This post is part 4 of the Write Less with More series of blog posts, focusing on new features of Oracle 12c that allow us developers to write less than in previous versions in order to achieve the same functionality. Each part is dedicated for one new feature that solves some common development task. For more details, including the setup of the examples used throughout the series and the list of development tasks that drive the series, please see part 1.

All the Series Parts

Published

Part 1 – write less configuration with SQL*Loader Express Mode
Part 2 – write less application code with Identity Columns
Part 3 – write less application code with In-Database Archiving
Part 4 – write less application code with Temporal Validity
Part 5 – write less code in SQL with Row Limiting
Part 6 – write less “inappropriately located” code with SELECT FROM Package-Level Collection Types
Part 7 – write less code in SQL with Lateral Inline Views
Part 8 – write less “inappropriately located” code with PL/SQL in the WITH Clause

Task #4

PROJECT_ASSIGNMENTS is too simplistic. At this point (after it was defined in part 1 and enhanced in Part 2) it only says which people are assigned to which projects, not when or for how long. Now we want to add this missing information.

A Pre-12c Solution

Before 12c we could add a pair of date/time columns to represent the active period boundaries, e.g. PERIOD_START and PERIOD_END:

alter table project_assignments add (
  period_start date,
  period_end   date
);

When dealing with time periods there are three important decisions to make:
One decision is how to represent periods that their start point or end point (or even both) are unlimited. I will use NULL for this.

I recommend reading Stew Ashton’s post about his preference of not using NULL to represent unlimited period boundaries. It doesn’t matter if you agree or not – reading about the reasons is the important part. Actually, I recommend reading everything Stew writes 🙂

Another decision is whether the period boundaries are inclusive or exclusive. I like to include the start point in the period and exclude the end point. In this way, if we want to assign someone to a project for the whole month of February 2015 (for example) we will use quite simple values:

insert into project_assignments (person_id,project_id,period_start,period_end)
values (101,1,date'2015-02-01',date'2015-03-01');

If I’d choose to include both the start point and the end point in the period, I would have to use a more complex value:

insert into project_assignments (person_id,project_id,period_start,period_end)
values (101,1,date'2015-02-01',timestamp'2015-02-28 23:59:59');

And a third decision is which data type to use for the start/end columns – DATE, TIMESTAMP or TIMESTAMP WITH TIME ZONE. This decision is derived from the functional requirements for period boundary granularity (and your data modeling standards, if you have any [I hope you do]).

After we added the columns to the table, we need to add conditions to all the relevant SQL statements in the application (or applications).
For example, if we want to get all the assignments that are active on a specific date:

select * from project_assignments a
where (a.period_start is null or a.period_start <= :my_date)
  and (a.period_end is null or a.period_end > :my_date);

And if we want to get all the assignments that are active sometime within a specific time range:

select * from project_assignments a
where (a.period_start is null or a.period_start < :my_end_date)
  and (a.period_end is null or a.period_end > :my_start_date);

A 12c Solution

Temporal Validity

In Oracle 12c a table can be defined with the PERIOD FOR clause, to associate it with one or more valid time dimensions. Each such dimension consists of a pair of date/time columns. These columns can be created either explicitly or implicitly. As a result, the data can become visible or invisible, based on statement- or session-level definitions.
In our case we can simply add a period called ASSIGNMENT_PERIOD to the table:
 

alter table project_assignments 
       add PERIOD FOR assignment_period;

This statement added two hidden TIMESTAMP WITH TIME ZONE columns – ASSIGNMENT_PERIOD_START and ASSIGNMENT_PERIOD_END – which can (and should) be set explicitly when records of PROJECT_ASSIGNMENTS are inserted and updated.
When we DESCRIBE the table, we see only the columns we explicitly defined:

> desc project_assignments
 Name                                         Null?    Type
 -------------------------------------------- -------- ----------------------
 ASSIGNMENT_ID                                NOT NULL NUMBER(38)
 PERSON_ID                                    NOT NULL NUMBER(38)
 PROJECT_ID                                   NOT NULL NUMBER(38)

But if we select from USER_TAB_COLS we discover the new columns – ASSIGNMENT_PERIOD_START and ASSIGNMENT_PERIOD_END (and one virtual column – ASSIGNMENT_PERIOD – that I will not discuss here):

> select column_name,data_type,nullable,hidden_column,virtual_column,user_generated
    from user_tab_cols
    where table_name = 'PROJECT_ASSIGNMENTS';

COLUMN_NAME                    DATA_TYPE                   NUL HIDDEN_CO VIRTUAL_C USER_GENE
------------------------------ --------------------------- --- --------- --------- ---------
ASSIGNMENT_ID                  NUMBER                      N   NO        NO        YES
PERSON_ID                      NUMBER                      N   NO        NO        YES
PROJECT_ID                     NUMBER                      N   NO        NO        YES
ASSIGNMENT_PERIOD_START        TIMESTAMP(6) WITH TIME ZONE Y   YES       NO        NO
ASSIGNMENT_PERIOD_END          TIMESTAMP(6) WITH TIME ZONE Y   YES       NO        NO
ASSIGNMENT_PERIOD              NUMBER                      Y   YES       YES       NO

Note: there is a major difference between the data dictionary views USER_TAB_COLUMNS and USER_TAB_COLS: while the former exposes only the explicitly defined columns, the latter includes also system-generated columns.

Update [17-Sep-2021] Connor McDonald shows how the period information is stored in the data dictionary – https://connor-mcdonald.com/2021/09/17/is-my-table-temporal/

In addition to the columns, Oracle also implicitly created a check constraint, to make sure the start time is always before the end time:

> select constraint_name,constraint_type,search_condition
  from user_constraints
  where table_name='PROJECT_ASSIGNMENTS'
    and constraint_name like 'ASSIGNMENT_PERIOD%';

CONSTRAINT_NAME                CON SEARCH_CONDITION
------------------------------ --- ------------------------------------------------------
ASSIGNMENT_PERIOD3366E0        C   ASSIGNMENT_PERIOD_START < ASSIGNMENT_PERIOD_END

Note that the condition is “less than” and not “less than or equal to”. The Temporal Validity feature uses the same rule I used in the pre-12c solution – the period includes the start time and excludes the end time.

Now, let’s fill the table with some data. We need to explicitly set values to the system-generated columns:

insert into project_assignments
    (person_id,
     project_id,
     assignment_period_start,
     assignment_period_end)
values
    (101,
     1,
     date'2015-09-16',
     date'2015-09-18');

I added some more records as a preparation for the next examples. This is the content of the table at this point:

> select person_id,
         project_id,
         assignment_period_start,
         assignment_period_end
  from   project_assignments
  order  by assignment_period_start,
            assignment_period_end;

                      PERIOD     PERIOD
 PERSON_ID PROJECT_ID START      END
---------- ---------- ---------- ----------
       101          1 16/09/2015 18/09/2015
       102          1 20/09/2015 22/09/2015
       101          1 22/09/2015 24/09/2015
       101          3 26/09/2015 27/09/2015
       101          2 27/09/2015 30/09/2015
       102          2 27/09/2015 30/09/2015
       101          1 28/09/2015 30/09/2015
       104          2 28/09/2015 30/09/2015
       103          2 28/09/2015 30/09/2015
       102          1 28/09/2015 30/09/2015
       201          2 29/09/2015 30/09/2015
       102          3 29/09/2015
       101          1 04/10/2015 06/10/2015
       102          1 06/10/2015 08/10/2015
       101          1 10/10/2015 12/10/2015
       102          1 14/10/2015 16/10/2015
       102          1 22/10/2015 24/10/2015
       103          1            03/10/2015

18 rows selected.

Nulls represent unlimited start time or end time.

Statement-Level Control

To get, in a specific statement, only records that are active on a specific date, we add the AS OF PERIOD clause.
For example, to get only the project assignments that are active today (when today is September 29th, 2015):

> select person_id,
         project_id,
         assignment_period_start,
         assignment_period_end
  from   project_assignments
  AS OF PERIOD FOR assignment_period SYSDATE;

                      PERIOD     PERIOD
 PERSON_ID PROJECT_ID START      END
---------- ---------- ---------- ----------
       101          1 28/09/2015 30/09/2015
       102          1 28/09/2015 30/09/2015
       103          1            03/10/2015
       101          2 27/09/2015 30/09/2015
       102          2 27/09/2015 30/09/2015
       103          2 28/09/2015 30/09/2015
       104          2 28/09/2015 30/09/2015
       201          2 29/09/2015 30/09/2015
       102          3 29/09/2015

9 rows selected.

Or to get only the project assignments that will be active on October 10th, 2015:

> select person_id,
         project_id,
         assignment_period_start,
         assignment_period_end
  from   project_assignments
  AS OF PERIOD FOR assignment_period date'2015-10-10';

                      PERIOD     PERIOD
 PERSON_ID PROJECT_ID START      END
---------- ---------- ---------- ----------
       101          1 10/10/2015 12/10/2015
       102          3 29/09/2015

If we want to get only records that are active in a specific time range, we use the VERSIONS PERIOD clause. For example:

> select person_id,
         project_id,
         assignment_period_start,
         assignment_period_end
  from   project_assignments
  VERSIONS PERIOD FOR assignment_period
    between date'2015-09-21' and date'2015-09-23';

                      PERIOD     PERIOD
 PERSON_ID PROJECT_ID START      END
---------- ---------- ---------- ----------
       101          1 22/09/2015 24/09/2015
       102          1 20/09/2015 22/09/2015
       103          1            03/10/2015

Session-Level Control

We can also filter non-active records out in the session-level, affecting all the statements without changing them. We do it by calling the procedure DBMS_FLASHBACK_ARCHIVE.ENABLE_AT_VALID_TIME. This procedure has 3 options.
The first option it to pass the value ‘CURRENT’ to the LEVEL parameter (the first parameter), which means that all the subsequent statements in the session will consider only records that are active (or valid) at the execution time, for all the tables that were defined with the PERIOD FOR clause.

> exec dbms_flashback_archive.enable_at_valid_time('CURRENT')

PL/SQL procedure successfully completed.

> select person_id,
         project_id,
         assignment_period_start,
         assignment_period_end
  from   project_assignments;

                      PERIOD     PERIOD
 PERSON_ID PROJECT_ID START      END
---------- ---------- ---------- ----------
       101          1 28/09/2015 30/09/2015
       102          1 28/09/2015 30/09/2015
       103          1            03/10/2015
       101          2 27/09/2015 30/09/2015
       102          2 27/09/2015 30/09/2015
       103          2 28/09/2015 30/09/2015
       104          2 28/09/2015 30/09/2015
       201          2 29/09/2015 30/09/2015
       102          3 29/09/2015

9 rows selected.

The second option is to pass the value ‘ASOF’ (read “as of”) to the LEVEL parameter. In this case we need to pass also a concrete timestamp value to the QUERY_TIME parameter (the second parameter).

> exec dbms_flashback_archive.enable_at_valid_time('ASOF',date'2015-10-10')

PL/SQL procedure successfully completed.

> select person_id,
         project_id,
         assignment_period_start,
         assignment_period_end
  from   project_assignments;

                      PERIOD     PERIOD
 PERSON_ID PROJECT_ID START      END
---------- ---------- ---------- ----------
       101          1 10/10/2015 12/10/2015
       102          3 29/09/2015

And the third option is to pass the value ‘ALL’ to the LEVEL parameter, which means no filtering occurs in subsequent statements.

> exec dbms_flashback_archive.enable_at_valid_time('ALL')

PL/SQL procedure successfully completed.

> select person_id,
         project_id,
         assignment_period_start,
         assignment_period_end
  from   project_assignments;

                      PERIOD     PERIOD
 PERSON_ID PROJECT_ID START      END
---------- ---------- ---------- ----------
       101          1 16/09/2015 18/09/2015
       101          1 22/09/2015 24/09/2015
       101          1 28/09/2015 30/09/2015
       101          1 04/10/2015 06/10/2015
       101          1 10/10/2015 12/10/2015
       102          1 20/09/2015 22/09/2015
       102          1 28/09/2015 30/09/2015
       102          1 06/10/2015 08/10/2015
       102          1 14/10/2015 16/10/2015
       102          1 22/10/2015 24/10/2015
       103          1            03/10/2015
       101          2 27/09/2015 30/09/2015
       102          2 27/09/2015 30/09/2015
       103          2 28/09/2015 30/09/2015
       104          2 28/09/2015 30/09/2015
       201          2 29/09/2015 30/09/2015
       101          3 26/09/2015 27/09/2015
       102          3 29/09/2015

18 rows selected.

Warning:
if the value you pass to the LEVEL parameter of dbms_flashback_archive.enable_at_valid_time is other than ‘CURRENT’, ‘ASOF’ or ‘ALL’, it will NOT change anything and the previous setting will still take effect.
The dangerous part is that the procedure does not throw any exception; it completes successfully (doing nothing).
And these 3 values are case-sensitive. Passing ‘current’ or ‘AsOf’, for example, will silently do nothing.

There are two sides to every coin

We’ve just seen that Temporal Validity adds hidden columns to tables and hidden predicates to SQL statements. This is cool, but also dangerous. When things are hidden, it is easy to forget them, but we shouldn’t. See what I’ve already written about that with regards to In-Database Archiving in part 3.

Conclusion

We saw in this part of the Write Less with More series that Temporal Validity allows us to write less application code.

The next post will be about another new Oracle 12c feature – Row Limiting.

This post is part 3 of the Write Less with More series of blog posts, focusing on new features of Oracle 12c that allow us developers to write less than in previous versions in order to achieve the same functionality. Each part is dedicated for one new feature that solves some common development task. For more details, including the setup of the examples used throughout the series and the list of development tasks that drive the series, please see part 1.

All the Series Parts

Published

Part 1 – write less configuration with SQL*Loader Express Mode
Part 2 – write less application code with Identity Columns
Part 3 – write less application code with In-Database Archiving
Part 4 – write less application code with Temporal Validity
Part 5 – write less code in SQL with Row Limiting
Part 6 – write less “inappropriately located” code with SELECT FROM Package-Level Collection Types
Part 7 – write less code in SQL with Lateral Inline Views
Part 8 – write less “inappropriately located” code with PL/SQL in the WITH Clause

Task #3

We don’t want to “really” delete obsolete PROJECTS records, as we may have PROJECT_ASSIGNMENTS records that reference them (via a foreign key constraint). We just want to “hide” them, or, as it is often called, “logically delete” them.

A Pre-12c Solution

We can add a column to the PROJECTS table – IS_DELETED – which contains either 0 (representing an active project) or 1 (an obsolete project):

ALTER TABLE projects ADD 
  is_deleted NUMBER(1) DEFAULT 0 NOT NULL 
  CHECK (is_deleted IN (0,1)); 

Now, since usually we want to hide the obsolete projects, we can rename the table to, say, ALL_PROJECTS, and create a view that exposes only the active records:

RENAME projects TO all_projects;
CREATE VIEW projects AS
  SELECT * 
  FROM all_projects 
  WHERE is_deleted=0;

By naming the view PROJECTS, just as the table’s original name, we make all the existing references to PROJECTS throughout the application (or applications) see only active projects. Now we only need to handle the rare cases where obsolete projects should be seen, by using ALL_PROJECTS in the relevant SQL statements.

A 12c Solution

In-Database Archiving

In our case we can recreate the table like this:

drop table projects cascade constraints;
create table projects (
       project_id integer not null constraint projects_pk primary key,
       project_name varchar2(100) not null,
       status number(1) not null,
       last_days_to_show_in_reports integer not null
)
ROW ARCHIVAL;

When we DESCRIBE the table, we see only the columns we explicitly defined:

> desc projects
 Name                                         Null?    Type
 -------------------------------------------- -------- ------------------------------
 PROJECT_ID                                   NOT NULL NUMBER(38)
 PROJECT_NAME                                 NOT NULL VARCHAR2(100)
 STATUS                                       NOT NULL NUMBER(1)
 LAST_DAYS_TO_SHOW_IN_REPORTS                 NOT NULL NUMBER(38)

But if we select from USER_TAB_COLS we discover one more column – ORA_ARCHIVE_STATE:

> select column_name, data_type, column_id, 
         hidden_column, char_length, data_default
  from user_tab_cols 
  where table_name = 'PROJECTS';

COLUMN_NAME                  DATA_TYPE   COLUMN_ID HIDDEN_CO CHAR_LENGTH DATA_DEFAULT
---------------------------- ---------- ---------- --------- ----------- ------------
PROJECT_ID                   NUMBER              1 NO                  0
PROJECT_NAME                 VARCHAR2            2 NO                100
STATUS                       NUMBER              3 NO                  0
LAST_DAYS_TO_SHOW_IN_REPORTS NUMBER              4 NO                  0
ORA_ARCHIVE_STATE            VARCHAR2              YES              4000 0

Note: there is a major difference between the data dictionary views USER_TAB_COLUMNS and USER_TAB_COLS: while the former exposes only the explicitly defined columns, the latter includes also system-generated columns.

Let’s fill the table with some records, without specifying values to the hidden ORA_ARCHIVE_STATE column:

insert into projects values (1,'Project A',1,2);
insert into projects values (2,'Project B',2,3);
insert into projects values (3,'Project C',1,4);
insert into projects values (4,'Project D',2,3);
commit;

Hidden columns are quite shy. By default, they are not shown:

> select * from projects;

                                      LAST_DAYS_TO
PROJECT_ID PROJECT_NAME     STATUS SHOW_IN_REPORTS
---------- ------------ ---------- ---------------
         1 Project A             1               2
         2 Project B             2               3
         3 Project C             1               4
         4 Project D             2               3

But if we explicitly select ORA_ARCHIVE_STATE we can see its value, which is, at this point, the default – ‘0’:

> select p.*,ORA_ARCHIVE_STATE from projects p;

                                      LAST_DAYS_TO
PROJECT_ID PROJECT_NAME     STATUS SHOW_IN_REPORTS ORA_ARCHIVE_STATE
---------- ------------ ---------- --------------- -----------------
         1 Project A             1               2 0
         2 Project B             2               3 0
         3 Project C             1               4 0
         4 Project D             2               3 0

Now, let’s “delete” two records. Not a “real” deletion, only a logical one. We do it by updating ORA_ARCHIVE_STATE to ‘1’ (or any other non-zero value):

> update projects set ORA_ARCHIVE_STATE='1' where project_id in (1,3);

2 rows updated.

And now, if we select again from the table, with no (explicit) condition, we’ll see only the “remaining” two records. Remember that we didn’t actually DELETE any records, and still:

> select * from projects; 

                                      LAST_DAYS_TO
PROJECT_ID PROJECT_NAME     STATUS SHOW_IN_REPORTS
---------- ------------ ---------- ---------------
         2 Project B             2               3
         4 Project D             2               3 

How did it happen? Let’s check the execution plan:

> select * from table(dbms_xplan.display_cursor());

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------
SQL_ID  dcthaywgmzra7, child number 1
-------------------------------------
select * from projects

Plan hash value: 2188942312

------------------------------------------------------------------------------
| Id  | Operation         | Name     | Rows  | Bytes | Cost (%CPU)| Time     |
------------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |          |       |       |     3 (100)|          |
|*  1 |  TABLE ACCESS FULL| PROJECTS |     4 |  8372 |     3   (0)| 00:00:01 |
------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   1 - filter("PROJECTS"."ORA_ARCHIVE_STATE"='0')

And here it is. We can see that Oracle added a hidden condition: ORA_ARCHIVE_STATE=’0′.

ROW ARCHIVAL VISIBILITY

We saw the default behavior. If we want to see all the records, including the archived ones, we should change the ROW ARCHIVAL VISIBILITY setting of the session to ALL:

alter session set ROW ARCHIVAL VISIBILITY = ALL;

If we repeat the query now, we’ll see all the records of the table. Oracle will not add the behind-the-scenes condition on ORA_ARCHIVE_STATE this time:

> select p.*,
         case ora_archive_state when '0' then 'Active' else 'Deleted' end is_active
  from projects p;

                                      LAST_DAYS_TO
PROJECT_ID PROJECT_NAME     STATUS SHOW_IN_REPORTS IS_ACTIVE
---------- ------------ ---------- --------------- ---------------------
         1 Project A             1               2 Deleted
         2 Project B             2               3 Active
         3 Project C             1               4 Deleted
         4 Project D             2               3 Active

> select * from table(dbms_xplan.display_cursor());

PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------
SQL_ID  bwkf66qbwkaqt, child number 0
-------------------------------------
select p.*,        case ora_archive_state when '0' then 'Active' else
'Deleted' end is_active from projects p

Plan hash value: 2188942312

------------------------------------------------------------------------------
| Id  | Operation         | Name     | Rows  | Bytes | Cost (%CPU)| Time     |
------------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |          |       |       |     3 (100)|          |
|   1 |  TABLE ACCESS FULL| PROJECTS |     4 |  8372 |     3   (0)| 00:00:01 |
------------------------------------------------------------------------------

To get back to the default behavior we should change the ROW ARCHIVAL VISIBILITY setting of the session to its default value – ACTIVE:

alter session set ROW ARCHIVAL VISIBILITY = ACTIVE;

> select * from projects; 


                                      LAST_DAYS_TO
PROJECT_ID PROJECT_NAME     STATUS SHOW_IN_REPORTS
---------- ------------ ---------- ---------------
         2 Project B             2               3
         4 Project D             2               3 

The Bad News

It seems that In-Database Archiving is not supported in PL/SQL at the moment (12.1.0.2).

> var rc refcursor
> begin
    open :rc for select p.*,ora_archive_state from projects p;
  end;
  /

PL/SQL procedure successfully completed.

> print rc

                                      LAST_DAYS_TO
PROJECT_ID PROJECT_NAME     STATUS SHOW_IN_REPORTS ORA_ARCHIVE_STATE
---------- ------------ ---------- --------------- -----------------
         1 Project A             1               2 1
         2 Project B             2               3 0
         3 Project C             1               4 1
         4 Project D             2               3 0

Update (20-Jun-2017): this bug has been fixed in the one-off patch 23080557, in patchset 12.1.0.2.170418 for Windows, and is marked as fixed in the future (as of now) version 12.2.0.218.1

There are two sides to every coin

We’ve just seen that In-Database Archiving adds hidden columns to tables and hidden predicates to SQL statements. This is cool, but also dangerous. When things are hidden, it is easy to forget them, but we shouldn’t.
For example, suppose that:

• there are many records in the PROJECTS table
• most of them are obsolete (i.e., their ORA_ARCHIVE_STATE != ‘0’)
• ROW ARCHIVAL VISIBILITY = ACTIVE (the default setting)

Since Oracle adds the condition ORA_ARCHIVE_STATE = ‘0’ to every query, and this filters most of the records out, we may want to create an index on ORA_ARCHIVE_STATE, or add ORA_ARCHIVE_STATE as an additional column to other indexes.

Why VARCHAR2(4000)?

As we saw, the column ORA_ARCHIVE_STATE is created as VARCHAR2(4000), which seems a bit extreme for storing Boolean values – representing that the row is either Active or Deleted. Since the condition that Oracle adds to filter “deleted” rows out is ORA_ARCHIVE_STATE = ‘0’, we can use any string other than ‘0’ to represent deleted rows. One benefit it gives is that we can specify some information regarding the deletion of each record (like the “deletion reason”). Still, I would prefer a NUMBER(1) column rather than VARCHAR2(4000).
This can also limit our ability to add ORA_ARCHIVE_STATE to indexes, as the maximum key length of indexes is limited. For example, suppose that the PROJECT_NAME column is defined as VARCHAR2(4000), and that we want to create a composite index on PROJECT_NAME and ORA_ARCHIVE_STATE. Using a standard block size of 8KB, the index creation will fail:

> alter table projects modify (project_name varchar2(4000));

Table altered.

> create index projects_i1 on projects (project_name,ora_archive_state);
create index projects_i1 on projects (project_name,ora_archive_state)
                            *
ERROR at line 1:
ORA-01450: maximum key length (6398) exceeded

Conclusion

We saw in this part of the Write Less with More series that In-Database Archiving allows us to write less application code.

The next post will be about another new Oracle 12c feature – Temporal Validity.

This post is part 2 of the Write Less with More series of blog posts, focusing on new features of Oracle 12c that allow us developers to write less than in previous versions in order to achieve the same functionality. Each part is dedicated for one new feature that solves some common development task. For more details, including the setup of the examples used throughout the series and the list of development tasks that drive the series, please see part 1.

All the Series Parts

Published

Part 1 – write less configuration with SQL*Loader Express Mode
Part 2 – write less application code with Identity Columns
Part 3 – write less application code with In-Database Archiving
Part 4 – write less application code with Temporal Validity
Part 5 – write less code in SQL with Row Limiting
Part 6 – write less “inappropriately located” code with SELECT FROM Package-Level Collection Types
Part 7 – write less code in SQL with Lateral Inline Views
Part 8 – write less “inappropriately located” code with PL/SQL in the WITH Clause

Task #2

The PROJECT_ASSIGNMENTS table has a surrogate* primary key – each record should have a unique number in the ASSIGNMENT_ID column. This number has no business meaning; the only requirement for such a key is that the number will be unique – that there will be no two records in the table with the same number.

* In a recent post about defining primary keys using Data Modeler, HeliFromFinland is nicely walking through the decision making process of choosing between natural and surrogate keys

 

A Pre-12c Solution

This is a very common scenario, and almost always involves a sequence in Oracle. This makes sense, as a sequence is an object that generates unique numbers.

Before version 12c of Oracle there is no built-in relationship between tables and sequences, although logically there is almost always a one-to-one relationship between a table and a sequence.

Note: with the lack of a built-in relationship between tables and sequences, it is a good practice to use naming convention to represent this relationship; for example, naming the sequence that is responsible for assigning numbers to the ABC table ABC_SEQ (this specific convention implies yet another rule – that the length of table names should not exceed 26 characters; otherwise, the related sequence name will exceed the maximum length allowed for Oracle objects – 30 characters). Obviously, if two or more columns in one table are populated from different sequences, a more complex naming convention will be used.

So, the first part of the solution for our task is creating a sequence:

CREATE SEQUENCE project_assignments_seq;

The second part is to use the sequence whenever a new record of PROJECT_ASSIGNMENTS is inserted. If we know for sure about all the places that insert records to the table (preferably a single place) we can use the sequence explicitly in those places. However, if we cannot guarantee this, we should make sure the sequence will be used (as stated specifically in the task requirement) implicitly. And the way to achieve this in a pre-12c solution is using a BEFORE INSERT trigger, like this:

CREATE TRIGGER project_assignments_bir_tr
  BEFORE INSERT ON project_assignments
  FOR EACH ROW
BEGIN
  :new.assignment_id := project_assignments_seq.nextval; -- assuming 11g; in earlier versions use SELECT...
END;
/

Note that with this solution the sequence is used for each and every new record. Even if a value is specified explicitly in the INSERT (or MERGE) statement, this value will be ignored and overridden by the trigger, using the next value generated by the sequence.

If the requirement would be weaker – allowing specifying values explicitly and using the sequence only when values are not specified explicitly – then the solution would be:

CREATE TRIGGER project_assignments_bir_tr
  BEFORE INSERT ON project_assignments
  FOR EACH ROW
  WHEN (new.assignment_id IS NULL)
BEGIN
  :new.assignment_id := project_assignments_seq.nextval;
END;
/

Note that this option may cause problems. If you use it, make sure to allocate different (non-overlapping) ranges of numbers to the sequence and to the explicitly specified values. Otherwise, conflicts may happen.

A 12c Solution

Identity Columns

The GENERATED ALWAYS Option

An identity column can behave in one of three different ways. In the first one – when the column is defined with the GENERATED ALWAYS AS IDENTITY option (or simply GENERATED AS IDENTITY) – the sequence is always used, and specifying explicit values is forbidden. Let’s use this option to implement our task:

drop table project_assignments purge;
create table project_assignments (
       assignment_id integer GENERATED AS IDENTITY constraint project_assignments_pk primary key,
       person_id integer not null constraint assignments_fk_people references people,
       project_id integer not null constraint assignments_fk_projects references projects
);

We’ll insert now two rows into the PROJECT_ASSIGNMENTS table, without specifying the ASSIGNMENT_ID column, and we’ll see that the rows get a unique ASSIGNMENT_ID value:

> insert into project_assignments (person_id,project_id) values (101,1);
1 row created.
> insert into project_assignments (person_id,project_id) values (102,2);
1 row created.
> select * from project_assignments ;

ASSIGNMENT_ID PERSON_ID  PROJECT_ID
------------- ---------- ----------
            1        101          1
            2        102          2

If we try to specify an ASSIGNMENT_ID value explicitly, we’ll get the following exception:

> insert into project_assignments (assignment_id,person_id,project_id) values (3,103,3);

ERROR at line 1:
ORA-32795: cannot insert into a generated always identity column

The GENERATED BY DEFAULT Option

If we do want the ability to specify the ASSIGNMENT_ID explicitly, and let Oracle assign a value automatically only when we do not specify it, we can use instead the GENERATED BY DEFAULT AS IDENTITY option. Let’s change the ASSIGNMENT_ID column to behave in this way:

> alter table project_assignments modify (
>   assignment_id generated BY DEFAULT as identity
> );
Table altered.

Retrying the last INSERT statement will succeed now:

> insert into project_assignments (assignment_id,person_id,project_id) values (3,103,3);
1 row created.

As I noted before, in the “A Pre-12c Solution” section, mixing manual and automatic assignments of values to the same column may cause conflicts, and unless there is a good reason to do it, I would refrain from it. As mentioned above, the implementation of identity columns uses sequences. In our example, this behind-the-scenes sequence has generated already the numbers 1 and 2, so the next time it will be used it will generate the number 3. However, in the third row that we inserted ASSIGNMENT_ID has the value 3, because we specified it explicitly. ASSIGNMENT_ID is the primary key of the table, so it cannot have the same value in two records. I feel an imminent conflict…

And indeed, trying to insert a record without specifying ASSIGNMENT_ID will fail on a unique constraint violation:

> insert into project_assignments (person_id,project_id) values (104,4);

ERROR at line 1:
ORA-00001: unique constraint (DEMO5.PROJECT_ASSIGNMENTS_PK) violated

Another attempt will succeed, as the sequence has already been incremented:

> insert into project_assignments (person_id,project_id) values (104,4);
1 row created.

 

The GENERATED BY DEFAULT ON NULL Option

Now, what will happen if we try to set ASSIGNMENT_ID to NULL during INSERT? An identity column is always mandatory – the column is defined as NOT NULL as part of the identity definition (just like it happens when we define a column as PRIMARY KEY) – and since we defined ASSIGNMENT_ID as GENERATED BY DEFAULT AS IDENTITY, the following statement will simply try to insert NULL to ASSIGNMENT_ID, and will fail:

> insert into project_assignments (assignment_id,person_id,project_id) values (null,201,1);

ERROR at line 1:
ORA-01400: cannot insert NULL into ("DEMO5"."PROJECT_ASSIGNMENTS"."ASSIGNMENT_ID")

We can use the third form of the identity clause – GENERATED BY DEFAULT ON NULL AS IDENTITY – which means that Oracle will generate values from the sequence whenever the identity column is not specified explicitly or it is specified explicitly with NULL. Let’s change the ASSIGNMENT_ID to behave in this way:

> alter table project_assignments modify (
>   assignment_id generated BY DEFAULT ON NULL as identity
> );

Table altered.

Retrying the last INSERT statement will succeed now:

> insert into project_assignments (assignment_id,person_id,project_id) values (null,201,1);
1 row created.

> select * from project_assignments;

ASSIGNMENT_ID PERSON_ID  PROJECT_ID
------------- ---------- ----------
            1        101          1
            2        102          2
            3        103          3
            4        104          4
            5        201          1

5 rows selected.

Note: DEFAULT ON NULL is actually an independent feature (new in 12c) – it can be used in the definition of any column, not only identity columns.

Note: It is also possible (as of 12c) to use a user-defined sequence as the default value of a numeric non-identity column. Identity columns, however, give a more complete solution.

Data Dictionary

Let’s see now what has changed in the data dictionary in order to represent identity columns. I’ll refer to the USER_* views, but the same changes are relevant to the corresponding ALL_* and DBA_* views as well.

USER_TABLES has a new column – HAS_IDENTITY – that contains YES if the table has an identity column and NO if not.

select table_name,HAS_IDENTITY from user_tables;

TABLE_NAME                     HAS_IDENTITY
------------------------------ ------------
PEOPLE                         NO
PROJECTS                       NO
PROJECT_ASSIGNMENTS            YES

The fact that USER_TABLES has such a Boolean attribute may hint that a table cannot have more than one identity column, and this is indeed one of the restrictions of this feature (which seems a very sensible restriction, in my opinion).

USER_TAB_COLUMNS has two new relevant columns: IDENTITY_COLUMN and DEFAULT_ON_NULL:

select column_name,data_type,nullable,
       column_id,IDENTITY_COLUMN,DEFAULT_ON_NULL 
from user_tab_columns
where table_name = 'PROJECT_ASSIGNMENTS'
order by column_id;

COLUMN_NAME   DATA_TYPE  NUL  COLUMN_ID IDENTITY_ DEFAULT_O
------------- ---------- --- ---------- --------- ---------
ASSIGNMENT_ID NUMBER     N            1 YES       YES
PERSON_ID     NUMBER     N            2 NO        NO
PROJECT_ID    NUMBER     N            3 NO        NO

A new view – USER_TAB_IDENTITY_COLS – contains the details of all the identity columns. In 12.1.0.1 this view contained TABLE_NAME, COLUMN_NAME, GENERATION_TYPE (which can be either ALWAYS or BY DEFAULT) and IDENTITY_OPTIONS (the configuration of the associated sequence). In 12.1.0.2 another column was added – SEQUENCE_NAME – which makes the relationship between the table column and the associated sequence quite formal.

select * From USER_TAB_IDENTITY_COLS;

                                  GENERATION
TABLE_NAME          COLUMN_NAME   TYPE       SEQUENCE_NAME IDENTITY_OPTIONS
------------------- ------------- ---------- ------------- -------------------------
PROJECT_ASSIGNMENTS ASSIGNMENT_ID BY DEFAULT ISEQ$$_111567 START WITH: 1, INCREMENT
                                                           BY: 1, MAX_VALUE: 9999999
                                                           999999999999999999999, MI
                                                           N_VALUE: 1, CYCLE_FLAG: N
                                                           , CACHE_SIZE: 20, ORDER_F
                                                           LAG: N

Let’s take a look at the objects we have in our schema at this point:

select object_type,object_name,object_id from user_objects;

OBJECT_TYPE  OBJECT_NAME                     OBJECT_ID
------------ ------------------------------ ----------
SEQUENCE     ISEQ$$_111567                      111568
TABLE        PEOPLE                             111530
INDEX        PEOPLE_PK                          111531
TABLE        PROJECTS                           111532
INDEX        PROJECTS_PK                        111533
TABLE        PROJECT_ASSIGNMENTS                111567
INDEX        PROJECT_ASSIGNMENTS_PK             111569

We can see the sequence that was created automatically by Oracle for supporting the ASSIGNMENT_ID identity column of the PROJECT_ASSIGNMENTS table. Note that the sequence_name contains the object_id of the associated table.

What happens when we drop the table?

drop table project_assignments;

select object_type,object_name,object_id from user_objects;

OBJECT_TYPE  OBJECT_NAME                     OBJECT_ID
------------ ------------------------------ ----------
INDEX        PROJECTS_PK                        111533
TABLE        PROJECTS                           111532
INDEX        PEOPLE_PK                          111531
TABLE        PEOPLE                             111530
SEQUENCE     ISEQ$$_111567                      111568

We can see that the table was dropped but the sequence was not. This may seem disappointing at first, but actually this is quite clever. The table was dropped but not purged, so it is still in the recycle bin (I’m using the default setting of the RECYCLEBIN parameter, which is “on”):

select object_name,original_name,operation,type from user_recyclebin;

OBJECT_NAME                    ORIGINAL_NAME          OPERATION TYPE
------------------------------ ---------------------- --------- ----------
BIN$lNxGd9cXTAuYpOa9kda26w==$0 PROJECT_ASSIGNMENTS    DROP      TABLE
BIN$qvScIgi+Rrel7veFMBxXsQ==$0 PROJECT_ASSIGNMENTS_PK DROP      INDEX

If we revive the table using FLASHBACK TABLE

flashback table project_assignments to before drop;

we can resume working with the table, and particularly with the identity column, because the sequence was not dropped:

insert into project_assignments (person_id,project_id) values (201,1);

select * from project_assignments;

ASSIGNMENT_ID  PERSON_ID PROJECT_ID
------------- ---------- ----------
            1        101          1
            2        102          2
            3        103          3
            4        104          4
            5        201          1
            6        201          1

If we drop the table and purge it (or purge the table from the recyclebin), then the associated sequence is dropped as expected:

drop table project_assignments purge;

select object_type,object_name,object_id from user_objects;

OBJECT_TYPE  OBJECT_NAME                     OBJECT_ID
------------ ------------------------------ ----------
TABLE        PEOPLE                             111530
INDEX        PEOPLE_PK                          111531
TABLE        PROJECTS                           111532
INDEX        PROJECTS_PK                        111533

select object_name,original_name,operation,type from user_recyclebin;

no rows selected

Sequence Configuration

When we create a sequence explicitly we can configure its attributes – from which number to start, the interval between numbers, the cache size, and more. When we create an identity column, we can similarly configure the implicitly-created sequence. For example, let’s recreate the PROJECT_ASSIGNMENTS table, but this time with a sequence that starts with 1000 and increments by 10:

create table project_assignments (
       assignment_id integer GENERATED AS IDENTITY 
                             (start with 1000 increment by 10)
                             constraint project_assignments_pk primary key,
       person_id integer not null  constraint assignments_fk_people references people,
       project_id integer not null constraint assignments_fk_projects references projects
);

insert into project_assignments (person_id,project_id) values (101,1);
insert into project_assignments (person_id,project_id) values (102,2);
insert into project_assignments (person_id,project_id) values (103,3);

select * from project_assignments;

ASSIGNMENT_ID  PERSON_ID PROJECT_ID
------------- ---------- ----------
         1000        101          1
         1010        102          2
         1020        103          3

DROP IDENTITY

A non-identity column cannot become an identity column. The opposite, however, is possible: an identity column can become a non-identity column, and its associated sequence is automatically dropped:

alter table project_assignments modify assignment_id DROP IDENTITY;
select * from user_tab_identity_cols;

no rows selected

select object_type,object_name,object_id from user_objects;

OBJECT_TYPE  OBJECT_NAME                     OBJECT_ID
------------ ------------------------------ ----------
TABLE        PEOPLE                             111530
INDEX        PEOPLE_PK                          111531
TABLE        PROJECTS                           111532
INDEX        PROJECTS_PK                        111533
TABLE        PROJECT_ASSIGNMENTS                111570
INDEX        PROJECT_ASSIGNMENTS_PK             111572

START WITH LIMIT VALUE

Let’s see one last option that is possible only because there is a real association between the table column and the sequence. I’ll drop the PROJECT_ASSIGNMENTS table and recreate it once more, with the GENERATED BY DEFAULT AS IDENTITY clause, and insert some records with explicit values in ASSIGNMENT_ID:

drop table project_assignments purge;
create table project_assignments (
       assignment_id integer GENERATED BY DEFAULT AS IDENTITY constraint project_assignments_pk primary key,
       person_id integer not null constraint assignments_fk_people references people,
       project_id integer not null constraint assignments_fk_projects references projects
);
insert into project_assignments (assignment_id,person_id,project_id) values (18,101,1);
insert into project_assignments (assignment_id,person_id,project_id) values (22,102,2);
insert into project_assignments (assignment_id,person_id,project_id) values (34,103,3);

select * from project_assignments;

ASSIGNMENT_ID  PERSON_ID PROJECT_ID
------------- ---------- ----------
           18        101          1
           22        102          2
           34        103          3

Suppose that now I want to convert the identity column to be GENERATED ALWAYS AS IDENTITY. Since the table already contains records, I need to prevent collisions between the existing values and the future to-be-automatically-generated values, and therefore I’d like to start the sequence from a number that is higher than all the existing ASSIGNMENT_ID values (in our case, 35).

We can make Oracle doing it automatically, by using the START WITH LIMIT VALUE clause:

alter table project_assignments modify assignment_id 
generated always as identity (START WITH LIMIT VALUE);

insert into project_assignments (person_id,project_id) values (104,4);
select * from project_assignments;

ASSIGNMENT_ID  PERSON_ID PROJECT_ID
------------- ---------- ----------
           18        101          1
           22        102          2
           34        103          3
           35        104          4

Conclusion

We saw in this part of the Write Less with More series that Identity Columns allow us to write less application code.

The next post will be about another new Oracle 12c feature – In-Database Archiving.