A Middle Ground Between Checked And Unchecked Exceptions

Monday 1 February 2016 at 08:00 GMT

I hear a lot of arguments against Java's checked exceptions. Normally, they run along these lines:

You have to propagate the exceptions everywhere. Too much noise, not enough signal.

And they're right. It can be very noisy. Personally, I think the noise is worth it so that exceptions are adequately handled, but it's a bit ridiculous.

If you want to write your code so that exceptions are only surfaced at run-time, go for it. I won't stop you. However, I take issue with this when it comes to interfaces. If your interfaces aren't able to throw exceptions, even in environments where that would be reasonable, it can be really unpleasant for your implementors.

Let's take java.lang.Runnable as an example. Here's the full source code, without the comments:

package java.lang;

@FunctionalInterface
public interface Runnable {
    public abstract void run();
}

My issue there is that run doesn't allow for any checked exceptions. However, we use this interface a lot with multithreaded code, for example with an ExecutorService, and exceptions, especially IOException, are common in these scenarios. This means we end up writing a lot of code like this:

executor.submit(() -> {
    try {
        doTheThing();
    } catch (IOException e) {
        throw new RuntimeException(e);
    }
});

The Sun/Oracle folks fixed this somewhat by allowing us to use Callable<V> instead. The call method does allow for exceptions, and so makes our lives easier.

package java.util.concurrent;

@FunctionalInterface
public interface Callable<V> {
    V call() throws Exception;
}

As the ExecutorService returns a Future<V> which will throw an ExecutionException on get if anything did go wrong, this is moot anyway. You have to handle that one, and that is checked.

When designing our own interfaces, we have a third option. Rather than throwing a checked or an unchecked exception, we can let the implementor decide.

Take a look at my Serializer interface from my Rekord library.

public interface Serializer<R, E extends Exception> {
    <T> R serialize(String name, FixedRekord<T> rekord) throws E;
}

This takes a FixedRekord and serializes it to an R. R can be an XML document, a JSON document, or maybe just a java.util.Map or a string. Now, serializing to JSON can cause exceptions, but converting a Rekord into a Map definitely won't throw an exception. If I had declared serialize as throwing Exception, you'd have to handle an exception in both cases, even when I could guarantee that one wouldn't be thrown in the latter case. Conversely, if I'd declared it without an exception, the JSON serialiser would have to wrap its exceptions in RuntimeException, which would mean the caller would be unaware of potential failure. Neither scenario is great.

Here's the JSON serializer, which uses Jackson:

public final class JacksonSerializer implements Serializer<Void, IOException> {
    private final Writer writer;

    private JacksonSerializer(Writer writer) {
        this.writer = writer;
    }

    @Override
    public <T> Void serialize(String name, FixedRekord<T> rekord) throws IOException {
        ...
    }

    ...
}

JacksonSerializer will throw an IOException on failure, and the caller will have to make sure they can handle this, either by propagating it or doing something about it.

Now let's look at my MapSerializer:

public final class MapSerializer implements SafeSerializer<Map<String, Object>> {
    @Override
    public <T> Map<String, Object> serialize(String name, FixedRekord<T> rekord) {
        ...
    }

    ...
}

No exception. But that's a different interface, right? Of course the SafeSerializer wouldn't throw an exception. Well, here it is:

public interface SafeSerializer<R> extends Serializer<R, ImpossibleException> {
    @Override
    <T> S serialize(String name, FixedRekord<T> rekord);
}

public final class ImpossibleException extends RuntimeException {
    private ImpossibleException() { }
}

The SafeSerializer is a Serializer, but parameterised with an exception that's impossible to construct, and so can never be thrown. That exception is a form of RuntimeException, and so isn't checked, and doesn't need to be handled by the caller. This means that when you invoke the serialize method on the MapSerializer, the compiler knows you don't need to handle the exception and won't force you to.

So there we have it. By parameterising the exception type, we can define an interface that is flexible enough to declare exceptional behaviour when it's present, but not force you to handle it when it's absent. The best of both worlds.