Oracle changes the default parameters of its DBMS packages from release to release. I remember back at the times of Oracle 9i a complete refresh would truncate the materialized view, thus the only work that the database was actually doing in a complete refresh, was just an INSERT after the TRUNCATE.
Now in Oracle 10g and Oracle 11g parameters have changed. When there is a COMPLETE materialized view refresh, for the purposes of data preservation, a DELETE is done instead of a TRUNCATE!
The reason for this is because Oracle "changed" the default parameter value of ATOMIC_REFRESH in the DBMS_MVIEW.REFRESH package. In earlier releases the parameter was set to FALSE by default but now it is set to TRUE, which forces a DELETE of the materialized view instead of TRUNCATE, making the materialized view more "available" at refresh time.
But this DELETE is an expensive operation in terms of refresh time it takes. A DELETE is always expensive and sometimes even impossible when we are talking about the complete refresh of materialized views with millions of rows. At COMPLETE refresh a DELETE command takes long time, as it has to make the materialized view available to the users and comply with the ACID theory and its transactional reasons, but who cares? What if we really want to get the refresh done quickly and we don't care about the availability of the materialized view, say in a data warehouse maintenance window?
I have tested how the refresh time improves when you set the ATOMIC_REFRESH => FALSE in the procedure on a 1 CPU home computer with Oracle 11g. If you have more CPUs it will even be quicker because of the PARALLEL parameter.
I just wanted to show the improvement you get in refresh times when ATOMIC_REFRESH is FALSE and Oracle does a TRUNCATE instead of a DELETE at complete refresh. A real time saver with big materialized views in data warehouses. Here is the test:
-- Enable parallel DML
ALTER SESSION FORCE PARALLEL DML
-- Cleanup
-- DROP TABLE BASE_TABLE;
-- Create the Test table
CREATE TABLE BASE_TABLE
(
id NUMBER(10) PRIMARY KEY NOT NULL,
X VARCHAR2(255),
Y VARCHAR2(255)
)
PARTITION BY RANGE
(
ID
)
INTERVAL
(
100000
)
(
PARTITION P1 VALUES LESS THAN (100000),
PARTITION P2 VALUES LESS THAN (200000)
);
-- Make the test table paralllel
ALTER TABLE BASE_TABLE PARALLEL 5
Load the test table with data from a dummy cartesian join so that you get millions of rows. Do you like my loop? Loops only once. Well I had it there to add more millions if I wanted to :-) >
BEGIN
FOR I IN 1 .. 1
LOOP
INSERT /*+ PARALLEL(BASE_TABLE, 5) */ INTO BASE_TABLE SELECT ROWNUM, a.object_name, a.status FROM all_objects a, emp b, dept c;
END LOOP;
COMMIT;
END;
3010392 rows inserted
Create a materialized view log for fast refresh
DROP MATERIALIZED VIEW LOG ON BASE_TABLE;
CREATE MATERIALIZED VIEW LOG ON BASE_TABLE;
Create the fast refresh materialized view
DROP MATERIALIZED VIEW MV_BASE_TABLE;
CREATE materialized VIEW mv_base_table parallel 5 refresh fast
AS
SELECT * FROM BASE_TABLE;
Update the Test table to simulate changing data
UPDATE BASE_TABLE SET Y='INVALID';
COMMIT;
3010392 rows updated
Now you are ready to do the materialized view refresh with the ATOMIC_REFRESH values set to TRUE and then to FALSE. Observe the refresh times. Refresh the materialized view with the two different values in the
ATOMIC_REFRESH parameter.
TRUE case with DELETE
EXEC DBMS_MVIEW.REFRESH(LIST => 'MV_BASE_TABLE', METHOD => 'C', ATOMIC_REFRESH => TRUE);
Elapsed 558.8 seconds
Now is time to do the test with the ATOMIC_REFRESH parameter set to FALSE.
FALSE case with TRUNCATE
Must update again to simulate change in the logs
UPDATE BASE_TABLE SET Y='VALID';
COMMIT;
EXEC DBMS_MVIEW.REFRESH(LIST => 'MV_BASE_TABLE', METHOD => 'C', ATOMIC_REFRESH => FALSE);
Elapsed 215.3 seconds
Half the time! Well if you have real computers it will be much faster. So again we see DELETE is really not good when you work with millions of rows. Nice to have options!
6 comments:
Would you expect to see a TRUNCATE on PCT refresh of an MV's partition with atomic_refresh=>false ? currently I only get deletes whether using true or false.
rnm1978,
I wouldn't expect it to truncate a partition. Probably is the right thing to do as well, as partitions are part of a table and truncating is asking it to do something like "truncate emp where empno=1234;" which doesn't make sense, as truncate is a irrevocable operation on a "table". But you can "drop" partitions if you don't need them. PCT is designed to be efficient tracking 'changes' on partitions, I would think that it is doing the right thing in refreshing with PCT and deleting. Besides, it has to keep the other partitions "up" as it refreshes so the "DELETE" would be the transactional of choice.
Hi,
Just to follow up on my previous comment - I did some more research into this, and in essence partition truncation of the MView *can* happen but there are several conditions to meet first.
I have blogged about it here: Materialised Views – PCT Partition Truncation.
Very good, I also used PARALLELISM =>'5' on 2 CPU based DB for 11G for Oarcle.
good to know this.
However, how do I change this if I'm creating MV with NEXT sysdate + 24?
Good One. Really Useful to know.
Post a Comment