New ACM paper, free-tier cloud, and open-source license

Lesson 9.1: The PERA model

Conceptual, logical, and physical models

In database design, there are normally three separate models for a database: the conceptual, logical, and physical models.

Conceptual model

Describes the data domain in human terms. It is entirely independent of software and hardware used to run the database, and captures the business logic of the domain in abstract form. The most widely adopted conceptual modeling paradigm is Chen’s Entity-Relationship model (ER model).

Logical model

Describes the implementation of the data domain in terms of the DBMS in use. It is independent of the hardware used to run the database. It is constructed by translating the conceptual model, and the business logic of the data domain must be modeled in terms of constraints provided by the DBMS. Some of the most common logical data modeling paradigms (i.e. database paradigms) are the relational model, the document model, the key-value model, the triplestore model, and the graph model.

Physical model

Describes how the data is stored on disk. It is tightly coupled to both the DBMS and hardware used to run the database. Modern DBMSs do not require physical models to be planned, which are implemented automatically by the DBMS based on the logical model. Properties of the physical model may sometimes be altered to optimise database performance, such as by adding indexes.

One of the problems with contemporary database paradigms is the need for translation between the conceptual and logical models. These translations are often not lossless, especially when dealing with polymorphic data, and leads to some of the hardest challenges in data engineering.

The conceptual PERA model

TypeDB avoids the translation between conceptual and logical models by implementing the Polymorphic Entity-Relation-Attribute model (PERA model), an extension of the ER model with dependent typing and polymorphic data structures. The PERA model is a conceptual data model, and there is no separate logical model in TypeDB: the conceptual model is the logical model. This allows database designers to model their data only once, at the highest and most abstract level, and avoid lossy translation processes. Because object models in applications are also conceptual in nature, it is possible to avoid any mismatch between the database model and application model.

Entities, relations, and attributes

A PERA data model comprises data-storing types and interface types (or simply "interfaces") between them, similar to classes and interfaces in OOP. The modeling paradigm is based on the theory of dependent types, and interfaces serve as abstractions of the dependencies between data-storing types. All data-storing types in the PERA model are either entity types, relation types, or attribute types. Each is used to model a different class of concepts in the data domain.

Entity types

Used to represent classes of independent objects. An entity might practically require other concepts to exist, such as a car that cannot exist without its parts, but can be conceptualized without reference to them: a car can be imagined without considering its parts.

Relation types

Used to represent classes of objects dependent on other objects. Every relation must depend on at least one other concept, and cannot be conceptualized without those dependencies: it is impossible to imagine a marriage without considering its spouses. The dependency of a relation type on another object type is described by an exposed role interface, and an object type that implements it is a roleplayer of that role. In the case of marriages, the role would be "spouse", but there could also be other roles such as "officiant" or "witness".

Attribute types

Used to represent classes of values dependent on objects. They represent properties of those objects, such as names of people, dates of marriages, and license plates of cars. Every attribute has a literal value, for instance a string, a number, or a boolean, and the type of this value is determined by the attribute type’s value type. The dependency of an attribute type on an object type is described by an exposed ownership interface, and an object type that implements it is an owner of that attribute type. Unlike relation types, which can have several role interfaces, every attribute type has only a single ownership interface.

We use the terms "entity type", "relation type" and "attribute type" to refer to types, and the terms "entity", "relation", and "attribute" to refer to respective instances of those types.

Comparing objects and values

A key difference between entity or relation types and attribute types is that entity and relation types contain objects whereas attribute types contain values. Entity and relation types can be collectively referred to as object types, as distinct from attribute types, similar to composite types and primitive types respectively in OOP. Much like their OOP analogs, objects types are freely instantiable, so it is possible to create two objects that have identical properties but are not themselves identical. Meanwhile, values are not freely instantiable, so two attributes with the same type and value are in fact the same attribute.

This difference also affects permitted interface implementations in the PERA model. Only object types are able to implement interfaces: own attribute types and play roles in relation types. It is not possible for an attribute type to own another attribute type, or to play a role in a relation type.

In TypeDB 2.x, attribute types can implement interfaces (own other attribute types and play roles in relation types) in the same way object types can, contrary to the PERA model. This behaviour is deprecated and will be removed in TypeDB 3.0. In some niche cases, usage of this feature can result in information loss due to attribute idempotency, so it is strongly advised against.

Polymorphic features

The biggest difference between the ER and PERA models is that the PERA model permits the use of polymorphic features like interfaces, inheritance, and abstraction. Interface implementations are used to define capabilities of object types, and an interface can be independently implemented by multiple object types. Meanwhile, inheritance allows types to be made subtypes of other types, and the interface implementations of an object type are inherited by its subtypes. Finally, abstraction is used to prevent instantiation of a type, which can only be instantiated through a concrete subtype.

Types and casting

The following table summarises the key properties of data-storing types in the PERA model. As we saw in Lesson 5.2, only an abstract attribute type can have subtypes, in order to prevent ambiguities in the interpretation of queries.

Entity types Relation types Attribute types













Can implement interfaces

Can be subtyped

if abstract

Can be abstract

In addition to these data-storing types, interface types and value types together make up the data-structuring types. In fact, all parts of the PERA model are types! This arises from the PERA model’s core philosophy:

The key difference between data-storing and data-structuring types is that only data-storing types can be instantiated directly, and their instances operated on. Meanwhile, the data-structuring types serve to define the behaviours of these data instances. The following diagram summarises the ways in which different kinds of types in the PERA model are associated.

typeql terminology

The upcasting of types into their supertypes enables the use of inheritance polymorphism in queries. Similarly, object types can be upcast into the interface types they implement, which enables interface polymorphism in queries. Finally, attribute types can also be upcast into their value types, which enables arithmetic expressions.

The TypeDB implementation

The PERA model is implemented in TypeDB through TypeQL, its type-theoretic and polymorphic query language. This allows us to define types and interfaces, declare implementations of interfaces, and make use of polymorphic features, as illustrated in the following excerpt from the bookstore schema.

book sub entity,
    owns title,
    owns price,
    plays order-line:item;
hardback sub book,
    owns stock;
paperback sub book,
    owns stock;
ebook sub book;
order sub entity,
    owns id,
    owns status,
    plays order-line:order;
order-line sub relation,
    relates order,
    relates item,
    owns quantity,
    owns price;
title sub attribute, value string;
price sub attribute, value double;
stock sub attribute, value long;
id sub attribute, value string;
status sub attribute, value string;
quantity sub attribute, value long;

Type definitions

A new type is defined using a sub statement. For example, in the above schema excerpt:

  • book sub entity; defines a new entity type with label book.

  • order-line sub relation; defines a new relation type with label order-line.

  • title sub attribute; defines a new attribute type with label title.

Interface definitions

A role interface is defined using a relates statement. The label of the created role interface is given by the label of the dependent relation followed by the name of the role, separated by a : delimiter. For example, in the above schema excerpt:

  • order-line relates item; defines a new role interface with label order-line:item, depended on by the relation type order-line.

Unlike roles, ownership interfaces are not explicitly defined in TypeQL. Because every attribute has only one ownership interface, an attribute’s ownership interface is created implicitly when the attribute is defined. As a result, an ownership does not have an explicitly referenceable label like a role does, but we can describe it with an implicit label comprising the label of the dependent attribute followed by an :OWNER suffix. For example, in the above schema excerpt:

  • title sub attribute; implicitly defines a new ownership interface with implicit label title:OWNER, depended on by the attribute type title.

Interface implementations

An implementation of a role interface is declared using a plays statement. For example, in the above schema excerpt:

  • book plays order-line:item; declares the object type book to implement the role interface order-line:item.

Meanwhile, an implementation of an ownership interface is declared using an owns statement. As ownership interfaces do not have explicit labels, the object of an owns statement is the label of the dependent attribute rather than the interface itself. For example, in the above schema excerpt:

  • book owns title; declares the object type book to implement the ownership interface title:OWNER.

Polymorphic features

Interface implementations are independent, allowing multiple object types to implement the same interfaces, even if they share no common supertypes. For example, in the above schema excerpt:

  • The object types paperback and hardback both implement the ownership interface stock:OWNER.

  • The object types book and order-line both implement the ownership interface price:OWNER.

A type hierarchy is defined using a sub statement. For example, in the above schema excerpt:

  • The types paperback, hardback, and ebook are all defined to be subtypes of book.

When an object type implements interfaces, those implementations are inherited by its subtypes. For example, in the above schema excerpt:

  • The object types paperback, hardback, and ebook all inherit the implementations of the title:OWNER, price:OWNER, and order-line:item interfaces from their supertype book.

Finally, a type if defined to be abstract using an abstract statement. For example, in the above schema excerpt:

  • The type book is declared to be abstract, and can only be instantiated through one of its concrete subtypes: paperback, hardback, and ebook.

Type inference

When TypeQL queries are executed by TypeDB, casting of types takes place automatically via type inference. Let’s consider some example constraints, starting with the following example of inheritance polymorphism.

$x isa book;

In this case, the variable $x is of type book. To resolve this constraint, TypeDB determines the list of types that can be upcast into book, and finds paperback, hardback, and ebook. Thus, instances of these types can be matched for $x. Now let’s consider an example of interface polymorphism.

$y has price $p;

In this case, the variable $y is of type price:OWNER. TypeDB determines that book and order-line can be upcast into price:OWNER, and so can have instances matched for $y. Finally, we’ll consider parametric polymorphism.

$z isa entity;

In this case, the variable $z is of type entity. Thus, TypeDB will match instances of any entity types, as they can all be upcast into entity.

Provide Feedback