Correct idiom for managing multiple chained resources in try-with-resources block?

The Java 7 try-with-resources syntax (also known as ARM block (Automatic Resource Management)) is nice, short and straightforward when using only one AutoCloseable resource. However, I am not sure what is the correct idiom when I need to declare multiple resources that are dependent on each other, for example a FileWriter and a BufferedWriter that wraps it. Of course, this question concerns any case when some AutoCloseable resources are wrapped, not only these two specific classes.

I came up with the three following alternatives:

1)

The naive idiom I have seen is to declare only the top-level wrapper in the ARM-managed variable:

static void printToFile1(String text, File file) {
    try (BufferedWriter bw = new BufferedWriter(new FileWriter(file))) {
        bw.write(text);
    } catch (IOException ex) {
        // handle ex
    }
}

This is nice and short, but it is broken. Because the underlying FileWriter is not declared in a variable, it will never be closed directly in the generated finally block. It will be closed only through the close method of the wrapping BufferedWriter. The problem is, that if an exception is thrown from the bw's constructor, its close will not be called and therefore the underlying FileWriter will not be closed.

2)

static void printToFile2(String text, File file) {
    try (FileWriter fw = new FileWriter(file);
            BufferedWriter bw = new BufferedWriter(fw)) {
        bw.write(text);
    } catch (IOException ex) {
        // handle ex
    }
}

Here, both the underlying and the wrapping resource are declared in the ARM-managed variables, so both of them will certainly be closed, but the underlying fw.close() will be called twice: not only directly, but also through the wrapping bw.close().

This should not be a problem for these two specific classes that both implement Closeable (which is a subtype of AutoCloseable), whose contract states that multiple calls to close are permitted:

Closes this stream and releases any system resources associated with it. If the stream is already closed then invoking this method has no effect.

However, in a general case, I can have resources that implement only AutoCloseable (and not Closeable), which doesn't guarantee that close can be called multiple times:

Note that unlike the close method of java.io.Closeable, this close method is not required to be idempotent. In other words, calling this close method more than once may have some visible side effect, unlike Closeable.close which is required to have no effect if called more than once. However, implementers of this interface are strongly encouraged to make their close methods idempotent.

3)

static void printToFile3(String text, File file) {
    try (FileWriter fw = new FileWriter(file)) {
        BufferedWriter bw = new BufferedWriter(fw);
        bw.write(text);
    } catch (IOException ex) {
        // handle ex
    }
}

This version should be theoretically correct, because only the fw represents a real resource that needs to be cleaned up. The bw doesn't itself hold any resource, it only delegates to the fw, so it should be sufficient to only close the underlying fw.

On the other hand, the syntax is a bit irregular and also, Eclipse issues a warning, which I believe is a false alarm, but it is still a warning that one has to deal with:

Resource leak: 'bw' is never closed

So, which approach to go for? Or have I missed some other idiom that is the correct one?

Solution 1:

Here's my take on the alternatives:

1)

try (BufferedWriter bw = new BufferedWriter(new FileWriter(file))) {
    bw.write(text);
}

For me, the best thing coming to Java from traditional C++ 15 years ago was that you could trust your program. Even if things are in the muck and going wrong, which they often do, I want the rest of the code to be on best behaviour and smelling of roses. Indeed, the BufferedWriter might throw an exception here. Running out of memory wouldn't be unusual, for instance. For other decorators, do you know which of the java.io wrapper classes throw a checked exception from their constructors? I don't. Doesn't do code understandability much good if you rely upon that sort of obscure knowledge.

Also there's the "destruction". If there is an error condition, then you probably don't want to be flushing rubbish to a file that needs deleting (code for that not shown). Although, of course, deleting the file is also another interesting operation to do as error handling.

Generally you want finally blocks to be as short and reliable as possible. Adding flushes does not help this goal. For many releases some of the buffering classes in the JDK had a bug where an exception from flush within close caused close on the decorated object not be called. Whilst that has been fixed for some time, expect it from other implementations.

2)

try (
    FileWriter fw = new FileWriter(file);
    BufferedWriter bw = new BufferedWriter(fw)
) {
    bw.write(text);
}

We're still flushing in the implicit finally block (now with repeated close - this gets worse as you add more decorators), but the construction is safe and we have to implicit finally blocks so even a failed flush doesn't prevent resource release.

3)

try (FileWriter fw = new FileWriter(file)) {
    BufferedWriter bw = new BufferedWriter(fw);
    bw.write(text);
}

There's a bug here. Should be:

try (FileWriter fw = new FileWriter(file)) {
    BufferedWriter bw = new BufferedWriter(fw);
    bw.write(text);
    bw.flush();
}

Some poorly implemented decorators are in fact resource and will need to be closed reliably. Also some streams may need to be closed in a particular way (perhaps they are doing compression and need to write bits to finish off, and can't just flush everything.

Verdict

Although 3 is a technically superior solution, software development reasons make 2 the better choice. However, try-with-resource is still an inadequate fix and you should stick with the Execute Around idiom, which should have a clearer syntax with closures in Java SE 8.

Solution 2:

The first style is the one suggested by Oracle. BufferedWriter doesn't throw checked exceptions, so if any exception is thrown, the program is not expected to recover from it, making resource recover mostly moot.

Mostly because it could happen in a thread, with the thread dieing but the program still continuing -- say, there was a temporary memory outage that wasn't long enough to seriously impair the rest of the program. It's a rather corner case, though, and if it happens often enough to make resource leak a problem, the try-with-resources is the least of your problems.

Solution 3:

Option 4

Change your resources to be Closeable, not AutoClosable if you can. The fact that the constructors can be chained implies it isn't unheard of to close the resource twice. (This was true before ARM too.) More on this below.

Option 5

Don't use ARM and code very carefully to ensure close() isn't called twice!

Option 6

Don't use ARM and have your finally close() calls in a try/catch themselves.

Why I don't think this problem is unique to ARM

In all these examples, the finally close() calls should be in a catch block. Left out for readability.

No good because fw can be closed twice. (which is fine for FileWriter but not in your hypothetial example):

FileWriter fw = null;
BufferedWriter bw = null;
try {
  fw = new FileWriter(file);
  bw = new BufferedWriter(fw);
  bw.write(text);
} finally {
  if ( fw != null ) fw.close();
  if ( bw != null ) bw.close();
}

No good because fw not closed if exception on constructing a BufferedWriter. (again, can't happen, but in your hypothetical example):

FileWriter fw = null;
BufferedWriter bw = null;
try {
  fw = new FileWriter(file);
  bw = new BufferedWriter(fw);
  bw.write(text);
} finally {
  if ( bw != null ) bw.close();
}

Solution 4:

I just wanted to build on Jeanne Boyarsky's suggestion of not using ARM but making sure the FileWriter is always closed exactly once. Don't think there are any problems here...

FileWriter fw = null;
BufferedWriter bw = null;
try {
    fw = new FileWriter(file);
    bw = new BufferedWriter(fw);
    bw.write(text);
} finally {
    if (bw != null) bw.close();
    else if (fw != null) fw.close();
}

I guess since ARM is just syntactic sugar, we can't always use it to replace finally blocks. Just like we can't always use a for-each loop to do something that is possible with iterators.

Solution 5:

To concur with earlier comments: simplest is (2) to use Closeable resources and declare them in order in the try-with-resources clause. If you only have AutoCloseable, you can wrap them in another (nested) class that just checks that close is only called once (Facade Pattern), e.g. by having private bool isClosed;. In practice even Oracle just (1) chains the constructors and doesn't correctly handle exceptions partway through the chain.

Alternatively, you can manually create a chained resource, using a static factory method; this encapsulates the chain, and handle cleanup if it fails part-way:

static BufferedWriter createBufferedWriterFromFile(File file)
  throws IOException {
  // If constructor throws an exception, no resource acquired, so no release required.
  FileWriter fileWriter = new FileWriter(file);
  try {
    return new BufferedWriter(fileWriter);  
  } catch (IOException newBufferedWriterException) {
    try {
      fileWriter.close();
    } catch (IOException closeException) {
      // Exceptions in cleanup code are secondary to exceptions in primary code (body of try),
      // as in try-with-resources.
      newBufferedWriterException.addSuppressed(closeException);
    }
    throw newBufferedWriterException;
  }
}

You can then use it as a single resource in a try-with-resources clause:

try (BufferedWriter writer = createBufferedWriterFromFile(file)) {
  // Work with writer.
}

The complexity comes from handling multiple exceptions; otherwise it's just "close resources that you've acquired so far". A common practice seems to be to first initialize the variable that holds the object that holds the resource to null (here fileWriter), and then include a null check in the cleanup, but that seems unnecessary: if the constructor fails, there's nothing to clean up, so we can just let that exception propagate, which simplifies the code a little.

You could probably do this generically:

static <T extends AutoCloseable, U extends AutoCloseable, V>
    T createChainedResource(V v) throws Exception {
  // If constructor throws an exception, no resource acquired, so no release required.
  U u = new U(v);
  try {
    return new T(u);  
  } catch (Exception newTException) {
    try {
      u.close();
    } catch (Exception closeException) {
      // Exceptions in cleanup code are secondary to exceptions in primary code (body of try),
      // as in try-with-resources.
      newTException.addSuppressed(closeException);
    }
    throw newTException;
  }
}

Similarly, you can chain three resources, etc.

As a mathematical aside, you could even chain three times by chaining two resources at a time, and it would be associative, meaning you would get the same object on success (because the constructors are associative), and same exceptions if there were a failure in any of the constructors. Assuming you added an S to the above chain (so you start with a V and end with an S, by applying U, T, and S in turn), you get the same either if you first chain S and T, then U, corresponding to (ST)U, or if you first chained T and U, then S, corresponding to S(TU). However, it would be clearer to just write out an explicit three-fold chain in a single factory function.