Is it possible to assign a foreign key to a json property in PostgreSQL? Here is an example what I would like to achieve, but it doesn't work:
CREATE TABLE Users (Id int NOT NULL PRIMARY KEY);
CREATE TABLE Data (
Id int NOT NULL PRIMARY KEY,
JsonData json NOT NULL, -- [{Id: 1, somedata: null},{Id: 2, somedata: null}, ...]
CONSTRAINT FK_Users_Data FOREIGN KEY (JsonData->Id) REFERENCES Users(Id) -- this constraint will fail
);
It is not possible, and may not ever be possible, to assign a foreign key to a json property. It'd be a major and quite complicated change to PostgreSQL's foreign key enforcement. I don't think it's impossible to do, but would face similar issues to those experienced by the foreign-keys-to-arrays patch.
With 9.4 it'll be possible to make a whole json object a foreign key as jsonb supports equality tests. In 9.3 you can't even do that.
Yes it is possible but you will have to store another value. If you change your schema to:
CREATE TABLE Users (Id int NOT NULL PRIMARY KEY);
CREATE TABLE Data (
Id int NOT NULL PRIMARY KEY,
JsonData json NOT NULL,
UserId int generated always as ((JsonData->>'Id')::int) stored references Users(Id)
);
INSERT INTO Users VALUES (1);
Foreign key that doesn't exist:
INSERT INTO Data VALUES (1, '{"Id": 3}');
Returns the error:
ERROR: insert or update on table "data" violates foreign key constraint "data_userid_fkey" DETAIL: Key (userid)=(3) is not present in table "users".
Foreign key that does work:
INSERT INTO Data VALUES (1, '{"Id": 1}');
Here's a little SPI function have_ids which I use for an integrity constraint on a one-to-many relationship with a jsonb column
CREATE TABLE foo (
id INTEGER NOT NULL
)
CREATE TABLE bar (
foo_ids pg_catalog.jsonb DEFAULT '[]'::jsonb NOT NULL,
CONSTRAINT bar_fooids_chk CHECK (have_ids ('foo', foo_ids))
)
With a couple of triggers on foo it's almost as good as a foreign key.
have_ids() function? Even with "bad performance", if it is generic and stable, will be a good proposal to PostgreSQL community.
SELECT COUNT(*) FROM query for every ID. It's as fast as the sum of these queries. The performance cost of the function itself should be neglible. I don't have an SQL version, neither I know how to write one. I think PosgreSQL will eventually have first-class support for array-based foreign keys (wiki.postgresql.org/wiki/Todo#Referential_Integrity), so this workaround shouldn't be needed in the long term.
COUNT(1) instead of COUNT(*)COUNT(*) is optimized away by PostgreSQL (e.g. the star is not expanded into any fields) and I tend to keep COUNT(*) for readability (POLA).COUNT(1) because it has the same readability (to me) and I've always heard it's more performant than counting column(s)The foreign key parameter must be a column name:
http://www.postgresql.org/docs/current/static/sql-createtable.html
You will have to normalize
create table user_data (
id int not null primary key,
user_id int not null,
somedata text,
constraint fk_users_data foreign key (user_id) references Users(Id)
);
