为什么在多态关联中不能有外键？

A foreign key must reference only one parent table. This is fundamental to both SQL syntax, and relational theory.

A Polymorphic Association is when a given column may reference either of two or more parent tables. There's no way you can declare that constraint in SQL.

The Polymorphic Associations design breaks rules of relational database design. I don't recommend using it.

There are several alternatives:

• Exclusive Arcs: Create multiple foreign key columns, each referencing one parent. Enforce that exactly one of these foreign keys can be non-NULL.

• Reverse the Relationship: Use three many-to-many tables, each references Comments and a respective parent.

• Concrete Supertable: Instead of the implicit "commentable" superclass, create a real table that each of your parent tables references. Then link your Comments to that supertable. Pseudo-rails code would be something like the following (I'm not a Rails user, so treat this as a guideline, not literal code):

   class Commentable < ActiveRecord::Base
  has_many :comments
  end
  
  
  class Comment < ActiveRecord::Base
  belongs_to :commentable
  end
  
  
  class Article < ActiveRecord::Base
  belongs_to :commentable
  end
  
  
  class Photo < ActiveRecord::Base
  belongs_to :commentable
  end
  
  
  class Event < ActiveRecord::Base
  belongs_to :commentable
  end
  

I also cover polymorphic associations in my presentation Practical Object-Oriented Models in SQL, and my book SQL Antipatterns Volume 1: Avoiding the Pitfalls of Database Programming.

Re your comment: Yes, I do know that there's another column that notes the name of the table that the foreign key supposedly points to. This design is not supported by foreign keys in SQL.

What happens, for instance, if you insert a Comment and name "Video" as the name of the parent table for that Comment? No table named "Video" exists. Should the insert be aborted with an error? What constraint is being violated? How does the RDBMS know that this column is supposed to name an existing table? How does it handle case-insensitive table names?

Likewise, if you drop the Events table, but you have rows in Comments that indicate Events as their parent, what should be the result? Should the drop table be aborted? Should rows in Comments be orphaned? Should they change to refer to another existing table such as Articles? Do the id values that used to point to Events make any sense when pointing to Articles?

These dilemmas are all due to the fact that Polymorphic Associations depends on using data (i.e. a string value) to refer to metadata (a table name). This is not supported by SQL. Data and metadata are separate.

I'm having a hard time wrapping my head around your "Concrete Supertable" proposal.

• Define Commentable as a real SQL table, not just an adjective in your Rails model definition. No other columns are necessary.

   CREATE TABLE Commentable (
  id INT AUTO_INCREMENT PRIMARY KEY
  ) TYPE=InnoDB;
  

• Define the tables Articles, Photos, and Events as "subclasses" of Commentable, by making their primary key be also a foreign key referencing Commentable.

   CREATE TABLE Articles (
  id INT PRIMARY KEY, -- not auto-increment
  FOREIGN KEY (id) REFERENCES Commentable(id)
  ) TYPE=InnoDB;
  
  
  -- similar for Photos and Events.
  

• Define the Comments table with a foreign key to Commentable.

   CREATE TABLE Comments (
  id INT PRIMARY KEY AUTO_INCREMENT,
  commentable_id INT NOT NULL,
  FOREIGN KEY (commentable_id) REFERENCES Commentable(id)
  ) TYPE=InnoDB;
  

• When you want to create an Article (for instance), you must create a new row in Commentable too. So too for Photos and Events.

   INSERT INTO Commentable (id) VALUES (DEFAULT); -- generate a new id 1
  INSERT INTO Articles (id, ...) VALUES ( LAST_INSERT_ID(), ... );
  
  
  INSERT INTO Commentable (id) VALUES (DEFAULT); -- generate a new id 2
  INSERT INTO Photos (id, ...) VALUES ( LAST_INSERT_ID(), ... );
  
  
  INSERT INTO Commentable (id) VALUES (DEFAULT); -- generate a new id 3
  INSERT INTO Events (id, ...) VALUES ( LAST_INSERT_ID(), ... );
  

• When you want to create a Comment, use a value that exists in Commentable.

   INSERT INTO Comments (id, commentable_id, ...)
  VALUES (DEFAULT, 2, ...);
  

• When you want to query comments of a given Photo, do some joins:

   SELECT * FROM Photos p JOIN Commentable t ON (p.id = t.id)
  LEFT OUTER JOIN Comments c ON (t.id = c.commentable_id)
  WHERE p.id = 2;
  

• When you have only the id of a comment and you want to find what commentable resource it's a comment for. For this, you may find that it's helpful for the Commentable table to designate which resource it references.

   SELECT commentable_id, commentable_type FROM Commentable t
  JOIN Comments c ON (t.id = c.commentable_id)
  WHERE c.id = 42;
  

Then you'd need to run a second query to get data from the respective resource table (Photos, Articles, etc.), after discovering from commentable_type which table to join to. You can't do it in the same query, because SQL requires that tables be named explicitly; you can't join to a table determined by data results in the same query.

Admittedly, some of these steps break the conventions used by Rails. But the Rails conventions are wrong with respect to proper relational database design.

Bill Karwin is correct that foreign keys cannot be used with polymorphic relationships due to SQL not really having a native concept polymorphic relationships. But if your goal of having a foreign key is to enforce referential integrity you can simulate it via triggers. This gets DB specific but below is some recent triggers I created to simulate the cascading delete behavior of a foreign key on a polymorphic relationship:

CREATE FUNCTION delete_related_brokerage_subscribers() RETURNS trigger AS $$
BEGIN
DELETE FROM subscribers
WHERE referrer_type = 'Brokerage' AND referrer_id = OLD.id;
RETURN NULL;
END;
$$ LANGUAGE plpgsql;


CREATE TRIGGER cascade_brokerage_subscriber_delete
AFTER DELETE ON brokerages
FOR EACH ROW EXECUTE PROCEDURE delete_related_brokerage_subscribers();




CREATE FUNCTION delete_related_agent_subscribers() RETURNS trigger AS $$
BEGIN
DELETE FROM subscribers
WHERE referrer_type = 'Agent' AND referrer_id = OLD.id;
RETURN NULL;
END;
$$ LANGUAGE plpgsql;


CREATE TRIGGER cascade_agent_subscriber_delete
AFTER DELETE ON agents
FOR EACH ROW EXECUTE PROCEDURE delete_related_agent_subscribers();

In my code a record in the brokerages table or a record in the agents table can relate to a record in the subscribers table.