Joseph D. Darcy's Sun Weblog

Thursday April 17, 2008

Kinds of Compatibility: Source, Binary, and Behavioral

Every change is an incompatible change. A risk/benefit analysis is always required.
—Martin Buchholz
Veteran JDK Engineer

When evolving the JDK, compatibility concerns are taken very seriously. However, different standards are applied to evolving various aspects of the platform. From a certain point of view, it is true that any observable difference could potentially cause some unknown application to break. Indeed, just changing the reported version number is incompatible in this sense because, for example, a JNLP file can refuse to run an application on later versions of the platform. Therefore, since not making any changes at all is clearly not viable for evolving the platform, changes need to be evaluated against and managed according to a variety of compatibility contracts.

For Java programs, there are three main categories of compatibility:

1. Source: Source compatibility concerns translating Java source code into class files.

2. Binary: Binary compatibility is defined in The Java Language Specification as preserving the ability to link without error.

3. Behavioral: Behavioral compatibility includes the semantics of the code that is executed at runtime.

Note that non-source compatibility is sometimes colloquially referred to as "binary compatibility." Such usage is incorrect since the JLS spends an entire chapter precisely defining the term binary compatibility; often behavioral compatibility is the intended notion instead.

There are many other observable aspects of the JDK not related to Java programs, such as file layout, etc. Those will not be further discussed in this note.

The basic challenge of compatibility is the difficulty of finding and modifying all the software and systems impacted by a change. In a closed-world scenario where all the clients of an API are known and can in principle be simultaneously changed, introducing "incompatible" changes is just a matter of being able to coordinate the engineering necessary to evaporate the liquid in a small body of water, perhaps only a puddle or pot on a stove. In contrast, for APIs that are used as widely as the JDK, rigorously finding all the possible programs impacted by an incompatible change is as impractical as boiling the oceans, so evolving such APIs is quite constrained by comparison.

Generally, we will consider whether a program P is compatible is some fashion (or not) with respect to two versions of a library L₁ and L₂ that differ in some way. (We will not consider the compatibility impact of such changes to independent implementers of L.) Sometimes only a particular program is of interest; is the change from L₁ to L₂ compatible with this program? When evaluating how the platform should evolve, a broader consideration of the programs of concern is used. For example, does the change from L₁ to L₂ cause a problem for any program that currently exists? If so, what fraction of existing programs is affected? Finally, the broadest consideration is does the change affect any program that could exist? Often once a platform version is released, the latter two notions are similar because imperfect knowledge about the set of actual programs means it can be more tractable to consider the worst possible outcome for any potential program rather than estimate the impact over actual programs. Stated more formally, depending on the change being considered, judging the change based on the worst possible outcome for any program is more appropriate than judging based on some other kind of norm of the disruption over the space of known programs.

Generally each kind of compatibility has both positive and negative aspects; that is, the positive aspect keeping things that "work" working and the negative aspect of keeping things that "don't work" not working. For example, the TCK tests for Java compilers include both positive tests of programs that must be accepted and negative tests of programs that must be rejected. In many circumstances, preserving or expanding the positive behavior is more acceptable and important than maintaining the negative behavior and we will focus on positive compatibility in this entry.

In terms of relative severity, source compatibility problems are usually the mildest since there are often straightforward workarounds, such as adjusting import statements or switching to fully qualified names. Gradations of source compatibility are identified and discussed below. Behavioral compatibility problems can have a range of impacts while true binary compatibility issues are problematic since linking is prevented.

Source Compatibility

The basic job of any linker or loader is simple: It binds more abstract names to more concrete names, which permits programmers to write code using the more abstract names. (Linkers and Loaders)

A Java compiler's job also includes mapping more abstract names to more concrete ones, specifically mapping simple and qualified names appearing in source code into binary names in class files. Source compatibility concerns this mapping of source code into class files, not only whether or not such a mapping is possible, but also whether or not the resulting class files are suitable. Source compatibility is influenced by changing the set of types available during compilation, such as adding a new class, as well as changes within existing types themselves, such as adding an overloaded method. There is a large set of possible changes to classes and interfaces examined for their binary compatibility impact. All these changes could also be classified according to their source compatibility repercussions, but only a few of kinds of changes will be analyzed below.

The most rudimentary kind of positive source compatibility is whether code that compiles against L₁ will continue to compile against L₂; however, that is not the entirety of the space of concerns since the class file resulting from compilation might not be equivalent. Java source code often uses simple names for types; using information about imports, the compiler will interpret these simple names and transform them into binary names for use in the resulting class file(s). In a class file, the binary name of an entity (along with its signature in the case of methods and constructors) serves as the unique, universal identifier to allow the entity to be referenced. So different degrees of source compatibility can be identified:

• Does the code still compile (or not compile)?

• If the code still compiles, do all the names resolve to the same binary names in the class file?

• If the code still compiles and the names do not all resolve to the same binary names, does a behaviorally equivalent class file result?

Whether or not a program is valid can also be affected by language changes. Usually previously invalid program are made valid, as when generics were added, but sometimes existing programs are rendered invalid, as when keywords were added (strictfp, assert, and enum). The version number of the resulting class file is also an external compatibility issue of sorts since that affects which platform versions the code can be run on.

Full source compatibility with any existing program is usually not achievable because of * imports. For example, consider L₁ with packages foo and bar where foo includes the class Quux. Then L₂ adds class bar.Quux. This program

import foo.*;
import bar.*;

public class HelloQuux {
    public static void main(String... args) {
	Object o = Quux.class;
	System.out.println("Hello " + o.toString());
    }
}

will compile under L₁ but not under L₂ since the name "Quux" is now ambiguous as reported by javac:

HelloQuux.java:6: reference to Quux is ambiguous, both class bar.Quux in bar and
 class foo.Quux in foo match
        Object o = Quux.class;
                   ^
1 error

An adversarial program could almost always include * imports that conflict with a given library.¹ Therefore, judging source compatibility by requiring all possible programs to compile is an overly restrictive criterion. However, when naming their types, API designers should not reuse "String", "Object", and other names of core classes from packages like java.lang and java.util to avoid this kind of annoying name conflict.

Due to the * import wrinkle, a more reasonable definition of source compatibility considers programs transformed to only use fully qualified names. Let FQN(P, L) be program P where each name is replaced by its fully qualified form in the context of libraries L. Call such a library transformation from L₁ to L₂ binary-preserving source compatible with source program P if FQN(P, L₁) equals FQN(P, L₂). This is a strict form of source compatibility that will usually result in class files for P using the same binary names when compiled against both versions of the library. Class files with the same binary names will result when each type has a distinct fully qualified name. Multiple types can have the same fully qualified name but differing binary names; those cases do not arise when the standard naming conventions are being followed.²

Adding overloaded methods has the potential to change method resolution and thus change the signatures of the method call sites in the resulting class file. Whether or not such a change is problematic with respect to source compatibility depends on what semantics are required and how the different overloaded methods operate on the same inputs, which interacts with behavioral equivalence notions. Assume class C originally has a method void m(T t) and then an overload void m(S s) is added. Some cases of interest include:

• S and T are both reference types:

  • If there is no typing relationship between S and T, overload resolution will not be affected.

  • If there is a typing relationship between S and T, such as T is a subtype of S, call sites in existing source may now resolve to the new method. Well-written programs will follow the Liskov substitution principle and C will do "the same" operation on the argument no matter which overloaded method is called. Less than well-written programs may fail to follow this principle.

• S and T are both primitive types: By extension, if a numerical value can be represented in multiple primitive types, overloaded methods taking a type with that value should usually perform an equivalent operation. However, the silent loss of precision in primitive widening conversion can affect the actual value that gets passed to an overloaded method.

  Concretely, consider class C with methods m(int) and m(double). The call site "m(123L)" will undergo primitive widening conversion, converting the argument value to double before m(double) is called. Now if m(long) is added to C, the call site will resolve to the new method. Even assuming each m method does an equivalent operation when passed a numerically equal value, there can still be differences after the third method is added since some long values lose precision when converted to double, for example, Long.MAX_VALUE. Therefore, a client when compiled against the two version of C can have different runtime behavior even if each m method behaves reasonably. This kind of subtle change in overloading behavior occurred with the addition of a BigDecimal constructor taking as long as part of JSR 13

• One of S and T is a reference type, the other is primitive: Before generics were added to the language, two methods which differed in the primitive/reference status of the i^th parameter could not possibly be applicable to the same arguments. But, along with generics came boxing and unboxing conversions that can map, for example, a value of an int primitive type to a java.lang.Integer object with a reference type, and vice versa. These mapping have the potential to introduce ambiguities in method resolution such that adding a method could introduce an ambiguity that prevented previously valid code from compiling; however, the rules for method invocation expressions were updated to avoid such potential ambiguities from boxing/unboxing as well as var-args.

If a new method cannot change resolution, then it is a binary-preserving source transformation. If a new method can change resolution, if the different class file that results has acceptably similar behavior, the change may still be acceptable, while changing resolution in such a way that does not preserve semantics is likely problematic. Changing a library in such a way that current clients no longer compile is seldom appropriate.

Binary Compatibility

JLSv3 §13.2 What Binary Compatibility Is and Is Not
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.

The JLS defines binary compatibility strictly according to linkage; it P links with L₁ and continues to link with L₂, the change made in L₂ is binary compatible. The runtime behavior after linking is not included in binary compatibility:

JLSv3 13.4.22 Method and Constructor Body
Changes to the body of a method or constructor do not break [binary] compatibility with pre-existing binaries.

As an extreme example, if the body of a method is changed to throw an error instead of compute a useful result, while the change is certainly a compatibility issue, it is not a binary compatibility issue since client classes would continue to link. Also, it is not a binary compatibility issue to add methods to an interface. Class files compiled against the old version of the interface will still link against the new interface despite the class not having an implementation of the new method. If the new method is called at runtime, an AbstractMethodError is thrown; if the new method is not called, the existing methods can be used without incident. (Adding a method to an interface is a source incompatibility that can break compilation though.)

A design requirement from the addition of generics via JSR 14 was migration compatibility. Migration compatibility requires that a library can be generified and existing (nongeneric) clients can continue to compile and link against the generic version. Meeting this constraint led to the use of erasure, a controversial aspect of the generics design. During JSR 14, it was not known how to add generics in a way that supported both reification and migration compatibility; future work might address this shortcoming.

Behavioral Compatibility

Intuitively, behavioral compatibility should mean that with the same inputs program P does "the same" or an "equivalent" operation under different versions of libraries or the platform. Defining equivalence can be a bit involved; for example, even just defining a proper equals method in a class can be nontrivial. In this case, to formalize this concept would require an operational semantics for the JVM for the aspects of the system a program was interested in. For example, there is a fundamental difference in visible changes between programs that introspect on the system and those that do not. Examples of introspection include calling core reflection, relying on stack trace output, using timing measurements to influence code execution, and so on. For programs that do not use, say, core reflection, changes to the structure of libraries, such as adding new public methods, is entirely transparent. In contrast, a (poorly behaved) program could use reflection to look up the set of public methods on a library class and throw an exception if any unexpected methods were present. A tricky program could even make decisions based on information like a timing side channel. For example, two threads could repeatedly run different operations and make some indication of progress, for example, incrementing an atomic counter, and the relative rates of progress could be compared. If the ratio is over a certain threshold, some unrelated action could be taken, or not. This allows a program to create a dependence on the optimization capabilities of a particular JVM, which is generally outside a reasonable behavioral compatibility contract.

The evolution of a library is constrained by the library's contract included in its specification; for final classes this contract doesn't usually include a prohibition of adding new public methods! While an end-user may not care why a program does not work with a newer version of a library, what contracts are being followed or broken should determine which party has the onus for fixing the problem. That said, there are times in evolving the JDK when differences are found between the specified behavior and the actual behavior (for example 4707389, 6365176). The two basic approaches to fixing these bugs are to change the implementation to match the specified behavior or to change the specification (in a platform release) to match the implementation's (perhaps long-standing) behavior; often the latter option is chosen since it has a lower de facto impact on behavioral compatibility.

Case Study

Consider two versions of a simple enum representing the crew of the USS Enterprise, one for the first season:

public enum StarTrekCast {
    JAMES_T_KIRK("Jim"),
    LEONARD_MCCOY("Bones"),
    JANICE_RAND("Yeoman Rand"),
    MONTGOMERY_SCOTT("Scotty"),
    SPOCK("Spock"),
    HIKARU_SULU("Sulu"),
    UHURA("Uhura"); // Any first name for Uhura is non-canon.

    private String nickname;
    StarTrekCast(String nickname) {
	this.nickname=nickname;
    }

    public String nickname() { return nickname;}
}

and another for the second season:

public enum StarTrekCast {
    JAMES_T_KIRK("Jim"),
    SPOCK("Spock"),
    MONTGOMERY_SCOTT("Scotty"),
    LEONARD_MCCOY("Bones"),
    /* JANICE_RAND("Yeoman Rand"), */ // Only in 8 episodes!
    HIKARU_SULU("Sulu"),
    PAVEL_CHEKOV("Chekov"), // Introduced in season 2.
    UHURA("Uhura"); // Any first name for Uhura is non-canon.

    private String nickname;
    StarTrekCast(String nickname) {
	this.nickname=nickname;
    }

    public String nickname() { return nickname;}
}

Compared to the first reason, the second season:

1. Deletes yeoman Janice Rand

2. Adds Pavel Chekov

3. Reorders Bones, Scotty, and Spock to better reflect the order of who commands the ship if the Captain and others are unavailable.

These changes have varying source, binary, and behavioral compatibility effects:

1. Deleting JANICE_RAND is source incompatible, able to break compilations. The deletion is also binary incompatible. Besides being observable via reflection, the deletion affects the behavior of various built-in methods on the enum, including values and valueOf. In addition, the deletion will break previously serialized streams with this constant.

2. Adding CHEKOV is binary-preserving source compatible. Likewise, the addition of a new public static final field is binary compatible. However, the addition of a new constant is visible to reflection and alters the behavior of built-in enum methods. Existing serialized instance continue to work after a new constant is added.

3. Reordering McCoy, Scotty, and Spock is a binary-preserving source compatible and binary compatible change, but the reordering changes the behavior of built-in methods, most notably compareTo.

JDK Platform and Update Release Compatibility Policies

The compatibility policies we apply to platform releases, like JDK 7, differ from those applied to maintenance and update releases, like JDK 6 updates. For both kinds of releases, binary compatibility must be maintained for JCP-managed APIs. Update releases must maintain source compatibility, but platform releases are able to break source compatibility given sufficient justification. In update releases, behavioral compatibility is regarded as very important; programs may be relying on specified-to-be-unspecified behavior of a particular implementation and switching to another update in the same release family should be seamless whenever possible. In contrast, platform releases have fewer restrictions on changing such behavior. So, for example, modifying the order of iteration of elements in a HashMap to allow faster hashing algorithms, would be quite appropriate for a platform release ("This class makes no guarantees as to the order of the map; in particular, it does not guarantee that the order will remain constant over time."), but would be much less suited to an update release.

Managing Compatibility

Original Preface to JLS
Except for timing dependencies or other non-determinisms and given sufficient time and sufficient memory space, a program written in the Java programming language should compute the same result on all machines and in all implementations.

The above statement from the original JLS could be regarded as vacuously true about any platform: except for the non-determinisms, a program is deterministic. The difference was that in Java, with programmer discipline, the set of deterministic programs was nontrivial and the set of predictable programs was quite large. In other words, the platform provider and the programmer both have responsibilities in making programs portable in practice; the platform should abide by the specification and conversely programs should tolerate any valid implementation of the specification.

To make continued evolution of the platform more tractable, it may be helpful to introduce more structured ways of tracking behavioral changes so that programs could in principle by audited for depending on aspects of the platform in ways that are not recommended. For example, potentially annotations could be used to:

1. Mark classes and methods whose specification has changed in a release (analogous to change bars in a written document).

2. Record stability information about a method's contract, deterministic, non-deterministic, volatile (expected to change over time), etc., for example whether the hashCode of a class is specified to return particular values or just obey the general contract.

3. Using com.sun.* annotations, annotate constructs whose implementations we have changed in our specific implementation in a particular release, such as HashMap ordering.

Annotation processing is a general purpose meta-programming framework, standardized as part of the platform as of JDK 6. Annotation processors, probably also using the tree API, could be written to check for usage of changed or problematic APIs in source code. The D compiler in DTrace can enforce analogous limits on the stability levels and dependency classes of D scripts.

While there would be considerable cost and complication to designing such a scheme and retrofitting it onto at least a subset of the JDK, the ability to define and then programmatically test policies for behavioral compatibility issues could enable platform providers and programmers to have a smoother joint stewardship of keeping applications running and Java usage growing.

Conclusion

Compatibility is a multifaceted concept, with nuances within each broad category. In the future, annotation processors or other program analyzers might help manage source, binary, and behavioral analysis by direct analysis or program markup.

Acknowledgments

Éamonn McManus gave useful feedback on a draft of this entry.

Notes

¹ There are some cases where such an adversarial program could be thwarted in practice. For example, when the Unicode version supported by JDK platform is upgraded previously illegal identifier strings are often allowed. A new JDK platform class could use the newly valid names not open to preexisting malicious clients; although new adversaries could afterward use the new name. This assumes the compatibility threat model only includes class files generated from Java sources. As of class file version 49.0 for JDK 5 and later, at the JVM level many more identifiers are legal than those accepted in Java source.

² Even code that always uses fully qualified names is not completely immune from ambiguities and unintended (or malicious) changes in the meaning of names stemming from changes in the library environment since distinct types can have the same fully qualified name. For example, the type name "a.b.C" could refer to:

• class C in package a.b:
  

  package a.b;
  public class C {}
  

• class C nested inside class b where class b is a member of package a:
  

  package a;
  public class b {
      public static class C{}
  }
  

• class C nested inside class b which is in turn nested inside class a where class a is a member of an unnamed package (unnamed packages are not Immortal):

  public class a {
      public static class b {
  	public static class C{}
      }
  }
  

These three classes cannot all be compiled together ("package a.b clashes with class of same name"); however, they can be compiled separately to the same output location and so can all appear on a classpath when another file is compiled. If all three are on the classpath together, when other code is compiled the qualified name "a.b.C" resolves to the doubly-nested class C in an unnamed package.

To avoid such name collisions, binary names use "$" instead of "." to separate the name of an enclosing class from a nested class, leading to the distinct binary names "a.b.C", "a.b$C", and "a$b$C", respectively, for the classes in question. Following the recommended naming conventions avoids such name clashes. Therefore, such name clashes should be rare in practice when compiling against libraries following the conventions, as JCP moderated java.* and javax.* APIs should do. As an extreme case, do not write this program:

public class java {
    public static class lang {
	public static class String {
	   String(Object o){}
	}
    }

    public static void main(String... args) {
	java.lang.String s = 
	    new java.lang.String("I don't think this means " +
				 "what you think it means.");
	if (!s.getClass().getName().equals("java.lang.String"))
	    System.out.println("Inconceivable!");
    }
}

In this perverse example, the nested class java.lang obscures the venerable java.lang package and the local java.lang.String declaration shadows the standard java.lang.String.

Further Reading

1. Evolving Java-based APIs, by Jim des Rivières.

2. Different kinds of compatibility by Alex Buckley.