Answer this in your head before trying it out or looking it up. Assume the following script is run in a single session on an Oracle database:
CREATE SEQUENCE seq START WITH 1 INCREMENT BY 1;
CREATE TABLE t (a INTEGER, b INTEGER);
-- Statement #1
INSERT INTO t VALUES (seq.NEXTVAL, seq.NEXTVAL);
-- Statement #2
INSERT INTO t VALUES (seq.CURRVAL, seq.NEXTVAL);
-- Statement #3
INSERT INTO t VALUES (seq.CURRVAL, seq.CURRVAL);
Which of the following is/are true?
- The inserted rows will be {1,2}, {2,3} and {3,3}.
- The inserted rows will be {1,1}, {2,2} and {2,2}.
- The inserted rows will be {1,1}, {1,2} and {2,2}.
- Statements #2 and #3 will each raise “ORA-08002: sequence SEQ.CURRVAL is not yet defined in this session”.
I came across this at dbdebunk (ON THE NOTHING THAT’S WRONG WITH NULLS with Hugh Darwen, Fabian Pascal). Couldn’t let that go so I tested it under 9i (9.1) and 10g (10.2) and got identical results. The last two statements are logically equivalent, but gives different results!
SQL> create type point as object (x real, y real);/
Type created.
SQL> create table t (p point);
Table created.
SQL> insert into t values (point(null, null));
1 row created.
SQL> insert into t values (point(1, null));
1 row created.
SQL> insert into t values (point(1, 2));
1 row created.
SQL> select * from t;
P(X, Y)
——————————————————————————–
POINT(NULL, NULL)
POINT(1, NULL)
POINT(1, 2)
SQL> select * from t where p=p or not p=p;
P(X, Y)
——————————————————————————–
POINT(1, 2)
SQL> select * from t where not p=p or p=p;
P(X, Y)
——————————————————————————–
POINT(1, NULL)
POINT(1, 2)
According to CM, it looks like a bug. The filter predicates seem to be applied incorrectly by the optimiser:
where ( p=p ) or ( not p=p );
filter(“T”.”SYS_NC00003$”=”T”.”SYS_NC00003$” AND
“T”.”SYS_NC00002$”=”T”.”SYS_NC00002$” OR T.”P”T.”P”)
where ( not p=p ) or ( p=p );
filter(T.”P”T.”P” OR “T”.”SYS_NC00002$”=”T”.”SYS_NC00002$”)
Jeffrey Kemp
11 February 2006
bug / SQL /
How many times have you seen ''
used for a NULL string?
Oracle 9i SQL Reference Release 2 (9.2) – “Basic Elements of Oracle SQL, 5 of 10: Nulls”
“…Oracle currently treats a character value with a length of zero as null. However, this may not continue to be true in future releases, and Oracle recommends that you do not treat empty strings the same as nulls.”
Simplify your SQL – Avoid Unnecessary Joins
I can’t count how many times I’ve seen SQL like the following:
SELECT b.some_column
FROM a, b
WHERE a.id = :p_id
AND b.id = a.id;
(where table a is unique on a.id, and table b has a referential integrity constraint to a, i.e. b.id references a.id)
I’ve seen many variations on this theme, where tables are unnecessarily referenced in a query. In this case, the query should have been formulated as:
SELECT b.some_column
FROM b
WHERE b.id = :p_id;
To be fair, this can often be blamed on maintenance by multiple programmers over time; maybe at some point some data from table a was needed, and only later removed; or we have a cut-and-paste artifact by a less experienced programmer.
The only problem with this is code simplicity (generally, the simpler the code, the easier it is to understand it, and therefore to maintain it); this will not usually be a performance issue. On Oracle, as long as you have a unique constraint on a.id, the optimiser will generally infer that :p_id can be used to directly access table b, and will then probably do a quick index unique scan on the a.id index (technically, the optimiser should know it doesn’t even have to do this, because of the foreign key constraint on b.id -> a.id and the unique constraint on a.id).
[Aside: Of course, if those constraints did NOT exist, then the two queries above would be logically different, and only one or the other would be correct for a given situation. The programmer must in either case know why they are or are not querying any particular table.]
Here’s a simple example using the SCOTT sample schema (run on Oracle 10.1.0.2, but should be pretty much the same on any version). I’ll issue two queries that are logically equivalent; the first one will unnecessarily access DEPT; the second one will access EMP directly.
==========================================================
SQL> select emp.ename from emp, dept where emp.deptno = dept.deptno and dept.deptno = 10;
ENAME
———-
CLARK
KING
MILLER
Elapsed: 00:00:00.01
Execution Plan
———————————————————-
0 SELECT STATEMENT Optimizer=ALL_ROWS (Cost=3 Card=5 Bytes=60)
1 0 NESTED LOOPS (Cost=3 Card=5 Bytes=60)
2 1 INDEX (UNIQUE SCAN) OF ‘PK_DEPT’ (INDEX (UNIQUE)) (Cost=0 Card=1 Bytes=3)
3 1 TABLE ACCESS (FULL) OF ‘EMP’ (TABLE) (Cost=3 Card=5 Bytes=45)
Statistics
———————————————————-
0 recursive calls
0 db block gets
9 consistent gets
0 physical reads
0 redo size
442 bytes sent via SQL*Net to client
508 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
3 rows processed
SQL> select emp.ename from emp where emp.deptno = 10;
ENAME
———-
CLARK
KING
MILLER
Elapsed: 00:00:00.01
Execution Plan
———————————————————-
0 SELECT STATEMENT Optimizer=ALL_ROWS (Cost=3 Card=5 Bytes=45)
1 0 TABLE ACCESS (FULL) OF ‘EMP’ (TABLE) (Cost=3 Card=5 Bytes=45)
Statistics
———————————————————-
1 recursive calls
0 db block gets
8 consistent gets
0 physical reads
0 redo size
442 bytes sent via SQL*Net to client
508 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
3 rows processed
==========================================================
As you can see, the second query is much simpler. The performance is pretty much identical.
What I haven’t included above is the predicates for the plans; if I had you’d see that the optimiser does actually apply the “deptno = 10” predicate directly to EMP (thanks to Connor for pointing this out).
In these simple queries it may not seem much of an issue; but for queries involving many tables with many predicates, every bit of unnecessary code makes it all the harder to read, understand and maintain.
Jeffrey Kemp
24 January 2006
SQL /
If you’re using the SAMPLE clause in Oracle 9i, in combination with a join, be aware of this small gotcha. I found a workaround, thankfully.
select * from (select * from dual), (select * from dual sample (10));
Expected result: should return zero or one row (more or less at random)
Actual result (tested with 9.2.0.6.0): “ORA-30561: SAMPLE option not allowed in statement with multiple table references”
Workaround: put the query with the SAMPLE clause first, i.e.
select * from (select * from dual sample (10)), (select * from dual);
Note: It works fine in 10g (tested under 10.1.0.2.0) since 10g does not restrict how many SAMPLE clauses are in a query.
Background: see my earlier posts Avoiding Lost Updates
and Avoiding Lost Updates #2: Updateable Views
I’ve discussed this problem with my colleagues and have come up with a number of potential solutions, one of which we believe is bulletproof.
The desired property of such a solution is to reduce the probability that an update of a view will leave the row with the same “version” number, to as close to zero as possible. A bulletproof solution would have a probability of exactly zero.
Solution #1:
Instead of taking the SUM of all the versions, shift the versions from the outer-joined tables (e.g. by multiplying by 10000, 100000, etc).
This reduces the likelihood of getting a false negative; but it doesn’t reduce it to zero, since it would still be theoretically possible to get an update that would cause compensating deletes and updates on the outer-joined tables.
Solution #2:
Use ORA_ROWSCN instead of version.
This also reduces the likelihood of a false negative; in our view, astronomically so; but with a bit of mental gymnastics we can think of a contrived scenario in which the total of ORA_ROWSCN stays the same.
Solution #3:
Cause an update on the view to always do at least one update of one of the non-outer-joined tables (even if it is updating a column to the same value); and don’t include versions from outer-joined tables in the summed version.
This means that version will now always increase whenever an update occurs; therefore, the probability is zero, and the solution is bullet-proof!
Thanks to Howard and Maciej for their help.
This is a followup to my earlier post “Avoiding Lost Updates: Protecting Data in a Multi-user environment”.
In this post I’ll talk about how this problem can be solved when you’re creating updateable views. Now, if you read the previous post, you’ll know the five options that I thought of (plus a sixth pointed out by Gary). For updateable views, if you want to protect against lost updates (in other words, “Last Update Wins” is not an acceptable solution), you can still use “Column Compare” and “Hash Compare” without any changes.
For the “Timestamp Compare”, “Version Compare” and Gary’s ORA_ROWSCN options, however, you’ll need to bring some data from the base tables up to the view’s SELECT clause.
For “Timestamp Compare” and ORA_ROWSCN, you need to get the latest of the timestamps/SCNs. You could use the GREATEST function for this (as long as ORA_ROWSCN is always greater for later transactions?).
It’s just as simple when you’re using “Version Compare”. It was pointed out by a colleague of mine that all you have to do is present the SUM of all the version columns from each base table that can be updateable via the view. This way, if any or all of the base tables experience changes, the “version” column in the view will increment by 1 or more. Most of the time, that is!
Let’s consider an example. We have the following imaginatively named base tables with a view on them:
CREATE TABLE table_a (
id NUMBER(12) PRIMARY KEY,
t VARCHAR2(10),
version NUMBER(38) NOT NULL);
CREATE TABLE table_b (
id NUMBER(12) PRIMARY KEY,
u VARCHAR2(30) NOT NULL,
version NUMBER(38) NOT NULL );
CREATE TABLE table_c (
id NUMBER(12) PRIMARY KEY,
v VARCHAR2(10) NOT NULL,
version NUMBER(38) NOT NULL );
CREATE TRIGGER table_a_ins_upd_row BEFORE INSERT OR UPDATE ON table_a FOR EACH ROW
BEGIN
IF inserting THEN :NEW.version := 1; ELSE :NEW.version := :OLD.version + 1; END IF;
END;
/
CREATE TRIGGER table_b_ins_upd_row BEFORE INSERT OR UPDATE ON table_b FOR EACH ROW
BEGIN
IF inserting THEN :NEW.version := 1; ELSE :NEW.version := :OLD.version + 1; END IF;
END;
/
CREATE TRIGGER table_c_ins_upd_row BEFORE INSERT OR UPDATE ON table_c FOR EACH ROW
BEGIN
IF inserting THEN :NEW.version := 1; ELSE :NEW.version := :OLD.version + 1; END IF;
END;
/
CREATE VIEW view_ab AS
SELECT table_a.id, table_a.t, table_b.u, table_c.v, table_a.version + table_b.version + table_c.version version
FROM table_a
LEFT OUTER JOIN table_b ON table_a.id = table_b.id
LEFT OUTER JOIN table_c ON table_a.id = table_c.id;
The view therefore shows all rows in table_a, as well as the corresponding rows in table_b and table_c if they exist.
Assume that the view_ab has instead-of-insert and update triggers that do the following:
Insert/update id and t on table_a.
If “u” is inserted or updated with a value, do the corresponding insert or update on table_b.
If “u” is cleared (set to NULL), delete it from table_b.
And the same for “v” in relation to table_c.
If I do this:
INSERT INTO view_ab (id, t, u, v) VALUES (1, 'A', 'B', 'C');
The view should have (1, ‘A’, ‘B’, ‘C’, 3). Version = 3 because we have one row in each of the three tables each with version = 1.
Now if I instead did:
INSERT INTO view_ab (id, t, u) VALUES (1, 'A', 'B');
The view should have the row (1, ‘A’, ‘B’, null, 2). Version = 2 because we have rows in table_a and table_b with version = 1, but no row in table_c.
The only purpose of the version column is to change if any insert, update or delete occurs on any of the base tables.
What if I do the following:
UPDATE view_ab SET u = NULL, v = 'C' where id = 1;
What will the view have? It will have (1, ‘A’, null, ‘C’, 2). Hang on there – I have done a change to the row, but the version has stayed the same! Why? Because the update trigger had to delete from table_b (reducing the total version to 1) and insert into table_c (increasing the total version back to 2).
How to solve this? This is answered in my next post.
There is a particular problem with Oracle and other databases where access to data is not serialized (by default), and there are a number of ways this problem is generally dealt with. The problem crops up when you have many users, some of whom may wish to update the same data at around the same time. The process that takes place when a user wishes to update a row is:
Session #1:
- Read the data from the database into the client memory.
- Make the changes in the client memory.
- Write the changes from the client memory to the database.
Between steps 1 and 2 another session (#2) might read the same data (unless you’re in a particularly poorly designed database in which writers block readers; e.g. one database I helped to refactor did just that – it issued exclusive locks on rows every time they were read into a form(!) – this was the first thing we got rid of in the new system), and modify it.
The row isn’t locked, so the other user makes their changes and clicks their “Save” button which commits the changes to the database. Meanwhile, session #1 finally gets to step 3 and saves their changes. If no checks happened, they would overwrite the changes made by session #2.
To illustrate:
- Session #1 reads the row “Joe Bloggs”.
- Session #2 reads the row “Joe Bloggs”.
- Session #2 changes the row to “Jane Bloggs” and saves.
– now, Session #1 is still looking at “Joe Bloggs” on their screen.
- Session #1 changes the row to “Joe Jones” and saves.
The problem is that Session #1 has unwittingly undone the change that Session #2 made.
There are a number of ways this problems is approached that I know of:
- “Last Update Wins” – just allow the update.
- “Column Compare” – all the columns to be updated are compared when updating, and if any are different to the values originally read, an error is returned (e.g. Oracle Forms does this and issues “FRM-40654: Record has been updated by another user. Re-query to see change.”).
- “Hash Compare” – this is a variant on the “Column Compare” method used by HTMLDB. It compares a hash of the combined data for the row.
- “Timestamp Compare” – the table has a “timestamp” column, which is automatically set to the last date/time when the row was inserted or most recently updated. An update will only succeed if the timestamp of the updated row is the same as the timestamp originally read.
- “Version Compare” – the table has a “version” column, which is set to 1 on insert and incremented on update. An update will only succeed if the version of the updated row is the same as the version originally read.
Note: all the “Compare” approaches above do not necessarily require a second query-and-lock on the row prior to the update; the best way to implement these is to add the compare as a predicate for the update – then, if no rows are updated, the caller can tell that the update failed (because the row had been updated or deleted by another session).
Comparison
- “Last Update Wins” – This approach may be acceptable for some applications.
- “Column Compare” – If only the updated columns are compared, this method means that two sessions can modify different columns in the same row independently.
- “Hash Compare” – This method stops two sessions modifying the same row, even if they’re touching different sets of columns.
- “Timestamp Compare” – This method requires the timestamp to have sufficient granularity so that a read by one session, followed by an update by another session, will not occur at exactly the same timestamp. Generally milliseconds or microseconds are needed at least, although even this granularity might not be sufficient for some situations. This method also stops two sessions modifying the same row, even if they’re touching different sets of columns.
- “Version Compare” – This is a favourite for Java developers and works very well. It is simple to code, computationally cheap, and doesn’t suffer from the timestamp granularity problem. This method also stops two sessions modifying the same row, even if they’re touching different sets of columns.
From the above comments you might think I prefer #5 (Version Compare) but I think it can depend on the requirements of the system you’re working with. Of course, if you’re using Forms or HTMLDB you don’t need to worry about it too much because it’s been largely taken care of for you.
In my next entry I describe how this works when you’re working with updateable views.
Further Reading
Came across this SQL problem in Eddie Awad’s blog, where he gives a solution to a problem posted to the Oracle-l mailing list.
I thought, this would be so much easier to solve if we were allowed to model the complete problem domain in the database, e.g.:
— create the data set
SQL> create table t (id number);
Table created.
SQL> insert into t values (1);
1 row created.
SQL> insert into t values (2);
1 row created.
SQL> select * from t;
ID
———-
1
2
— create the query set
SQL> create table u (id number);
Table created.
SQL> insert into u values (1);
1 row created.
SQL> insert into u values (2);
1 row created.
SQL> insert into u values (3);
1 row created.
SQL> select * from u;
ID
———-
1
2
3
SQL> select * from t
2 where id in (select id from u)
3 and not exists (select id from u minus select id from t);
no rows selected
— correct: the complete query set was not found in the data set
SQL> insert into t values (3);
1 row created.
SQL> select * from t
2 where id in (select id from u)
3 and not exists (select id from u minus select id from t);
ID
———-
1
2
3
— correct: the complete query set was found in the data set
The query could be further improved by changing the “where id in ()” to a “where exists ()”.
To make the query table usable by multiple sessions simultaneously, it can be created as a global temporary table (and populated whenever a query is needed), or a “group” distinguisher could be added to it.
Jeffrey Kemp
12 October 2005
SQL /
In software development, quite often the problem you encounter has virtually nothing to do with the thing you’re trying to do.
I’m creating some views which will provide an interface to a database. The views will have Instead-of triggers so that all DML will be passed to my own (packaged) procedures.
I’ve got a particular view which is a join between a base table, filtered by some criteria, and a code/description mapping view. I’ve implemented instead-of-insert, update and delete triggers that do the appropriate DML on the base table.
Inserts and updates work fine, but when deleting from the view I’m getting “TM contention” wait events. When deleting across a database link the session gets “ORA-02049: timeout: distributed transaction waiting for lock” after a long delay. If the delete is issued within the same database as the view, the session waits indefinitely for the lock.
I had a look at the view and noticed that it is a key-preserved view, so I theorised that maybe the session was locking the row before running my instead-of-delete trigger that does the delete; the trigger tries to delete the row that has already been locked. Doesn’t make sense. When I remove the instead-of-delete trigger however, I get “ORA-01752: cannot delete from view without exactly one key-preserved table”. So I need the trigger to handle the delete.
Looked up the Oracle doc “Instance Tuning Using Performance Views” about “TM enqueue”: “The most common reason for waits on TM locks tend to involve foreign key constraints where the constrained columns are not indexed. Index the foreign key columns to avoid this problem.” I saw this ref come up in a search early on but I initially dismissed it because my problem wasn’t related to performance. I’m glad I took a second look.
Righty-o. Looked at the table I was deleting from, and lo and behold there is another table with a fk to my table with no index on the fk column (mind you, there was an index that contained that column along with another column, obviously the database wasn’t going to use that to avoid the table lock).
When I created an index on the fk column, the problem was resolved. Brilliant!
The application never needed to delete from that table before, so this had never been encountered. In this case adding the index solved the problem. It’s on a very small, rarely-used table so the index won’t have a negative impact. Took about half an hour to work out, but I learned a bit more about Oracle, so that’s good.