多个但相互排斥的外键 - 这是要走的路吗？

Question

I have three tables: Users, Companies and Websites. Users and companies have websites, and thus each user record has a foreign key into the Websites table. Also, each company record has a foreign key into the Websites table.

Now I want to include foreign keys in the Websites table back into their respective "parent" records. How do I do that? Should I have two foreign keys in each website record, with one of them always NULL? Or is there another way to go?

Answer 1

If we look into the model here, we will see the following:

1. A user is related to exactly one website
   • A company is related to exactly one website
   • A website is related to exactly one user or company

The third relation implies existence of a "user or company" entity whose PRIMARY KEY should be stored somewhere.

To store it you need to create a table that would store a PRIMARY KEY of a website owner entity. This table can also store attributes common for a user and a website.

Since it's a one-to-one relation, website attributes can be stored in this table too.

The attributes not shared by users and companies should be stored in the separate table.

To force the correct relationships, you need to make the PRIMARY KEY of the website composite with owner type as a part of it, and force the correct type in the child tables with a CHECK constraint:

CREATE TABLE website_owner (
    type INT NOT NULL,
    id INT NOT NULL,
    website_attributes,
    common_attributes,
    CHECK (type IN (1, 2)) -- 1 for user, 2 for company
    PRIMARY KEY (type, id)
)

CREATE TABLE user (
    type INT NOT NULL,
    id INT NOT NULL PRIMARY KEY,
    user_attributes,
    CHECK (type = 1),
    FOREIGN KEY (type, id) REFERENCES website_owner
)

CREATE TABLE company (
    type INT NOT NULL,
    id INT NOT NULL PRIMARY KEY,
    company_attributes,
    CHECK (type = 2),
    FOREIGN KEY (type, id) REFERENCES website_owner
)

Answer 2

you don't need a parent column, you can lookup the parents with a simple select (or join the tables) on the users and companies table. if you want to know if this is a user or a company website i suggest using a boolean column in your websites table.

Answer 3

Why do you need a foreign key from website to user/company at all? The principle of not duplicating data would suggest it might be better to scan the user/company tables for a matching website id. If you really need to you could always store a flag in the website table that denotes whether a given website record is for a user or a company, and then scan the appropriate table.

Answer 4

The problem I have with the accepted answer (by Quassnoi) is that the object relationships are the wrong way around: company is not a sub-type of a website owner; we had companies before we had websites and we can have companies who are website owners. Also, it seems to me that website ownership is a relationship between a website and either a person or a company ie we should have a relationship table (or two) in the schema. It may be an acceptable approach to keep personal website ownership separate from corporate website ownership and only bring them together when required eg via VIEW s:

CREATE TABLE People
(
 person_id CHAR(9) NOT NULL UNIQUE,  -- external identifier
 person_name VARCHAR(100) NOT NULL
);

CREATE TABLE Companies
(
 company_id CHAR(6) NOT NULL UNIQUE,  -- external identifier
 company_name VARCHAR(255) NOT NULL
);

CREATE TABLE Websites
(
 url CHAR(255) NOT NULL UNIQUE
);

CREATE TABLE PersonalWebsiteOwnership
(
 person_id CHAR(9) NOT NULL UNIQUE
    REFERENCES People ( person_id ),
 url CHAR(255) NOT NULL UNIQUE
    REFERENCES Websites ( url )
);

CREATE TABLE CorporateWebsiteOwnership
(
 company_id CHAR(6) NOT NULL UNIQUE
    REFERENCES Companies( company_id ),
 url CHAR(255) NOT NULL UNIQUE
    REFERENCES Websites ( url )
);

CREATE VIEW WebsiteOwnership AS
SELECT url, company_name AS website_owner_name
  FROM CorporateWebsiteOwnership
       NATURAL JOIN Companies
UNION
SELECT url, person_name AS website_owner_name
  FROM PersonalWebsiteOwnership
       NATURAL JOIN People;

The problem with the above is there is no way of using database constraints to enforce the rule that a website is either owned by a person or a company but not both.

If we can assuming the DBMS enforces check constraints (as the accepted answer does) then we can exploit the fact that a (human) person and a company are both legal persons and employ a super-type table ( LegalPersons ) but still retain relationship table approach ( WebsiteOwnership ), this time using the VIEW s to separate personal website ownership from separate from corporate website ownership but this time with strongly typed attributes:

CREATE TABLE LegalPersons
(
 legal_person_id INT NOT NULL UNIQUE,  -- internal artificial identifier
 legal_person_type CHAR(7) NOT NULL
    CHECK ( legal_person_type IN ( 'Company', 'Person' ) ),
 UNIQUE ( legal_person_type, legal_person_id )
);

CREATE TABLE People
(
 legal_person_id INT NOT NULL
 legal_person_type CHAR(7) NOT NULL
    CHECK ( legal_person_type = 'Person' ),
 UNIQUE ( legal_person_type, legal_person_id ),
 FOREIGN KEY ( legal_person_type, legal_person_id )
     REFERENCES LegalPersons ( legal_person_type, legal_person_id ),
 person_id CHAR(9) NOT NULL UNIQUE,  -- external identifier
 person_name VARCHAR(100) NOT NULL
);

CREATE TABLE Companies
(
 legal_person_id INT NOT NULL
 legal_person_type CHAR(7) NOT NULL
    CHECK ( legal_person_type = 'Company' ),
 UNIQUE ( legal_person_type, legal_person_id ),
 FOREIGN KEY ( legal_person_type, legal_person_id )
     REFERENCES LegalPersons ( legal_person_type, legal_person_id ),
 company_id CHAR(6) NOT NULL UNIQUE,  -- external identifier
 company_name VARCHAR(255) NOT NULL
);

CREATE TABLE WebsiteOwnership
(
 legal_person_id INT NOT NULL
 legal_person_type CHAR(7) NOT NULL
 UNIQUE ( legal_person_type, legal_person_id ),
 FOREIGN KEY ( legal_person_type, legal_person_id )
     REFERENCES LegalPersons ( legal_person_type, legal_person_id ),
 url CHAR(255) NOT NULL UNIQUE
    REFERENCES Websites ( url )
);

CREATE VIEW CorporateWebsiteOwnership AS 
SELECT url, company_name
  FROM WebsiteOwnership
       NATURAL JOIN Companies;

CREATE VIEW PersonalWebsiteOwnership AS
SELECT url, person_name
  FROM WebsiteOwnership
       NATURAL JOIN Persons;

What we need are new DBMS features for 'distributed foreign keys' ("For each row in this table there must be exactly one row in one of these tables") and 'multiple assignment' to allow the data to be added into tables thus constrained in a single SQL statement. Sadly we are a far way from getting such features!

Answer 5

First of all, do you really need this bi-directional link? It is a good practice to avoid it unless absolutely needed.

I understand it that you wish to know whether the site belongs to a user or to a company. You can achieve that by having a simple boolean field in the Website table - [BelongsToUser]. If true, then you look up a user, if false - you look up a company.

Answer 6

A bit late, but all the existing answers seemed to fall somewhat short of the mark:

• Owner to website is a 1:Many relation
• Website to owner is a 1:1 relation
• Users and Companies tables should not have a foreign key into the Websites table
• None of the website data, common to users and companies or not, should be in the Users or Companies tables
• None of the owner's information, common or not, should be in the Websites table
• MySQL ignores, silently, CHECK constraints on tables (no enforcement of referential integrity)
• The DBMS ought to handle the 'relation' logic, not the application using the database

Some of this is recognized in the answer from onedaywhen , yet that answer still missed the opportunity to make MySQL do the heavy lifting and enforce the referential integrity.

---

A website can only have one owner, legally, anyway. A person, or company, can have any number of websites, including none. A link in the database from owner to website can only be 1:1 at any level of normalization. In reality the relation is 1:Many , and would require having multiple table entries for each owner that happens to own more than one website. A link from website to owner is 1:1 in both database terms and in reality. Having the link from website to owner represents the model better. With an index in the website table, doing the 1:Many lookup for a given owner becomes reasonably efficient.

The CHECK attribute in SQL would be an excellent solution, if MySQL didn't happen to silently ignore it.

_{MySQL Docs 13.1.20 CREATE TABLE Syntax}

The CHECK clause is parsed but ignored by all storage engines.

MySQL's functionality does offer two solutions as work-arounds to implement the behavior of CHECK and keep the referential integrity of the data. Triggers with stored procedures is one, and works well with all manner of constraints. Easier to implement, though less versatile, is using a VIEW with a WITH CHECK OPTION clause, which MySQL will implement.

_{MySQL Docs 24.5.4 The View WITH CHECK OPTION Clause}

The WITH CHECK OPTION clause can be given for an updatable view to prevent inserts to rows for which the WHERE clause in the select_statement is not true. It also prevents updates to rows for which the WHERE clause is true but the update would cause it to be not true (in other words, it prevents visible rows from being updated to nonvisible rows).

The MySQLTUTORIAL site gives a good example of both options in their Introduction to the SQL CHECK constraint tutorial. (You have to think around the typos, but good otherwise.)

---

Having found this question while trying to resolve a similar mutually exclusive foreign key split and developing a solution, with hints generated by the answers, it seems only proper to share my solution in return.

Recommended Solution

For the minimum impact to the existing schema, and the application accessing the data, retain the Users and Companies tables as they are. Rename the Websites table and replace it with a VIEW named Websites which the application can continue to access. Except when dealing with the ownership information, all the old queries to Websites should still work. So:

The setup

-- Keep the `Users` table about "users"
CREATE TABLE `Users` (
    `id` INT SERIAL PRIMARY KEY,
    `name` VARCHAR(180),
    -- user_attributes
);

-- Keep the `Companies` table about "companies"
CREATE TABLE `Companies` (
    `id` SERIAL PRIMARY KEY,
    `name` VARCHAR(180),
    -- company_attributes
);

-- Attach ownership information about the website to the website's record in the `Websites` table, renamed to `WebsitesData`
CREATE TABLE `WebsitesData` (
    `id` SERIAL PRIMARY KEY,
    `name` VARCHAR(255),
    `is_personal` BOOL,
    `owner_user` BIGINT UNSIGNED DEFAULT NULL,
    `owner_company` BIGINT UNSIGNED DEFAULT NULL,
    website_attributes,
    FOREIGN KEY `WebsiteOwner_User` (`owner_user`)
        REFERENCES `Users` (`id`)
            ON DELETE RESTRICT ON UPDATE CASCADE,
    FOREIGN KEY `WebsiteOwner_Company` (`owner_company`)
        REFERENCES `Companies` (`id`)
            ON DELETE RESTRICT ON UPDATE CASCADE,
);

-- Create a new `VIEW` with the original name of `Websites` as the gateway to the website records which can enforce the constraints you need
CREATE VIEW `Websites` AS
SELECT * FROM `WebsitesData` WHERE
    (`is_personal`=TRUE AND `owner_user` IS NOT NULL AND `owner_company` IS NULL) OR
    (`is_personal`=FALSE AND `owner_user` IS NULL AND `owner_company` IS NOT NULL)
WITH CHECK OPTION;

---

Usage

-- Use the Websites VIEW for the INSERT, UPDATE, and SELECT operations as you normally would and leave the WebsitesData table in the background.
INSERT INTO `Websites` SET
    `is_personal`=TRUE,
    `owner_user`=$userID;
INSERT INTO `Websites` SET
    `is_personal`=FALSE,
    `owner_company`=$companyID;

-- Or, using different field lists based on the type of owner
INSERT INTO `Websites` (`is_personal`,`owner_user`, ...)
    VALUES (TRUE, $userID, ...);
INSERT INTO `Websites` (`is_personal`,`owner_company`, ...)
    VALUES (FALSE, $companyID, ...);

-- Or, using a common field list, and placing NULL in the proper place
INSERT INTO `Websites` (`is_personal`,`owner_user`,`owner_company`,...)
    VALUES (TRUE, $userID, NULL, ...);
INSERT INTO `Websites` (`is_personal`,`owner_user`,`owner_company`,...)
    VALUES (FALSE, NULL, $companyID, ...);

-- Change the company that owns a website
-- Will ERROR if the site was owned by a User.
UPDATE `Websites` SET `owner_company`=$new_companyID;

-- Force change the ownership from a User to a Company
UPDATE `Websites` SET
    `owner_company`=$new_companyID,
    `owner_user`=NULL,
    `is_personal`=FALSE;

-- Force change the ownership from a Company to a User
UPDATE `Websites` SET
    `owner_user`=$new_userID,
    `owner_company`=NULL,
    `is_personal`=TRUE;

-- Selecting the owner of a site without needing to know if it is personal or not
(SELECT `Users`.`name` AS `Owner`
    FROM `Websites`
        JOIN `Users` ON `Websites`.`owner_user`=`Users`.`id`
    WHERE `is_personal`=TRUE AND `Websites`.`id`=$siteID)
UNION
(SELECT `Companies`.`name` AS `Owner`
    FROM `Websites`
        JOIN `Companies` ON `Websites`.`owner_company`=`Companies`.`id`
    WHERE `is_personal`=FALSE AND `Websites`.`id`=$siteID);

-- Selecting the sites owned by a User
SELECT `name` FROM `Websites`
    WHERE `is_personal`=TRUE AND `id`=$userID;
SELECT `Websites`.`name`
    FROM `Websites`
        JOIN `Users` ON `Websites`.`owner_user`=`Users`.$userID
    WHERE `is_personal`=TRUE AND `Users`.`name`="$user_name";

-- Selecting the sites owned by a Company
SELECT `name` FROM `Websites` WHERE `is_personal`=FALSE AND `id`=$companyID;
SELECT `Websites`.`name`
    FROM `Websites`
        JOIN `Comnpanies` ON `Websites`.`owner_company`=`Companies`.$userID
    WHERE `is_personal`=FALSE AND `Companies`.`name`="$company_name";

-- Listing all websites and their owners
(SELECT `Websites`.`name` AS `Website`,`Users`.`name` AS `Owner`
    FROM `Websites`
        JOIN `Users` ON `Websites`.`owner_user`=`Users`.`id`
    WHERE `is_personal`=TRUE)
UNION ALL
(SELECT `Websites`.`name` AS `Website`,`Companies`.`name` AS `Owner`
    FROM `Websites`
        JOIN `Companies` ON `Websites`.`owner_company`=`Companies`.`id`
    WHERE `is_personal`=FALSE)
ORDER BY Website, Owner;

-- Listing all users or companies which own at least one website
(SELECT `Websites`.`name` AS `Website`,`Users`.`name` AS `Owner`
    FROM `Websites`
        JOIN `Users` ON `Websites`.`owner_user`=`Users`.`id`
    WHERE `is_personal`=TRUE)
UNION DISTINCT
(SELECT `Websites`.`name` AS `Website`,`Companies`.`name` AS `Owner`
    FROM `Websites`
        JOIN `Companies` ON `Websites`.`owner_company`=`Companies`.`id`
    WHERE `is_personal`=FALSE)
GROUP BY `Owner` ORDER BY `Owner`;

---

Normalization Level Up

As a technical note for normalization, the ownership information could be factored out of the Websites table and a new table created to hold the ownership data, including the is_normal column.

CREATE TABLE `Websites` (
    `id` SERIAL PRIMARY KEY,
    `name` VARCHAR(255),
    `owner` BIGINT UNSIGNED DEFAULT NULL,
    website_attributes,
    FOREIGN KEY `Website_Owner` (`owner`)
        REFERENCES `WebOwners` (id`)
            ON DELETE RESTRICT ON UPDATE CASCADE
);

CREATE TABLE `WebOwnersData` (
    `id` SERIAL PRIMARY KEY,
    `is_personal` BOOL,
    `user` BIGINT UNSIGNED DEFAULT NULL,
    `company` BIGINT UNSIGNED DEFAULT NULL,
    FOREIGN KEY `WebOwners_User` (`user`)
        REFERENCES `Users` (`id`)
            ON DELETE RESTRICT ON UPDATE CASCADE,
    FOREIGN KEY `WebOwners_Company` (`company`)
        REFERENCES `Companies` (`id`)
            ON DELETE RESTRICT ON UPDATE CASCADE,
);

CREATE VIEW `WebOwners` AS
SELECT * FROM WebsitesData WHERE
    (`is_personal`=TRUE AND `user` IS NOT NULL AND `company` IS NULL) OR
    (`is_personal`=FALSE AND `user` IS NULL AND `company` IS NOT NULL)
WITH CHECK OPTION;

I believe, however, that the created VIEW, with its constraints, prevents any of the anomalies that normalization aims to remove, and adds complexity that is not needed in the situation. The normalization process is always a trade off anyway.