2021 Call for Code Awards: Live from New York, with SNL’s Colin Jost! Learn more

Dynamic, typesafe queries in JPA 2.0

The Java developer community has welcomed JPA since its introduction in 2006. The next major update of the specification — version 2.0 (JSR 317) — will be finalized later in 2009. One of the key features introduced in JPA 2.0 is the Criteria API, which brings a unique capability to the Java language: a way to develop queries that a Java compiler can verify for correctness at compile time. The Criteria API also includes mechanisms for building queries dynamically at run time.

This article introduces the Criteria API and the closely associated metamodel concept. You will learn how to use the Criteria API to develop queries that a Java compiler can check for correctness to reduce run-time errors, in contrast to string-based Java Persistence Query Language (JPQL) queries. And through example queries that use database functions or match a template instance, I’ll demonstrate the added power of programmatic query-construction mechanics compared to JPQL queries that use a predefined grammar. The article assumes you have a basic familiarity with Java language programming and common JPA usage such as EntityManagerFactory or EntityManager.

What’s wrong with this JPQL query?

JPA 1.0 introduced JPQL, a powerful query language that’s considered a major reason for JPA’s popularity. However, JPQL — being a string-based query language with a definite grammar — has a few limitations. To understand one of the major limitations, consider the simple code fragment in Listing 1, which executes a JPQL query to select the list of Persons older than 20 years:

Listing 1. A simple (and wrong) JPQL query
EntityManager em = ...;
String jpql = "select p from Person where p.age > 20";
Query query = em.createQuery(jpql);
List result = query.getResultList();

This basic example shows the following key aspects of the query-execution model in JPA 1.0:

  • A JPQL query is specified as a String (line 2).
  • EntityManager is the factory that constructs an executable query instance given a JPQL string (line 3).
  • The result of query execution consists of the elements of an untyped java.util.List.

But this simple example has a serious error. Effectively, the code will compile happily, but it will fail at run time because the JPQL query string is syntactically incorrect. The correct syntax for the second line of Listing 1 is:

String jpql = "select p from Person p where p.age > 20";

Unfortunately, the Java compiler has no way to detect such an error. The error will be encountered at run time at line 3 or line 4 (depending on whether the JPA provider parses a JPQL string according to JPQL grammar during query construction or execution).

How does a typesafe query help?

One of the major advantages of the Criteria API is that it prohibits the construction of queries that are syntactically incorrect. Listing 2 rewrites the JPQL query in Listing 1 using the CriteriaQuery interface:

Listing 2. Basic steps of writing a CriteriaQuery
EntityManager em = ...
CriteriaBuilder qb = em.getCriteriaBuilder();
CriteriaQuery<Person> c = qb.createQuery(Person.class);
Root<Person> p = c.from(Person.class);
Predicate condition = qb.gt(p.get(Person_.age), 20);
TypedQuery<Person> q = em.createQuery(c); 
List<Person> result = q.getResultList();

Listing 2 illustrates the Criteria API’s core constructs and demonstrates its basic usage:

  • Line 1 obtains an EntityManager instance by one of the several available means.
  • In line 2, EntityManager creates an instance of CriteriaBuilder. CriteriaBuilder is the factory for CriteriaQuery.
  • In line 3, the CriteriaBuilder factory constructs a CriteriaQuery instance. A CriteriaQuery is generically typed. The generic type argument declares the type of result this CriteriaQuery will return upon execution. You can supply various kinds of result-type arguments — from a persistent entity such as Person.class to a more free-form one such as Object[]— while constructing a CriteriaQuery.
  • In line 4, query expressions are set on the CriteriaQuery instance. Query expressions are the core units or nodes that are assembled in a tree to specify a CriteriaQuery. Figure 1 shows the hierarchy of query expressions defined in the Criteria API:

    Figure 1. Interface hierarchy of query expressions
    Interface hierarchy of query expressions

    To begin with, the CriteriaQuery is set to query from Person.class. As a result, a Root<Person> instance p is returned. Root is a query expression that denotes the extent of a persistent entity. Root<T> essentially says: “Evaluate this query across all instances of type T.” It is similar to the FROM clause of a JPQL or SQL query. Also notice that Root<Person> is generically typed. (Actually, every expression is.) The type argument is the type of the value the expression evaluates to. So Root<Person> denotes an expression that evaluates to Person.class.

  • Line 5 constructs a Predicate. Predicate is another common form of query expression that evaluates to either true or false. A predicate is constructed by the CriteriaBuilder, which is the factory not only for CriteriaQuery, but also for query expressions. CriteriaBuilder has API methods for constructing all kinds of query expressions that are supported in traditional JPQL grammar, plus a few more. In Listing 2, CriteriaBuilder is used to construct an expression that evaluates whether the value of its first expression argument is numerically greater than the value of the second argument. The method signature is:

    Predicate gt(Expression<? extends Number> x, Number y);
    This method signature is a fine example of how a strongly typed language such as the Java language can be judiciously used to define an API that allows you to express what is correct and prohibits what is not. The method signature specifies that it is possible to compare an expression whose value is a Number only to another Number (and not, for example, to a String):
    Predicate condition = qb.gt(p.get(Person.age), 20);
    But there is more to line 5. Notice the qb.gt() method’s first input argument: `p.get(Person.age),wherepis theRootexpression obtained previously.p.get(Person.age)` is a _path expression. A path expression is the result of navigation from a root expression via one or more persistent attribute(s). Hence p.get(Person_.age) denotes a path expression by navigating from the root expression p by the age attribute of Person. You may wonder what Person_.age is. For the time being, assume it is a way to denote the age attribute of Person. I’ll elaborate on the meaning of Person_.age when I discuss the new Metamodel API introduced in JPA 2.0.

    As I mentioned earlier, every query expression is generically typed to denote the type of the value the expression evaluates to. The path expression p.get(Person_.age) evaluates to an Integer if the age attribute in Person.class is declared to be of type Integer(or int). Because of the type safety inherent in the API, the compiler itself will raise an error for a meaningless comparison, such as:

    Predicate condition = qb.gt(p.get(Person_.age, "xyz"));

  • Line 6 sets the predicate on the CriteriaQuery as its WHERE clause.
  • In line 7, EntityManager creates an executable query given an input CriteriaQuery. This is similar to constructing an executable query given a JPQL string as input. But because the input CriteriaQuery carries richer type information, the result is a TypedQuery that is an extension of the familiar javax.persistence.Query. The TypedQuery, as the name suggests, knows the type it returns as a result of its execution. It is defined as:
    public interface TypedQuery<T> extends Query {
                 List<T> getResultList();
    As opposed to corresponding untyped super-interface:
    public interface Query {
    List getResultList();
    Naturally, the TypedQuery result has the same Person.class type specified during the construction of the input CriteriaQuery by a CriteriaBuilder (line 3).
  • In line 8, the type information that is carried throughout shows its advantage when the query is finally executed to get a list of results. The result is a typed list of Persons that saves the developer of trouble of an extra (and often ugly) cast (and minimizes the risk of ClassCastException error at run time) while iterating through the resultant elements.

To summarize the basic aspects of the simple example in Listing 2:

  • CriteriaQuery is a tree of query-expression nodes that are used to specify query clauses such as FROM, WHERE, and ORDER BY in a traditional string-based query language. Figure 2 shows the clauses related to a query:
    Figure 2. CriteriaQuery encapsulates the clauses of a traditional query
    Interface hierarchy of query expressions
  • The query expressions are generically typed. A few typical expressions are:
    • Root<T>, which is equivalent to a FROM clause.
    • Predicate, which evaluates to a Boolean value of true or false. (In fact, it is declared as interface Predicate extends Expression<Boolean>.)
    • Path<T>, which denotes a persistent attribute navigated from a Root<_?_>Root<?> expression. Root<T> is a special Path<T> with no parent.
  • CriteriaBuilder is the factory for CriteriaQuery and query expressions of all sorts.
  • CriteriaQuery is transferred to an executable query with its type information preserved so that the elements of the selected list can be accessed without any run-time casting.

Metamodel of a persistent domain

The discussion of Listing 2 points out an unusual construct: Person_.age, which is a designation for the persistent age attribute of Person. Listing 2 uses Person_.age to form a path expression navigating from a Root<Person> expression p by p.get(Person_.age). Person_.age is a public static field in the Person_ class, and Person_ is the static, instantiated, canonical metamodel class corresponding to the original Person entity class.

A metamodel class describes the meta information of a persistent class. A metamodel class is canonical if the class describes the meta information of a persistent entity in the exact manner stipulated by the JPA 2.0 specification. A canonical metamodel class is static in the sense all its member variables are declared static (and public). The Person_.age is one such static member variable. You instantiate the canonical class by generating a concrete Person_.java at a source-code level at development time. Through such instantiation, it is possible to refer to persistent attributes of Person at compile time, rather than at run time, in a strongly typed manner.

This Person_ metamodel class is an alternative means of referring to meta information of Person. This alternative is similar to the much-used (some may say, abused) Java Reflection API, but with a major conceptual difference. You can use reflection to obtain the meta information about an instance of a java.lang.Class, but meta information about Person.class cannot be referred to in a way that a compiler can verify. For example, using reflection, you’d refer to the field named age in Person.class with:

Field field = Person.class.getField("age");

However, this technique is fraught with a limitation similar to the one you observed in the case of the string-based JPQL query in Listing 1. The compiler feels happy about this piece of code but cannot verify whether it will work. The code can fail at run time if it includes even a simple typo. Reflection will not work for what JPA 2.0’s typesafe query API wants to accomplish.

A typesafe query API must enable your code to refer to the persistent attribute named age in a Person class in a way that a compiler can verify at compile time. The solution that JPA 2.0 provides is the ability to instantiate a metamodel class named Person_ that corresponds to Person by exposing the same persistent attributes statically.

Any discussion about meta or meta-meta information often induces somnolence. So I’ll present a concrete example of a metamodel class for a familiar Plain Old Java Object (POJO) entity class —domain.Person, shown in Listing 3:

Listing 3. A simple persistent entity
package domain;
public class Person {
  private long ssn;
  private string name;
  private int age;

  // public getter/setter methods
  public String getName() {...}

This is a typical definition of a POJO, with annotations — such as @Entity or @Id— that enable a JPA provider to manage instances of this class as persistent entities.

The corresponding static canonical metamodel class of domain.Person is shown in Listing 4:

Listing 4. Canonical metamodel for a simple entity
package domain;
import javax.persistence.metamodel.SingularAttribute;


public class Person_ {
  public static volatile SingularAttribute<Person,Long> ssn;
  public static volatile SingularAttribute<Person,String> name;
  public static volatile SingularAttribute<Person,Integer> age;

The metamodel class declares each persistent attribute of the original domain.Person entity as a static public field of SingularAttribute<person,_?_>SingularAttribute<Person,?> type. Making use of this Person_metamodel class, I can refer to the persistent attribute of domain.Person named age— not via the Reflection API, but as a direct reference to the static Person_.age field — at compile time. The compiler can then enforce type checking based on the declared type of the attribute named age. I’ve already cited an example of such a restriction: CriteriaBuilder.gt(p.get(Person_.age), "xyz") will cause a compiler error because the compiler can determine from the signature of CriteriaBuilder.gt(..) and type of Person_.age that a Person‘s age is a numeric field and cannot be compared against a String.

A few other key points to notice are:

  • The metamodel Person_.age field is declared to be of type javax.persistence.metamodel.SingularAttribute. SingularAttribute is one of the interfaces defined in the JPA Metamodel API, which I’ll describe in the next section. The generic type arguments of a SingularAttribute<Person, Integer> denote the class that declares the original persistent attribute and the type of the persistent attribute itself.
  • The metamodel class is annotated as @StaticMetamodel(domain.Person.class) to designate it as a metamodel class corresponding to the original persistent domain.Person entity.

The Metamodel API

I’ve defined a metamodel class as a description of a persistent entity class. Just as the Reflection API requires other interfaces — such as java.lang.reflect.Field or java.lang.reflect.Method — to describe the constituents of java.lang.Class, so the JPA Metamodel API requires other interfaces, such as SingularAttribute, PluralAttribute, to describe a metamodel class’s types and their attributes.

Figure 3 shows the interfaces defined in the Metamodel API to describe types:

Figure 3. Interface hierarchy for persistent types in the Metamodel API
Interface hierarchy for persistent types in the Metamodel API

Figure 4 shows the interfaces defined in the Metamodel API to describe attributes:

Figure 4. Interface hierarchy of persistent attributes in the Metamodel API
Interface hierarchy of persistent attributes in the Metamodel API

The interfaces of JPA’s Metamodel API are more specialized than those of the Java Reflection API. This finer distinction is required to express rich meta information about persistence. For example, the Java Reflection API represents all Java types as java.lang.Class. That is, no special distinction is made via separate definitions among concepts such as class, abstract class, and an interface. Of course, you can ask a Class whether it is an interface or if it is abstract — but that is not the same as representing the concept of interface differently from an abstract class via two separate definitions.

The Java Reflection API was introduced at the inception of the Java language (and was quite a pioneering concept at that time for a common general-purpose programming language), but awareness of the use and power of strongly typed systems has progressed over the years. The JPA Metamodel API harnesses that power to introduce strong typing for persistent entities. For example, persistent entities are semantically distinguished as MappedSuperClass, Entity, and Embeddable. Before JPA 2.0, this semantic distinction was represented via corresponding class-level annotations in the persistent-class definition. JPA Metamodel describes three separate interfaces — MappedSuperclassType, EntityType, and EmbeddableType — in the javax.persistence.metamodel package to bring their semantic specialties into sharper focus. Similarly, the persistent attributes are distinguished at type-definition level via interfaces such as SingularAttribute, CollectionAttribute, and MapAttribute.

Aside from the aesthetics of description, these specialized metamodel interfaces have practical advantages that help to build typesafe queries and reduce the chance of run-time errors. You’ve seen some of these advantages in the earlier examples, and you will see more of them when I describe examples of joins using CriteriaQuery.

Run-time scope

Broadly speaking, one can draw some parallels between the traditional interfaces of the Java Reflection API and the interfaces of javax.persistence.metamodel specialized to describe persistence metadata. To further the analogy, an equivalent concept of run-time scope is needed for the metamodel interfaces. The java.lang.Class instances are scoped by java.lang.ClassLoader at run time. A set of Java class instances that reference one another must all be defined under the scope of a ClassLoader. The set boundaries are strict or closed in the sense that if a class A defined under the scope of ClassLoader L tries to refer to class B, which is not under the scope of ClassLoader``L, the result is a dreaded ClassNotFoundException or NoClassDef FoundError (and often sleep deprivation for a developer or deployer for environments with multiple ClassLoaders).

This notion of run-time scope as a strict set of mutually referable classes is captured in JPA 1.0 as a persistence unit. The scope of a persistence unit in terms of its persistent entities is enumerated in the <class> clause of a META-INF/persistence.xml file. In JPA 2.0, the scope is made available to the developer at run time via the javax.persistence.metamodel.Metamodel interface. The Metamodel interface is the holder of all persistent entities known to a specific persistence unit, as illustrated in Figure 5:

Figure 5. Metamodel interface is the container of types in a persistence unit
Metamodel interface is the container of types in a persistence unit

This interface lets the metamodel elements be accessed by their corresponding persistent-entity class. For example, to obtain a reference to the persistent metadata for a Person persistent entity, you can write:

EntityManagerFactory emf = ...;
Metamodel metamodel = emf.getMetamodel();
EntityType<Person> pClass = metamodel.entity(Person.class);

This is analogous, with slightly different style and idioms, to obtaining a Class by its name via ClassLoader:

ClassLoader classloader =  Thread.currentThread().getContextClassLoader();
Class<?> clazz = classloader.loadClass("domain.Person");

EntityType<Person> can be browsed at run time to get the persistent attributes declared in the Person entity. If the application invokes a method on pClass such as pClass.getSingularAttribute("age", Integer.class), it will return a SingularAttribute<Person, Integer> instance that is effectively the same as the static Person_.age member of the instantiated canonical metamodel class. Essentially, the persistent attribute that the application can refer to at run time via the Metamodel API is made available to a Java compiler by instantiation of the static canonical metamodel Person_class.

Apart from resolving a persistent entity to its corresponding metamodel elements, the Metamodel API also allows access to all the known metamodel classes (Metamodel.getManagedTypes()) or access to a metamodel class by its persistence-specific information — for example embeddable(Address.class), which returns a EmbeddableType<Address> instance that is a subinterface of ManagedType<>.

In JPA, the meta information about a POJO is further attributed with persistent specific meta information — such as whether a class is embedded or which fields are used as primary key — with source-code level annotations (or XML descriptors). The persistent meta information falls into two broad categories: for persistence (such as @Entity) and for mapping (such as @Table). In JPA 2.0, the metamodel captures the metadata only for persistence annotations —not for the mapping annotation. Hence, with the current version of the Metamodel API, it’s possible to know which fields are persistent, but it’s not possible to find out which database columns they are mapped to.

Canonical vs. non-canonical

Although the JPA 2.0 specification stipulates the exact shape of a canonical static metamodel class (including the metamodel class’s fully qualified name and the names of its static fields), it is entirely possible for an application to write these metamodel classes as well. If the application developer writes the metamodel classes, they are called non-canonical metamodel. Currently, the specification for non-canonical metamodel is not very detailed, and support for non-canonical metamodel can be nonportable across the JPA providers. You may have noticed that the public static fields are only declared in canonical metamodel but not initialized. The declaration makes it possible to refer to these fields during development of a CriteriaQuery. But they must be assigned a value to be useful at run time. While it is the responsibility of the JPA provider to assign values to these fields for canonical metamodel, similar warranty is not extended for non-canonical metamodel. Applications that use non-canonical metamodel must either depend on specific vendor mechanisms or devise their own mechanics to initialize the field values to metamodel attributes at run time.

Annotation processing and metamodel generation

Quite naturally, if you have many persistent entities you will not be inclined to write the metamodel classes yourself. The persistence provider is expected to generate these metamodel classes for you. The specification doesn’t mandate such a facility or the generation mechanics, but an implicit understanding among JPA providers is that they’ll generate the canonical metamodel using the Annotation Processor facility integrated in the Java 6 compiler. Apache OpenJPA provides a utility to generate these metamodel classes either implicitly when you compile the source code for persistent entities or by explicitly invoking a script. Prior to Java 6, an annotation processor tool called apt was available and widely used — but with Java 6, the coupling between compiler and Annotation Processor is defined as part of the standard.

The process of generating these metamodel classes in OpenJPA as your persistence provider is as simple as compiling the POJO entity with OpenJPA class libraries in the compiler’s classpath:

$ javac domain/Person.java

The canonical metamodel Person_ class will be generated, written in the same source directory as Person.java, and compiled as a side-effect of this compilation.

Writing queries in a typesafe manner

So far, I’ve established the components of CriteriaQuery and its associated metamodel classes. Now I’ll show you how to develop some queries with the Criteria API.

Functional expressions

Functional expressions apply a function to one or more input arguments to create a new expression. The functional expression’s type depends on the nature of the function and type of its arguments. The input arguments themselves can be expressions or literal values. The compiler’s type-checking rules coupled with the API’s signature govern what constitutes legitimate input.

Consider a single-argument expression that applies averaging on its input expression. Listing 5 shows the CriteriaQuery to select the average balance of all Accounts:

Listing 5. Functional expression in CriteriaQuery
CriteriaQuery<Double> c = cb.createQuery(Double.class);
Root<Account> a = c.from(Account.class);


An equivalent JPQL query would be:

String jpql = "select avg(a.balance) from Account a";

In Listing 5, the CriteriaBuilder factory (represented by the variable cb) creates an avg() expression and uses that expression in the query’s select() clause.

The query expression is a building block that can be assembled to define the final selection predicate for the query. The example in Listing 6 shows a Path expression created by navigating to the balance of Account, and then the Path expression is used as an input expression in a couple of binary functional expressions — greaterThan() and lessThan() — both of which result in a Boolean expression or simply a predicate. The predicates are then combined via an and() operation to form the final selection predicate to be evaluated by the query’s where() clause:

Listing 6. where() predicate in CriteriaQuery
CriteriaQuery<Account> c = cb.createQuery(Account.class);
Root<Account> account = c.from(Account.class);
Path<Integer> balance = account.get(Account_.balance);
       (cb.greaterThan(balance, 100), 
        cb.lessThan(balance), 200)));

An equivalent JPQL query would be:

"select a from Account a where a.balance>100 and a.balance<200";

Complex predicates

Certain expressions — such as in()— apply to a variable number of expressions. Listing 7 shows an example:

Listing 7. Multivalued expression in CriteriaQuery
CriteriaQuery<Account> c = cb.createQuery(Account.class);
Root<Account> account = c.from(Account.class);
Path<Person> owner = account.get(Account.owner);
Path<String> name = owner.get(Person.name);

This example navigates from Account via two steps to create a path expression representing the name of an account’s owner. Then it creates an in() expression with the path expression as input. The in() expression evaluates if its input expression equals any of its variable number of arguments. These arguments are specified through the value() method on the In<T> expression, which has the following method signature:

In<T> value(T value);

Notice how Java generics are used to specify that an In<T> expression can be evaluated for membership only with values of type T. Because the path expression representing the Account owner’s name is of type String, the only valid comparison is against String-valued arguments, which can be either a literal or another expression that evaluates to String.

Contrast the query in Listing 7 with the equivalent (correct) JPQL:

"select a from Account a where a.owner.name in ('X','Y','Z')";

A slight oversight in the JPQL will not only be undetected by the compiler but also will produce an unintended outcome. For example:

"select a from Account a where a.owner.name in (X, Y, Z)";

Joining relationships

Although the examples in Listing 6 and Listing 7 use expressions as the building blocks, the queries are based on a single entity and its attributes. But often queries involve more than one entity, which requires you to join two or more entities. CriteriaQuery expresses joining two entities by typed join expressions. A typed join expression has two type parameters: the type you are joining from and the bindable type of the attribute being joined. For example, if you want to query for the Customers whose one or more PurchaseOrder(s) are not delivered yet, you need to express this by an expression that joins Customer to PurchaseOrders, where Customer has a persistent attribute named orders of type java.util.Set<PurchaseOrder>, as shown in Listing 8:

Listing 8. Joining a multivalued attribute
CriteriaQuery<Customer> q = cb.createQuery(Customer.class);
Root<Customer> c = q.from(Customer.class);
SetJoin<Customer, PurchaseOrder> o = c.join(Customer_.orders);

The join expression created from the root expression c and the persistent Customer.orders attribute is parameterized by the source of join — that is, Customer— and the bindable type of the Customer.orders attribute, which is PurchaseOrder and not the declared type java.util.Set<PurchaseOrder>. Also notice that because the original attribute is of type java.util.Set, the resultant join expression is SetJoin, which is a specialized Join for an attribute of declared type java.util.Set. Similarly, for other supported multivalued persistent attribute types, the API defines CollectionJoin, ListJoin, and MapJoin. ( Figure 1 shows the various join expressions.) There is no need for an explicit cast in the third line in Listing 8 because CriteriaQuery and the Metamodel API recognize and distinguish attribute types that are declared as java.util.Collection or List or Set or Map by overloaded methods for join().

The joins are used in queries to form a predicate on the joined entity. So if you want to select the Customers with one or more undelivered PurchaseOrder(s), you can define a predicate by navigating from the joined expression o via its status attribute, comparing it with DELIVERED status, and negating the predicate as:

Predicate p = cb.equal(o.get(PurchaseOrder_.status), Status.DELIVERED)

One noteworthy point about creating a join expression is that every time you join from an expression, it returns a new join expression, as shown in Listing 9:

Listing 9. Every join creates a unique instance
SetJoin<Customer, PurchaseOrder> o1 = c.join(Customer.orders);
SetJoin<Customer, PurchaseOrder> o2 = c.join(Customer.orders);
assert o1 == o2;

The assertion in Listing 9 for equality of two join expressions from the same expression c with the same attribute will fail. Thus, if a query involves a predicate for PurchaseOrders that are not delivered and whose value is more than $200, then the correct construction is to join PurchaseOrder with the root Customer expression only once, assign the resultant join expression to a local variable (equivalent to a range variable in JPQL terminology), and use the local variable in forming the predicate.

Using parameters

Take a look back at this article’s original JPQL query (the correct one):

String jpql = "select p from Person p where p.age > 20";

Though queries are often written with constant literals, it is not a good practice. The good practice is to parameterize a query, which allows the query to be parsed or prepared only once, cached, and reused. So a better way to write the query is to use a named parameter:

String jpql = "select p from Person p where p.age > :age";

A parameterized query binds the value of the parameter before the query execution:

Query query = em.createQuery(jpql).setParameter("age", 20);
List result = query.getResultList();

In a JPQL query, the parameters are encoded in the query string as either named (preceded by a colon — for example, :age) or positional (preceded by a question mark — for example, ?3). In CriteriaQuery, the parameters themselves are query expressions. Like any other expression, they are strongly typed and constructed by the expression factory — namely, CriteriaBuilder. The query in Listing 2, then, can be parameterized as shown in Listing 10:

Listing 10. Using parameters in a CriteriaQuery
ParameterExpression<Integer> age = qb.parameter(Integer.class);
Predicate condition = qb.gt(p.get(Person_.age), age);
TypedQuery<Person> q = em.createQuery(c); 
List<Person> result = q.setParameter(age, 20).getResultList();

To contrast the usage of parameters to that of JPQL: the parameter expression is created with explicit type information to be an Integer and is directly used to bind a value of 20 to the executable query. The extra type information is often useful for reducing run-time errors, because it prohibits the parameter from being compared against an expression of an incompatible type or being bound with a value of an inadmissible type. Neither form of compile-time safety is warranted for the parameters of a JPQL query.

The example in Listing 10 shows an unnamed parameter expression that is directly used for binding. It is also possible to assign a name to the parameter as the second argument during its construction. In that case, you can bind the parameter value to the query using that name. What is not possible, however, is use of positional parameters. Integral position in a (linear) JPQL query string makes some intuitive sense, but the notion of an integral position cannot be carried forward in the context of CriteriaQuery, where the conceptual model is a tree of query expressions.

Another interesting aspect of JPA query parameters is that they do not have intrinsic value. A value is bound to a parameter in the context of an executable query. So it is perfectly legal to create two separate executable queries from the same CriteriaQuery and bind two different integer values to the same parameter for these executable queries.

Projecting the result

You’ve seen that the type of result a CriteriaQuery will return upon execution is specified up front when a CriteriaQuery is constructed by CriteriaBuilder. The query’s result is specified as one or more projection terms. There are two ways to specify the projection term on the CriteriaQuery interface:

CriteriaQuery<T> select(Selection<? extends T> selection);
CriteriaQuery<T> multiselect(Selection<?>... selections);

The simplest and often used projection term is the candidate class of the query itself. It can be implicit, as shown in Listing 11:

Listing 11. CriteriaQuery selects candidate extent by default
CriteriaQuery<Account> q = cb.createQuery(Account.class);
Root<Account> account = q.from(Account.class);
List<Account> accounts = em.createQuery(q).getResultList();

In Listing 11, the query from Account does not explicitly specify its selection term and is the same as explicitly selecting the candidate class. Listing 12 shows a query that uses an explicit selection term:

Listing 12. CriteriaQuery with explicit single selection term
CriteriaQuery<Account> q = cb.createQuery(Account.class);
Root<Account> account = q.from(Account.class);
List<Account> accounts = em.createQuery(q).getResultList();

When the projected result of the query is something other than the candidate persistent entity itself, several other constructs are available to shape the result of the query. These constructs are available in the CriteriaBuilder interface, as shown in Listing 13:

Listing 13. Methods to shape query result

<Y> CompoundSelection<Y> construct(Class<Y> result, Selection<?>... terms);
    CompoundSelection<Tuple> tuple(Selection<?>... terms);

The methods in Listing 13 build a compound projection term composed of other selectable expressions. The construct() method creates an instance of the given class argument and invokes a constructor with values from the input selection terms. For example, if CustomerDetails — a nonpersistent entity — has a constructor that takes String and int arguments, then a CriteriaQuery can return CustomerDetails as its result by creating instances from name and age of selected Customer — a persistent entity — instances, as shown in Listing 14:

Listing 14. Shaping query result into instances of a class by construct()
CriteriaQuery<CustomerDetails> q = cb.createQuery(CustomerDetails.class);
Root<Customer> c = q.from(Customer.class);
              c.get(Customer.name), c.get(Customer.age));

Multiple projection terms can also be combined into a compound term that represents an Object[] or Tuple. Listing 15 shows how to pack the result into an Object[]:

Listing 15. Shaping query result into an Object[]
CriteriaQuery<Object[]> q = cb.createQuery(Object[].class);
Root<Customer> c = q.from(Customer.class);
q.select(cb.array(c.get(Customer.name), c.get(Customer.age));
List<Object[]> result = em.createQuery(q).getResultList();

This query returns a result list in which each element is an Object[] of length 2, the zero-th array element is Customer‘s name, and the first element is Customer‘s age.

Tuple is a JPA-defined interface to denote a row of data. A Tuple is conceptually a list of TupleElements — where TupleElement is the atomic unit and the root of all query expressions. The values contained in a Tuple can be accessed by either a 0-based integer index (similar to the familiar JDBC result), an alias name of the TupleElement, or directly by the TupleElement. Listing 16 shows how to pack the result into a Tuple:

Listing 16. Shaping query result into Tuple
CriteriaQuery<Tuple> q = cb.createTupleQuery();
Root<Customer> c = q.from(Customer.class);
TupleElement<String> tname = c.get(Customer.name).alias("name");
q.select(cb.tuple(tname, c.get(Customer.age).alias("age");
List<Tuple> result = em.createQuery(q).getResultList();
String name = result.get(0).get(name);
String age  = result.get(0).get(1);

This query returns a result list each element of which is a Tuple. Each tuple, in turn, carries two elements — accessible either by index or by the alias, if any, of the individual TupleElements, or directly by the TupleElement. Two other noteworthy points in Listing 16 are the use of alias(), which is a way to attach a name to any query expression (creating a new copy as a side-effect), and a createTupleQuery() method on CriteriaBuilder, which is merely an alternative to createQuery(Tuple.class).

The behavior of these individual result-shaping methods and what is specified as the result type argument of the CriteriaQuery during construction are combined into the semantics of the multiselect() method. This method interprets its input terms based on the result type of the CriteriaQuery to arrive at the shape of the result. To construct CustomerDetails instances as in Listing 14 using multiselect(), you need to specify the CriteriaQuery to be of type CustomerDetails and simply invoke multiselect() with the terms that will compose the CustomerDetails constructor, as shown in Listing 17:

Listing 17. multiselect() interprets terms based on result type
CriteriaQuery<CustomerDetails> q = cb.createQuery(CustomerDetails.class);
Root<Customer> c = q.from(Customer.class);
q.multiselect(c.get(Customer.name), c.get(Customer.age));

Because the query result type is CustomerDetails, multiselect() interprets its argument projection terms as the constructor argument to CustomerDetails. If the query were specified to return a Tuple, the multiselect() method with the exact same arguments would create Tuple instances instead, as shown in Listing 18:

Listing 18. Creating Tuple instances with multiselect()
CriteriaQuery<Tuple> q = cb.createTupleQuery();
Root<Customer> c = q.from(Customer.class);
q.multiselect(c.get(Customer.name), c.get(Customer.age));

The behavior of multiselect() gets more interesting with Object as result type or if no type argument is specified. In such cases, if multiselect() is used with a single input term, then the return value is the selected term itself. But if multiselect() contains more than one input term, the result is an Object[].

Advanced features

So far, I have mainly emphasized the strongly typed nature of the Criteria API and the fact that it helps to minimize syntactic errors that can creep into string-based JPQL queries. The Criteria API is also a mechanism for building queries programmatically and so is often referred to as a dynamic query API. The power of a programmable query construction API is limited only by the inventiveness of its user. I’ll present four examples:

  • Using a weakly typed version of the API to build dynamic queries
  • Using a database-supported function as a query expression to extend the grammar
  • Editing a query for search-within-result functionality
  • Query-by-example — a familiar pattern popularized by the object-database community

Weak typing and dynamic query building

The Criteria API’s strong type checking is based on the availability of instantiated metamodel classes at development time. However, for some use cases, the entities to be selected can only be determined at run time. To support such usage, the Criteria API methods provide a parallel version in which persistent attributes are referred by their names (similar to the Java Reflection API) rather than by reference to instantiated static metamodel attributes. This parallel version of the API can support truly dynamic query construction by sacrificing the type checking at compile time. Listing 19 rewrites the example in Listing 6 using the weakly typed version:

Listing 19. Weakly typed query

Class<Account> cls =Class.forName("domain.Account");
Metamodel model = em.getMetamodel();
EntityType<Account> entity = model.entity(cls); 
CriteriaQuery<Account> c = cb.createQuery(cls);
Root<Account> account = c.from(entity);
Path<Integer> balance = account.<Integer>get("balance");
       (cb.greaterThan(balance, 100), 
        cb.lessThan(balance), 200)));

The weakly typed API, however, cannot return correct generically typed expressions, thereby generating a compiler warning for an unchecked cast. One way to get rid of these pesky warning messages is to use a relatively rare facility of Java generics: parameterized method invocation, as shown in Listing 19‘s invocation of the get() method to obtain a path expression.

Extensible datastore expressions

A distinct advantage of a dynamic query-construction mechanism is that the grammar is extensible. For example, you can use the function() method in the CriteriaBuilder interface to create an expression supported by the database:

<T> Expression<T> function(String name, Class<T> type, Expression<?>...args);

The function() method creates an expression of the given name and zero or more input expressions. The function() expression evaluates to the given type. This allows an application to create a query that evaluates a database function. For example, the MySQL database supports a CURRENT_USER() function that returns the user-name and host-name combination as a UTF-8 encoded string for the MySQL account that the server used to authenticate the current client. An application can use the CURRENT_USER() function, which takes no argument, in a CriteriaQuery, as Listing 20 demonstrates:

Listing 20. Using a database-specific function in a CriteriaQuery
CriteriaQuery<Tuple> q = cb.createTupleQuery();
Root<Customer> c = q.from(Customer.class);
Expression<String> currentUser = 
    cb.function("CURRENTUSER", String.class, (Expression<?>[])null);
q.multiselect(currentUser, c.get(Customer.balanceOwed));

Notice that an equivalent query is simply not possible to express in JPQL, because it has a defined grammar with a fixed number of supported expressions. A dynamic API is not strictly limited by a fixed set of expressions.

Editable query

A CriteriaQuery can be edited programmatically. The clauses of the query — such as its selection terms, the selection predicate in a WHERE clause, and ordering terms in an ORDER BY clause — can all be mutated. This editing capability can be used in a typical “search-within-result”-like facility whereby a query predicate is further refined in successive steps by adding more restrictions.

The example in Listing 21 creates a query that orders its result by name and then edits the query to order also by postal (ZIP) code:

Listing 21. Editing a CriteriaQuery
CriteriaQuery<Person> c = cb.createQuery(Person.class);
Root<Person> p = c.from(Person.class);
List<Person> result = em.createQuery(c).getResultList();
// start editing
List<Order> orders = c.getOrderList();
List<Order> newOrders = new ArrayList<Order>(orders);
List<Person> result2 = em.createQuery(c).getResultList();

The setter methods on CriteriaQueryselect(), where(), or orderBy() — erase the previous values and replace them with new arguments. The list returned by the corresponding getter methods such as getOrderList() is not live — that is, adding or removing elements on the returned list does not modify the CriteriaQuery; furthermore, some vendors may even return an immutable list to prohibit inadvertent usage. So a good practice is to copy the returned list into a new list before adding or removing new expressions.


Another useful facility in a dynamic query API is that it can support query-by-example with relative ease. Query-by-example (developed by IBM® Research in 1970) is often cited as an early example of end-user usability for software. The idea of query-by-example is that instead of specifying the exact predicates for a query, a template instance is presented. Given the template instance, a conjunction of predicates — where each predicate is a comparison for a nonnull, nondefault attribute value of the template instance — is created. Execution of this query evaluates the predicate to find all instances that match the template instance. Query-by-example was considered for inclusion in the JPA 2.0 specification but is not included. OpenJPA supports this style of query through its extended OpenJPACriteriaBuilder interface, as shown in Listing 22:

Listing 22. Query-by-example using OpenJPA’s extension of CriteriaQuery
CriteriaQuery<Employee> q = cb.createQuery(Employee.class);

Employee example = new Employee();

q.where(cb.qbe(q.from(Employee.class), example);

As this example shows, OpenJPA’s extension of the CriteriaBuilder interface supports the following expression:

public <T> Predicate qbe(From<?, T> from, T template);

This expression produces a conjunction of predicates based on the attribute values of the given template instance. For example, this query will find all Employees with a salary of 10000 and rating of 1. The comparison can be further controlled by specifying an optional list of attributes to be excluded from comparison and the style of comparison for String-valued attributes. (See Resources for a link to the Javadoc for OpenJPA’s CriteriaQuery extensions.)


This article has introduced the new Criteria API in JPA 2.0 as a mechanism for developing dynamic, typesafe queries in the Java language. A CriteriaQuery is constructed at run time as a tree of strongly typed query expressions whose use the article has illustrated with a series of code examples.

This article also establishes the critical role of the new Metamodel API and shows how instantiated metamodel classes enable the compiler to verify the correctness of the queries, thereby avoiding run-time errors caused by syntactically incorrect JPQL queries. Besides enforcing syntactic correctness, JPA 2.0’s facilities for programmatic construction of queries can lead to more powerful usage, such as query-by-example, using database functions, and — I hope — many other innovative uses of these powerful new APIs that this article’s readers will devise.


I acknowledge Rainer Kwesi Schweigkoffer for his careful review of this article and valuable suggestions, and fellow members of JPA 2.0 Expert Group for explaining the finer points of this powerful API. I also thank Fay Wang for her contribution, and Larry Kestila and Jeremy Bauer for their support during development of the Criteria API for OpenJPA.