to sleep within the exception handler before retrying the select.

DCookie
A: 

I'm not entirely sure, but I think that if you set both transactions to SERIALIZABLE you would get an error in the second one so that you would know that something was wrong.

a_horse_with_no_name
+5  A: 

Do you have another table that holds one row per AppId, enforced via unique or primary key constraint? If so use select for update on the parent table to serialize access per AppId.

Create the tables:

session_1> create table parent (AppId number primary key);

Table created.

session_1> create table child (AppId number not null references Parent(AppId)
  2      , status varchar2(1) not null check (status in ('A', 'I'))
  3      , InsertedAt date not null)
  4  /

Table created.

Insert beginning values:

session_1> insert into Parent values (123);

1 row created.

session_1> insert into child values (123, 'I', sysdate);

1 row created.

session_1> insert into child values (123, 'A', sysdate);

1 row created.

session_1> commit;

Commit complete.

Begin the first transaction:

session_1> select AppId from Parent where AppId = 123 for update;

     APPID
----------
       123

session_1> update Child set Status = 'I' where AppId = 123 and Status = 'A';

1 row updated.

session_1> insert into child values (123, 'A', sysdate);

1 row created.

Prior to commit, in a second session, make sure we are only seeing the first rows:

session_2> select * from Child;

     APPID S INSERTEDAT
---------- - -------------------
       123 I 2010-08-16 18:07:17
       123 A 2010-08-16 18:07:23

Start the second transaction:

session_2> select AppId from Parent where AppId = 123 for update;

Session 2 is now blocked, waiting on session 1. And will not proceed. Commiting session 1 will unblock Session

session_1> commit;

Commit complete.

Session 2 we now see:

     APPID
----------
       123

Complete the second transaction:

session_2> update Child set Status = 'I' where AppId = 123 and Status = 'A';

1 row updated.

session_2> insert into child values (123, 'A', sysdate);

1 row created.

session_2> commit;

Commit complete.

session_2> select * from Child;

     APPID S INSERTEDAT
---------- - -------------------
       123 I 2010-08-16 18:07:17
       123 I 2010-08-16 18:07:23
       123 I 2010-08-16 18:08:08
       123 A 2010-08-16 18:13:51

EDIT Technique cribbed from Expert Oracle Database Architecture second edition by Thomas Kyte, pages 23-24. http://www.amazon.com/Expert-Oracle-Database-Architecture-Programming/dp/1430229462/ref=sr_1_2?ie=UTF8&s=books&qid=1282061675&sr=8-2

EDIT 2 I would also recommend implementing Patrick Merchand's answer to this question for a constraint that enforces the rule that an AppId can only have one active record. So the final solution would have two parts, this answer for how to do the updates in a way that gets what you want, and Patrick's to make sure that table conforms to the requirements to protect the integrity of the data.

Shannon Severance
+1, clever approach, as Mr. Kyte's solutions usually are ;-)
DCookie
if I do not have parent table, does this select for update trick still work? thx!
Ding
I don't know how to make it work without a parent table. `select for update` only locks rows that exists. So with just one table, you don't have anything to lock on that will keep the multiple inserts from happening in seperate transactions. You might test `select * from YourTable where AppId = :AppId for update` It might work, but will be much more expensive then locking the table, especially as `YourTable` grows. You might consider adding a parent table for this purpose.
Shannon Severance
select AppId from child where AppId = 123 for update; it is not lock the whole table, right?
Ding
I think this answer is wrong.
Shannon Severance
@Ding: `select AppId from child where AppId = 123 for update` will not lock the table, just those rows that match the where clause.
Shannon Severance
@Shannon, if this is the case, use parent table, lock 1 record, use child table, lock multi-records. then what is the benefits using a parent table? thanks!
Ding
@Ding: Parent table is not needed in this case. Neither is the `select for update`. Please see my other answer. I am going to delete this answer shortly because it is wrong.
Shannon Severance
Rob van Wijk
A: 

"the 3rd and 4th one both are active which is wrong."

A simple unique index can prevent that at the database level.

create table rec (id number primary key, app_id number, status varchar2(1));
create unique index rec_uk_ix on rec (app_id, case when status = 'N' then id end);
insert into rec values (1,123,'N');
insert into rec values (2,123,'N');
insert into rec values (3,123,'N');
insert into rec values (4,123,'Y');
insert into rec values (5,123,'Y');

The unique index ensures that there can only by one record for any app with a status other than 'N'.

Obviously the application has to catch the error and know what to do with it (re-try or inform the user the data has changed).

Gary
i dont think this is right. all the N records would also fail the unique index...
Randy
Nope, because the 'N' records in the index don't have N, they have the id (primary key) which is always unique.In 11g you can use a virtual column, 10g you need the function based index.
Gary
+3  A: 

If you want to make sure you never can have more than one "active" records in the db for a given id, here's a cool (credit goes here): http://asktom.oracle.com/pls/apex/f?p=100:11:0::::P11_QUESTION_ID:1249800833250

It takes advantage of the fact that Oracle doesn't store completely NULL index entries and will guarantee that a particular id can't have more than one "active" record:

drop table test
/

create table test (a number(10), b varchar2(10))
/

CREATE UNIQUE INDEX unq ON test (CASE WHEN b = 'INACTIVE' then NULL ELSE a END)
/

these inserts work fine:

insert into test (a, b) values(1, 'INACTIVE');
insert into test (a, b) values(1, 'INACTIVE');
insert into test (a, b) values(1, 'INACTIVE');
insert into test (a, b) values(1, 'ACTIVE');
insert into test (a, b) values(2, 'INACTIVE');
insert into test (a, b) values(2, 'INACTIVE');
insert into test (a, b) values(2, 'INACTIVE');
insert into test (a, b) values(2, 'ACTIVE');

these inserts fail:

insert into test values(1, 'ACTIVE');

ORA-00001: unique constraint (SAMPLE.UNQ) violated

insert into test values(2, 'ACTIVE');

ORA-00001: unique constraint (SAMPLE.UNQ) violated

Patrick Marchand
+1 for making it absolutely impossible to have two active records
Rob van Wijk
+1  A: 

Yesterday I created a test case to reproduce the problem described. Today I found that the test case was flawed. I did not understand the problem, as such, I believe the answer I gave yesterday is is incorrect.

There are two possible issues:

  1. There is a commit happening between the update and insert.

  2. This is only an issue for new AppIds.

Test case:

Create the test table and insert two rows:

session 1 > create table test (TestId number primary key
  2             , AppId number not null
  3             , Status varchar2(8) not null 
  4                 check (Status in ('inactive', 'active'))
  5  );

Table created.

session 1 > insert into test values (1, 123, 'inactive');

1 row created.

session 1 > insert into test values (2, 123, 'active');

1 row created.

session 1 > commit;

Commit complete.

Begin first transaction:

session 1 > update test set status = 'inactive'
  2         where AppId = 123 and status = 'active';

1 row updated.

session 1 > insert into test values (3, 123, 'active');

1 row created.

Begin second transaction:

session 2 > update test set status = 'inactive'
  2         where AppId = 123 and status = 'active';

Now session 2 is blocked, waiting to get a row lock on row 2. Session 2 can not proceed until the transaction in session 1 either commits or rollsback. Commit session 1:

session 1 > commit;

Commit complete.

Now session 2 is unblocked and we see:

1 row updated.

When session 2 was unblocked, the update statement restarted, saw the changes in session 1, and updated row 3.

session 2 > select * from test;

    TESTID      APPID STATUS
---------- ---------- --------
         1        123 inactive
         2        123 inactive
         3        123 inactive

Complete the transaction in session 2:

session 2 > insert into test values (4, 123, 'active');

1 row created.

session 2 > commit;

Commit complete.

Check the results (using session 1):

session 1 > select * from test;

    TESTID      APPID STATUS
---------- ---------- --------
         1        123 inactive
         2        123 inactive
         3        123 inactive
         4        123 active

The only way for the two updates to not block each other is for there to be a commit or rollback between one and the other. There may be an implicit commit hidden somewhere in the software stack you are using. I don't know enough about .NET to advise on tracking that down.

However, the same problem will happen if AppId is brand new to the table. Test using a new AppId of 456:

session 1 > update test set status = 'inactive'
  2         where AppId = 456 and status = 'active';

0 rows updated.

No locks are taken because no rows are written to.

session 1 > insert into test values (5, 456, 'active');

1 row created.

Start second transaction for the same new AppId:

session 2 > update test set status = 'inactive'
  2          where AppId = 456 and status = 'active';

0 rows updated.

Session 2 does not see row 5, so it will not attempt to aquire a lock on it. Continue session 2:

session 2 > insert into test values (6, 456, 'active');

1 row created.

session 2 > commit;

Commit complete.

Commit session 1 and view results:

session 1 > commit;

Commit complete.

session 1 > select * from test;

    TESTID      APPID STATUS
---------- ---------- --------
         1        123 inactive
         2        123 inactive
         3        123 inactive
         4        123 active
         5        456 active
         6        456 active

6 rows selected.

To fix, use the function based index from Patrick Marchand (http://stackoverflow.com/questions/3497065/oracle-transaction-isolation/3498372#3498372):

session 1 > delete from test where AppId = 456;

2 rows deleted.

session 1 > create unique index test_u
  2         on test (case when status = 'active' then AppId else null end);

Index created.

Start first transaction of new AppId:

session 1 > update test set status = 'inactive'
  2         where AppId = 789 and status = 'active';

0 rows updated.

session 1 > insert into test values (7, 789, 'active');

1 row created.

Again session 1 does not take any locks with the update. There is a write lock on row 7. Start the second transaction:

session 2 > update test set status = 'inactive'
  2         where AppId = 789 and status = 'active';

0 rows updated.

session 2 > insert into test values (8, 789, 'active');

Again, session 2 does not see row 7, so it does not attempt to take a lock on it. BUT the insert is trying to write to the same slot on the function based index, and blocks on the write lock held by session 1. Session 2 will now wait for session 1 to commit or rollback:

session 1 > commit;

Commit complete.

And is session 2 we see:

insert into test values (8, 789, 'active')
*
ERROR at line 1:
ORA-00001: unique constraint (SCOTT.TEST_U) violated

At which point your client could retry the entire transaction. (Both the update and the insert.)

Shannon Severance
Hi Shannon, I have tested select for update in my machine, and it worked. Would you mind explain why your previous answer was wrong? thanks for the detailed reply!
Ding
@Ding: Because the `update` statement should cause the sessions to block and serialize on AppId, unless AppId is new to the table (or AppId has no `'active'` rows. If it is new, then `select for update` won't find any rows to lock on. (Without a parent table, with that AppId.) So it doesn't add anything. Essentially, I can't reproduce your senario as describe.
Shannon Severance