Creating large tables in Oracle with dynamic partitions to allow for instant reads

by MikeHogg29. September 2011 14:55

A lot of my work in Oracle was with tables between a million rows and a billion, and so indexing and query performance was always part of writing the system.  In this case I knew my table was going to grow a quarter to a half million rows a day, and so to retain immediate selects, I needed partitions.  While it is easy to build a table with the initial partitions, you have to maintain the scheme with a nightly or a weekly job.  Here is the initial table creation:

createtable SOME_CONTRACT
(
  ID                             NUMBER not null,
  CONTRACT_IDVARCHAR2(150) not null,
MONTH                          NUMBER not null,
  CUSTOMER_IDVARCHAR2(150),
  BUSUNIT_IDVARCHAR2(25),
  SEGMENT_IDVARCHAR2(25),
  PRODUCT_IDVARCHAR2(100),
  <snip some hundred columns>
  ZONE_MARKERVARCHAR2(50),
  ACTIVE_FG                      NUMBER,
  SNAPSHOT_DATEDATE,
  AUDIT_DATEDATE
)
tablespace SOME_TBS01
  pctfree 10
  initrans 1
  maxtrans 255
  storage
  (
    initial 64K
    minextents 1
    maxextents unlimited
  )
partition by range (snapshot_date)( 
partition P20110101 values less than (TO_DATE(' 2011-02-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110201 values less than (TO_DATE(' 2011-03-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110301 values less than (TO_DATE(' 2011-04-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110401 values less than (TO_DATE(' 2011-05-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110501 values less than (TO_DATE(' 2011-06-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110601 values less than (TO_DATE(' 2011-07-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110701 values less than (TO_DATE(' 2011-08-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110801 values less than (TO_DATE(' 2011-09-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110901 values less than (TO_DATE(' 2011-09-02 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN'))    tablespace CER_ABM_TBS01    pctfree 0    initrans 1    maxtrans 255    storage    (      initial 64K      minextents 1      maxextents unlimited    ),
partition P20110902 values less than (TO_DATE(' 2011-09-03 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110905 values less than (TO_DATE(' 2011-09-06 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110906 values less than (TO_DATE(' 2011-09-07 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110907 values less than (TO_DATE(' 2011-09-08 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110908 values less than (TO_DATE(' 2011-09-09 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110909 values less than (TO_DATE(' 2011-09-10 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110912 values less than (TO_DATE(' 2011-09-13 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110913 values less than (TO_DATE(' 2011-09-14 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110914 values less than (TO_DATE(' 2011-09-15 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition P20110915 values less than (TO_DATE(' 2011-09-16 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) ,
  partition PMAX values less than (TO_DATE(' 2099-12-31 00:00:00', 'SYYYY-MM-DD HH24:MI:SS', 'NLS_CALENDAR=GREGORIAN')) 
)
-- Create/Recreate primary, unique and foreign key constraints 
alter table SOME_CONTRACT
add constraint PK_SOMECONTRACTS_ID primary key (ID)
using index
  tablespace SOME_TBS01
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 64K
    minextents 1
    maxextents unlimited
  );
-- Create/Recreate indexes 
create unique index UK_SOME_PK_CONTR_ACTIVE on SOME_CONTRACT (DECODE(ACTIVE_FG,1,CONTRACT_ID||TO_CHAR(MONTH)))
  tablespace SOME_TBS01
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 64K
    minextents 1
    maxextents unlimited
  );

So new rows get inserted into PMAX each day in a daily process.  Partitions in Oracle are like separate tables, and in our case, most of our select queries will always be on snapshot_date, so it is like we are querying a single table.  We can also add indices if we want.  I think in my case here I was using an active flag and so I added a composite index with the PK and the column Active_fg.

To maintain this scheme, you have to periodically add new partitions, and split up PMAX.  I did that with a scheduled job, that ran a Stored Proc each night.

procedure Maintain_Partitions is
cursor c_nightly_deletes is
 select sdate
 from (select distinct snapshot_date sdate, count(distinct snapshot_date) over (partition by month) total
 from SOME_contract_nightly
 where month = to_char(  sysdate  ,'YYYYMM') 
 order by snapshot_date ) 
 where rownum < total - 3;-- keep three latest in any given month
 
cursor c_compress_ddl (p_table_name varchar2 ) is
select 'alter table ' || p_table_name || ' move partition ' || partition_name || ' compress' ddl, partition_name
from user_tab_partitions
where table_name = p_table_name
 and partition_name not like '%MAX%'
 and (compression <> 'ENABLED'
 or -- recompress NIGHTLYs since those older than 3 days are being deleted
                 table_name = 'SOME_CONTRACT_NIGHTLY' and partition_name like 'P' || to_char(sysdate,'YYYYMM') ||'%'
                 );
 
cursor c_truncate_ddl ( p_table varchar2  ) is
select 'alter table ' || table_name || ' truncate partition ' || partition_name || ' update global indexes' as ddl, partition_name
from user_tab_partitions
where table_name = p_table
 and partition_name not like '%MAX%'
 and to_date( trim( substr(partition_name, 2) ), 'YYYYMMDD' )   <= (add_months(sysdate,-24));
 
begin
for r_nightly_deletes in c_nightly_deletes loop
delete from SOME_contract_nightly
where snapshot_date = r_nightly_deletes.sdate;
      Log_This('PKG_SOME_UI.Maintain_Partitions','SOME_CONTRACT_NIGHTLY snapshot_date' ||
        to_char(r_nightly_deletes.sdate,'DD-MON-YY') || ' deleted: ' || sql%rowcount);
end loop;
for each_table in (select distinct table_name from user_tab_partitions where table_name like 'SOME%') loop
-- create new partitions
      SplitWeekdayPartitions( each_table.TABLE_NAME);
-- compress (today's) data
for r_compress_ddl in c_compress_ddl ( each_table.table_name ) loop
execute immediate r_compress_ddl.ddl;
        Log_This('PKG_SOME_UI.Maintain_Partitions',each_table.table_name || ' partition compressed: ' || r_compress_ddl.partition_name);
end loop;
-- remove any partitions older than two years
for r_truncate_ddl in c_truncate_ddl ( each_table.table_name  ) loop
 execute immediate r_truncate_ddl.ddl;
          Log_This('PKG_SOME_UI.Maintain_Partitions', each_table.table_name || ' Partition Truncated: ' || r_truncate_ddl.partition_name);
end loop;
end loop;
    Rebuild_Indices();
-- GATHER STATISTICS FOR NEW PARTITIONS
for x in (  select t.owner, t.table_name, p.partition_name
 from all_tables t
 join all_tab_partitions p
 on t.table_name = p.table_name
where p.table_name like 'SOME%'
 and (p.last_analyzed is null
 or
                  p.table_name = 'SOME_CONTRACT_NIGHTLY'
 and p.partition_name like 'P' || to_char(sysdate,'YYYYMM') ||'%' ) )
    loop
              dbms_stats.gather_table_stats(
                      degree=> dbms_stats.AUTO_DEGREE,
                      ownname => x.owner,
                      tabname => x.table_name,
                      partname => x.partition_name,
 cascade => dbms_stats.AUTO_CASCADE );
              Log_This('PKG_SOME_UI.Maintain_Partitions','Statistics gathered for table.partition: ' || 
                                    x.table_name || '.' || x.partition_name);
end loop;
exception when others then
      Rebuild_Indices(); -- do not leave in unuseable state
      raise; -- notify user to investigate and rerun
end Maintain_Partitions;

 

In this particular case, I was deleting old partitions, and using compression to keep the total database size down.   The SplitWeekdayPartitions created the new ones:

--------- PARTITIONING
PROCEDURE SplitWeekdayPartitions( p_table_name in varchar2 ) AS
  v_latest_date date;
  v_new_date date;
  v_part_name varchar2(8);
  v_upper_bound varchar2(8);
BEGIN
SELECT to_date(substr(MAX(partition_name) KEEP (DENSE_RANK FIRST ORDER BY partition_position desc ), 2),'YYYYMMDD')
into v_latest_date
FROM (SELECT partition_name, partition_position
 FROM user_tab_partitions
 WHERE table_name = upper(p_table_name) 
 AND partition_name <> 'PMAX' );
 
if v_latest_date < sysdate
then
            v_new_date := v_latest_date + 1;
 while to_char(v_new_date, 'd') in (1,7) loop -- skip weekend dates
                v_new_date := v_new_date + 1;
 end loop;
            v_part_name := to_char(v_new_date,'YYYYMMDD');
            v_upper_bound := to_char(v_new_date + 1,'YYYYMMDD');
 execute immediate
            'ALTER TABLE ' || p_table_name || ' SPLIT PARTITION PMAX AT (to_date(' || v_upper_bound ||
                ',''YYYYMMDD'')) INTO ' || '( PARTITION P' || v_part_name || 
                ', PARTITION PMAX ) UPDATE GLOBAL INDEXES';
 
 
            Log_This('PKG_SOME_UI.SplitWeekdayPartitions', 'New partition P' || v_part_name || ' created for ' || p_table_name);
 
            SplitWeekdayPartitions( p_table_name );-- do we need to catch up to today?
end if;
END SplitWeekdayPartitions;

It’s interesting that I picked a recursive call here instead of a while loop.

… and in the MaintainPartitions proc, RebuildIndices was necessary afterwards because the alter table commands and truncating didn’t trigger index updates.

procedure Rebuild_Indices is
begin
-- LOCAL INDICES   (not using any now but maybe in the future...)
for x in ( SELECT UIP.INDEX_NAME, UIP.PARTITION_NAME, UIP.COMPRESSION
 FROM USER_IND_PARTITIONS UIP
 WHERE UIP.STATUS <> 'USABLE' )
    loop
 execute immediate
                          'ALTER INDEX ' || x.INDEX_NAME || ' REBUILD PARTITION ' || x.PARTITION_NAME ;
            Log_This('PKG_SOME_UI.Maintain_Partitions','Partition Index Rebuilt: ' || x.index_name);
end loop;
-- GLOBAL INDICES
for x in ( SELECT INDEX_NAME
 FROM USER_INDEXES
 WHERE table_name like 'SOME%' and STATUS not in ( 'VALID', 'N/A' ) )
    loop
 execute immediate
                        'ALTER INDEX ' || x.INDEX_NAME || ' REBUILD ' ;
             Log_This('PKG_SOME_UI.Maintain_Partitions','Global Index Rebuilt: ' || x.index_name);
end loop;
end;
and that’s it.  Neat feature to plug into your next large table.

About Mike Hogg

Mike Hogg is a c# developer in Brooklyn.

More Here

Favorite Books

This book had the most influence on my coding style. It drastically changed the way I write code and turned me on to test driven development even if I don't always use it. It made me write clearer, functional-style code using more principles such as DRY, encapsulation, single responsibility, and more.amazon.com

This book opened my eyes to a methodical and systematic approach to upgrading legacy codebases step by step. Incrementally transforming code blocks into testable code before making improvements. amazon.com

More Here