In SQL, how to get MERGE to update relevant row(s) with a single row from a grouped set of returned results - sql

Question

I am using MERGE (Oracle) to do updates to records that match on criteria specified in the ON clause joined on a virtual table create by a subquery. Form of the statement is:
MERGE INTO table1 t1 USING SELECT (t2.f21, MAX(t2.f22), t3.f31, t3.f32
from
table2 t2, table3 t3
where
{... various join/filter criteria ...}
group by t2.f21, t3.f31, t3.f32) MATCHDATA
ON (t1.f11 = MATCHDATA.f21)
where t1.f12 = 'something';
Now the rub: MATCHDATA will return multiple rows because the "... criteria ..." will by nature return multiple groups of matching records. So the 'group by' along with the use of 'MAX()' is not buying me any guarantees; on the contrary, if I added:
where rownum = 1
after MATCHDATA after wrapping the MATCHDATA result in a another SELECT statement that simply repeated the returned field names, I would then be limiting myself to being able to update only the one record in the one group of records that needs updating that has the highest value as determined by MAX(). Instead, I need to have the records in table1 that match on the join field in each MAX() record for their group of records updated. I started on Fri. down the PARTITION BY path and am new to that one, so didn't make much headway. But it looks promising and maybe tomorrow will yield better results. As it is, when I try to use it without for example limiting the returned recordset in MATCHDATA to one record via use of "rownum = 1", I get that familiar "could not return a stable set of records" error that MERGE proponents (like myself) must smile sheepishly at when their colleagues come to them for advice on this "better-than-correlated-subqueries"-evangelized nouveau SQL command as they face this same error.
As you can see, I am treating MERGE as the more successful brother of the correlated subquery. But is this a case where I should be looking back to the lesser of two weevils (i.e., use a correlated subquery instead) to get the job done? Or is the fix to be found in PARTITION BY or another modification to the above?
Thanks to all who take the time to offer their advice, I appreciate it.

Answer 1

I get that familiar "could not return a stable set of records" error
Because the join key you have used in the ON clause is not enough to make the row unique to perform the WHEN MATCHED THEN UPDATE statement.
You must include more keys in the ON clause until the matched rows are unique and thus returning a stable set of records.
Let's see a test case:
Set up
SQL> CREATE TABLE source_table (
2 col1 NUMBER,
3 col2 VARCHAR2(10),
4 col3 VARCHAR2(10)
5 );
Table created.
SQL>
SQL> INSERT INTO source_table (col1, col2, col3) VALUES (1, 'a', 'p');
1 row created.
SQL> INSERT INTO source_table (col1, col2, col3) VALUES (1, 'b', 'q');
1 row created.
SQL> INSERT INTO source_table (col1, col2, col3) VALUES (2, 'c', 'r');
1 row created.
SQL> INSERT INTO source_table (col1, col2, col3) VALUES (3, 'c', 's');
1 row created.
SQL>
SQL> COMMIT;
Commit complete.
SQL>
SQL> CREATE TABLE target_table (
2 col1 NUMBER,
3 col2 VARCHAR2(10),
4 col3 VARCHAR2(10)
5 );
Table created.
SQL>
SQL> INSERT INTO target_table (col1, col2, col3) VALUES (1, 'b', 'p');
1 row created.
SQL> INSERT INTO target_table (col1, col2, col3) VALUES (3, 'd', 'q');
1 row created.
SQL>
SQL> COMMIT;
Commit complete.
SQL>
SQL> SELECT * FROM source_table;
COL1 COL2 COL3
---------- ---------- ----------
1 a p
1 b q
2 c r
3 c s
SQL> SELECT * FROM target_table;
COL1 COL2 COL3
---------- ---------- ----------
1 b p
3 d q
SQL>
Error reproduce
SQL> MERGE INTO target_table trg
2 USING source_table src
3 ON (trg.col1 = src.col1) -- Not Unique
4 WHEN MATCHED THEN UPDATE SET
5 trg.col2 = src.col2,
6 trg.col3 = src.col3
7 WHEN NOT MATCHED THEN INSERT
8 (
9 col1,
10 col2,
11 col3
12 )
13 VALUES
14 (
15 src.col1,
16 src.col2,
17 src.col3
18 );
USING source_table src
*
ERROR at line 2:
ORA-30926: unable to get a stable set of rows in the source tables
SQL>
So, as expected we get the error ORA-30926: unable to get a stable set of rows in the source tables
Let's make the ON clause unique.
SQL> MERGE INTO target_table trg
2 USING source_table src
3 ON (trg.col1 = src.col1
4 AND
5 trg.col2 = src.col2) -- Unique
6 WHEN MATCHED THEN UPDATE SET
7 trg.col3 = src.col3
8 WHEN NOT MATCHED THEN INSERT
9 (
10 col1,
11 col2,
12 col3
13 )
14 VALUES
15 (
16 src.col1,
17 src.col2,
18 src.col3
19 );
4 rows merged.
SQL> SELECT * FROM target_table;
COL1 COL2 COL3
---------- ---------- ----------
1 b q
3 d q
2 c r
3 c s
1 a p
SQL>
Problem solved!
Remember, you cannot update the columns which are referenced in the ON clause.

Answer 2

Let's say we have this table T2:
C1 C2 AMOUNT UF
-- -- ---------- ----------
A X 12 101
A Y 3 102
A Y 12 103
B X 7 104
B Y 9 105
I need to have the records in table1 that match on the join field in
each MAX() record for their group of records updated. I started on
Fri. down the PARTITION BY path and am new to that one, so didnt make
much headway.
This is good path and you can do this using function rank():
select * from (
select t2.*, rank() over (partition by c1 order by amount desc) rn from t2 )
where rn=1
C1 C2 AMOUNT UF RN
-- -- ---------- ---------- --
A X 12 101 1
A Y 12 103 1
B Y 9 105 1
But if your joining field for merge is only 'C1' then this set of records is not stable, because for C1='A'
we have two rows and Oracle looks sheepishly, it does not know which one interests you.
To resolve this you can use row_number()
instead of rank() - if it's all the same. But if this matters you need something more in order clause, for instance:
select * from (
select t2.*, rank() over (partition by c1 order by amount desc, c2) rn from t2 )
where rn = 1
C1 C2 AMOUNT UF RN
-- -- ---------- ---------- --
A X 12 101 1
B Y 9 105 1
This set of rows is stable, because for C1 there are no duplicates and you can use it in your merge.
merge into t1
using (
select * from (
select t2.*, rank() over (partition by c1 order by amount desc, c2) rn from t2 )
where rn=1) md
on (md.c1 = t1.c1)
when matched then update set t1.uf = md.uf
when not matched then insert (t1.c1, t1.uf)
values (md.c1, md.uf)