For example, if I have two tables - Parent and Child - where Child records are owned by Parent records, which table needs the ON DELETE CASCADE?

ON DELETE CASCADE is an optional clause in a foreign key declaration. So it goes with the foreign key declaration. (Meaning, in the "child" table.)

...it could mean delete the Parent record when the Child record is deleted, or it could mean delete the Child record when the Parent is deleted. So which is it?

One way to interpret a foreign key declaration is, "All valid values for this column come from 'that_column' in 'that_table'." When you delete a row in the "child" table, nobody cares. It doesn't affect data integrity.

When you delete a row from the "parent" table--from "that_table"--you remove a valid value from the possible values for the "child" table. To maintain data integrity, you have to do something to the "child" table. Cascading deletes is one thing you can do.

Chapter and verse, from PostgreSQL docs.

Restricting and cascading deletes are the two most common options. RESTRICT prevents deletion of a referenced row. NO ACTION means that if any referencing rows still exist when the constraint is checked, an error is raised; this is the default behavior if you do not specify anything. (The essential difference between these two choices is that NO ACTION allows the check to be deferred until later in the transaction, whereas RESTRICT does not.) CASCADE specifies that when a referenced row is deleted, row(s) referencing it should be automatically deleted as well. There are two other options: SET NULL and SET DEFAULT. These cause the referencing column(s) in the referencing row(s) to be set to nulls or their default values, respectively, when the referenced row is deleted. Note that these do not excuse you from observing any constraints. For example, if an action specifies SET DEFAULT but the default value would not satisfy the foreign key constraint, the operation will fail.

SQL:2011 Spec

There are five options for ON DELETE, and ON UPDATE that can apply to the FOREIGN KEY. These are called <referential actions>, directly from the SQL:2011 spec

• ON DELETE CASCADE: if a row of the referenced table is deleted, then all matching rows in the referencing table are deleted.
• ON DELETE SET NULL: if a row of the referenced table is deleted, then all referencing columns in all matching rows of the referencing table to be set to null.
• ON DELETE SET DEFAULT: if a row of the referenced table is deleted, then all referencing columns in all matching rows of the referencing table to be set to the column’s default value.
• ON DELETE RESTRICT: it is prohibited to delete a row of the referenced table if that row has any matching rows in the referencing table.
• ON DELETE NO ACTION (the default): there is no referential delete action; the referential constraint only specifies a constraint check.

The foreign key establishes the dependent relationship. The <referential action> determines what happens when the relationship is dissolved.

Example / Metaphor / Explanation

For this example, we will accept the common model of society and economy: where every business is a company that maintains a relationship to the bourgeoisie through a fatcat_owner.

CREATE TABLE bourgeoisie(
  fatcat_owner varchar(100) PRIMARY KEY
);
INSERT INTO bourgeoisie(fatcat_owner) VALUES
  ( 'Koch Brothers' );

CREATE TABLE business (
  name         varchar(100),
  fatcat_owner varchar(100) REFERENCES bourgeoisie
);
INSERT INTO business(name, fatcat_owner)
  VALUES ('Georgia-Pacific', 'Koch Brothers');

If all businesses are directly affected by bourgeoisie by way of their fatcat_owner then what do you do after the workers' revolution when that you purge the fatcat_owners and have a classless society?

-- Viva la revolución 
BEGIN;
  DELETE FROM bourgeoisie;
END;

You have a few options here,

• Stop the revolution. In SQL parlance, RESTRICT. Some people believe this is the lesser evil, but they're usually wrong.

• Allow it to go on. If so when the revolution happens SQL gives you four options,

  • SET NULL -- leave it blank. Who knows, maybe capitalism is restored the bourgeoisie comes up and oligarchs fill the roll of the fatcat_owners. Important note, the column must be NULLABLE (not NOT NULL) or this can never happen.
  • SET DEFAULT -- perhaps you had a DEFAULT that handled this? A DEFAULT can call a function. Maybe your schema is revolution-ready, already.
  • CASCADE -- there is no damage control. If the bourgeoisie goes, so does the business. If a business must have a fatcat_pig, then sometimes it makes more sense to lose the data rather than have a non-business in a business table.

  • NO ACTION -- this is essentially a method of delaying the check, in MySQL it's no different than RESTRICT, but in PostgreSQL, you'd be able to do

    -- Not a real revolution.
    -- requires constraint be DEFERRABLE INITIALLY DEFERRED
    BEGIN;
      SET CONSTRAINTS ALL DEFERRED;
      DELETE FROM bourgeoisie;
      INSERT INTO bourgeoisie VALUES ( 'Putin' );
      UPDATE business SET fatcat_pig = 'Putin';
    END;
    

    In such a system, the constraint is validated only before the transaction commits. This may result in stopping the revolution, but you can recover in the transaction -- for some degree of "recover."

A simple mnemonic would be

ON DELETE of parent CASCADE [by deleting] here

That tells you which deletes (deletes of the parent) get cascaded, where the ON DELETE CASCADE statement goes (on the child), and what gets deleted (the child).

well, perhaps we can rationalize the syntax. Let's take a Python example:

class Parent(self):
    # define parent's fields

class Child(self):    
    # define child's fields
    parent_pk_is_childs_foreign_key = models.ForeignKey(Parent, on_delete=models.CASCADE)

what this line says is on_delete of the Parent (which is accidentally mentioned in the statement), please cascade the deletion onto the child. That is why the CASCADE statement is defined at the child level, it marks those children that need to be deleted

For example if you had another class

class GrownUpChild(self):    
        # define grown up child's fields
        parent_pk_is_childs_foreign_key = models.ForeignKey(Parent, on_delete=models.DO_NOTHING)

this structure would clearly show which of the children need to be removed (Child) and which are to stay (GrownUpChild) albeit orphaned

[Edit: Given the context of the discussion, specifically in cases of on_delete=models.CASCADE etc, ] in fact it is often a desired behavior to leave children of a deleted parent, due to auditing and reporting reasons, as well as recovering accidental deletions. [ of course enterprise level software will be built around such behavior and will flag deleted records as deleted=1 instead of actually deleting them and also will not include them in any queries for the front end, minus some specially designed reports. In addition it will have a function of purging the deleted==1 records from the database, which will usually be executed by the UI administrator, often avoiding any involvement from the database administrator's side.]