Oracle FAQ Your Portal to the Oracle Knowledge Grid

Home -> Community -> Mailing Lists -> Oracle-L -> RE: Deadlock problem

RE: Deadlock problem

From: Mercadante, Thomas F \(LABOR\) <>
Date: Mon, 8 May 2006 13:50:33 -0400
Message-ID: <ABB9D76E187C5146AB5683F5A07336FF6811D9@EXCNYSM0A1AJ.nysemail.nyenet>


There were two theories concerning where a commit statement should be issued from.

The first theory I heard from Oracle when PL/SQL first came out was that packages should not have commit statements in them - that the application should issue the commit when all of the pieces of work were completed. It was thought that the application would better know when a commit should be issued.

The other theory was to put all of the work in the PL/SQL packages and let it control everything and either report back success (commit) or failure (rollback) to the application.

Today, either way works just fine in my view.

As for your problem, dig a little deeper. Most deadlocks that I've seen are caused by foreign key constraints and missing indexes. So look at the tables involved and look for the table being updated being referenced by another table via a FK. Simply adding indexes to the foreign key columns solves this problem.

And remember - this is an application problem. Somebody might have to fix some code!

Good Luck.


-----Original Message-----

[] On Behalf Of Alessandro Vercelli Sent: Monday, May 08, 2006 1:34 PM
To: Oracle
Subject: Deadlock problem

Hi all,
I'm trying to solve an ora-4020 (deadlock) issue; the trace file (sorry if messed) shows:

ksqded1: deadlock detected via did
Current SQL statement for this session:
Update <TABLE> set <FIELD1>='<VALUE>' where <FIELD2> like '<STRING>%' The following deadlock is not an ORACLE error. It is a deadlock due to user error in the design of an application or from issuing incorrect ad-hoc SQL. The following information may aid in determining the deadlock: Deadlock graph:


Resource Name process session holds waits process session holds waits TX-00040015-0000305b 13 11 X 10 14 X TX-0007000b-0000309e 10 14 X 13 11 X session 11: DID 0001-000D-00000001 session 14: DID 0001-000A-00000001 session 14: DID 0001-000A-00000001 session 11: DID
Rows waited on:
Session 14: obj - rowid = 0000147E - AAABR+AAKAAAAJIAAH Session 11: obj - rowid = 0000147E - AAABR+AAKAAAzEeAAH

The trace file shows clearly that session 11 and 14 are blocking each other.

Note that <STRING> can be very long, but it's almost certain that this is not causing the problem.

Database version is 8.0.5 on Solaris 8 sparc.

So, I'm looking at the piece of source containing the affected code (I'm not the developer neither a skilled one) and I have seen something strange, that is a sql package containing many procedures with insert/update statements and none of these insert/update was followed by a commit; I asked the developer for this matter and she said that a commit would prevent a possible rollback of database transaction.

Now, my questions are:
1. Is it correct an insert/update without a commit into a sql package? If yes, when are the inserted/updated data commited? 2. Would this the possible cause of the deadclock, as the table indexes could be locked by a large number of records inserted/updated? 3. Is this the correct way to get the choice of performing a rollback?

Thanks for you help,



-- Received on Mon May 08 2006 - 12:50:33 CDT

Original text of this message