Are composite primary keys bad practice?

To say that the use of "Composite keys as PRIMARY KEY is bad practice" is utter nonsense!

Composite PRIMARY KEYs are often a very "good thing" and the only way to model natural situations that occur in everyday life!

Think of the classic Databases-101 teaching example of students and courses and the many courses taken by many students!

Create tables course and student:

CREATE TABLE course
(
  course_id SERIAL,
  course_year SMALLINT NOT NULL,
  course_name VARCHAR (100) NOT NULL,
  CONSTRAINT course_pk PRIMARY KEY (course_id)
);


CREATE TABLE student
(
  student_id SERIAL,
  student_name VARCHAR (50),
  CONSTRAINT student_pk PRIMARY KEY (student_id)
);

I'll give you the example in the PostgreSQL dialect (and MySQL) - should work for any server with a bit of tweaking.

Now, you obviously want to keep track of which student is taking which course - so you have what's called a joining table (also called linking, bridging, many-to-many or m-to-n tables). They are also known as associative entities in more technical jargon!

1 course can have many students.
1 student can take many courses.

So, you create a joining table

CREATE TABLE registration
(
  cs_course_id INTEGER NOT NULL,
  cs_student_id INTEGER NOT NULL,

  -- now for FK constraints - have to ensure that the student
  -- actually exists, ditto for the course.

  CREATE CONSTRAINT cs_course_fk  FOREIGN KEY (cs_course_id)
    REFERENCES course  (course_id),
  CREATE CONSTRAINT cs_student_fk FOREIGN KEY (cs_student_id) 
    REFERENCES student (student_id)
);

Now, the only way to sensibly give the registration table a PRIMARY KEY is to make that KEY a combination of course and student. That way, you can't get:

• a duplicate of student and course combination

• a course can only have the same student enrolled once, and

• a student can only enroll in the same course one time only

• you also have a ready made search KEY on course per student - AKA a covering index,

• it is trivial to find courses without students and students who are taking no courses!

  -- The db-fiddle example has the PK constraint folded into the CREATE TABLE -- It can be done either way. I prefer to have everything in the CREATE TABLE statement.

ALTER TABLE registration
ADD CONSTRAINT registration_pk 
PRIMARY KEY (cs_course_id, cs_student_id);

Now, you could, if you were finding that searches for student by course were slow, use a UNIQUE INDEX on (sc_student_id, sc_course_id).

ALTER TABLE registration 
ADD CONSTRAINT course_student_sc_uq  
UNIQUE (cs_student_id, cs_course_id);

There is no silver bullet for adding indexes - they will make INSERTs and UPDATEs slower, but at the great benefit of greatly decreasing SELECT times! It's up to the developer to decide to index given their knowledge and experience, but to say that composite PRIMARY KEYs are always bad is just plain wrong.

In the case of joining tables, they are usually the only PRIMARY KEY that make sense! Joining tables are also very frequently the only way of modelling what happens in business or nature or in virtually every sphere I can think of!

This PK is also of use as a covering index which can help speed up searches. In this case, it would be particularly useful if one were searching regularly on (course_id, student_id) which, one would imagine, can often be the case!

This is just a small example of where a composite PRIMARY KEY can be a very good idea, and the only sane way to model reality! Off the top of my head, I can think of many many more.

An example from my own work!

Consider a flight table containing a flight_id, a list of departure and arrival airports and the relevant times and then also a cabin_crew table with crew members!

The only sane way this can be modelled is to have a flight_crew table with the flight_id and the crew_id as attibutes and the only sane PRIMARY KEY is to use the composite key of the two fields!

My half-educated take: a "primary key" doesn't have to be the only unique key used to look up data in the table, although data management tools will offer it as default selection. So for choosing whether to have a composite of two columns or a random (probably serial) generated number as the table key, you can have two different keys at once.

If data values include a suitable unique term that can represent the row, I'd rather declare that as "primary key", even if composite, than use a "synthetic" key. The synthetic key may perform better for technical reasons, but my own default choice is to designate and use the real term as primary key, unless you really need to go the other way to make your service work.

A Microsoft SQL Server has the distinct but related feature of the "clustered index" that controls the physical storage of data in index order, and also is used inside other indexes. By default, a primary key is created as a clustered index, but you can choose non-clustered instead, preferably after creating the clustered index. So you can have an integer identity generated column as clustered index, and, say, file name nvarchar(128 characters) as primary key. This may be better because the clustered index key is narrow, even if you store the file name as the foreign key term in other tables - although this example is a good case for also not doing that.

If your design involves importing tables of data that include an inconvenient primary key to identify related data, then you're pretty much stuck with that.

https://www.techopedia.com/definition/5547/primary-key describes an example of choosing whether to store data with a customer's social security number as the customer key in all the data tables, or to generate an arbitrary customer_id when you register them. Actually, this is a grave abuse of SSN, aside from whether it works or not; it is a personal and confidential data value.

So, an advantage of using a real-world fact as the key is that without joining back to the "Customer" table, you can retrieve information about them in other tables - but it's also a data security issue.

Also, you're in trouble if the SSN or other data key was recorded incorrectly, so you have the wrong value in 20 constrained tables instead of in "Customer" only. Whereas the synthetic customer_id has no external meaning so it can't be a wrong value.