PostgreSQL is the world’s most advanced open source database, and per the PostgreSQL Wikipedia page it is an object-relational database management system (ORDBMS) with an emphasis on extensibility and standards compliance.

In this article, we try to understand why would PostgreSQL be named an object-relational thing. What is Object Oriented Programming and how does that apply to a database system?

Object Orientation in Progamming Languages

Erik Naggum was a Norwegian computer programmer recognized for his work in the fields of SGML, Emacs and Lisp. He had strong opinions and shared them on Usenet. One of my favorite piece from Erik Naggum is titled The Long, Painful History of Time and dives into our calendars history, and how to write clever handling of dates and timestamps and time zone when you have enough historical background to actually understand what you’re dealing with.

Spoiler:

The Gregorian calendar improved on the quadrennial leap years in the Julian calendar by making only every fourth centennial a leap year, […] but the simplicity of the scheme is quite amazing: a 400-year cycle not only starts 2000-03-01 (as it did 1600-03-01), it contains an even number of weeks: 20,871. This means that we can make do with a single 400-year calculation for all time within the Gregorian calendar with respect to days of week, leap days, etc.

Today though, I wanted to use Naggum’s rant about the Common Lisp Object System and how it compares to other ones, such as the ones in Smalltalk or in the Java, C++, Python, Ruby, and other mainstream programming languages. The usenet article Re: How much use of CLOS? is a must read, and mentions that there are two real approaches to object-orientation:

  • Message-Passing

    The first is known as message-passing. You send an object a message and ask it to deal with it. […] The meaning of the message is local to the object, which inherits it from the class of which it is an instance, which may inherit it from superclasses of that class. In other words, you have no idea what happens when you send a message to an object, how many arguments it needs to be happy or anything. […]

    This can get very messy, and it is therefore deemed appropriate to “clean up” this mess by adding compiler checks that that message on that object really takes that argument list. This is the core mistake in the message-passing model. Smalltalk did not make this core mistake, which means that people from the non-OO-“OO”-language camps get all uneasy about Smalltalk.

  • Generic Functions

    The second approach is generic functions. A generic function has one definition of its semantics, its argument list, and is only specialized on particular types of arguments. It can be specialized on any argument or any number of arguments in the argument list, on any type each. […]

    You do not get to make the class implement methods on generic functions. There is no way to add “method” definitions to a defclass. […]

Naguum then writes the following summary:

What does this mean with respect to when you think you use OO? In the message-passing paradigm, you need to define a class to get any methods at all and you have to use those methods on instances of that class. In the generic function paradigm, you can define generic functions without ever defining any classes. This tends to blow “OO” people’s mind.

If you’re interested in generic functions in the Common Lisp Object System, one of the best resource available is the book Practical Common Lisp from Peter Seibel, and in particular the chapters Object Reorientation: Generic Functions and Object Reorientation: Classes.

If the only Object Model you’ve ever worked with looks like the Java, C++, PHP, and Python one, prepare for your mind to be blown.

Object Orientation in PostgreSQL

PostgreSQL is the first RDBMS to have put emphasis on extensible data types and plugability. As a result, 30 years later, we have many awesome extensions available, such as PostGIS, Citus, prefix or ip4r, to name but a few.

In the PostgreSQL design, when it comes to data types, functions, operators and index support, nothing is hardcoded. Even with the simplest queries, the planner has to lookup the catalogs to understand what’s going on.

In the following query, for example, PostgreSQL’s planner has to lookup the catalog tables in order to know what you mean, and any extension could be providing new data types, operators, functions and index support.

select x, 
       1 + x as "1+",
       '127.0.0.1'::inet + x as "ip address",
       date 'today' + x as date
  from (values (0), (1), (2), (3)) as t(x);

The query has the following result, where you can see that the + operator implementation obviously depends on the datatype:

 x │ 1+ │ ip address │    date    
═══╪════╪════════════╪════════════
 0 │  1 │ 127.0.0.1  │ 2018-03-22
 1 │  2 │ 127.0.0.2  │ 2018-03-23
 2 │  3 │ 127.0.0.3  │ 2018-03-24
 3 │  4 │ 127.0.0.4  │ 2018-03-25
(4 rows)

How does that work? Well, the PostgreSQL catalogs contains useful hints to answer that question:

  SELECT CASE WHEN o.oprkind='l' THEN NULL
              ELSE pg_catalog.format_type(o.oprleft, NULL)
          END AS "Left arg type",
       
         CASE WHEN o.oprkind='r' THEN NULL
              ELSE pg_catalog.format_type(o.oprright, NULL)
          END AS "Right arg type",
       
         pg_catalog.format_type(o.oprresult, NULL) AS "Result type",
         
         oprcode::regprocedure as function
  
    FROM pg_catalog.pg_operator o
         LEFT JOIN pg_catalog.pg_namespace n ON n.oid = o.oprnamespace
         LEFT JOIN pg_catalog.pg_type l ON l.oid = o.oprleft
         
   WHERE o.oprname = '+'
     AND l.typname in ('integer', 'inet', 'date')
     AND pg_catalog.pg_operator_is_visible(o.oid)
    
ORDER BY 1, 2, 3, 4;

At query planning time, PostgreSQL does a catalog lookup that is equivalent to our catalog query above. For optimisation purposes, the lookup is done in-memory, using an efficient cache of the catalog contents, though.

Here’s the result of our manual catalog search for the operator + and the data types named integer, inet, and date:

 Left arg type │     Right arg type     │         Result type         │                 function                 
═══════════════╪════════════════════════╪═════════════════════════════╪══════════════════════════════════════════
 date          │ integer                │ date                        │ date_pli(date,integer)
 date          │ interval               │ timestamp without time zone │ date_pl_interval(date,interval)
 date          │ time with time zone    │ timestamp with time zone    │ datetimetz_pl(date,time with time zone)
 date          │ time without time zone │ timestamp without time zone │ datetime_pl(date,time without time zone)
 inet          │ bigint                 │ inet                        │ inetpl(inet,bigint)
(5 rows)

Each operator is then associated with a PostgreSQL function that embeds the actual implementation of the expected behavior. Function themselves have support for polymorphism, much as the generic functions mentionned above. See the documentation part about Function Overloading.

In PostgreSQL the function overloading is adapted to the SQL engine static typing needs: the type of a SQL query result must be computed entirely before we run the query. Given that, a PostgreSQL function can be overloaded only in such a way that every specialized implementation must return the same data type.

That’s the reason why each implementation of the + operator is using a different function: we have a second level of overloading that takes place at the operator level.

Included in PostgreSQL is the function date_part that implements the EXTRACT standard syntax. As every specific implementation of the date_part function returns the same result type, we can see function overloading here:

                                   List of functions
   Schema   │   Name    │ Result data type │        Argument data types        │  Type  
════════════╪═══════════╪══════════════════╪═══════════════════════════════════╪════════
 pg_catalog │ date_part │ double precision │ text, abstime                     │ normal
 pg_catalog │ date_part │ double precision │ text, date                        │ normal
 pg_catalog │ date_part │ double precision │ text, interval                    │ normal
 pg_catalog │ date_part │ double precision │ text, reltime                     │ normal
 pg_catalog │ date_part │ double precision │ text, time with time zone         │ normal
 pg_catalog │ date_part │ double precision │ text, time without time zone      │ normal
 pg_catalog │ date_part │ double precision │ text, timestamp with time zone    │ normal
 pg_catalog │ date_part │ double precision │ text, timestamp without time zone │ normal
(8 rows)

Here’s an example of calling that function, where the expression date 'today' is of type date and the expression now() is of type timestamp with time zone.

select date_part('dow', date 'today') as dow,
       date_part('month', now()) as month;

Which, today, gives:

 dow │ month 
═════╪═══════
   4 │     3
(1 row)

As a user of PostgreSQL you can write functions that are using that overloading mechanism, much like you do when writing Java, C++, Python, PHP, or Ruby code, to list just some of the languages having polymorphism features.

Object-Relational Database Management System

PostgreSQL is an Object-Relational Database Management System. It’s possible to relate its object orientation to having implemented table inheritance, which kind of looks like object inheritance in some ways.

This article shows that object systems also can be defined in terms of generic functions that can be implemented separately depending on the type of arguments used at run time.

Conclusion

PostgreSQL provides a complete implementation of function overloading and operator overloading and uses it a basis for advanced indexing support. A PostgreSQL extension can define new data types and provide for their indexing support, all without having to hack on PostgreSQL source code, thanks to the extensibility support of the system. And several successful PostgreSQL extension do just that, our best example being PostGIS.

PostgreSQL truely is the world’s most advanced open source database. PostgreSQL is YeSQL!