Question

I came across this example on SO which gives a solution to create a unique index by ignoring nulls. However, I want to extend it and I am not able reach to a solution.

I have a composite index for 3 columns of a table (there are other 10 columns in the table). These columns are not part of PK. Of these 3 columns, 2 will always hold some value and 3rd may be NULL. I have huge test data, and there are many inserts with 2 of the columns with same value and 3rd column NULL. These all inserts work well for PostgreSQL, however Oracle complains. For my test cases to work, I think the simplest solution that I think is to try a unique index for Oracle that would work as it works in PostgreSQL.

Precisely: I want a construct of following kind, not sure how to combine col1 + col2 + col3

create unique index tbl_idx on tbl (nvl2(col3, col1 + col2, col1 + col2 + col3))

I am using liquibase. Index is created in following way –

<changeSet dbms="postgresql,oracle" author="abc" id="222">
    <createIndex indexName="Index_7" schemaName="ss" tableName="Users" unique="true">
        <column name="idOrganization"/>
        <column name="strUsername"/>
        <column name="strNotDeleted"/>
    </createIndex>
</changeSet>

I am using liquibase to create my test data, here are two insert statements

<insert schemaName="ss" tableName="Users">
    <column name="strUsername" value="user1" />
    <column name="idUser" valueNumeric="20741" />
    <column name="idOrganization" valueNumeric="4" />
    <column name="strFirstName" value="user" />
    <column name="strLastName" value="one" />
    <column name="strEmail" value="[e-mail]" />
    <column name="strNotDeleted" />
</insert>
<insert schemaName="ss" tableName="Users">
    <column name="strUsername" value="user1" />
    <column name="idUser" valueNumeric="20771" />
    <column name="idOrganization" valueNumeric="4" />
    <column name="strFirstName" value="user" />
    <column name="strLastName" value="one" />
    <column name="strEmail" value="[e-mail]" />
    <column name="strNotDeleted" />
</insert>

These 2 inserts work fine for PostgreSQL, however fail for Oracle with error “Index_7 constraint violation”.

Answer 1

If the goal is only to prevent duplicates where strNotDeleted is set to a non-NULL value, then you want a function-based index like this

SQL> create table users(
  2    idOrganization number,
  3    strUsername    varchar2(100),
  4    strNotDeleted  varchar2(3)
  5  );

Table created.


SQL> create unique index idx_users
  2      on users( (case when strNotDeleted is not null
  3                      then idOrganization
  4                      else null
  5                  end),
  6                (case when strNotDeleted is not null
  7                      then strUsername
  8                      else null
  9                 end) );

Index created.

This allows the two rows you mention in your question to be inserted

SQL> insert into users values( 4, 'user', null );

1 row created.

SQL> insert into users values( 4, 'user', null );

1 row created.

You can insert one row where the strNotNull column is set to a non-NULL value

SQL> insert into users values( 4, 'user', 'Yes' );

1 row created.

But you then can’t insert a second such row

SQL> insert into users values( 4, 'user', 'Yes' );
insert into users values( 4, 'user', 'Yes' )
*
ERROR at line 1:
ORA-00001: unique constraint (SCOTT.IDX_USERS) violated

Behind the scenes, an Oracle b*-tree index does not index completely NULL entries. The two CASE statements ensure that the index only has entries for idOrganization and strUsername if strNotDeleted is not NULL. If strNotDeleted is NULL, then both CASE statements evaluate to NULL and no entry is made in the index. Conceptually, it’s similar to a partial index in other databases which allows you to specify a WHERE clause on your index so that you only index “interesting” rows.