5

This seems like a pretty common scenario: several types that all compose the same child type.

This could typically look like so:

-- 'name' is unique per parent record
CREATE TABLE sometype (
  sometype_id serial PRIMARY KEY,
  name        text
);

CREATE TABLE foo (foo_id serial PK);
CREATE TABLE bar (bar_id serial PK);

CREATE TABLE foo_sometype (
  foo_id      int4,
  sometype_id int4
);

CREATE TABLE bar_sometype (
  bar_id      int4,
  sometype_id int4
);

Which is fine, but cumbersome to query. I'm thinking this could be cleaner:

-- 'name' is unique per parent record
CREATE TABLE sometype (
  name        text
);

CREATE TABLE foo_sometype (
  foo_id      int4
) INHERITS(sometype);

CREATE TABLE bar_sometype (
  bar_id      int4,
) INHERITS(sometype);

What I like about this:

  • simple to join (with USING)
  • no need to add a surrogate key to 'sometype', it is explicitly a component of 'foo' and 'bar'

Seems like an atypical use of inheritance, though.

Any reason not to do this?


On the "deficiencies" of inheritance

Note that the standard caveats to Pg inheritance are only relevant when table inheritance is used to directly model class inheritance, which is not what I am doing here. In fact, for this to work I need inheritance to behave the way it does.

I almost wish they called it something other than "inheritance", since the behavior is quite logical, and the "shortcomings" are only relevant to one use-case.


Benefits over manually duplicated table structure

As Evan points out, I could just manually create 'foo_sometype' and 'bar_sometype' that look exactly like what I describe, but I think there are a couple of significant benefits to the inherited structure:

  • The inherits relationship explicitly defines 'foo_sometype' and 'bar_sometype' as being of the same type, not just two tables that happen to have the same columns.
  • Making future schema changes through the parent table lessens the chance of accidental divergences (with a little work, this could actually be enforced).
  • More importantly, client code can be generated against the parent table, and applied to the children with only a change in the table name, with (again) confidence that the table structure is enforced.

So, as a contrived example, Foo and Bar could have a HasSomeTypeList trait, which abstracts all 'sometype' operations, and knows that both tables can be mapped to the SomeType class.

Representing the Foo/Bar relationship, whether it's modeled as a trait or as inheritance, is the ultimate goal here.

Incidentally, to the naive user/query-writer - who is not expected to make schema changes - these two ways will look the same.

  • 1
    Not having a PK is good until you are alone adding stuff there. If this is not the case, it leads to confusion. Also, I don't know how you use your DB (like doing many ad-hoc queries), but in most cases writing queries a few characters shorter is not really a good argument for choosing any design. – dezso Dec 6 '16 at 20:46
  • @dezso: should rephrase, no additional generated key, it would still have PRIMARY KEY(foo_id/bar_id, name), which is more obvious than the link table. It's not about characters, but query complexity; the inheritance version eliminates a join and an extra key, which quickly adds up for complex queries. Especially if both child tables are used in the same query, and you have multiple 'sometype_id' columns now. – Dmitri Dec 6 '16 at 21:07
  • 1
    You might want to add the subtypes tag to this question. – Walter Mitty Dec 6 '16 at 21:43
  • @Dmitri please check if the edit says what you mean – dezso Dec 6 '16 at 22:43
  • @dezso, thanks for the correction; I clarified a bit more - I was talking about the key on 'sometype', I don't think there is any scenario where the '*_sometype' tables would need a surrogate key. – Dmitri Dec 8 '16 at 1:04
8

Inheritance is one of those features that I wouldn't touch. AFAIK, it's used internally for replication and partitioning in some capacity. I'm not sure if it was even designed with the intent to be used by the end-user.

Concrete Technical Drawbacks

Drawbacks on UNIQUE and REFERENCES

The docs covers some of the drawbacks in the CAVEAT section (below is important).

  • If we declared parent.name to be UNIQUE or a PRIMARY KEY, this would not stop the child table from having rows with names duplicating rows in parent. And those duplicate rows would by default show up in queries from parent. In fact, by default child would have no unique constraint at all, and so could contain multiple rows with the same name. You could add a unique constraint to child, but this would not prevent duplication compared to parent.
  • Similarly, if we were to specify that parent.name REFERENCES some other table, this constraint would not automatically propagate to child. In this case you could work around it by manually adding the same REFERENCES constraint to child.
  • Specifying that another table's column REFERENCES parent(name) would allow the other table to contain parent names, but not child names. There is no good workaround for this case.

Slow progress developing INHERITs

These deficiencies were first mentioned in the docs to 7.3 released in 1996 though they existed since inheritance was implemented

This deficiency will probably be fixed in some future release.

And the only change was to make the deficiencies more explicit and verbose in the docs to 8.0 released in 2010.

These deficiencies will probably be fixed in some future release, but in the meantime considerable care is needed in deciding whether inheritance is useful for your problem.

Good luck waiting for that some future release. And, some of the things you talk about features just aren't unique to composition,

Saving a "key" is moot

  • no surrogate key on 'sometype', it's explicitly a composition

How is that different from making sometype an attribute list, and linking directly to it?

CREATE TABLE sometype (sometype_name text PRIMARY KEY);
CREATE TABLE foo (foo_id serial PRIMARY KEY);
CREATE TABLE foo_sometype (
  foo_id int REFERENCES foo,
  sometype_name text REFERENCES sometype,
  PRIMARY KEY ( foo_id, sometype_name )
);

Now you don't even have to join foo_sometype to sometype to get sometype.sometype_name.

Table Partitioning

All of those problems aside, it gets even worse with the upcoming PostgreSQL 10 release of table partitioning

Multiple inheritance is not allowed, and partitioning and inheritance can't be mixed

So you want inheritance? Forgo partitioning, which actually has real planner advantages.

ALTER TABLE

Alas, ALTER TABLE has quite a few drawbacks listed in its notes as well,

If a table has any descendant tables, it is not permitted to add, rename, or change the type of a column, or rename an inherited constraint in the parent table without doing the same to the descendants. That is, ALTER TABLE ONLY will be rejected. This ensures that the descendants always have columns matching the parent. [...] A recursive DROP COLUMN operation will remove a descendant table's column only if the descendant does not inherit that column from any other parents and never had an independent definition of the column. A nonrecursive DROP COLUMN (i.e., ALTER TABLE ONLY ... DROP COLUMN) never removes any descendant columns, but instead marks them as independently defined rather than inherited. [...] The TRIGGER, CLUSTER, OWNER, and TABLESPACE actions never recurse to descendant tables; that is, they always act as though ONLY were specified. Adding a constraint recurses only for CHECK constraints that are not marked NO INHERIT.

Conclusion

I don't think many people use inheritance. I've never seen it in the wild. Inheritance in the db adds to the learning curve and some features are just better left alone. You don't have to find an application for them.

You may find this post on Stack Overflow useful, "When to use inherited tables in PostgreSQL?".

  • Thanks for the write-up. I added a couple of explanatory sections to expand on this. I do understand the "it's not used much" argument, but I was really more curious if anyone can come up with concrete technical drawbacks that I am overlooking. As you say, table inheritance has been in pg for 20+ years, it's well-understood, and it's not going anywhere. Sometimes it's also ok to try new things :) – Dmitri Dec 8 '16 at 2:25
  • There are a lot of concrete technical drawbacks in my answer. And, I just added another one. – Evan Carroll Dec 8 '16 at 16:31
  • 1
    Can you explain why those drawbacks apply here? As far as I can tell, they are completely irrelevant in this case, but would like to know if I'm misunderstanding something. – Dmitri Dec 8 '16 at 18:41
  • I have no idea what you value. Clearly not data integrity or scaling. That said, I'm content with my answer to your question and hope others find it useful. – Evan Carroll Dec 11 '16 at 7:24
2

Many reasons to avoid, and all boil down to violating the principles of the relational model. Most obviously, the child tables are bags, not relations, and the lack of foreign keys violate the information principle, which requires all information to be in form of attribute values in tuples.

The biggest prectical consequences are loss of identity integrity in child tables (duplicates have a way of getting inserted when you would least expect) and loss of transparence in the model: you will sorely regret when you have to explain this to the next user of the model, and no, you can never be sure you will be the sole user for life.

As for the relational construct (with foreign keys) being cumbersome to query, it is not. First, one can always create views (derived relations); and second, it is better that the user understands what he is querying anyway, than have automagical behaviour.

2

The mentioned deficiencies are not a reason not to use inheritance! Inheritance works here similar to Class-Inheritance with independent Objects. You may have a class/table 'fruits' and a class/table 'apples' and 'oranges'. Since apples and oranges inherit their meta definitions from fruits, you can fetch them via fruits. However they are independent classes, with independent enumeration, and what else you can expect....

In case you really need to prevent collisions: define a trigger, or a check/foreign key on main table (with NOT in).

But please, just because you are dissatisfied with the way inheritance works in PostgreSQL- do not discourage people from using it! Independent subclasses or tables are wonderful! And if you need some dependency - implement it the way you need it. There are many good tutorials, including official Documentation. Here is another example: Inheritance – One more reason to love PostgreSQL.

I am also using inheritance in one of my projects, despite the possible primary key collisions (my parent table does NOT have primary keys defined, and my model does not require it).

There are also ways to use inheritance to improve performance, not only the ORM.

Difficult is not the same as bad.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.