Pre-creating Interval Partitions

One of the major problems with interval-based partitioning is the statistics. Partitions appear dynamically as-needed based upon data being inserted or udpated, and any partition magically appearing isn’t going to have any statistics.

This is generally a stability issue as you will, at best, be using dynamic statistics for your optimizations. So how do we deal with it? My preferred method is to pre-create the partitions and copy statistics from a good partition into the pre-created partition. But how do we get the partitions to appear? Here’s 2 options:

  1. Insert data into the row with the partition key for the partition we wish to create, and rollback. This can be tricky, especially with tables containing many NOT NULL columns, and is subject to failure based upon table evolution.
  2. Lock the relevant partition in shared mode using the commandLOCK TABLE .x. PARTITION FOR <partition-key> IN SHARE MODE;

    This will place a shared lock on the non-existant partition, which Oracle will create so it can lock it. A much less messy solution, and not one I had thought of until shown the light by Dan Jankowski.

So does option 2 work? Here’s a quick example:

10:51:55 NEIL @ ORCL01 > CREATE TABLE interval_table (id NUMBER, created_date DATE)
10:51:55   2             PARTITION BY RANGE (created_date) INTERVAL (NUMTOYMINTERVAL(1,'MONTH'))
10:51:55   3           ( PARTITION part_01 values LESS THAN (TO_DATE('01-JAN-2015','DD-MON-YYYY')))
10:51:55   4  /
Table created.

10:51:55 NEIL @ ORCL01 > select table_name, partition_name,high_value from user_tab_partitions order by 1,2;
TABLE_NAME                     PARTITION_NAME                 HIGH_VALUE
------------------------------ ------------------------------ --------------------------------------------------------------------------------
INTERVAL_TABLE                 PART_01                        TO_DATE(' 2015-01-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA


use a shared lock to generate a new partition

10:51:55 NEIL @ ORCL01 > lock table interval_table partition for (to_date('01-JAN-2016','DD-MON-YYYY')) in share mode;
Table(s) Locked.

10:51:55 NEIL @ ORCL01 > select table_name, partition_name,high_value from user_tab_partitions order by 1,2;
TABLE_NAME                     PARTITION_NAME                 HIGH_VALUE
------------------------------ ------------------------------ --------------------------------------------------------------------------------
INTERVAL_TABLE                 PART_01                        TO_DATE(' 2015-01-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA
INTERVAL_TABLE                 SYS_P647                       TO_DATE(' 2016-02-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA

10:51:55 NEIL @ ORCL01 > lock table interval_table partition for (to_date('01-FEB-2016','DD-MON-YYYY')) in share mode;
Table(s) Locked.
10:51:55 NEIL @ ORCL01 > lock table interval_table partition for (to_date('01-MAR-2016','DD-MON-YYYY')) in share mode;
Table(s) Locked.
10:51:55 NEIL @ ORCL01 > lock table interval_table partition for (to_date('01-APR-2016','DD-MON-YYYY')) in share mode;
Table(s) Locked.
10:51:55 NEIL @ ORCL01 > select table_name, partition_name,high_value from user_tab_partitions order by 1,2;
TABLE_NAME                     PARTITION_NAME                 HIGH_VALUE
------------------------------ ------------------------------ --------------------------------------------------------------------------------
INTERVAL_TABLE                 PART_01                        TO_DATE(' 2015-01-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA
INTERVAL_TABLE                 SYS_P647                       TO_DATE(' 2016-02-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA
INTERVAL_TABLE                 SYS_P648                       TO_DATE(' 2016-03-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA
INTERVAL_TABLE                 SYS_P649                       TO_DATE(' 2016-04-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA
INTERVAL_TABLE                 SYS_P650                       TO_DATE(' 2016-05-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIA

10:51:55 NEIL @ ORCL01 > -- and release the locks... the rollback doesn't rollback the creation of the new partitions.
10:51:55 NEIL @ ORCL01 > rollback;
Rollback complete.

So now what? To get the stats right I follow the following rule set:

Firstly, lock the table stats when you create the table and have a dedicated, focused stats job. This stop the general stats job from getting in the way of this this important partitioned table.

  • Create a new partition
  • Copy stats using DBMS_STATS.COPY_TABLE_STATS from your fullest partition (with FORCE=TRUE to override the locked stats option). Always try to pretend you have more data than you really have if faking it – it’s safer as an optimized plan with larger expected data sets processing a smaller data set tends to work much better than the stats assuming a small data set and being surprised by lots of data. Consider using SCALE_FACTOR if you have a growing dataset. Don’t reply upon Optimizer magic, such as Adaptive Execution Plans to dig out of a hole.
  • Once a partition becomes “old” (like a no-longer used date-range partition), gather actual stats to get all your stats for that partition accurate. That may even become your new baseline stats copy partition. You will possibly never need to gather stats again for that partition.
  • Be careful if you are copying stats when you have an object-level change. For example, if you put a new index on, you need to re-copy stats post change to any pre-created partitions. Empty pre-created partitions will have accurate (i.e. empty) stats and that’s really not what you want.
  • Repeat as often as you pre-create a new partition

SQL Plan Management – 12C dumb feature

In Oracle 11G, Oracle introduced SQL Plan Management (SPM). It is excellent (I love it to bits). It allows you to create Baselines against SQL which lock-down the SQL execution plan. No more plan flips. More consistency. Perfect**.

Whenever some Baselined SQL is ran, Oracle still parses it and compares the parsed output to the accepted (Evolved) baselines. If the newly parsed plan is better, a new baseline is added to DBA_SQL_PLAN_BASELINES but is NOT accepted. This means that you need to spend time manually accepting the baseline; running the command DBMS_SPM.EVOLVE_SQL_BASELINE plan and checking the new plan.

If you want it, and/or Oracle evaluates that is it a better plan for that particular set of bind variables, the plan is accepted and becomes a candidate to be used by future execution of your SQL. Complete control over your execution plans.

So, Oracle, what’s all this about in Oracle 12C, eh?

In Oracle 12C there’s a new SPM Evolve advisor task. “By default, SYS_AUTO_SPM_EVOLVE_TASK runs daily in the scheduled maintenance window” – So, it runs every night and by default it runs DBMS_SPM.EVOLVE_SQL_BASELINE for all new baselines created today and automatically accepts the new plans.

BY DEFAULT? NO! NO! NO!

That is precisely what I don’t want from baselines – Oracle making it’s own mind up about plans without any input from me. I’m using baselines to stop Oracle changing its mind. To explicitly limit the number of paths allowed by the Optimizer to ones I know about and with which I am comfortable. Don’t introduce functionality to do the opposite.

So, immediately following the installation of 12C, I would recommend running (you need to be SYS for this):

SELECT PARAMETER_NAME, PARAMETER_VALUE AS "VALUE"
FROM   DBA_ADVISOR_PARAMETERS
WHERE  TASK_NAME = 'SYS_AUTO_SPM_EVOLVE_TASK' AND
         PARAMETER_NAME in ('ACCEPT_PLANS','TIME_LIMIT')
ORDER BY 1;

PARAMETER_NAME            VALUE
------------------------- ----------
ACCEPT_PLANS              TRUE
TIME_LIMIT                3600

Then run:

BEGIN
  DBMS_SPM.SET_EVOLVE_TASK_PARAMETER('SYS_AUTO_SPM_EVOLVE_TASK',
    'ACCEPT_PLANS', 'false');
END;
/

OK, back where we were, with any baselines fixed in place and doing what I want them to do! Not change.

 

**Perfect? No. But Baselines are good and, as long as your DB structure does not change, they should keep working. If they don’t, raise an SR with Oracle as it’s probably a bug.

UPDATE 2015-11-25: This is still as true in 12.1.0.2.5 as it was in 12.0 Grrrr!!!

%d bloggers like this: