The ORACLE-BASE Blog

So you want to use pipelined table functions and cursor_sharing in 10.1.0. Check this out…

First create this test package:

CREATE OR REPLACE PACKAGE ptf_test AS
TYPE t_ptf_row IS RECORD (
dummy dual.dummy%TYPE
);

TYPE t_ptf_tab IS TABLE OF t_ptf_row;

FUNCTION get_data_1
RETURN t_ptf_tab PIPELINED;

FUNCTION get_data_2 (p_in IN VARCHAR2)
RETURN t_ptf_tab PIPELINED;

END;
/
SHOW ERRORS

CREATE OR REPLACE PACKAGE BODY ptf_test AS
FUNCTION get_data_1
RETURN t_ptf_tab PIPELINED
AS
l_row t_ptf_row;
BEGIN
SELECT dummy
INTO l_row
FROM dual;

PIPE ROW(l_row);
RETURN;
END;

FUNCTION get_data_2 (p_in IN VARCHAR2)
RETURN t_ptf_tab PIPELINED
AS
l_row t_ptf_row;
BEGIN
SELECT dummy
INTO l_row
FROM dual;

PIPE ROW(l_row);
RETURN;
END;
END;
/
SHOW ERRORS

Then turn on cursor_sharing and query the pipelined table functions:

SQL> ALTER SESSION SET CURSOR_SHARING=SIMILAR;

Session altered.

SQL> SELECT *
2  FROM TABLE(ptf_test.get_data_1);

D
-
X

1 row selected.

SQL> SELECT *
2  FROM TABLE(ptf_test.get_data_2('x'));
FROM TABLE(ptf_test.get_data_2('x'))
*
ERROR at line 2:
ORA-22905: cannot access rows from a non-nested table item

SQL>

Now turn it off and try again:

SQL> ALTER SESSION SET CURSOR_SHARING=EXACT;

Session altered.

sys@db10g> SELECT *
2  FROM TABLE(ptf_test.get_data_1);

D
-
X

1 row selected.

SQL> SELECT *
2  FROM TABLE(ptf_test.get_data_2('x'));

D
-
X

1 row selected.

SQL>

So you can’t use cursor_sharing and pipelined table functions with parameters together. It took me a while to figure out the problem, but when I did it wasn’t long before support came back with an unfixed bug relating to it (Bug 4178323).

The solution? Don’t use cursor_sharing 🙂

Cheers

Tim…

Note. Script changed to use ALTER SESSION to prevent anyone leaving their test instance in an dodgy state. Thanks for the heads-up Gary 😉

I used a stored outline in a production environment for the first time yesterday. I can’t say I’ve ever seen the need to do this before, but this time it masked a pretty serious problem. Normally I try to let the optimizer do its own thing. I even avoid optimizer hints if I can, but in this case the stored outline was the only solution.

The issue was that an SQL statement generated by a finder method in a J2EE application was acting differently depending on JDBC driver being used to make the database connection. When the connection was made via the thin driver the SQL statement used the correct execution plan and worked real quick. When the OCI driver was used the same SQL statement was using a ridiculous index, causing it to trawl through loads of data, making it take around 4 seconds to complete. In this case the impact was massive as this same statement is executed many times throughout the system.

This begs two questions:

• Why is it happening?
• Is there a way to force a good execution plan until we can solve the real issue?

The answer to the first questions is still under investigation, but my answer to the second questions was to use optimizer hints! Nice idea, but it won’t work! Unfortunately the SQL is generated at runtime by the OC4J container so we have little influence over the actual statement produced. Then I thought *stored outlines*. A couple of minutes later I had created an outline for the query and both drivers were giving the same performance!

I knocked up a quick article on Stored Outlines and Plan Stability on my website.

In this case the stored outline saved my bacon by hiding a problem until we can resolve it, but they kinda scare me. When you add a hint to an SQL statement it’s there for the whole world to see. People will only know the outline exists if I tell them about it. It’s going to be really easy for everyone to forget about it, then get confused when another environment doesn’t react in the same way.

I guess I’ll put stored outlines in the same category as triggers. Useful at times, but avoid if possible because sooner or later they are going to screw you up 🙂

My list of things to do now includes:

• Find out why the same statement acts differently with the two drivers.
• Try to convince everyone that not every SQL problem we’ve ever had can be fixed using stored outlines.
• Convince the world that Java is a great language, but J2EE sucks. Runtime generation of SQL is certainly fuel for the fire 🙂

Cheers

Tim…

PS. The latest IOUG SELECT magazine has one of my articles in it. I even made the front page. If you’re bored turn to page 6 and check it out. Next stop Rolling Stone…

I’ve just been asked a question on mass loading of images into the database (Optimizing BLOBs for SQL*Loader) and to be perfectly honest I’m not sure what the best solution is.

I’ve worked on a number of projects where images are stored in the database, but in all cases the images have been loaded on a one-by-one basis as they arrived. Apart from exports and imports between schemas I can’t remember ever doing mass loads of images into the database.

Anyone got any suggestions, preferably based on experience of this?

Cheers

Tim…

I’ve just noticed that production versions of 10g release 2 are now available for Windows, Solaris, AIX and HP-UX. See Oracle Database 10g Downloads…

Cheers

Tim…

Some Oracle features are so useful you quickly get to a point where you can’t remember what life was like before them. Three that spring to mind today are the DBA_SCHEDULER_JOB_RUN_DETAILS view, the DBMS_MONITOR package and the DBMS_METADATA package.

When using the new 10g scheduler, the DBA_SCHEDULER_JOB_RUN_DETAILS view provides a history of previous job runs. It’s cool to query this and see how the run duration of particular jobs varies during the course of the day, or between days, allowing you to do trend analysis over time. It’s not rocket science to store this information in a table yourself, but it’s nice that it happens straight out of the box.

The DBMS_MONITOR package is part of the end-to-end tracing available in 10g. It allows you to initiate tracing for one or more sessions, where the sessions are identified using combinations of criteria including: service, module, action, client_identifier or just the plain old sid. The really cool thing about this is that it doesn’t matter how many sessions match the criteria, or which RAC nodes they are running on, they all get traced. You can then consolidate all the trace files into a single file using the trcsess utility so you can use TKPROF as normal. This is way cool!

The DBMS_METADATA package was introduced in Oracle9i to allow you to extract DDL or XML definitions of database objects. Most tools like PL/SQL Developer and TOAD allow you to do this easily, but from the server it was always a pain. I used to spend ages writing and maintaining scripts to do this, and invariably they didn’t work for some features. This package threw all that out of the window. Today a friend asked me how he could dump out a schema creation script and I was in the middle of explaining about DBMS_METADATA when I remembered he works on 8i. Bummer…

I use these things all the time so the thought of moving back to 8i or 9i fills me with dread!

Someone mentioned recently that Oracle 11(xyz) is scheduled for release in November next year. That sounds a little early, but if it’s true I wonder how long it will be before I’m saying, “I’d rather quit than work with Oracle 10g!” 🙂

Cheers

Tim…