Alan Gutierrez

Strata is a B+Tree that I’ve written in Java. It implements Java serialization. Not to store the leaves of the tree, but to store the definition of the tree itself.

BentoStorage.Creator newStorage = new BentoStorage.Creator();

newStorage.setSize(8 + 8);
newStorage.setReader(new MyRecordReader());
newStorage.setWriter(new MyRecordWriter());

Strata.Creator newIndex = new Strata.Creator();

newIndex.setStorage(newStorage.create());
newIndex.setExtractor(new MyFieldExtractor());

Strata strata = newStrata.create();

Bento.Creator newBento = new Bento.Creator();

FileOutputStream outFile = new FileOutputStream(new File("index"));
ObjectOutputStream out = new ObjectOutputStream(outFile);
out.writeObject(strata);
out.close();

Here is how MyReader might be defined.

public final class MyReader
implements BentoStorage.Reader, Serializable
{
    private final static int serialVersionID = 20070602L;

    public Object read(ByteBuffer bytes)
    {
        return new MyRecord(bytes.getLong(), bytes.getLong());
    }
}

What if I felt like using this from Groovy and taking advantage of closures?

BentoStorage.Creator newStorage = new BentoStorage.Creator();

newStorage.size = 8 + 8
newStorage.writer = { object, bytes ->
    bytes.putLong(object.key)
    bytes.putLong(object.version)
}
newStorage.reader = { bytes ->
    return new MyRecord(bytes.getLong(), bytes.getLong())
}

Strata.Creator newIndex = new Strata.Creator()

newIndex.storage = newStorage.create()
newIndex.listExtractor = { txn, object ->
    Person person = txn.heap.get(txn.unmarshaller, object.key)
    return [ person.lastName, person.firstName ]
}

Strata index = newIndex.create()

The problem is that Groovy closures cannot be serialized. Now I cannot serialize Strata. This struck me as a setback for Groovy support, and a problem specific to Groovy, until I stubbed my toe on annonymous inner classes.

Strata.Creator newIndex = new Strata.Creator();

newIndex.setStorage(newStorage.create());
newIndex.setExtractor(new FieldExtractor()
{
    public Comparable[] getFields(Object txn, Object object)
    {
        MyRecord record = (MyRecord) object;
        MyDatabase database = (MyDatabase) txn;
        Person person = database.getHeap()
            .get(database.getUnmarshaller(), record.getKey());
        return new Comparable[] {  person.getLastName(),
                                   person.getFirstName() };
    }
});

Strata index = newIndex.create()

Even if I make the anonymous inner class Serializable, perhaps by deriving from an interface that mixes FieldExtractor and Serializable the class that defines the method that builds the Strata must also be serializable. I’m serializing the hidden reference to the containing class. Unnecessary code. Pointless code. Not really part of the data structure.

This must be a conundrum faced any library that could be built using anonymous inner classes.

I wanted my creational pattern to follow the notion of creatation of the object through construction using the Creator classes once. Thereafter, the Strata is serialized and deserialized.

I do not have to get rid of this pattern, but in order to implement it, people will have to avoid the use of anonymous inner classes to define their readers, writers, and extactors. The should always create static classes that implement Serializable.

If they do want to use anonyomous inner classes for readers, writers and extractors, they can simply repeat the creation of the object through construction using the Creator classes. Same goes for the use of Groovy closures.

This muddles an assumption that I had higher in my application stack. I am working on an object database that I call Depot.

The object database I’m building based on Strata stores the Strata B+Tree definitions in a heap. I created a bridge interface for the FieldExtractor, where the object database fishes the object out of the heap, and so only has one parameter, the object found in the heap, and does not pass in the application specific transaction context object.

public class Strata
{
    public interface FieldExtractor
    {
        Comparable[] getFields(Object txn, Object object);
    }
}

public class Depot
{
    public interface FieldExtractor
    {
        Comparable[] getFields(Object object);
    }
}

I’d written example code to imagine how the API would work. With Strata I’d always made Strata.FieldExtractor static, because it was always so complicated, turning a key into an object. With Depot that was taken care of, so the application only had to crack the object and return the array of @Comparable@S upon which an alterate index would sort a set of objects.

Furthermore, I’d decided that for a desktop application, it would be nice to keep the definition, trees and heap in one file. Not thinking, I’d made so that the initial version of Depot demands that the Depot.FieldExtractor be Serializable. My code examples of how clever I am, to define an index in a few short lines of code, did not work out.

It is not so much a conundrum. I merely need to make this one file format optional. The definition could be deserialized from the one file, or it could simply be rebuilt. It’s an application developers choice. If the application developer choses to make her @Depot.FieldExtractor@S static and serializable, then I can provide a static function that will create a file store for use as a heap, that tucks that definition into a longish header field.

I’m sure you could serialize a compound Swing view after you’ve built it, but you could also build it each time and only serialize the user preference settings, like the position of a window splitter.

If anyone out thinks that I misunderstand the trade off, please let me know. It caught me off gaurd. I was going to trouble the Groovy listserv with some questions, but this understanding came to me while I was getting around to that.