Checked exception

From APIDesign

Revision as of 11:56, 4 April 2016

Checked exceptions are Java invention and many like to argue that they are the worst invention ever. I like exceptions and I like Checked exceptions. One day I'll explain why.

Don't Forget to Catch Me

There is a really nice thing on checked exceptions: if a method declares that it throws a checked exception, the caller of the method has to handle it. This is a really nice language feature, if used at the appropriate place. What is such appropriate place? If one reads a file one shall be ready for an input/output error - e.g. forcing people to catch IOException seems like the right thing to do.

Thus in certain situations having checked exceptions is beneficial. On the other hand, throwing checked exceptions in cases where the recovery is unlikely - a frequently mentioned example is ParserConfigurationException - is just going to pollute the client code with useless catch statements.

Can't Decide Upfront

There is however one more example: Imagine an exception that needs to be caught when thrown from certain methods, but when it is thrown from other methods, it should behave as as RuntimeException - e.g. propagate silently. We have seen an example of this recently in our Truffle project. The InteropException should smoothly propagate through many calls, but if invoked via the ForeignAccess's send method, we want every caller to handle it. What are our API design options?

Duplicate the Exceptions

We could have one RuntimeException subclass and one Exception subclass. Sometimes that might work, but in this case there are four subclasses of InteropException currently (and the number is expected to grow in the future), and having a duplicated set of classes is clearly annoying. Such solution would only support the argument of some that checked exceptions are the worst invention ever!

Wrap the Exceptions

We could introduce one RuntimeException subclass and let it carry the real checked InteropException exception. That would probably work on the catch-side:

try {
   // some interop code
} catch (InteropRuntimeException ex) {
  throw (InteropException)ex.getCause();
}

and even on the throw side the code wouldn't be that bad, imagine throwing UnknownIdentifierException (a subclass of InteropException) exception:

throw new InteropRuntimeException(new UnknownIdentifier("name"));
// which could be simplified and hidden into a factory method:
throw UnknownIdentifierException.raise("name");

However this still suffers from the duality of exceptions. In situations where needs to be sure, one needs to catch both exceptions InteropRuntimeException as well as InteropException, which is again a reason for few to claim that checked exceptions are bad.

Unchecking Checked Exception

Luckily, if one knows the difference between source compatibility and binary compatibility one can realize that the JVM doesn't know anything about the difference between RuntimeException and checked Exception - it is all just a Java language construct. Other languages built on top of JVM may ignore it. And that is what we decided to do.

We designed a checked exception InteropException (and its subclasses) and added a raise method to throw the exception as unchecked one. The usage is simple:

throw UnknownIdentifierException.raise("name");

E.g. the usage on the throwing side is exactly the same as in case of wrapping of the exceptions. There is just one difference - there is no wrapping. The checked exception UnknownIdentifierException is really being thrown and one can use

try {
  // the interop code
} catch (UnknownIdentifierException ex) {
  System.out.println("Error " + ex.getUnknownIdentifier());
}

The Lamda Problem

JDK8 is facing similar problem as well. The code using closures/lambdas has been greatly simplified - unless it is dealing with exceptions. If you have an instance of Executor and you want to perform a file read in its execute method - you'll be forced deal with the exception somehow. The usual thing won't compile:

static void execRead(Executor secureSyncExecutor, Path path) throws IOException {
    secureSyncExecutor.execute(() -> Files.readAllLines(path));
}

The exception raised by readAllLines is very hard to be propagated outside of the lamda. If the checked exception was converted to a hidden one by reraise method as shown in following example:

static void execRead(Executor secureExecutor, Path path) throws IOException {
    secureExecutor.execute(reraise(() -> Files.readAllLines(path)));
}
 
static Runnable reraise(Callable<?> runWithException) {
    return () -> {
        try {
            runWithException.call();
        } catch (Exception ex) {
            throw silenceException(RuntimeException.class, ex);
        }
    };
}
 
@SuppressWarnings({"unchecked", "unused"})
static <E extends Exception> RuntimeException silenceException(Class<E> type, Exception ex) throws E {
    throw (E) ex;
}

then the usage of checked exceptions would be much easier in modern Java code. All that is needed is to add the raise method into Exception, so everyone can use it and have the reraise method somewhere or (better) have some support in Javac to automatically generate the conversion if some annotation is used in the code.

Drawbacks

Of course, throwing checked exceptions in such a hidden way may yield some surprises. However, at least from my perspective, the benefits seem to outweigh them.

Throw before Return

One problem is the return type of the raise method. The typical recommended usage is

void silentCheckedThrow() {
  throw UnknownIdentifierException.raise("name");
}

and that means the raise method must return a RuntimeException or Error. But there is no such object - there is just an instance of checked exception. That means the method raise mustn't return - it must throw the exception before returning. Thus one could also write:

void silentCheckedThrow() {
  UnknownIdentifierException.raise("name");
}

and the effect would be the same. Yes, it would be unless the method returns some value. The following only compiles with throw and as such the throw style is the recommended one:

int silentCheckedThrowWithReturn() {
  throw UnknownIdentifierException.raise("name");
}

but, of course, one can still add a never-to-be-reached return statement just as one would after calling System.exit() method:

int silentCheckedThrowWithReturn() {
  UnknownIdentifierException.raise("name");
  return -1;
}

Harder to Catch

Other problems may appear when catching the exception. Per Java specification it is not allowed to write

void uselessCatch() {
  try {
    Integer.valueOf("10");
  } catch (IOException ex) {
    // never generated
  }
}

as the compiler knows that the IOException is not generated in the body of the Integer parse method and refuses to compile such code. This isn't a problem in Truffle API, as we want people to only catch the InteropException when one of ForeignAccess sendXYZ methods is called - and all of them declare they throw the InteropException. Still, there is a way to fool the compiler when necessary:

void uselessCatch() {
  try {
    Integer.valueOf("10");
    fakeThrow();
  } catch (IOException ex) {
    // never generated
  }
}
void fakeThrow() throws IOException {
}

and with this little change everything compiles without problems. Moreover this situation shouldn't be that common - we expect methods that assume handling of the checked exception to really declare it.

Broken Catch

One more problem we have noticed so far is related to wrong assumption in a generic catch. Following code may throw ClassCastException:

try {
  // do something
} catch (Throwable t) {
  if (t instanceof Error) {
    throw (Error)t;
  } else {
    throw (RuntimeException)t;
  }
}

However this program has always been broken from a JVM perspective. The caught Throwable can really be anything - not only subclass of RuntimeException or Error - just normal Java programs don't expose the error. A fix is to rewrite the code to

try {
  // do something
} catch (Error e) {
  throw e;
} catch (RuntimeException re) {
  throw re;
}

Summary

We have demonstrated that checked exceptions are beneficial. Not only when dealing with input and output - e.g. IOException, but also in many other cases. We have shown that the biggest problem of checked exceptions - e.g. that one needs to decide upfront whether an exception is or isn't checked at declaration time isn't problem at all. The JVM doesn't make difference between normal and checked exceptions and thus one can easily, and without significant drawbacks use checked exceptions as unchecked ones with a simple raise method. We demonstrated that similar problems are now well recognized the the JDK itself and demonstrated few simple JDK API extensions that would provide standard solution for the checked vs. unchecked problem for once and ever.

TBD