Connecting to the EIS
ConnectionFactory is the application's gateway to the EIS. It's instantiated at the time of deployment by the container and a reference to it, or its serialized representation, is bound to the JNDI namespace. Later the application components do a JNDI lookup to get hold of the factory to create connection objects to the EIS. The ConnectionFactory does not represent the actual connection repository, nor is the connection the actual physical connection to the EIS. Both are client-side proxies that nevertheless maintain handles to the actual factory and connection objects, respectively. The container plays the role of an intermediary between the real objects and their proxies, and the mechanism allows the container to inject value-added services by intercepting calls made on the proxies before they're actually delegated to the real objects.

When a ConnectionFactory encounters a getConnection request, it delegates it to the ConnectionManager instance.

public Connection getConnection()
throws ResourceException {

return (Connection)connMgr.allocateConnection
(mngdConnFactory, new ConnectionRequestInfoImpl());
}

In a managed environment the ConnectionManager implementation is provided by the container. The ConnectionManager maintains a pool of connections corresponding to each factory. The ConnectionManager checks whether it can service the request from the pool. If not, it requests the real factory (ManagedConnectionFactory) to create a new connection.

How does the container know about the availability status of a connection?

The Connector Architecture prescribes an elaborate mechanism of listeners to implement callbacks. Whenever a new connection is created, the container registers a listener with it. The connection raises event notifications to intimate the container about the happenings on the connection. When a close is called on the connection proxy it terminates its association with the physical connection. The physical connection in turn generates an event. All the listeners registered on the connection can react to the event. The container on its part uses the event to change the availability status of the connection from in-use to available, i.e., if the connection is not participating in a transaction. If the connection is participating in a transaction, the container waits for the transaction to commit before it can make the connection available (see Listing 1).

Invoking EIS Functions
An InteractionSpec implementation holds properties for driving an interaction with an EIS instance. In our case each property maps to a File Operation that may be performed.

public static final int CREATE = 10;
public static final int UPDATE = 20;
public static final int DELETE = 30;
public static final int MOVE = 40;
public static final int SKIP = 50;
public static final int READ = 60;

Since CCI is EIS independent it cannot use any of the EIS data structures. To get over this, CCI introduces the concept of a record. A record is a generic representation of data that's exchanged with the EIS. More specific implementations to represent hierarchical, tabular data collections can also be implemented. XAFileConnector extends the MappedRecord, which is a key value representation of record elements for both input and output records (see Listing 2).

CCI defines an interaction interface that allows a client to interact with EIS. An interaction represents a single communication with the EIS, an EIS function call. An interaction instance is obtained from a connection. The interaction instance is required to maintain its association with the connection instance. The interaction delegates its execution to the connection, which in turn delegates its execution to the associated ManagedConnection (see Listing 3).

Transaction Contract
A transaction as we know it is a set of operations performed in such a way that the state of the system after the transaction is either the cumulative effect of state changes due to successful individual operations or is the initial state before the commencement of the transaction in case one or more operations happens to fail. Transactions can be classified into two types based on the number of resources participating.
1.  A local transaction performs operations on a single RM. The local transactions can be further classified based on the manner of demarcating transaction boundaries.

2.  Global/distributed transactions - work in this type of transaction can span multiple RMs. The transaction manager demarcates the transaction boundaries. A TM coordinates the activities on the participating RMs using the XAResource interface. An RM knows only about the work it does for its transaction branch. The XAResource interface implements the two-phase commit protocol between the RMs and the TM. The XAResource interface allows for a one-phase optimization in case only one resource is participating.

With this brief introduction to transactions we'll start with how we can incorporate transactional ability into our file system adapter.

For a file system adapter to be able to roll back a transaction, i.e., revert to state prior to the start of a transaction, an adapter must either memorize the original state or it must be able to get back to the original state by performing certain anti-operations that reverse the effect of the operations. To ensure feasibility and consistency at all times, the anti-operations have to be performed in the order that is exactly the reverse order in which the operations were performed. Similarly, committing a transaction should flush all transactional logs to free all memory.

For each file operation on the connection that modifies the state of the file system (nonread only call), XAFileConnector adds a WorkItem to the associated transaction's work list. If there's no active transaction associated with the connection, it autocommits the work and no WorkItem is created. The WorkItem stores enough information about the associated operation to be able to reverse its effect in case of a rollback or to do a cleanup of the backup in case of a commit. A work area is designated to store the information needed for restoration in the event of a rollback (see Listing 4).

Note how MOVE and CREATE file operation do not have any associated cleanup during operation commit while UPDATE and DELETE operations delete the backups.

Depending on the kind of transaction running, javax.resource.cci.LocalTransaction, javax.resource.spi. LocalTransaction, or XAResource keeps track of the work being done in the transaction impending completion.

CCI Local Transaction implementations are obtained directly from the connection, without the container having any role to play. Since the container is not directly involved, it is not intimated of the state of the transaction, so CCI implementations of local transactions differ from SPI implementations in the sense that they need to raise event notifications to apprise the container of transaction life-cycle events. The events help the container in connection pool management. Since there's only a marginal difference in functionality, the same class is used to implement both forms of local transactions with a flag identifying the type of transaction the current instance represents.

Our implementation model exhibits the following relationships shown in Figure 1.

 
Figure 1

Each ManagedConnection can provide an XAResource implementation for distributed transaction management. The beginTransaction call associates the application component's thread of control with a global transaction. The TM generates a globally unique XID and calls start on all open RMs that are linked with the thread. The RMs are enlisted with the transaction. To guarantee scalability of the Connector Architecture, it recommends that a physical connection and hence the associated XAResource can participate in multiple transactions, but at any point only one of these transactions can be active. To ensure this, the start call first checks that the associated connection isn't running a local transaction nor is the XAResource instance associated with an active transaction. Once assured, it associates the passed XID with the XAResource instance and activates the transaction. All work done henceforth on this physical connection, until the transaction is suspended or completed, accumulates against the active transaction.

The start call is accompanied by the following flags:

The application component calls commit to make the effects of the transaction permanent. The TM first calls end for each involved RM, from the AP's thread of control, to dissociate the thread from the global transaction. The TM then executes the two-phase commit protocol.

STAGE 1 is the prepare stage or the voting stage. TM calls prepare for each RM that was associated with the global transaction. RMs express their readiness to commit the transaction. A single negative vote ensures rollback. A prepare attempt on a suspended transaction would throw back a protocol exception. Similarly it throws back an exception, albeit a different one, when the transaction has been marked earlier for rollback due to internal errors. Prepare stage also offers a two-phase optimization. A transaction that has not changed the state of the RM doesn't need to go through the second phase if prepare returns XA_RDONLY (see Listing 6).

Stage 2 is the commit stage. If all RMs return success from prepare, the TM records a decision to commit the transaction and calls commit for each RM. The XA specification allows for an optimization in this procedure. If the TM has dealt with only one subordinate RM in the global transaction, it can omit Phase 1 by directly calling commit with the onePhase flag set to true (see Listing 7).

Deployment Process
Before we can start using the adapter, we need to deploy it. All adapter interfaces, classes, native libraries, etc., along with the Deployment Descriptor (DD), are packaged in a .rar archive. In situations where multiple J2EE applications need to share the XAFileConnector, it can be deployed directly into an application server as a standalone unit. However, the resource adapter module may also be bundled with a J2EE application, if it is needed only by a single application.

The Deployment Descriptor provides a lot of information to the container, including:

To be able to successfully deploy an adapter, a better understanding of the deployment process is essential. We will briefly walk through the responsibilities of the deployment code and how it interacts with the adapter code.

public Object createConnectionFactory(ConnectionManager cxManager)
throws ResourceException
{
return new ConnectionFactoryImpl(this, cxManager);
}

Note how the ConnectionManager instance is passed around while creating the ConnectionFactory. The mechanism helps associate the two factories and the ConnectionManager. The ConnectionManager is the ConnectionFactory's interface to the container.

Writing a Client
Clients needing to interact with the file system will have to use the CCI interface. They cannot use the java.io package to access the file system if they want the transactional support. While the approach may look nonintuitive and tiresome, which in fact it is, it is not as big a pain as you would imagine it to be.

By this time we would have understood how to acquire a connection to the file system.

Interaction interaction = conn.createInteraction();

Interaction maintains an association with the connection from which it was created and executes on the same connection.

RecordFactory rf = cf.getRecordFactory();
MappedRecord in = rf.createMappedRecord("IN");
MappedRecord out = rf.createMappedRecord("OUT");

in.put("SOURCE_DIR", "E:\\bea\\weblogic700\\FILES1");
in.put("SOURCE_FILENAME", "create.txt");

interaction.execute(new
InteractionSpecImpl(InteractionSpecImpl.CREATE), in, out);

What Further?
The proposed draft connector specification makes provisions for inbound communication through a Message Inflow Contract and also includes a Work Management Contract. The inclusion of the two would make it feasible for the adapter to work as a poller.

The Work Management Contract would allow the adapter to monitor folders for incoming files. The activity can be performed by submitting work to the container. The contract would save the adapter the task of creating its own threads thus allowing the container to exercise better control over its runtime environment.

The Message Inflow Contract would allow the adapter to trigger action in the event of receiving a file. The situation can be thought of as analogous to a JMS situation, with the adapter instead of an MDB delivering messages by polling on a folder instead of a queue.

Better concurrency control can also be built into the adapter using the new improved I/O support found in newer Java runtimes, 1.4 onwards.