Revision as of 12:00, 1 October 2007

Evolving Java-based APIs (part 2)

Achieving API Binary Compatibility

"[A]n object-oriented model must be carefully designed so

that class-library transformations that should not break already compiled applications, indeed, do not break such applications."
---Ira Forman, Michael Conner, Scott Danforth, and Larry Raper, "Release-to-Release Binary Compatibility in SOM", in Proceedings of OOPSLA '95.

Achieving API binary compatibility depends in part on the Java language's notion of binary compatibility:

"A change to a type is binary compatible with (equivalently, does not break binary compatibility with) preexisting binaries if preexisting binaries that previously linked without error will continue to link without error." (JLS3, 13.2)

Reference: Gosling, Joy, Steele, and Bracha, The Java Language Specification, Third Edition, Addison-Wesley, 2005; chapter 13 Binary Compatibility.

The tables in the following sections summarize which kinds of changes break API binary compatibility.

Bear in mind that many changes will have effects in several places. For example, defining a new public interface with one public method and adding that interface as the superinterface of an existing interface has the following ramifications:

• a new public API interface is added to an API package
• the superinterface set of the existing API interface has expanded
• a new public API method is added to the existing API interface

Each of these individual net effects could break binary compatibility. Use the tables to determine whether the net effects preserve or break compatibility.

Evolving API packages

It is always possible to evolve the Component API to include a new API package. However, once introduced in a release, an API package cannot easily be withdrawn from service. When an API package becomes obsolete, its API classes and API interfaces should continue to work but be marked as deprecated. After a couple of releases, it may be possible to phase out an obsolete API package.

The names of non-public (non-API) types in API packages do not appear in Client source code or binaries. Non-API types can be added or deleted without jeopardizing binary compatibility. However, once made public in a release, these types are part of the API and cannot easily be withdrawn from service without breaking existing Clients. When an API type becomes obsolete, it should continue to work but be marked as deprecated.
 

(1) API class-interface gender changes break binary compatibility, even in cases where the class/interface is used by, but not implemented by, Clients. This is because the Java VM bytecodes for invoking a method declared in an interface are different from the ones used for invoking a method declared in a class. More generally, all gender changes involving classes, enums, interfaces, and annotation types break binary compatibility for one reason or another.

Evolving API Interfaces

Evolving API interfaces is somewhat more straightforward than API classes since all methods are public and abstract, all fields are public static and final, all type members are public and static, and there are no constructors. Annotation types (@interface) , which are a form of interface, are also covered.

(0) Although adding a new method to an API interface which need not be reimplemented by Clients does not break binary compatibility, a pre-existing Client subclass of an existing implementation might still provide a pre-existing implementation of a method by this name. See #Example 4 - Adding an API method in the preceding section for why this breaks API contract compatibility.

(1) Adding a new method to an API interface that is implemented by Clients (e.g., a callback, listener, or visitor interface) breaks compatibility because hypothetical pre-existing implementations do not implement the new method.

(2) Adding an API field to an API interface that is implemented by Clients (e.g., a callback, listener, or visitor interface) breaks binary compatibility in a different way. A field added to a superinterface of C may hide an instance field inherited from a superclass of C, causing linking errors to be detected. Because of this fact, it is important to distinguish between API interfaces that Clients should implement from those that Clients should merely use. API interfaces that Clients should implement should not include fields.

(3) Shrinking the set of API interfaces that a given API interfaces extends (either directly or inherited) breaks compatibility because some casts between API interfaces in hypothetical pre-existing Client code between will no longer work. However, non-API superinterfaces can be removed without breaking binary compatibility.

(4) Altering the type parameters of a parameterized type breaks compatibility. However, adding type parameters to a previously unparameterized type retains compatibility because of Java's special treatment of legacy references (raw types).

(5) Existing annotations would not have a value for the new element, causing an exception (IncompleteAnnotationException) to be thrown when the annotation is read.

(6) Existing annotations that mention the deleted element will cause an exception (AnnotationTypeMismatchException) to be thrown when the annotation is read.

Evolving API interfaces - API methods

All methods in an API interface are implicitly public and abstract, and are therefore all considered API methods. The same is true for method declarations defining the elements of an API annotation type.

(1) Adding and deleting checked exceptions declared as thrown by an API method does not break binary compatibility; however, it breaks contract compatibility (and source code compatibility).

(2) Adding type parameters to an unparameterized method is a compatible change owing to Java's story for interfacing with non-generic legacy code.

(3) A variable arity method declaration such as "void foo(String... y)" is compiled as if it had been written "void foo(String[] y)".

(4) Although existing binaries will continue to work, existing invocations in source code may not compile because the compiler no longer automatically bundles up the extra arguments into an array.

(5) Defaults are applied dynamically at the time annotations are read. Changing the default value may affect annotations in all classes, including ones compiled before the change was made.

Evolving API interfaces - API fields

All fields in an API interface are implicitly public, static, and final; they are therefore all considered API fields.

Because of binary compatibility problems with fields, the Java Language Specification recommends against using API fields. However, this is not always possible; in particular, enumeration constants to be used in switch statements must be defined as API fields.

(1) All field type changes break binary compatibility, even seemingly innocuous primitive type widenings like turning a short into an int.

(2) Java compilers always inline the value of constant fields (ones with compile-time computable values, whether primitive or String type). As a consequence, changing the value of an API constant field does not affect pre-existing Clients. Invariably, this fails to meet the objective for changing the API field's value in the first place.

Evolving API interfaces - API type members

All type members in an API interface are implicitly public and static; they are therefore considered API type members. The rules for evolving an API type member are basically the same as for API classes and interfaces declared at the package level.

Evolving API Classes

Evolving API classes is somewhat more complex than API interfaces due to the wider variety of modifiers, including protected API members. Enums, which are a form of class, are also covered.

(0) Although adding a new method to an API class which need not be reimplemented by Clients does not break binary compatibility, a pre-existing subclass might still provide a pre-existing implementation of a method by this name. See #Example 4 - Adding an API method in the preceding section for why this breaks API contract compatibility.

(1) Adding a new method to an API class that must be reimplemented by Clients breaks compatibility because pre-existing subclasses would not provide any such implementation.

(2) Adding the first constructor to an API class causes the compiler to no longer generate a default (public, 0 argument) constructor, thereby breaking compatibility with pre-existing code that invoked this API constructor. To avoid this pitfall, API classes should always explicitly declare at least one constructor.

(3) Adding a new field to an API class that is subclassed by Clients breaks binary compatibility. A field in a superinterface of C may hide an added field inherited from a superclass of C, causing linking errors to be detected when a static field hides an instance field. Apart from the binary compatibility issues, it is generally good software engineering practice that API classes should not expose any fields.

(4) Shrinking an API class's set of API superclasses and superinterfaces (either directly or inherited) breaks compatibility because some casts in pre-existing Client code will now longer work. However, non-API superclasses and superinterfaces can be removed without breaking binary compatibility.

(5) Pre-existing binaries that attempt to create new instances of the API class will fail with a link-time or runtime error.

(6) Pre-existing binaries that subclass the API class will fail with a link-time error.

(7) Adding type parameters to an unparameterized type is a compatible change owing to Java's story for interfacing with non-generic legacy code.

(8) Client code can use the values() method to determine the ordinal positions of the enum constants. So although this is a binary compatible change, it may break contractual compatibility.

Evolving API classes - API methods and constructors

(1) Adding and deleting checked exceptions declared as thrown by an API method does not break binary compatibility; however, it breaks contract compatibility (and source code compatibility).

(2) Perhaps surprisingly, the binary format is defined so that changing a member or constructor to be more accessible does not cause a linkage error when a subclass (already) defines a method to have less access.

(3) Pre-existing binaries that invoke the method will fail with a runtime error.

(4) Pre-existing binaries that reimplement the method will fail with a link-time error.

(5) Adding or removing the synchronized modifier also has a bearing on the method's behavior in a multi-threaded world, and may therefore raise a question of contract compatibility.

(6) Adding type parameters to an unparameterized type is a compatible change owing to Java's story for interfacing with non-generic legacy code.

(7) A variable arity method declaration such as "void foo(int x, String... y)" is compiled as if it had been written "void foo(int x, String[] y)".

(8) Although existing binaries will continue to work, existing invocations in source code may not compile because the compiler no longer automatically bundles up the extra arguments into an array.

Evolving API classes - API fields

Because of binary compatibility problems with fields, the Java Language Specification recommends against using API fields. However, this is not always possible; in particular, enumeration constants to be used in switch statements must be defined as API constant fields.
 

(1) All field type changes break binary compatibility, even seemingly innocuous primitive type widenings link turning a short into an int.

(2) Java compilers always inline the value of constant fields (ones with compile-time computable values, whether primitive or String type). As a consequence, changing the value of an API constant field does not affect pre-existing Clients. Invariably, this does not meet the objective for changing the API field's value.

(3) Java compilers always inline the value of constant fields (ones with a compile-time computable values, whether primitive or Stringtype). As a consequence, changing an API constant field into a non-final one does not propagate to pre-existing Clients. Invariably, this does not meet the objective for making the API field non-final.

(4) Making an API field final breaks compatibility with pre-existing binaries that attempt to assign new values to the field.

(5) Changing whether an API field is declared static or not results in link-time errors where the field is used by a pre-existing binary which expected a field of the other kind.

Evolving API classes - API type members

The rules for evolving an API type member are basically the same as for API classes and interfaces declared at the package level, with these additional rules for changing access modifiers:
 

Evolving non-API packages

The names of non-API packages, classes, and interfaces do not appear in Client source code or binaries. Consequently, non-API packages, classes, and interfaces can be added or deleted without jeopardizing binary compatibility. However, when non-API classes and interfaces containing public or protected members are among the superclass or superinterface sets of API classes and interfaces, non-API changes may have ramifications to API methods, fields, and constructors.

Turning non-generic types and methods into generic ones

Generic types and methods were added to the Java language in Java SE 5 (aka JDK 1.5) along with a special story for how legacy non-generic code can continue to use types and methods that have been "generified". The prime example of this is the java.util Collections Framework, which was upgraded to make use of generics while remaining compatible with code that uses collections in the old way.

The key concepts behind Java's special compatibility mechanism are raw types and erasures.

A raw type is a use of a generic type without the normal type arguments. For example, "List" in the declaration statement "List x = null;" is a raw type since List is a generic type declared "public interface List<E> ..." in JDK 1.5. Contrast this to a normal use of List which looks like "List<String> x = null;" or "List<?> x = null;" where a type augument ("String") or wildcard is specified.

The term erasure is suggestive. Imagine going through your code and literally erasing the type parameters from the generic type declaration (e.g., erasing the "<E>" in "public interface List<E> ...") to get a non-generic type declaration, and replacing all occurrence of the deleted type variable with Object. For type parameters with type bounds (e.g., "<E extends T1 & T2 & T3 & ...>"), the leftmost type bound ("T1"), rather than Object, is substituted for the type variable. The resulting declaration is known as the erasure.

According to the special compatibility story, the Java compiler treats a raw type as a reference to the type's erasure. An existing type can be evolved into a generic type by adding type parameters to the type declaration and judiciously introducing uses of the type variables into the signatures of its existing methods and fields. As long as the erasure looks like the corresponding declaration prior to generification, the change is binary compatible with existing code.

As a case study of how to generify an existing API, carefully compare the Java 1.4 Collections Framework[1] classes with their counterparts in the Java 1.5 Collections Framework[2]. You will see that the erasures of the 1.5 versions looks just like the 1.4 versions.

Variable arity methods were also introduced in 1.5, also with a special story for how legacy code can continue to invoke or override a method that has been upgraded. A variable arity method declaration such as "void foo(int x, String... y)" is compiled as if it had been written "void foo(int x, String[] y)". This provides a consistent interpretation for old-style invocations of the form "foo(int 5, new String[]{"a","b"})". The class Arrays.asList[3] is an example of a method that became both generic and variable arity in 1.5.

Two final notes. Regarding whether existing APIs ought to embrace generics, the Java Language Specification says only this:</p>

"The use of raw types is allowed only as a concession to compatibility of legacy code. The use of raw types in code written after the introduction of genericity into the Java programming language is strongly discouraged. It is possible that future versions of the Java programming language will disallow the use of raw types." (JLS3, 4.8[4])

The implication is that code that uses the Collections Framework should use types like "List<?>" instead of the raw type "List".

But, also bear in mind that there are severe constraints on how a type or method that already is generic can be compatibly evolved with respect to its type parameters (see the tables above). So if you plan to generify an API, remember that you only get one chance (release), to get it right. In particular, if you change a type in an API signature from the raw type "List" to "List<?>" or "List<Object>", you will be locked into that decision. The moral is that generifying an existing API is something that should be considered from the perspective of the API as a whole rather than piecemeal on a method-by-method or class-by-class basis.

Evolving annotations on API elements

Annotations were added to the Java language in Java SE 5 (aka JDK 1.5). Annotation can be used on most any named Java element, including packages, types, methods, and fields. It's a natural question to ask whether adding, deleting, or changing an annotation on a API element is a compatible or a breaking change.

On one hand, adding or removing annotations has no effect on the correct linkage of class files by the Java virtual machine. On the other hand, annotations exist to be read via reflective APIs for manipulating annotations. So there is no uniform answer as to what will happen if a given annotation is or is not present on an API element (or non-API element, for that matter). It depends entirely on the specifics of the annotation and the mechanisms that are processing those annotations.

Parties that declare annotation types should try to provide helpful guidance for their customers.

Copyright © 2000, 2007 IBM Corporation