It can be frustrating at times to examine an execution plan for a given query against one or more large tables only to see this dreaded step:
MERGE JOIN CARTESIAN
Oracle ACE Director Jonathan Lewis posted such a problem on his blog referring to a customer of his experiencing that exact problem. As a solution Jonathan thought that finding the distinct combinations of fields the query was looking for would reduce the data set considerably providing a different and more efficient path to the data. The problem resulted from the matching criteria between the two tables; all eight attributes are mandatory in the large table, the attribute columns in the types table can be null but they both can’t all be null in the same record. Also, since the attribute columns aren’t indexed this will result in a full table scan for both tables. Oracle, in this case, chose a MERGE JOIN CARTESIAN to generate the result set, which needed to be fixed. A slight re-work of the example of the fix he posted is shown below, along with the plan it generated (run on Oracle 11.2.0.4):
SQL> with alldat as (
2 select
3 /*+ materialize */
4 my_id, my_val, my_attr1, my_attr2, my_attr3, my_attr4, my_attr5, my_attr6, my_attr7, my_attr8
5 from
6 largetab
7 ),
8 unq_data as (
9 select
10 /*+ materialize */
11 distinct my_attr1, my_attr2, my_attr3, my_attr4, my_attr5, my_attr6, my_attr7, my_attr8
12 from alldat
13 )
14 select
15 ad.my_id, ad.my_val, tt.my_cat, tt.my_rel_val
16 from
17 unq_data ud,
18 ttpe_tab tt,
19 alldat ad
20 where
21 nvl(tt.my_attr1(+), ud.my_attr1) = ud.my_attr1
22 and nvl(tt.my_attr2(+), ud.my_attr2) = ud.my_attr2
23 and nvl(tt.my_attr3(+), ud.my_attr3) = ud.my_attr3
24 and nvl(tt.my_attr4(+), ud.my_attr4) = ud.my_attr4
25 and nvl(tt.my_attr5(+), ud.my_attr5) = ud.my_attr5
26 and nvl(tt.my_attr6(+), ud.my_attr6) = ud.my_attr6
27 and nvl(tt.my_attr7(+), ud.my_attr7) = ud.my_attr7
28 and nvl(tt.my_attr8(+), ud.my_attr8) = ud.my_attr8
29 and ad.my_attr1 = ud.my_attr1
30 and ad.my_attr2 = ud.my_attr2
31 and ad.my_attr3 = ud.my_attr3
32 and ad.my_attr4 = ud.my_attr4
33 and ad.my_attr5 = ud.my_attr5
34 and ad.my_attr6 = ud.my_attr6
35 and ad.my_attr7 = ud.my_attr7
36 and ad.my_attr8 = ud.my_attr8
37 ;
Execution Plan
----------------------------------------------------------
Plan hash value: 2183901041
---------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
---------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 8 | 2080 | 24 (5)| 00:00:01 |
| 1 | TEMP TABLE TRANSFORMATION | | | | | |
| 2 | LOAD AS SELECT | SYS_TEMP_0FD9D6676_627FFB3 | | | | |
| 3 | TABLE ACCESS FULL | BIG_TABLE | 8 | 288 | 5 (0)| 00:00:01 |
| 4 | LOAD AS SELECT | SYS_TEMP_0FD9D6677_627FFB3 | | | | |
| 5 | HASH UNIQUE | | 8 | 192 | 3 (34)| 00:00:01 |
| 6 | VIEW | | 8 | 192 | 2 (0)| 00:00:01 |
| 7 | TABLE ACCESS FULL | SYS_TEMP_0FD9D6676_627FFB3 | 8 | 288 | 2 (0)| 00:00:01 |
|* 8 | HASH JOIN | | 8 | 2080 | 16 (0)| 00:00:01 |
| 9 | NESTED LOOPS OUTER | | 8 | 1008 | 14 (0)| 00:00:01 |
| 10 | VIEW | | 8 | 832 | 2 (0)| 00:00:01 |
| 11 | TABLE ACCESS FULL | SYS_TEMP_0FD9D6677_627FFB3 | 8 | 192 | 2 (0)| 00:00:01 |
|* 12 | TABLE ACCESS FULL | TYPES | 1 | 22 | 2 (0)| 00:00:01 |
| 13 | VIEW | | 8 | 1072 | 2 (0)| 00:00:01 |
| 14 | TABLE ACCESS FULL | SYS_TEMP_0FD9D6676_627FFB3 | 8 | 288 | 2 (0)| 00:00:01 |
---------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
8 - access("MD"."ATT1"="DD"."ATT1" AND "MD"."ATT2"="DD"."ATT2" AND "MD"."ATT3"="DD"."ATT3"
AND "MD"."ATT4"="DD"."ATT4" AND "MD"."ATT5"="DD"."ATT5" AND "MD"."ATT6"="DD"."ATT6" AND
"MD"."ATT7"="DD"."ATT7" AND "MD"."ATT8"="DD"."ATT8")
12 - filter("DD"."ATT1"=NVL("TY"."ATT1"(+),"DD"."ATT1") AND
"DD"."ATT2"=NVL("TY"."ATT2"(+),"DD"."ATT2") AND "DD"."ATT3"=NVL("TY"."ATT3"(+),"DD"."ATT3") AND
"DD"."ATT4"=NVL("TY"."ATT4"(+),"DD"."ATT4") AND "DD"."ATT5"=NVL("TY"."ATT5"(+),"DD"."ATT5") AND
"DD"."ATT6"=NVL("TY"."ATT6"(+),"DD"."ATT6") AND "DD"."ATT7"=NVL("TY"."ATT7"(+),"DD"."ATT7") AND
"DD"."ATT8"=NVL("TY"."ATT8"(+),"DD"."ATT8"))
Notice that the query rewrite did eliminate the MERGE JOIN CARTESIAN and replaced it with a nested loops hash join of the smaller table and the reduced result set of distinct attribute values. Since there were eight attribute values to match having a list of distinct combinations made Oracle’s job much easier.
This works because the number of distinct combinations of the searched columns was small compared to the entire set of table data; if the number of distinct combinations was small relative to the entire data set; if the distinct combinations approach the size of the total rows in the source table it’s likely this type of re-write won’t provide a better plan. But, when faced with a MERGE JOIN CARTESIAN and large data sets such a technique could ‘save the day’ by reducing the work Oracle must do to return the desired results.
