This repo provides utilities to for modeling variant types in Postgres and provides a demo schema, as well. To try it out, first use Git to obtain the relevant code:

:;  git clone git@github.com:solidsnack/pg-sql-variants.git
:;  cd pg-sql-variants/
:;  git submodule update --init --recursive

Then load the utilities and the sample schema in Postgres:

:;  psql
Line style is unicode.
Expanded display is used automatically.
Null display is "\N".
Timing is on.
psql (12.1)
Type "help" for help.

--# thelyfsoshort@[local]/~
\i init.psql 
BEGIN
...
COMMIT
...
BEGIN
...
COMMIT
...
BEGIN
...
COMMIT
...

The sample schema helps us to demonstrate a simple polymorhpic datatype: an animal type with concrete cat, dog and walrus subtypes.

--# thelyfsoshort@[local]/~
SELECT tablename FROM pg_tables WHERE schemaname = 'inetorg';
 tablename
───────────
 cat
 walrus
 dog
 animal
(4 rows)

Let's setup the variant relationship between the types in the inetorg namespace with the variant() function from the variants namespace:

--# thelyfsoshort@[local]/~
SET search_path TO inetorg, variants, "$user", public;

--# thelyfsoshort@[local]/~
SELECT * FROM variant('animal', 'cat');
SELECT * FROM variant('animal', 'walrus');
SELECT * FROM variant('animal', 'dog');

We can see that there are no animals and there are no cats:

--# thelyfsoshort@[local]/~
SELECT * FROM animal;
 ident 
───────
(0 rows)

--# thelyfsoshort@[local]/~
SELECT * FROM cat;
 license │ responds_to │ doglike
─────────┼─────────────┼─────────
(0 rows)

The variants.variant() function is basically a SQL macro; it sets up several triggers every time it is called. Let's add a cat:

--# thelyfsoshort@[local]/~ 
INSERT INTO cat VALUES ('00000000-0000-0000-0000-000000000001', 'felix', FALSE);
INSERT 0 1

The triggers ensure that records are added to the animal table, as well.

--# thelyfsoshort@[local]/~
SELECT * FROM cat;
               license                │ responds_to │ doglike
──────────────────────────────────────┼─────────────┼─────────
 00000000-0000-0000-0000-000000000001 │ felix       │ f
(1 row)

--# thelyfsoshort@[local]/~
SELECT * FROM animal;
                ident                 
──────────────────────────────────────
 00000000-0000-0000-0000-000000000001
(1 row)

In addition to the triggers, variants.variant() also maintains a join table, with one column for each variant type. In this case, the join table is named animal*:

--# thelyfsoshort@[local]/~
SELECT * FROM "animal*";
Time: 0.285 ms
─[ RECORD 1 ]──────────────────────────────────────────
ident  │ 00000000-0000-0000-0000-000000000001
type   │ cat
cat    │ (00000000-0000-0000-0000-000000000001,felix,f)
walrus │ \N
dog    │ \N
(1 row)

The join table illustrates a cool Postgres features: columns with row types.

What good is a cat? Better delete while we're not sure:

--# thelyfsoshort@[local]/~
DELETE FROM cat;
DELETE 1

No more cats, no more animal*s:

--# thelyfsoshort@[local]/~
SELECT * FROM cat;
 license │ responds_to │ doglike
─────────┼─────────────┼─────────
(0 rows)

--# thelyfsoshort@[local]/~
SELECT * FROM "animal*";
 ident │ type │ cat │ walrus │ dog
───────┼──────┼─────┼────────┼─────
(0 rows)

Typed variants, case classes, tagged unions, algebraic data types or just enums: variant types are a feature common to many programming languages but are an awkward fit for SQL.

The fundamental difficulty is that foreign keys can reference columns of only one other table. By combining VIEWs, triggers and Postgres's JSON data-type, we can group related types like cat and walrus under a tagged union like animal, allowing other tables to create foreign keys that reference animal.

CREATE TABLE cat (
  license       uuid PRIMARY KEY,
  responds_to   text NOT NULL,
  doglike       boolean DEFAULT TRUE
);

CREATE TABLE walrus (
  registration  uuid PRIMARY KEY,
  nickname      text,
  size          text NOT NULL DEFAULT 'big' CHECK (size IN ('small', 'big')),
  haz_bucket    boolean NOT NULL DEFAULT FALSE
);

CREATE TABLE animal (
  ident         uuid PRIMARY KEY
);

The process for forming the foreign key and triggers is completely formulaic and we capture it in a stored procedure, variant (in variants.sql) that allows one to put cat and walrus together under animal:

SELECT * FROM variant('animal', 'cat');
SELECT * FROM variant('animal', 'walrus');

Changes to keys are propagated bidirectionally between animal and its variants. A DELETE against a cat's UUID in animal will remove the row from cat; and a delete against cat will remove the row from animal.

Imagine for a moment the data loaded in your app. There are Cats and Walruses of class Animal; there are Strings, Integers, StructTimes... But how would you go about searching and sorting these objects? One could say this is a bad question with a bad answer.

It's a bad question because most of the time we have the objects we need ready to hand, assigned to variables in the right place in our program -- we don't need to find them. We don't ever sort "all" integers, just the relevant ones.

The answer is bad because one would search and sort all the objects of a given type by walking the heap. This might be facilitated by the runtime (Ruby's ObjectSpace.each_object(<cls>) comes to mind) or it might not; but one is in for a linear scan either way; and there is potential for conflict with other threads of execution -- either preventing them from running, or tripping over inconsistencies they introduce.

In a database, however, we do not rely on having the right context to find an object. Whereas in a programming context we use the objects to get the fields, in a storage context we use the fields to find the objects; there is no notion of identity apart from field values. This is the heart of the object-relational (or struct-enum-relational or ADT-relational) mismatch.

The two models overlap when we consider global, concurrent data structures like event buses or concurrent maps. In SQL terms, each concurrent map would be a relation, and in SQL each relation is a distinct type. A database is what you would get if each of the types in your language were automatically associated with a concurrent map.

There are two abstractions relating to types which become strange in this all-types-backed-with-maps model:

• Inheritance, abstract base classes, and traits
• Generics (in the Java sense) or templates (in the C++ sense)

With regards to inheritance, one wonders what it would mean to insert an orange in the fruit table or the citrus table. Clearly, inserting it in any one of them should insert it in all of them. It is an ambiguity that gives the author pause.

With regards to templated definitions, it stands to reason that these can have no "live" representation in the database. Tables are there, or they aren't. Perhaps a database's SQL dialect could support template expansion; but this feature would have no impact on the nature of queries or relationships between tables.

Typed variants -- or tagged unions -- are a minimal way to expand SQL's support for polymorphism that is helpful to object-oriented languages, and languages like Haskell, Rust and Go which provide products or sums of products as well as traits.

In our approach, the types which are part of the union are all themselves physical tables with primary keys that are type compatible. We create a new table for the union, the only columns of which are the columns of the primary key and, through triggers, we ensure that inserts, updates and deletes to any of the variant tables are also propagated to the union.

The union table ensures that the key spaces of the variants are disjoint and allows for other tables to declare foreign keys that references the union. Normal database validation logic takes over from there. The alternative would be to have constraint triggers on each client table for each table in the variant and to have triggers on each variant table for each client. This would both be less expressive and a likely source of errors.

SQL tagged unions in this style support "composition" instead of inheritance for polymorphism. For example, in the case where we have a type of letters and would like be able handle Swiss letter, Spanish letter, Egyptian letter and more, the modeller is tasked with breaking out the common fields into a letter table which would reference national_variant which is a union of egyptian, ethiopian, etruscan and so forth.