It happened to me several times to being asked about the mysterious full table scan in CONNECT BY operations. In this post I would like to share with you some of the information I wrote about it in my book (pages 233 to 236) .
The operation CONNECT BY WITH FILTERING is used to process hierarchical queries. It is characterized by two child operations. The first one is used to get the root of the hierarchy, and the second one is executed once for each level in the hierarchy.
Here is a sample query and its execution plan. Note that the execution plan was generated on Oracle Database 11g (the reason will be explained later).
SELECT level, rpad('-',level-1,'-')||ename AS ename, prior ename AS manager FROM emp START WITH mgr IS NULL CONNECT BY PRIOR empno = mgr --------------------------------------------------------------------- | Id | Operation | Name | Starts | A-Rows | --------------------------------------------------------------------- |* 1 | CONNECT BY WITH FILTERING | | 1 | 14 | |* 2 | TABLE ACCESS FULL | EMP | 1 | 1 | | 3 | NESTED LOOPS | | 4 | 13 | | 4 | CONNECT BY PUMP | | 4 | 14 | | 5 | TABLE ACCESS BY INDEX ROWID| EMP | 14 | 13 | |* 6 | INDEX RANGE SCAN | EMP_MGR_I | 14 | 13 | --------------------------------------------------------------------- 1 - access("MGR"=PRIOR "EMPNO") 2 - filter("MGR" IS NULL) 6 - access("MGR"=PRIOR "EMPNO")
To help you understand the execution plan with a hierarchical query more easily, it is useful to look at the data returned by the query as well:
LEVEL ENAME MANAGER ---------- ---------- ---------- 1 KING 2 -JONES KING 3 --SCOTT JONES 4 ---ADAMS SCOTT 3 --FORD JONES 4 ---SMITH FORD 2 -BLAKE KING 3 --ALLEN BLAKE 3 --WARD BLAKE 3 --MARTIN BLAKE 3 --TURNER BLAKE 3 --JAMES BLAKE 2 -CLARK KING 3 --MILLER CLARK
The execution plan carries out the operations as follows:
- Operation 1 has two children (2 and 3), and operation 2 is the first of them in ascending order. Therefore, the execution starts with operation 2.
- Operation 2 scans the table EMP, applies the filter predicate “MGR” IS NULL, and returns the root of the hierarchy to its parent operation (1).
- Operation 3 is the second child of operation 1. It is therefore executed for each level of the hierarchy—in this case, four times. The first child, operation 4, is executed, and for each row it returns, the inner loop (composed of operation 5 and its child operation 6) is executed once. Notice, as expected, the match between the column A-Rows of operation 4 with the column Starts of operations 5 and 6.
- For the first execution, operation 4 gets the root of the hierarchy through the operation CONNECT BY PUMP. In this case, there is a single row (KING) at level 1. With the value in the column mgr, operation 6 does a scan of the index EMP_MGR_I by applying the access predicate “MGR”=PRIOR “EMPNO”, extracts the rowids, and returns them to its parent operation (5). Operation 5 accesses the table EMP with the rowids and returns the rows to its parent operation (3).
- For the second execution of operation 4, everything works the same as for the first execution. The only difference is that the data from level 2 (JONES, BLAKE, and CLARK) is passed to operation 4 for the processing.
- For the third execution of operation 4, everything works like in the first one. The only difference is that level 3 data (SCOTT, FORD, ALLEN, WARD, MARTIN, TURNER, JAMES, and MILLER) is passed to operation 4 for the processing.
- For the fourth and last execution of operation 4, everything works like in the first one. The only difference is that level 4 data (ADAMS and SMITH) is passed to operation 4 for the processing.
- Operation 3 gets the rows passed from its children and returns them to its parent operation (1).
- Operation 1 applies the access predicate “MGR”=PRIOR “EMPNO” and sends the 14 rows to the caller.
The execution plan generated on Oracle Database 10g is slightly different. As can be seen, the operation CONNECT BY WITH FILTERING has a third child (operation 8). In this case, it was not executed, however. The value in the column Starts for operation 8 confirms this. Actually, the third child is executed only when the CONNECT BY operation uses temporary space. When that happens, performance might degrade considerably. This problem, which is fixed as of version 10.2.0.4, is known as bug 5065418.
--------------------------------------------------------------------- | Id | Operation | Name | Starts | A-Rows | --------------------------------------------------------------------- |* 1 | CONNECT BY WITH FILTERING | | 1 | 14 | |* 2 | TABLE ACCESS FULL | EMP | 1 | 1 | | 3 | NESTED LOOPS | | 4 | 13 | | 4 | BUFFER SORT | | 4 | 14 | | 5 | CONNECT BY PUMP | | 4 | 14 | | 6 | TABLE ACCESS BY INDEX ROWID| EMP | 14 | 13 | |* 7 | INDEX RANGE SCAN | EMP_MGR_I | 14 | 13 | | 8 | TABLE ACCESS FULL | EMP | 0 | 0 | ---------------------------------------------------------------------
Christian, thanks for this helpful page. I was at my wit’s end trying to speed up an 86 line
connect-by query that takes 3 hour to finish on a 10.2.0.1. THANKS for mentioning that it is a bug.
Thank you for reporting this. I very likely hit this bug and verified it has been fixed in patch set 3.
Bye :-)
[…] an interesting post on this topic on Christian Antognini’s blog (which I’ve also referenced from the OTN posting) […]
I’ve been always wondering what is that third child operation which always shows “table access full” and never gets executed. Finally I got the answer!
Thank you for this insightfull and very well written post!
[…]9. Operation 1 applies the access predicate “MGR”=PRIOR “EMPNO” and sends the 14 rows to the caller.[…]
Why do we need the access predicate in operation 1 ? I think the whole algorithm ensures that only “valid” rows get reported to operation 1. Why is it needed to apply the predicate once more before sending them to the caller?
Hi Todor
Sorry, but I don’t know why this double check is needed. My best guess is the following: since it is an access predicate (not a filter predicate), this probably means that data is placed in a memory structure (kind-of hash table) and, to access it, the key is MGR.
Cheers,
Chris
[…] of data. So please test for performance – with a data volume comparable to the real one. Here is a very good detailed explanation by Christian Antognini of how CONNECT_BY is executed internally […]
Hi Chris,
Thank you for your good article.But I have one question about ‘For the second execution of operation 4, everything works the same as for the first execution. The only difference is that the data from level 2 …is passed to operation 4 for the processing. ‘
My question is which operation passes data from level 2 to operation 4?
Can I understand like this,the first execution has generated the level 2 data from level 1 and CONNECT BY PUMP gets these data and then launches the subseque loop?
Best regards,
Leon
Hi Leon
> My question is which operation passes data from level 2 to operation 4?
Operation 1, CONNECT BY WITH FILTERING
> Can I understand like this,the first execution has generated the level 2 data from
> level 1 and CONNECT BY PUMP gets these data and then launches the subseque loop?
There are 4 executions:
– The first one extracts the data for level 1 and, therefore, provides the data to access level 2.
– The second execution extract the data for level 2 and provides the data to access level 3
– …
In general: execution “n” extracts data for level “n” and provides the data to access level “n+1”.
HTH
Chris
Hi Chris,
Thank you for your reply.
But since ‘CONNECT BY WITH FILTERING’ just starts once, how can this operation identify all the levels of data?
And it seems that your answer aslo implies that operation for extracting data will start 4 times. That’s why I asked whether it’s operation 4 passes n level data to level n+1 instead of operation 1.
Thank you very much.
Leon
Hi Leon
> But since ‘CONNECT BY WITH FILTERING’ just starts once, how can this operation identify all the levels of data?
The execution is not atomic. During the execution this operation does several things… like getting the data of level “n” and starting the search of level “n+1”.
You can imagine that there is a kind of pipe between operation 1 (CONNECT BY WITH FILTERING) and operation 4 (CONNECT BY PUMP). So, every time operation 1 gets the data for level “n” it sends into it the ID to find the data for level “n+1”.
> And it seems that your answer aslo implies that operation for extracting data will start 4 times.
Also the Starts column of the dbms_xplan output confirms that the extraction of data is executed 4 times.
HTH
Chris
Hi Chris,
Thank you very much. I am now more clear.
Best regards,
Leon
[…] Die rekursive Variante benötigt zwei Full Scans auf “emp”, die alte Variante nur einen (das war nicht immer so) […]