=head1 NAME

Tangram::Relational::Mappings - Mapping inheritance

=head1 DESCRIPTION

There are many ways of representing inheritance relationships in a relational
database. This document describes three popular ways and how Tangram supports
them.

=head1 STRATEGIES FOR MAPPING INHERITANCE

Inheritance is a concept that has no equivalent in the relational
world. However, it is possible to implement it by using strict
disciplines and a combination of relational features like tables and
foreign keys.

One of the paramount issues about mapping inheritance is how well the
mapping supports polymorphism.  Any Object-Oriented persistence
facility that deserves its name needs to allow the retrieval of all
the Fruits, and return a heterogeneous collection of Apples, Oranges
and Bananas.  Also, it must perform this operation in an efficient
manner.  In particular, polymorphic retrieval should not cost one
SELECT per retrieved object.

A secondary - yet important - issue is how well the mapping plays by
the rules of orthogonal orthodoxy.

Another issue we'll examine is how well the mapping supports 'complex'
queries, that is, queries that involve several objects.

Three strategies are in common use, that go by the name Vertical,
Horizontal and Filtered mapping.  They all have advantages and
disadvantages.

The following sections describe the three strategies in details. They
make use of a simple object model to illustrate the mappings.


                          +---------------------+
                          |        Person       |
                          |      {abstract}     |
     +---------<------- 1 +---------------------+
     |                    | name: string        |
     |                    +---------------------+
     |                               |
     |                               ^
     |                               |
     |            +------------------+---------------------+
     |            |                                        |
     |   +---------------+                        +-----------------+
     V   | NaturalPerson |                        |    LegalPerson  |
     |   +---------------+                        +-----------------+
     |   |  age: integer |                        | form: string    |
     |   +---------------+                        +-----------------+
     |
     |
     |
     |                    +---------------------+
     +-------->-------- * |        Vehicle      |
                          |       {abstract}    |
                          +---------------------+
                          | make: string        |
                          +---------------------+
                                     |
                                     ^
                                     |
                  +------------------+-------------------+
                  |                                      |
         +---------------+                      +-----------------+
         |      Car      |                      |      Plane      |
         +---------------+                      +-----------------+
         | plate: string |                      | ident: string   |
         +---------------+                      +-----------------+


=head1 Horizontal Mapping

=head1 description

Each I<concrete> class is mapped onto a single table. Each row in the
table describes the persistent state of one object.

The attributes are mapped onto columns, usually one column per
attribute but not necessarily. For example, collections may be stored
elsewhere (for example on a link table) and thus require no column on
the class' table.

In effect, the database looks like this:

         +---------------+
         | NaturalPerson |
         +------+--------+-------+------+
         |  id  |  name          | age  |
         ================================
         |  17  | Bill Gates     |  46  |
         +------+----------------+------+
         |  23  | Georges Bush   |  50  |
         +------+----------------+------+


         +-------------+
         | LegalPerson |
         +------+------+---------+------+
         |  id  |  name          | form |
         ================================
         |  36  |  Microsoft     |  Inc |
         +------+----------------+------+


         +------+
         | Car  |
         +------+-------+----------------+--------+
         |  id  | owner	|  make          | plate  |
         ==========================================
         |  12  |  17  	| Saab           | BILL-1 |
         +------+-------+----------------+--------+
         |  50  |  36  	| Miata          | MS-001 |
         +------+-------+----------------+--------+
         |  51  |  36  	| Miata          | MS-002 |
         +------+-------+----------------+--------+


         +-------+
         | Plane |
         +------++-----+----------------+--------+
         |  id  | owner|  make          | ident  |
         =========================================
         |  29  |  23  |  Boeing        | AF-001 |
         +------+------+----------------+--------+

=head2 advantages

Polymorphic retrieval costs one SELECT per concrete conforming class;
retrieving all the Persons costs two SELECTs. These SELECTs, however,
don't use joins - an expensive operation. In our example, retrieving
all the Persons requires the following two SELECTs:

   SELECT id, name, age FROM NaturalPerson
   SELECT id, name, form FROM LegalPerson

=head2 disadvantages

This mapping is reasonable with regard to relational orthodoxy, but
not perfect: the 'name' column is present on two different tables,
with the same semantic.

The biggest drawback, however, happens when you try to perfrom complex
queries.  Suppose oyu want to retrieve all the Persons (Natural- or
Legal-) that own a Vehicle of make 'Saab' (be it a Car or a
Plane). Sticking with equijoins, the cost of the operation is four
SELECTs:

   SELECT NaturalPerson.id, NaturalPerson.name, NaturalPerson.age
      FROM NaturalPerson, Car
      WHERE Car.owner = NaturalPerson.id

   SELECT NaturalPerson.id, NaturalPerson.name, NaturalPerson.age
      FROM NaturalPerson, Plane
      WHERE Plane.owner = NaturalPerson.id

   SELECT LegalPerson.id, LegalPerson.name, LegalPerson.form
      FROM LegalPerson, Car
      WHERE Car.owner = LegalPerson.id

   SELECT LegalPerson.id, LegalPerson.name, LegalPerson.form
      FROM LegalPerson, Plane
      WHERE Plane.owner = LegalPerson.id

When the depth of the hierarchies increase, the combinatory explosion
makes complex queries prohibitive.

=head1 Vertical Mapping

=head2 description

Each class has its corresponding table, which contains only the class'
direct fields. In other words, the table doesn't store the inherited
fields. Both concrete and abstract classes get a table. The state of
an object is thus scattered over several tables.

For example:

         +--------+
         | Person |
         +------+-+------+-------+
         |  id  |  name          |
         =========================
         |  17  | Bill Gates     |
         +------+----------------+
         |  23  | Georges Bush   |
         +------+----------------+
         |  36  | Microsoft      |
         +------+----------------+


         +---------------+     +-------------+
         | NaturalPerson |     | LegalPerson |
         +------+--------+     +-------+-----++
         |  id  | age    |     |  id   | form |
         =================     ================
         |  17  |  46    |     |  36   |  Inc |
         +------+--------+     +-------+------+
         |  23  |  50    |
         +------+--------+

         +---------+
         | Vehicle |
         +------+--+----+----------------+
         |  id  | owner |  make          |
         =================================
         |  12  | 17    | Saab           |
         +------+-------+----------------+
         |  29  | 23    | AF-001         |
         +------+-------+----------------+
         |  50  | 36    | Miata          |
         +------+-------+----------------+
         |  51  | 36    | Miata          |
         +------+-------+----------------+

         +------+              +-------+
         | Car  |              | Plane |
         +------++--------+    +-------+--------+
         |  id   | plate  |    |  id   | ident  |
         ==================    ==================
         |  12   | BILL-1 |    |  29   | AF-001 |
         +-------+--------+    +-------+--------+
         |  50   | MS-001 |
         +-------+--------+
         |  51   | MS-002 |
         +-------+--------+

Polymorphic retrieval is achieved by issuing one SELECT per concrete
conforming class; retrieving In our example, retrieving all the
Persons requires the following two SELECTs:

   SELECT Person.id, Person.name, NaturalPerson.age
      FROM Person, NaturalPerson
      WHERE Person.id = NaturalPerson.id

   SELECT Person.id, Person.name, LegalPerson.form
      FROM Person, LegalPerson
      WHERE Person.id = LegalPerson.id

This mapping sometimes needs an extra column that carries a type
identifier. In our example, we take the very resonable assumption that
Person is an abstract class. Had we decided to allow 'pure' Persons,
we would have been faced with the following problem: the Person table
would contain rows that describe pure Persons, but also rows that
describe the Person part of Natural- and LegalPersons. We would need
to filter those incomplete objects out when retrieving the pure
Persons. Thus the Person table would look like this:

         +--------+
         | Person |
         +-----+--+---+----------------+
         | id  | type |  name          |
         ===============================
         | 13  | 1    | Pure Person    |
         +-----+------+----------------+
         | 17  | 2    | Bill Gates     |
         +-----+------+----------------+
         | 23  | 2    | Georges Bush   |
         +-----+------+----------------+
         | 36  | 3    | Microsoft      |
         +-----+------+----------------+

In this case, we need an extra SELECT for retrieving pure Persons:

   SELECT Person.id, Person.name
      FROM Person
      WHERE Person.type IN (1)

=head2 advantages

From the relational point of view, this mapping is excellent: the
resulting database is in third normal form.

This mapping also supports complex queries very well. Take the Saab
owners example again: we don't need to involve the Car nor Plane
tables in the query. As a result, two SELECTs suffice:

   SELECT Person.id, Person.name, NaturalPerson.age
      FROM Person, NaturalPerson, Vehicle
      WHERE Person.id = NaturalPerson.id AND Vehicle.owner = Person.id

   SELECT Person.id, Person.name, LegalPerson.form
      FROM Person, LegalPerson, Vehicle
      WHERE Person.id = LegalPerson.id AND Vehicle.owner = Person.id


=head2 disadvantages

The mapping potentially has the highest performance cost: it requires
multiple SELECTs like the horizontal mapping, but in addition, these
SELECTs use joins.


=head1 Filtered Mapping

=head2 description

Entire hierarchies are mapped onto a single table. Two rows may
describe objects of different types, maybe completely unrelated. The
set of columns is the uperset of all the columns needed by all the
attributes of any of the classes involved in the mapping.

A special 'type' column contains an value that uniquely identifies the
concrete class of the object described by the row.

All the columns related to attributes that don't occur in all the
classes must be declared as NULLABLE. Indeed, the table may contain
mostly NULL values.

In our example, the database may look either like this:

         +---------+
         | Persons |
         +-----+---+--+----------------+------+------+
         | id  | type |  name          | age  | form |
         =============================================
         | 17  |  1   | Bill Gates     |  46  | NULL |
         +-----+------+----------------+------+------+
         | 23  |  1   | Georges Bush   |  50  | NULL |
         +-----+------+----------------+------+------+
         | 36  |  2   | Microsoft      | NULL |  Inc |
         +-----+------+----------------+------+------+

         +---------+
         | Persons |
         +-----+---+--+----------------+------+------+
         | id  | type |  name          | age  | form |
         =============================================
         | 17  |  1   | Bill Gates     |  46  | NULL |
         +-----+------+----------------+------+------+
         | 23  |  1   | Georges Bush   |  50  | NULL |
         +-----+------+----------------+------+------+
         | 36  |  2   | Microsoft      | NULL |  Inc |
         +-----+------+----------------+------+------+
         | 36  |  2   | Microsoft      | NULL |  Inc |
         +-----+------+----------------+------+------+

         +----------+
         | Vehicles |
         +-----+----+-+-------+----------------+--------+--------+
         | id  | type | owner |  make          | plate  | ident  |
         =========================================================
         | 12  |  3   |  17   | Saab           | BILL-1 | NULL   |
         +-----+------+-------+----------------+--------+--------+
         | 29  |  4   |  23   | Boeing         | NULL   | AF-001 |
         +-----+------+-------+----------------+--------+--------+
         | 50  |  3   |  36   | Miata          | MS-001 | NULL   |
         +-----+------+-------+----------------+--------+--------+
         | 51  |  3   |  36   | Miata          | MS-002 | NULL   |
         +-----+------+-------+----------------+--------+--------+

Retrieving all the Persons requires only one SELECT:

   SELECT id, name, age, form FROM Persons 

When retrieving NaturalPersons we must take care to filter out the
rows that belog to LegalPersons:

   SELECT id, name, age FROM Persons WHERE type = 1

We may even decide to place unrelated hierarchies on the same table:

   +---------+
   | Objects |
   +-----+---+--+---------------+------+------+--------+--------+--------+
   | id  | type |  name         | age  | form | make   | plate  | ident  |
   =======================================================================
   | 17  |  1   | Bill Gates    |  46  | NULL | NULL   | NULL   | NULL   |
   +-----+------+---------------+------+------+--------+--------+--------+
   | 23  |  1   | Georges Bush  |  50  | NULL | NULL   | NULL   | NULL   |
   +-----+------+---------------+------+------+--------+--------+--------+
   | 36  |  2   | Microsoft     | NULL | Inc  | NULL   | NULL   | NULL   |
   +-----+------+---------------+------+------+--------+--------+--------+
   | 12  |  3   | NULL          | NULL | NULL | Saab   | BILL-1 | NULL   |
   +-----+------+---------------+------+------+--------+--------+--------+
   | 29  |  4   | NULL          | NULL | NULL | Boeing | NULL   | AF-001 |
   +-----+------+---------------+------+------+--------+--------+--------+
   | 50  |  3   | NULL          | NULL | NULL | Miata  | MS-001 | NULL   |
   +-----+------+---------------+------+------+--------+--------+--------+
   | 51  |  3   | NULL          | NULL | NULL | Miata  | MS-002 | NULL   |
   +-----+------+---------------+------+------+--------+--------+--------+

=head2 advantages

Polymorphic retrieval costs exactly one SELECT, regardless of the
number of conforming types. Thus this mapping potentially is the most
efficient.

=head2 disadvantages

This mapping is very questionable according to relational
orthodoxy. Even if one decides to forgo these rules, using such a
mapping takes away many of the interesting features offered by modern
RDBM systems. Because nearly all the columns must allow NULL values,
we cannot take advantage of features like referential integrity
constraints, domain constraints, indexes, etc.

Also, as the table becomes cluttered with NULL values, the relative
number of significant columns in any given row tends towards zero: we
may end up retrieving rows consisting of a little information swimming
in a sea of NULLs.

In effect, this mapping may end up hindering performance instead of
improving it in presence of deep hierarchies with many attributes.

=head1 MAPPINGS SUPPORTED BY TANGRAM

Tangram supports both vertical mapping and filtered mapping, and any
hybrid of the two.

The 'table' attribute in the class description in the Schema can be
used to put the state of several classes on the same table. The table
name defaults to the class name, resulting in a vertical mapping.

For example, the following schema:

   Tangram::Relational->schema( {
        classes =>
          [ Person =>
            {
             table => 'Persons',
             fields => { string => [ qw( name ) ] }
            },
            
            NaturalPerson =>
            {
             table => 'Persons',
             fields => { int => [ qw( age ) ] }
            },
            
            LegalPerson =>
            {
             table => 'Persons',
             fields => { string => [ qw( form ) ] }
            }
          ] } );

...specifies a pure filtered mapping for the Person hierarchy:

   CREATE TABLE Persons
   (
     id INTEGER NOT NULL,
     PRIMARY KEY( id ),
     type INTEGER NOT NULL,
     form VARCHAR(255) NULL,
     age INT NULL,
     name VARCHAR(255) NULL
   );

The following schema:
   
   Tangram::Relational->schema( {
        classes =>
          [ Person =>
            {
             table => 'Person',
             fields => { string => [ qw( name ) ] }
            },
            
            NaturalPerson =>
            {
             table => 'NaturalPerson',
             fields => { int => [ qw( age ) ] }
            },
            
            LegalPerson =>
            {
             table => 'Person',
             fields => { string => [ qw( form ) ] }
            }
          ] } );

...gives NaturalPerson its own table, but LegalPerson shares the
Person table:

   CREATE TABLE Person
   (
     id INTEGER NOT NULL,
     PRIMARY KEY( id ),
     type INTEGER NOT NULL,
     form VARCHAR(255) NULL,
     name VARCHAR(255) NULL
   );

   CREATE TABLE NaturalPerson
   (
     id INTEGER NOT NULL,
     PRIMARY KEY( id ),
     type INTEGER NOT NULL,
     age INT NULL
   );