Table Types Supplied by Oracle

This is a list of all the table types I’ve found in Oracle-supplied packages, e.g. OWA_UTIL in 10g supplies the type:
TYPE datetype IS TABLE OF varchar2(10) INDEX BY binary_integer;

I find them handy for quick one-off scripts.

Oracle_supplied_table_types


PL/SQL Maintenance Nightmares (Learning to avoid…)

I’ve worked on some new PL/SQL packages and Forms which work quite well. They are efficient in their use of resources, and are easy for me to debug and maintain. After some other developers have had to do some changes (some from changes to requirements, some from bugs I didn’t find), I’ve learned that the highly modular style of code I use is difficult to modify without introducing new bugs.

The modular style is great while developing; it reduces duplication and provides useful abstractions. When it comes to maintaining that code, however, a new developer has to read all that code and follow its tortuous logic around in order to debug it. It’s much easier to just patch the existing code to force it to work.

This is obviously not the best outcome; more new code is added, instead of the existing code being fixed. The module thus becomes more complex and unmaintainable. After some time, a new developer confronts a hodge-podge of duplicated functionality and cut-and-paste monstrosities, and exclaims, “Who wrote this rubbish?” – perhaps with justification.

What I’ve learned is that after developing a new module, I must go back and refactor it to improve maintainability. This includes a number of steps (this list is not intended to be exhaustive):

  1. Delete simple wrappers for library functions/procedures, and replace all calls to them with calls to the original library functions/procedures.
  2. Don’t use constants for internal codes (e.g. names of items and blocks) – use literals instead.
  3. If a fairly simple function or procedure is only called once throughout the final code – delete it and move the code to where it was called from; the exception is if this would make the calling code more difficult to read or understand.
  4. Code that generates dynamic SQL should set the SQL in entirety in single statements if possible, instead of building it procedurally. Another developer should be able to easily find the exact SQL that will be generated just by looking at the code, instead of having to mentally build it procedurally, or having to printline the SQL and run it to see what it generates.
  5. Add comments that document the philosophy of how the code has been laid out.

Some examples to illustrate the above:

  1. We have a library procedure called lk_item.lp_show_item (pc_item, pn_visible, pn_enabled) which is used to show, hide, enable, and/or disable a form item. I wanted to write code like the following: lk_item.lp_show_item(‘my_item’, some_boolean_expression, another_boolean_expression) but I couldn’t because the parameters require PROPERTY_TRUE or PROPERTY_FALSE, which are numbers. My first cut included the following wrapper for lk_item.lp_show_item:
    PROCEDURE cp_show_hide
    (pc_item IN VARCHAR2
    ,pb_visible IN BOOLEAN
    ,pb_enabled IN BOOLEAN) IS
    FUNCTION CF_PROPERTY_TF
    (pb_value IN BOOLEAN) RETURN NUMBER IS
    BEGIN
    IF pb_value THEN
    RETURN PROPERTY_TRUE;
    ELSE
    RETURN PROPERTY_FALSE;
    END IF;
    END;
    BEGIN
    lk_item.lp_show_item(pc_item,
    CF_PROPERTY_TF(pb_visible),
    CF_PROPERTY_TF(pb_enabled));
    END;

    In the end, however, I only called this procedure two times, and within a single procedure. So I deleted the wrapper procedure, and called lk_item.lp_show_item directly. I kept my Boolean expressions, but wrapped them in a locally declared function that did the translation.
    FUNCTION CF_PROPERTY_TF
    (pb_value IN BOOLEAN) RETURN NUMBER IS
    BEGIN
    IF pb_value THEN
    RETURN PROPERTY_TRUE;
    ELSE
    RETURN PROPERTY_FALSE;
    END IF;
    END;

    lk_item.lp_show_item(‘ITEM_NAME’,
    CF_PROPERTY_TF(some expression),
    CF_PROPERTY_TF(some other expression));

  2. I had a procedure that populated a large number of items in the block in response to the user selecting a new record. Because the user can select a new record using any of several different methods (effectively, different search criteria), I used a parameter called “search mode” that is set to various constant strings, e.g. “GNL_ID”, “OFF_ID_DETAILS”, “COLLAR_NO”. Some of these strings were named after significant database table columns, some were generic – they all, however, simply differentiated slightly different conditions under which the procedure must operate.

    Initially I had a set of constants that took these values; however, because this code passed these values to a database package as well, the constants were defined in the database package as well. In the end the code was much simpler and easier to read by removing all the constants and just encoding the codes as literal strings.

  3. (no example)

  4. A first cut of the code might look like (this is a very simplified example):
    cp_lexical := ‘SELECT aaa, bee, cee ‘;
    IF (some expression) THEN
    cp_lexical := cp_lexical || ‘, dee FROM bla ‘;
    ELSE
    cp_lexical := cp_lexical || ‘FROM dee, eff ‘;
    END IF;
    IF (some other expression) THEN
    cp_lexical := cp_lexical
    || ‘, gee WHERE eee = eff AND gee = oh ‘;
    ELSE
    cp_lexical := cp_lexical
    || ‘, hat WHERE eee = oh AND ii = jay ‘;
    END IF;
    cp_lexical := cp_lexical
    || ‘ AND kay = ell AND emm = enn’;

    The final, more easily maintained code would look something like:
    IF (some expression)
    AND (some other expression) THEN
    cp_lexical := ‘SELECT aaa, bee, cee, dee ‘
    || ‘FROM bla, gee ‘
    || ‘WHERE eee = eff ‘
    || ‘AND gee = oh ‘
    || ‘AND kay = ell ‘
    || ‘AND emm = enn’;
    ELSIF (some expression) THEN
    cp_lexical := ‘SELECT aaa, bee, cee, dee ‘
    || ‘FROM bla, hat ‘
    || ‘WHERE eee = oh ‘
    || ‘AND ii = jay ‘
    || ‘AND kay = ell ‘
    || ‘AND emm = enn’;
    ELSIF (some other expression) THEN
    cp_lexical := ‘SELECT aaa, bee, cee ‘
    || ‘FROM dee, eff, gee ‘
    || ‘WHERE eee = eff ‘
    || ‘AND gee = oh ‘
    || ‘AND kay = ell ‘
    || ‘AND emm = enn’;
    ELSE
    cp_lexical := ‘SELECT aaa, bee, cee ‘
    || ‘FROM dee, eff, hat ‘
    || ‘WHERE eee = oh ‘
    || ‘AND ii = jay ‘
    || ‘AND kay = ell ‘
    || ‘AND emm = enn’;
    END IF;

    This doesn’t solve all the maintenance problems (i.e. if someone makes a change to one part of the code they might not know that they should make the same change everywhere else that code has been duplicated). However, it does make maintenance a bit easier because the code is easy to follow.

  5. I generally follow the following pattern when encapsulating logic in procedures and packages:
    • Code that populates data in items (e.g. in response to changes in other items) is encapsulated in procedures or packages named POPULATE_xxx.
    • Code that changes item properties (but not their values) is encapsulated in procedures or packages named SETUP_xxx.
    • Code that checks data entry errors is encapsulated in procedures or packages named VALIDATE_xxx.

    These conventions are documented so that another developer can more easily read the code; for example, a when-validate-item trigger might have code like this:
    cp_validate_something;
    cp_populate_something;
    cp_setup_something;

    The developer might still need to follow through all the code to debug it, but they might more easily see where the code should be. Hopefully they’ll recognise that validation code should go in cp_validate_something, rather than cp_populate_something.


Fun with copy-and-paste code

Came across this in a form (6i) to be run on a 9i db. Not only is this code about 33 lines of code too long and issues any number of unnecessary database queries, its name is quite unrelated to its intended function. Needless to say it was easily replaced with a single call to INSTR.

PROCEDURE alpha_check
(ref_in IN VARCHAR2
,ref_out OUT VARCHAR2) IS
-- Procedure included to distinguish
-- ref_in between ID or reference.
  l_alpha_char  VARCHAR2 (1);
  l_alpha_pos   NUMBER;
  l_found_pos   NUMBER;
  l_search_string VARCHAR2 (100) := ' ';
 

  CURSOR cur_get_next_alpha(N NUMBER) IS
  SELECT SUBSTR(l_search_string,N,1)
  FROM dual;

  CURSOR cur_check_for_alpha(C VARCHAR2)IS
  SELECT INSTRB(ref_in,C, 1)
  FROM dual;

BEGIN
  IF ref_in IS NULL THEN
    ref_out := 'X';
    RETURN;
  END IF;

  FOR I IN 1..LENGTH(l_search_string) LOOP
    OPEN cur_get_next_alpha(I);
    FETCH cur_get_next_alpha
    INTO l_alpha_char;
    CLOSE cur_get_next_alpha;

    FOR J IN 1..LENGTH(ref_in) LOOP
      OPEN cur_check_for_alpha(l_alpha_char);
      FETCH cur_check_for_alpha
      INTO l_found_pos;
      CLOSE cur_check_for_alpha;

      IF l_found_pos > 0 THEN
        ref_out := 'N';
        RETURN;
      END IF;
    END LOOP;
  END LOOP;

  ref_out := 'Y';
EXCEPTION
  WHEN OTHERS THEN
    pc_ref_out := 'X';
END;

Looks like it may have been copied from the same source as “As bad as it gets”.


Generating test data that matches existing data

I’ve had to create test data a number of times, including data for tables that had mandatory foreign keys to existing tables. It was not feasible to just create new master rows for my test data; I wanted to refer to a random sample of existing data; but the code that generates the test data had to perform reasonably well, even though it had to pick out some random values from a very large table.

Solution? A combination of the new 10g SAMPLE operator, and DBMS_RANDOM. To illustrate:


(create a “very large table”)
SQL> create table t as
2 select rownum n, dbms_random.string(‘a’,30) v
3 from all_objects;

Table created.

SQL> select count(*) from t;

COUNT(*)
———-
40981

(get a random sample from the table)
SQL> select n, substr(v,1,30) from t sample(0.01)
2 order by dbms_random.value;

N SUBSTR(V,1,30)
———- ——————————
11852 xSsdmFtGqkymbKCFoZwUzNxpJAPwaV
8973 RGyNjqMfVayKdiKFGvLYuAFYUpIbCw
25295 eJJtoieSWtzUTIZXCbOLzmdmWHHPOy
297 hiTxUPYKzWKAjFRYTTfJSSCuOwGGmG
1924 yZucJWgkFviAIeXiSCuNeUuDjClvxt
40646 wMTumPxfBMoAcNtVMptoPchILHTXJa

6 rows selected.

SQL> set serveroutput on

(Get a single value chosen at random)
SQL> declare
2 cursor cur_t is
3 select n from t sample(0.01)
4 order by dbms_random.value;
5 l number;
6 begin
7 open cur_t;
8 fetch cur_t into l;
9 dbms_output.put_line(l);
10 close cur_t;
11* end;
SQL> /
21098

PL/SQL procedure successfully completed.


My test code would open the cursor, fetch as many values as it needed, and then close it. If the cursor ran out of values (e.g. the sample was too small for the desired amount of test data, which varied), my code just re-opened the cursor to fetch another set of random values from the large table.

The reason I sort the sample by dbms_random.value is so that if I only want one value, it is not weighted towards rows found nearer the start of the table.

Note: If I didn’t really care about the sample being picked at random from throughout the table, I could have just selected from the table “where rownum < n".



Is this code actually unreachable?

PL/SQL User’s Guide and Reference (10.2): “4 Using PL/SQL Control Structures – Using the NULL Statement”

“…Note that the use of the NULL statement might raise an unreachable code warning if warnings are enabled.”

Example 4-23 Using NULL as a Placeholder When Creating a Subprogram

CREATE OR REPLACE PROCEDURE award_bonus (emp_id NUMBER, bonus NUMBER) AS
BEGIN — executable part starts here
NULL; — use NULL as placeholder, raises “unreachable code” if warnings enabled
END award_bonus;
/

Indeed, when I compile the above in 10.2 with PL/SQL warnings on, I get PLW-06002 as expected (due to bug 3680132 I get “Message 6002 not found; No message file for product=plsql, facility=PLW” but at least I can look it up in the reference).

“PLW-06002: Unreachable code”
“Cause: Static program analysis determined that some code on the specified line would never be reached during execution.”

I agree that a PL/SQL warning would be desirable in the case where a procedure has nothing but a NULL in it (probably a stub). Correct me if I’m wrong, but if I were to call award_bonus, surely the NULL is “executed” – therefore, it is reachable! A more appropriate warning would be something like “function/procedure does nothing”, or “get back to work you silly mug, you’ve forgotten to finish the code”. Maybe they just couldn’t be bothered making up another warning code.


Did you know that NULL is a “Boolean value”? Hmmm…

PL/SQL User’s Guide and Reference (9.2 and 10.2): “NULL Statement”

“The NULL statement and Boolean value NULL are unrelated.”

I can understand that this is to draw a distinction between “NULL” as a procedural statement and “NULL” as a literal. But why is NULL specifically identified as Boolean? So, the NULL statement is somehow related to NULL strings, NULL numbers, and NULL dates?

(I won’t even mention the problem with calling NULL a value, something which is prevalent throughout the literature, including the SQL standard.)


User-named locks with DBMS_LOCK

The application I’m working on involves a number of views that provide an interface between two databases that have quite different models. The views implement inserts, updates and deletes via Instead-of triggers.

On one of these views, the update trigger needs to actually do an update and an insert on the source table: it must update the existing row to mark it as “closed”, and then insert a new row to take its place. In this way the view hides the fact that the table keeps a history of changes.

Now, a logical constraint is that there can only be one “open” record for a given entity. Therefore, there can be zero or more “closed” records, and there must be a single “open” record. This constraint, however, has not been implemented on the database (naughty naughty) because our code is perfect and never makes a mistake 😉

This worked perfectly well until our thorough Java guy decided to do some load testing, in which a process would hammer away inserting, updating and deleting the same row over and over again, in multiple concurrent sessions.

This had the result that we ended up seeing lots of “open” records for the entity, which violates the logical constraint described above. “That’s not allowed” I said, and soon realised that the fact that multiple sessions are updating the same row means that even a trigger-based constraint check will not solve the problem, since the triggers cannot see updates from other sessions.

So, we need to have a lock of some sort, so that when one session wants to update an entity, other sessions cannot start updating it as well until the first one is finished. I can’t just lock the row I want to update, because after I’ve updated it, I still need the lock while I insert the new row.

The solution is to use a package supplied with Oracle calls DBMS_LOCK, with which I can request “user-named locks”. This package gives a lot of options but I’m only interested in simple Exclusive locks, so I’ll also create some wrapper procedures.

Easy, right? Well, yeah once you get past a few gotchas.

Try #1:

-- wrapper function to Request a lock; returns a lock handle
FUNCTION request_lock (lock_name IN VARCHAR2) RETURN VARCHAR2 IS
  lock_status NUMBER;
  lock_handle VARCHAR2(128);
BEGIN
  DBMS_LOCK.ALLOCATE_UNIQUE (
    lockname => lock_name,
    lockhandle => lock_handle,
    expiration_secs => 864000); -- 10 days
  lock_status := DBMS_LOCK.REQUEST(
    lockhandle => lock_handle,
    lockmode => DBMS_LOCK.X_MODE, -- eXclusive
    timeout => DBMS_LOCK.MAXWAIT, -- wait forever
    release_on_commit => TRUE);
  IF lock_status > 0 THEN
    RAISE_APPLICATION_ERROR(-20000,'request_lock failed: ' || lock_status);
  END IF;
RETURN lock_handle;
END request_lock;

-- wrapper to release a lock; call with the lock handle obtained previously
PROCEDURE release_lock (lock_handle IN VARCHAR2) IS
  lock_status NUMBER;
BEGIN
  lock_status := DBMS_LOCK.RELEASE(
    lockhandle => lock_handle);
  IF lock_status > 0 THEN
    RAISE_APPLICATION_ERROR(-20000,'release_lock failed: ' || lock_status);
  END IF;
END release_lock;

--example calling code
...
DECLARE
  lock_handle VARCHAR2(128);
BEGIN
  lock_handle := request_lock('MYLOCK1');
  -- do some stuff
  release_lock(lock_handle);
END;
...

The above code “worked” ok (although I hadn’t yet tested to see if it was actually getting the lock) until I started calling it from triggers on the views, when I started getting “ORA-04092: cannot commit in a trigger”. What the? I wasn’t doing any commits in my triggers. I couldn’t find any references to the DBMS_LOCK functions doing commits in the 10g documentation, but I thought, ok let’s make these wrappers Autonomous, so they can commit in peace:

Try #2:

FUNCTION request_lock (lock_name IN VARCHAR2) RETURN VARCHAR2 IS
  PRAGMA AUTONOMOUS_TRANSACTION;
  lock_status NUMBER;
  lock_handle VARCHAR2(128);
BEGIN
  DBMS_LOCK.ALLOCATE_UNIQUE (
    lockname => lock_name,
    lockhandle => lock_handle,
    expiration_secs => 864000); -- 10 days
  lock_status := DBMS_LOCK.REQUEST(
    lockhandle => lock_handle,
    lockmode => DBMS_LOCK.X_MODE, -- eXclusive
    timeout => DBMS_LOCK.MAXWAIT, -- wait forever
    release_on_commit => TRUE);
  IF lock_status > 0 THEN
    RAISE_APPLICATION_ERROR(-20000,'request_lock failed: ' || lock_status);
  END IF;
  RETURN lock_handle;
END request_lock;

PROCEDURE release_lock (lock_handle IN VARCHAR2) IS
  PRAGMA AUTONOMOUS_TRANSACTION;
  lock_status NUMBER;
BEGIN
  lock_status := DBMS_LOCK.RELEASE(
    lockhandle => lock_handle);
  IF lock_status > 0 THEN
    RAISE_APPLICATION_ERROR(-20000,'release_lock failed: ' || lock_status);
  END IF;
END release_lock;

No good: I got “ORA-06519: active autonomous transaction detected and rolled back”! This didn’t make much sense at the time, but I put a COMMIT in anyway. But of course, this wouldn’t work with the triggers. Anyway, I started experimenting and soon realised that even if I got the wrappers to work, the locks weren’t being made anyway, and I was able to request locks multiple times, which I expected to be impossible if the wrapper was working correctly.

Finally, I logged in as a user with some DBA privileges and read the specification of the SYS.DBMS_LOCK package. The one thing that jumped out at me was a little comment under ALLOCATE_UNIQUE:

“This routine will always do a commit.”

There was my answer. It was the ALLOCATE_UNIQUE that was doing a commit, not REQUEST or RELEASE. And, because I was setting release_on_commit => TRUE in the call to REQUEST, whenever I committed the lock was being released, straight away.

So I changed release_on_commit => FALSE, and moved ALLOCATE_UNIQUE to a separate function (set up as an Autonomous Transaction). Having read a bit more I realised I could make this API nicer to the programmers by allowing the release_lock to accept the lock name instead of requiring them to manage the lock handle. I also improved the error messages.

Try #3:

-- internal function to get a lock handle
-- (private for use by request_lock and release_lock)
FUNCTION get_handle (lock_name IN VARCHAR2) RETURN VARCHAR2 IS
  PRAGMA AUTONOMOUS_TRANSACTION;
  lock_handle VARCHAR2(128);
BEGIN
  DBMS_LOCK.ALLOCATE_UNIQUE (
    lockname => lock_name,
    lockhandle => lock_handle,
    expiration_secs => 864000); -- 10 days
  RETURN lock_handle;
END get_handle;

PROCEDURE request_lock (lock_name IN VARCHAR2) IS
  lock_status NUMBER;
BEGIN
  lock_status := DBMS_LOCK.REQUEST(
    lockhandle => get_handle(lock_name),
    lockmode => DBMS_LOCK.X_MODE, -- eXclusive
    timeout => DBMS_LOCK.MAXWAIT, -- wait forever
    release_on_commit => FALSE);
  CASE lock_status
    WHEN 0 THEN NULL;
    WHEN 2 THEN RAISE_APPLICATION_ERROR(-20000,'deadlock detected');
    WHEN 4 THEN RAISE_APPLICATION_ERROR(-20000,'lock already obtained');
    ELSE RAISE_APPLICATION_ERROR(-20000,'request lock failed: ' || lock_status);
  END CASE;
END request_lock;

-- wrapper to release a lock
PROCEDURE release_lock (lock_name IN VARCHAR2) IS
  lock_status NUMBER;
BEGIN
  lock_status := DBMS_LOCK.RELEASE(
    lockhandle => get_handle(lock_name));
  IF lock_status > 0 THEN
    RAISE_APPLICATION_ERROR(-20000,'release lock failed: ' || lock_status);
  END IF;
END release_lock;

--example calling code
...
BEGIN
  request_lock('MYLOCK1');
  -- do some stuff
  release_lock('MYLOCK1');
END;
...

This worked perfectly, even from triggers. I even tried it out with multiple sessions:

Session 1:

exec request_lock('JEFF');

Session 2:

exec request_lock('JEFF');

At this point, session 2 just hung – it was waiting for Session 1 to release the lock.

Session 1:

exec release_lock('JEFF');

Now session 2 returned and had the lock. The same thing happened if I just closed down session 1.

If I changed request_lock to set timeout to some value (e.g. 3 seconds), session 2 would wait for that amount of time, then raise my error message “request lock failed – 1” (1 is the status code when a lock request times out).

I tried it with a deadlock situation, with timeout = MAXWAIT:

Session 1:

exec request_lock('MYLOCKA');

Session 2:

exec request_lock('MYLOCKB');

Session 1:

exec request_lock('MYLOCKB');

At this point session 1 hangs as it waits for session 2 to release the lock.

Session 2:

exec request_lock('MYLOCKA');

This is a deadlock. Both sessions are waiting for a lock held by the other session. After a few seconds Oracle automatically detects the deadlock and picks one session (at random, apparently) and it returns the lock status 2 (my code raises the error message “deadlock detected”). The other session stayed hung, however – it was still waiting, but there was no deadlock.

I tried the above scenario again, but with timeout = 5 seconds. If I executed the request_locks quick enough, I got “deadlock detected” on one session, and the other session eventually timed out.

So, everything works as expected, and the Java guys can hammer away at my views with impunity.