What's New in Enterprise Objects Framework

This chapter describes changes made to the Enterprise Objects Framework (EOF) between release 3.0 and 4.5. It describes changes made to existing features and describes new features you may want to start using in your applications.

Schema Synchronization

In EOF 4.5, EOModeler supports synchronization of a database schema with the current state of a model.

To initiate the process of synchronizing a model and schema, select "Synchronize Schema" from the Model menu. Note that before synchronizing, you need to save your model. Because the operation cannot be undone, the "Synchronize Schema" menu item is only enabled when the model has no unsaved changes.

After starting the synchroniztion process, EOModeler reverse engineers the database's schema and assembles the adaptor operations necessary to synchronize it with the model. These operations are presented to the user for confirmation before execution.

For example, if you add simple attributes (attributes representing database columns) to an entity, schema synchronization adds the columns to the underlying table. Similarly new columns or tables created in the database can be selected for incorporation into the model. External type changes for attributes and the renaming of columns and tables are also supported.

Related API Changes

EOModeler makes use of the schema synchronization API to synchronize your database with your model. You don't need to use the API yourself unless you're implementing the API for a custom adaptor.

If you do need to use or implement this API (very unlikely), see the related documentation in the following class and interface/protocol specifications:

• EOAdaptor
• EOAdaptorChannel
• EOAdaptorChannel Delegate
• EOSQLExpression

Event Logging

WebObjects 4.5 introduces event logging. The goal for this feature is to allow the measurement of how long certain operations in EOF and WebObjects take. Measurements allow you to profile an application and optimize its execution time. For this, the EOF and WebObjects frameworks instrument key portions of their code to measure the elapsed time of functions and methods.

To support this feature, EOF adds two new classes: EOEvent and EOEventCenter. An EOEvent keeps information (such as duration) about a logged event, and EOEventCenter manages the events. EOEvent is an abstract class whose subclasses are responsible for defining the events they track. For example, there are (private) subclasses for Sybase adaptor events, editing context events, WOApplication events, and so on.

To enable event logging in an application, simply open the WOEventSetup page as described in "WOEventSetup page" and enable logging for the event classes you want to see.

In addition to the framework support, the WOExtensions framework provides components for using the feature. WOEventSetup is a page you use to configure event logging, and WOEventDisplay is a page the displays event information. Both pages can be accessed in any WebObjects 4.5 application with a direct action, as described in the following sections.

WOEventSetup page

The page used to set up the logging properties is accessed through a direct action named "WOEventSetup". So for example, you can access the WOEventSetup page for an application named "MyApp" with a URL such as the following:

http://myhost:aPort/cgi-bin/WebObjects/MyApp.woa/wa/WOEventSetup

On the WOEventSetup page, you can see all families of events that are registered for the application. Since the event classes are registered dynamically as the program executes, it is a good idea to "warm up" an application before accessing WOEventSetup.

The page lists the registered event classes, their subcategories, and a description of the kinds of events that can be logged. For instance, the EOEditingContext event class logs events for the saveChanges and objectsWithFetchSpecification: methods. Logging for each class can be enabled and disabled with the corresponding check box; it isn't possible to disable individual subcategories of an event class.

WOEventDisplay page

The page that displays collected events, WOEventDisplay, is also accessed through a direct action. For example, you can access the WOEventSetup page for an application named "MyApp" with a URL such as the following:

http://myhost:aPort/cgi-bin/WebObjects/MyApp.woa/wa/WOEventDisplay

On this page, you can view events in four different ways:

• Raw root events. This view, the most verbose, displays all events at the root level (events without an encompassing event). WOEventDisplay shows each event individually, which means that its possible for an event to appear multiple times if the thread of execution crosses its point more than once.
  
  
• Aggregated root events. This view is similar to the raw root event view, except that multiple identical events are aggregated or grouped in a single entry, and their combined time is displayed. In addition, the "Calls" column shows how many times an event was executed (in other words, how many events contributed to the displayed aggregate event).
  
  
• Events grouped by page and component. In this view, the first level of display shows only page names. By expanding a page, you get a list of components in that page. Expanding a component shows all the events within that component. This means that even events which were collected "deep" within a component are shown immediately below the component name. All identical events are aggregated as in the aggregated root event view for easier reading. It's possible to traverse the component event hierarchy by expanding the hyperlinks within a component.
  
  Note that since a page is also a component, a page with no dynamic subcomponents seems as if it's nested one level too deep. This is the correct behavior.
  
  
• Events grouped by page only. This display is similar to the grouped by page and component view, except the events do not have a by-component subgrouping.

In any of these displays, if an event or event group has subevents, it can be expanded by clicking the hyperlink or triangle image.

Each view orders events by duration (in milliseconds) from the longest to the shortest. Aggregation reduces rounding errors, which are a maximum of 1ms per event. In other words, an aggregate event consisting of ten events has at most 1ms deviation from the actual run time; however, manually adding ten individual events as displayed in the table might have up to a 10ms deviation. Therefore, any displayed sum is always more accurate than adding up the durations of individual events. Also note that the sub-events of an event branch doesn't necessarily add up to the duration of the branch event-the branch event's duration might be larger. This because the parent event generally consists of more than just calling the methods causing the sub-events.

Event System User Defaults

The event system provides three defaults for configuring its behavior:

EOEventLoggingEnabled

A boolean value that determines whether event logging is enabled. The default is NO, logging isn't enabled.
You can enable logging with EOEventLoggingEnabled, but the WOEventSetup page gives you more flexibility. The EOEventLoggingEnabled default enables logging for every class, whereas with WOEventSetup you can enable logging on a class by class basis.

EOEventLoggingLimit

An integer value that sets the maximum number of events the event system logs. The default logging limit is 200,000 per thread.
When the logging limit is reached, the event system attempts to purge old events before logging new ones. If the system is unable to purge old events, event logging is aborted.

EOEventLoggingOverflowDisplay

A boolean value that determines whether the event system logs a message when the event logging limit is reached. If enabled, the system logs messages when the event center truncates the log and also when event logging is aborted due to overflow.

Event Logging Questions and Answers

Question

What happens to an EOEvent if an exception is raised before the event is completed?

Answer

As soon as you close another event, the system detects that a previous event was not closed properly and closes it for you. All events logged between an unclosed event (due to an exception or improper coding) and the closing of another event are logged at the wrong place in the event hierarchy, but they are logged.
An improperly closed event can have another negative side effect: the improperly closed event can't be pruned with the automatic memory manager. This is virtually the only thing that can completely abort event logging.
If you have reason to believe that you might raise while an event is in progress, you should cancel the event in an exception handler.

Question

What's the overhead of enabling event logging?

Answer

The logging mechanism is extremely fast and memory efficient. A standard 300 MHz G3 system can log more than 300,000 events per second. Thus, the creation and logging of events is negligible compared to the time required to generate dynamic web pages. The only expensive operations are tree pruning when memory overflows (which takes about as long as logging ten events), and handling exceptions (which is linear to the depth of the tree-rarely more than 5 levels deep under normal circumstances). Therefore, the overhead is not really measurable.

Question

Can I enable event logging for a single-user application in production?

Answer

You shouldn't, because it would use a lot of memory: about 4 MB per thread for the default event log limit (200,000 events). Also, throughout the lifetime of an application, the system must continuously prune the event tree to keep the log size under the limit. As stated above, pruning the tree is a relatively expensive operation.

Question

Is performance impacted by the size of the event log? Does performance degrade if logging isn't reset periodically?

Answer

No. As stated above, the size of the event log is limited. Once you warm up the application and the logging framework, event logging overhead is constant-not per event, but over an average of a series of events.

Custom Event Logging

To define and log custom events, you create an event class, define the event's categories and subcategories, register the event class with the WOEvent center, and instrument the portions of code you want to log. This section describes these steps.

To create a custom event:

1. Create a subclass of EOEvent or an appropriate subclass.
   
   For example, to log events for a custom adaptor you've written, say MyAdaptor, create an EOEvent subclass named MyAdaptorEvent.
   
   Your subclass doesn't usually have to override any of the inherited methods, but you can customize the default behavior.



2. Create a description file for your event and add it to your project's Resources folder.
   
   An event's description file defines the event categories and subcategories used in the WOEventDisplay page. The file's contents is a dictionary in plist format. For the MyAdaptorEvent class, the file's name is MyAdaptorEvent.description, and it might look like the following:

   {
       EOEventGroupName = "MyAdaptor Event";
       connect = "Connect";
       openChannel = "Open Channel";
       evaluateExpression = "Evaluate Expression";
       fetchRow = "Fetch Row";
       commitTransaction = "Commit Transaction";
   }

   
   The EOEventGroupName entry is mandatory. It describes the family of events logged by the event class. Any other keys are self defined by the event class. In this example, the other keys (connect, openChannel, and so on) are the names of the events MyAdaptorEvent logs.



3. Register the event class with the EOEventCenter.
   
   Typically you register the event class in the initialize method of the class whose code you're instrumenting-MyAdaptor in this example.

   static Class MyAdaptorEventLoggingClass = Nil;
   static NSString *connectEvent = @"connect";
   static NSString *openChannelEvent = @"openChannel";
   static NSString *evaluateExpressionEvent = @"evaluateExpression";
   static NSString *fetchRowEvent = @"fetchRow";
   static NSString *commitTransactionEvent = @"commitTransaction";
   
   + (void)initialize {
       [EOEventCenter registerEventClass:[MyAdaptorEvent class]
               classPointer:&MyAdaptorEventLoggingClass];
   }

   
   As in this example, you might want to define string constants for the keys in your event's description dictionary.



4. Instrument the methods.
   
   In any method you want to instrument, add the following code, substituting the appropriate event key. This code instruments the "connect" event of MyAdaptorEvent.

   MyAdaptorEvent *event=nil;
   
   // Setup and start logging
   if(MyAdaptorEventLoggingClass) {
       event = EONewEventOfClass(MyAdaptorEventLoggingClass, connectEvent);
       EOMarkStartOfEvent(event, nil);
   }
   
   // Code to be timed goes here.
   
   // Finish logging.
   if(event) {
       EOMarkEndOfEvent(event);

   
   The second argument to EONewEventOfClass is an event key corresponding with an entry in the .description file. The corresponding value is used in the Title column of the WOEventDisplay page. If the argument isn't a key in the description dictionary, EONewEventOfClass uses the argument instead.
   
   For more information on the methods used in this example, see the class descriptions for EOEvent and EOEventCenter. To see a complete example of timing events, refer to the ODBC Adaptor source code that's distributed with WebObjects.

Related API Changes

Three new classes and one interface have been added to support event logging. They are:

• EOEventCenter (EOControl/EOEventCenter.h)
• EOEvent (EOControl/EOEvent.h)
• EOAggregateEvent (EOControl/EOAggregateEvent.h)
• EOEventCenter.EventRecordingHandler (Java) or EOEventRecordingHandler (Objective-C; EOControl/EOEventCenter.h)

For more information, see the corresponding class or interface specifications.

Object Sharing

EOF 4.5 introduces a new technique for sharing read-only enterprise objects. The new subclass of EOEditingContext, EOSharedEditingContext, defines a mechanism that allows editing contexts to share enterprise objects for reading. This mechanism can reduce both the number of fetches an application makes and the amount of redundant data it requires.

As an example, consider the FeeType entity in the samle Rentals model that ships with EOF 4.5. A FeeType enterprise objects describes a type of fee that a video store can charge its customers-"Rental" and "Late", are the two FeeTypes in the sample database. It is very uncommon to add or remove FeeTypes, and it's perhaps even more uncommon to modify an existing FeeType (to rename it, for example). For the most part, FeeTypes are read-only.

With 4.5, you can fetch read-only objects such as FeeTypes into a shared editing context once, when an application starts, and all the application's sessions can share those objects. For example, objects in any session can create relationships to the shared FeeType objects even though the FeeTypes are in a different editing context from the source objects. Using previous releases, you would have to make local copies of the read-only FeeTypes in each of the editing contexts that use them.

How It Works

The idea behind shared editing contexts is to load read-only (or read-mostly) objects into a central context that all sessions have transparent access to. It works like this.

1. A model file identifies any objects to be shared.
   
   Models identify shared objects by defining shared object fetch specifications, which define criteria for fetching objects that are to be shared. For information on creating shared object fetch specifications, see "Setting Up Object Sharing" .
   
   The first time your application accesses a model's entities, it checks the model for shared object fetch specifications. If any are found, they are evaluated and the corresponding fetched objects are loaded into the default shared editing context. Any existing editing contexts that don't have any registered objects begin using the default shared editing context. Similarly, any editing contexts subsequently created use the default shared editing context.
   
   Note that if you don't specify any shared fetch specifications, a shared editing context is never created and no object sharing occurs. Conversely, if you do specify shared object fetch specifications, a shared editing context is automatically created and object sharing is enabled for all standard editing contexts.
   
   
2. The application's shared editing context is created and populated the first time a model containing shared object fetch specifications is accessed.
   
   Generally an application's shared editing context is initialized when the first editing context attempts to access the database. At this time, all the application's models are loaded, and any shared object fetch specifications are detected. If any of the application's models have shared object fetch specifications, a shared editing context is created and populated and is set as the shared editing context for all empty editing contexts.
   
   You can disable object sharing on individual editing contexts as described in "Inserting, Updating, and Deleting Shared Objects" or you can disable object sharing altogether as described in "Disabling Sharing During Development" .
   
   
3. Standard editing contexts use objects in the shared editing context as if they were local.
   
   When a standard editing context fetches, any relationships to shared objects are automatically resolved to the shared editing context's objects. Similarly, an object in any standard editing context can create a relationship to a shared editing context's object.
   
   This all works transparently. EOEditingContext's implementations of objectForGlobalID and faultForGlobalID look for an object in the shared editing context when a standard editing context doesn't find the object locally. If the methods finds the object in the shared editing context, they return the shared object as they would if the object were local.

To allow object sharing to work, EOF makes the following assumptions:

• Shared objects are read-only (or read-mostly).
• Objects must be unique in their editing contexts and their editing context's shared editing context.

The following sections describe why these assumptions are necessary and how they are enforced.

Shared Objects Are Read-Only

If you could update a shared object, all of an application's users would see the changes immediately since all sessions share the exact same object. This behavior is undesirable. You only want users to see committed changes to objects. For example, suppose you make a change to a shared object in a web application, but you haven't yet saved it to the database. The changes are written to the object as soon as the request-response loop begins, and every other web user sees the change you made, even if you undo them later or make further changes before saving.

For this reason EOF enforces the read-only quality of shared objects. Shared objects can't be inserted, updated, or deleted in a shared editing context the same way that normal enterprise objects can be inserted, updated, or deleted when they're in a standard editing context. EOSharedEditingContext overrides EOEditingContext methods that mutate data (takeValueForKey, saveChanges, deleteObject, and insertObject for example) to raise exceptions. Correspondingly, methods that report on changes in a shared editing context return either null/nil or an empty array.

It would also be undesirable if you could delete a shared object. In Objective-C, it would be bad if a shared object were released while objects in standard editing contexts had relationships to it. Objects in a shared editing context are always retained. You are guaranteed that no object in a shared editing context will be destroyed while the application is running (a shared object can become invalid, but it won't go away).

Shared Objects Are Uniqued

EOF uses object uniquing to ensure that a single editing context never has more than one object with the same global ID. Because every standard editing context has access to the objects in a shared editing context as if the objects were local, none of a shared editing context's objects can have the same global ID as any object in any standard editing context.

EOF does most of the work for you. A shared editing context sends out notifications when it initializes new objects. Standard editing contexts listen for these notifications and raise exceptions if they have local objects with the same global ID as the new shared object. However, you have to take some precautions not to fetch into a shared editing context objects that have already been fetched into a standard editing context:

• Don't explicitly fetch into a standard editing context an object that is already shared unless you disable sharing, as described in "Inserting, Updating, and Deleting Shared Objects" .
• If a shared object has relationships to entities that aren't shared, remove the relationships from the model. Otherwise, when the faults fire, the destination objects are fetched into the shared editing context.

Setting Up Object Sharing

EOModeler has a new entity inspector for specifying the objects you want to fetch into a shared editing context when an application starts. To use it, select the entity whose objects you want to share, and open the Shared Objects Inspector.

To share all the entity's objects, select "Share all objects." If the entity has an unqualified fetch specification, that fetch specification is selected. Otherwise, EOModeler creates an unqualified fetch specification and selects it.

To share only some objects, define fetch specifications that select the objects you want to share. In the Shared Objects Inspector, select "Share objects fetched with" and select your fetch specifications.

You can also prefetch relationships into the shared editing context. For example, suppose that Product has a relationship to ProductPlatforms and that both entities should be shared. You can create a Product fetch specification that has a hint to prefetch the Product's ProductPlatforms relationship. Using this fetch specification to fetch Products into the shared editing context also causes ProductPlatforms to be loaded into the shared editing context.

Accessing Shared Objects

Shared editing contexts maintain dictionaries of all the objects they have fetched, both by entity name and by the fetch specification name they were retrieved with. This makes it easy to use shared objects for populating user interface controls. The two methods for accessing these dictionaries are:

objectsByEntityName
objectsByEntityNameAndFetchSpecificationName

These two methods make it easy to use key-value coding and key paths to bind data to WebObjects elements directly in a WebObjects .wod file.

For example: Say you have a model with the entity "Products" and an attribute called "isDiscontinued", which indicates whether an item is unavailable. You could set up a fetch specification in EOModeler named "regularProducts" that fetches only objects that are available (the qualifier would specify that "isDiscontinued = 0"). To use the returned set of "regular products" in a WebObjects page, create a user interface control (such as a WOPopup or WOBrowser), and set its list attribute like this:

list =
session.defaultSharedEditingContext.objectsByEntityNameAndFetchSpecificationName.Products.regularProducts

The user interface control's list is automatically filled with data from shared objects.

Inserting, Updating, and Deleting Shared Objects

Generally shared objects are read-mostly. For example, an application could provide an administration mode that allows administrative users to insert, update, and delete otherwise read-only shared objects. You can still use a shared editing context in such an application, but you have to write special code. Since insertions, updates, and deletions of shared objects are typically very infrequent, the performance benefit of sharing objects can still be significant.

To insert new objects, create and insert a new object in a standard editing context. Once you do this, you need to fetch the new object into the shared editing context as described in "Refreshing the Shared Editing Context" .

Updating and deleting them is a little trickier. To modify shared objects, you need to disable object sharing for a particular editing context, as follows:

mySession.defaultEditingContext().setSharedEditingContext(null); // Java
[[mySession defaultEditingContext] setSharedEditingContext:nil]; // ObjC

A session that does this doesn't have access to shared objects and can explicitly fetch objects that would otherwise be shared. After fetching the objects, they are local, and you can update or delete them. When you save the changes, the shared editing context receives ObjectsChangedInStoreNotifications. In response to the notifications, the shared editing context refaults updated objects and removes deleted objects from its objectsByEntityName and objectsByEntityNameAndFetchSpecificationName dictionaries.

Refreshing the Shared Editing Context

You might want to refresh the shared editing context after an administrative user inserts a new shared object or to periodically synchronize with a database. To refresh a shared editing context, use the method bindObjectsWithFetchSpecification (Java) or bindObjectsWithFetchSpecification:toName: (Objective-C). For example:

EOModelGroup modelGroup = EOModelGroup.defaultGroup();
EOSharedEditingContext sharedEC =
    EOSharedEditingContext.defaultSharedEditingContext(); 
    
EOFetchSpecification fs =
    modelGroup.fetchSpecificationNamed("regularProducts", "Product"); 
    
if (fs == null) {
    System.out.println("Couldn't get regularProducts fetch specification.");    
} else {
    sharedEC.bindObjectsWithFetchSpecification(fs, "regularProducts");
}

This has the effect of refetching all the shared objects and binding them to the objectsByEntityName and objectsByEntityNameAndFetchSpecificationName dictionaries.

Disabling Sharing During Development

Initializing a shared editing context does increase the amount of time it takes for an application to start up. During development when you starting your application frequently, the additional start up time can become a nuisance. To make debugging passes quicker, you can essentially disable object sharing using the static method setSharedObjectLoadingEnabled (or the equivalent class method in Objective-C).

This method specifies whether EOF looks for shared fetch specifications when it loads models. While a shared editing context can still be created and referenced, it won't contain any objects unless you programmatically fetch them into it. The advantage of disabling shared object loading is that the application starts up more quickly.

Performance

Using shared editing contexts can have the following positive performance impact on your application:

• Standard editing contexts don't have their own copies of shared objects, so the memory footprint is smaller.
• Fetching read-only data into an application is a lot simpler, so fewer database trips may be needed.

However, there are potentially negative performance effects as well. They are:

• Every editing context has to hash twice: once to look up objects in their own tables, and again to look up objects in the shared tables. Hashing is very fast, but with nested contexts it can add up.
• On Windows NT, locking a thread is significantly slower than it is on other platforms with similar hardware. Because shared editing contexts generate a lot of locking activity, sharing objects can actually degrade the performance of an application on NT. If you plan to deploy on NT, test your server's performance carefully.
• Fetching or firing faults in a shared editing context causes all other threads to block when they try to get shared data. To avoid this problem, try to fetch everything when the application starts up.

Multithreaded Access and Locking

WebObjects applications take care of most multithreading issues for you. This is also the case with applications that use shared editing contexts. Here are the basics of how EOF locks shared editing contexts:

• Shared editing contexts use a new lock, EOMultiReaderLock, which allows multiple contexts to read data from the shared editing context concurrently. When the shared editing context needs to update objects (because of fetching or firing faults), the context waits for all reader locks to be surrendered and then issues a writer lock. There can only be one writer at a time.
• Whenever an EOEditingContext lock method is called, it also obtains a reader lock for its shared editing context.
• When an EOEditingContext is about to call its parent object store, it surrenders its shared reader lock until the call is complete (this prevents deadlocks with the object store lock).

From a WebObjects perspective, everything works automatically as long as you interact with a shared editing context from within the context of a session. This is because when a web session awakes, it locks its editing context, which reader-locks the shared editing context. If you need to interact with the shared editing context outside the context of a session, you should access it through a locked EOEditingContext.

If you invoke methods on a shared editing context directly, you must obey a strict lock ordering protocol to avoid deadlocks and unsafe multithreaded access. If you directly invoke any method besides objectsWithFetchSpecification, objectForGlobalID, or faultForGlobalID on a shared editing context, you need to take the a reader or writer lock yourself before calling the method.

Related API Changes

A new subclass of EOEditingContext, EOSharedEditingContext, has been added to support object sharing. For more information on this class, see the corresponding class specification.

Additionally, new API has been added to existing classes as summarized in the following tables.

Subclassing EOGenericRecord

EOF 4.5 adds a new option for creating custom enterprise objects: rather than creating a subclass of EOCustomObject (Java) or NSObject (Objective-C), you can now subclass EOGenericRecord.

This feature is most significant in applications that use the Java bridge. By default, a subclass of EOGenericRecord stores its properties in a dictionary on the Objective-C side of the bridge instead of in individual instance variables on the Java side. This allows EOF to access enterprise object properties with many fewer trips across the bridge, which reduces memory usage and improves performance.

More specifically, subclassing EOGenericRecord provides the following advantages over subclassing EOCustomObject (Java):

• Java enterprise objects don't incur a performance penalty relative to Objective-C. Previously, each call to an EOKeyValueCoding method (such as valueForKey) resulted in a bridged method invocation and possibly the creation of a new object. Because EOGenericRecord provides the storage for an enterprise object's values, the values don't need to cross the bridge during a fetch. This results in a dramatic performance improvement in EOKeyValueCoding-bound operations such as fetching, snapshotting, and validation.
• Java enterprise objects don't use more memory than Objective-C enterprise objects. Since an enterprise object's values were previously stored in its instance variables, each value object had to be morphed between Java and ObjC when interacting with the EOF frameworks. This meant that the values couldn't be shared between an enterprise object and its snapshots in the EOEditingContext and EODatabase. Consequently, many more objects were resident in memory than would be the case in the non-bridged case.
• Enterprise object classes are more flexible and easier to maintain. As you edit your model by adding or removing keys, you only have to update your enterprise object class if you want to add or remove custom accessors or validation methods. You don't have to maintain instance variables since property values are stored in a dictionary.
  Also, you can use a specific subclass of EOGenericRecord for multiple entities as you can when you use EOGenericRecord. In previous releases, a single enterprise object class (other than EOGenericRecord) could correspond to only one entity. Subclasses of EOGenericRecord can now correspond to multiple entities the same way EOGenericRecord can. This might be useful for providing common behavior such as validation logic. Note that a "generic" subclass of EOGenericRecord (that is, a subclass which is shared among entities) must be instantiated using EOClassDescription's createInstanceWithEditingContext method in the same way that EOGenericRecord must. If it isn't, the resulting instance won't be properly initialized.
• Subclasses of EOGenericRecord automatically use deferred faulting, which increases performance and decreases memory usage even further. For more information on deferred faulting, see "Deferred Faulting" .

Given all the advantages of subclassing EOGenericRecord, you might wonder if there's ever a reason to subclass EOCustomObject. Subclassing EOGenericRecord usually yields better performance. Even if it doesn't yield better performance, you can design it to be at least as fast and to have the same level of functionality as a subclass of EOCustomObject by storing properties in instance variables. Therefore, you really don't ever need to subclass EOCustomObject.

Property Storage: Dictionary or Instance Variables

A subclass of EOGenericRecord can take three approaches to storing property values:

1. Store them in the EOGenericRecord dictionary.
2. Store them in instance variables.
3. Store them in a combination of the above.

Generally the first approach is the best because it reduces the number of trips across the Java bridge, improving performance and reducing memory usage. Enterprise objects are accessed from Objective-C much more frequently than from Java. A typical enterprise object is populated from Objective-C when it's fetched from the database, snapshotted from Objective-C each time the object is changed, and accessed from Objective-C to provide values for bindings in WebObjects components.

On the other hand, access from Java is typically limited to custom validation (implemented in the enterprise object class) and to accesses by other enterprise objects to obtain property values explicitly (for example, to check the department budget before validating a new salary).

The second approach might be appropriate if your application accesses and manipulates enterprise objects from Java quite frequently, however this case is probably rare.

The third approach is to create instance variables for some of your enterprise object's properties and leave the rest in the EOGenericRecord dictionary. This approach is the most difficult to maintain, because the implementation of accessor methods depends on the way the property is stored.

Creating a Subclass

EOModeler has been updated with templates that take advantage of this feature, so new enterprise object classes are automatically subclasses of EOGenericRecord. To re-parent existing business objects, perform the following steps in your source files:

1. Set the enterprise object class's superclass to EOGenericRecord.
2. Delete all EOKeyValueCoding-related instance variables (that is, instance variables that store values for attribute and relationship keys).
3. (Optional) For maximum performance, delete the three-argument constructor if you don't use the arguments. If you need custom logic in your constructor but don't need the three arguments, implement a custom default constructor. If you don't need a custom constructor, don't implement one at all; the compiler inserts the default constructor for you. If the three-argument constructor is present, EOF uses it; otherwise the default constructor is invoked.
4. (Optional) For maximum performance, delete all accessors that don't contain custom logic. If you want strong typing to simplify business logic, implement accessors using storedValueForKey, such as:

   public NSGregorianDate dateReleased() {
       return (NSGregorianDate)storedValueForKey("dateReleased");
   }

   
   

In lieu of (or in addition to) adding custom accessors, it might be convenient to add and use String constants for an enterprise object class's keys, as shown in the following:

public class Movie extends EOGenericRecord {
    public static final String Title = "title";
    public static final String DateReleased = "dateReleased";
    public static final String Studio = "studio";

    public NSGregorianDate dateReleased() {
        return (NSGregorianDate)storedValueForKey(DateReleased);
    }
}

Deferred Faulting

EOF uses faults as stand-ins for objects whose data has not yet been fetched. Although fault creation is much faster than fetching, fault instantiation still takes time. To improve performance, EOF 4.5 has the ability to use deferred faults (which are more efficient) for enterprise object classes that enable the feature.

In an object whose class enables deferred faulting, the object's relationships are initially set to deferred faults. For a particular relationship, a single deferred fault is shared between all instances of an enterprise object class. This sharing of deferred faults can significantly reduce the number of faults that need to be created, and usually reduces the overhead of fault creation during a fetch.

For example, consider a Movie class with a studio relationship. Assuming the worst case in which each movie has a different studio, without deferred faulting, during a fetch of twenty Movie objects, twenty faults are created for the studio relationship-one fault for each movie. With deferred faulting, only one fault is created-a deferred fault that is shared by all the movies.

Deferred faults have a special fault handler, which knows how to replace the deferred fault with a standard fault. Once the deferred fault is replaced with a normal fault, the normal faulting behavior applies.

In 4.5, EOGenericRecord enables deferred faulting; you get the behavior without making any changes to existing code. Non-EOGenericRecord enterprise object classes don't enable deferred faulting by default. To enable it on a custom class, implement the class method useDeferredFaultCreation to return true/YES (the default is false/NO). Additionally, invoke the method willReadRelationship before accessing a relationship that might be a deferred fault. The willReadRelationship method allows the special fault handler to instantiate a normal fault before it is accessed.

Deferred Faulting and Inheritance

In Java programming, there's an additional benefit of deferred faulting: Deferred faulting allows you to have to-one relationships into an inheritance hierarchy. Without deferred faulting, a to-one relationship to a non-leaf entity is impossible without implementing workarounds (because of strong typing in the Java language). The workarounds are not necessary if you use deferred faulting. For more information on the inheritance limitation in Java and the workarounds, refer to the Enterprise Objects Framework Developer's Guide.

Related API Changes

New API added to support the deferred faulting feature is summarized in this section.

Snapshot Reference Counting

Snapshot reference counting is a new feature that removes snapshots from an EODatabase when they are no longer used by any enterprise objects in an application. This feature reduces the memory footprint of WebObjects applications.

The reference count on a snapshot is implicitly incremented when you create an enterprise object, either by fetching it in a batch or by faulting it in. An EOEditingContext decrements the reference count automatically when it forgets or refaults an enterprise object.

To support this feature, the method editingContextDidForgetObjectWithGlobalID/ editingContext:didForgetObjectWithGlobalID: has been added to EOObjectStore. This method is not intended to be called directly by your application.

As a developer you should not worry about this feature; it should just work without any additional code on your part. There are only a few cases where it could be incompatible with pre-EOF 4.5 applications:

• When the snapshots are removed because their reference counts fall to zero, the globalID associated with these snapshots are also released. If your application is written in Objective-C and you cache globalIDs in your application, make sure you retain them. If you don't, your application will crash with "Message sent to freed object" errors when you try to access globalIDs that have been freed. Previously GlobalIDs were never released; they were kept with the snapshots. This is not a problem if you use Java as your development language, since Java automatically retains all referenced objects.
• If your application assumes that snapshots for fetched objects are always present, you could have a problem. This is no longer the case, and asking for a snapshot might now return null/nil.

Sometimes you want a snapshot to stay around for the lifetime of your application. If you need this functionality there are three ways to force selected snapshots to stay around:

• Use the method incrementSnapshotCountForGlobalID on EODatabase to force the snapshot to stay around. If you use this approach, make sure to call decrementSnapshotCountForGlobalID when you don't need the snapshot anymore.
• Fetch the objects whose snapshots you need. (Note that if your application is written in pure Objective-C and it's not multithreaded, the EOEditingContext doesn't retain its enterprise objects-you have to retain them yourself.)
• Turn on entity caching for the entity. This automatically increments the reference count for all the snapshots your entity caches.

You can turn the whole feature off and revert to the previous behavior of retaining snapshots forever by invoking the class method disableSnapshotRefCounting on EODatabase. Make sure to call this method early in your application initialization code, before the creation of any EODatabase instances.

Related API Changes

The following tables summarize new or changed API to support the snapshot reference counting feature.

Snapshot Timestamping

EOF caches database snapshots and uses the cached values to initialize objects. This significantly improves performance, since using the cached values is much faster than making round-trips to the database for fetches. However, this behavior can lead to staleness of the cached data; and it sometimes produces confusing results, especially when new values available from a fetch are ignored in favor of the stale snapshots.

A new snapshot timestamping feature updates snapshots appropriately when fetching and allows an editing context to request that the snapshots used to build enterprise objects are no older than the editing context's fetchTimestamp. The default value for the fetchTimestamp of a new editing context is one hour earlier than the creation time of the editing context. So any snapshots that are less than an hour old are acceptable to the editing context; older cached values are ignored. The default "lag" can be adjusted using the static method setDefaultFetchTimestampLag (or the corresponding Objective-C class method), and the fetchTimestamp for a specific editing context can be set directly with setFetchTimestamp.

Note that the fetchTimestamp is significant only when fetching data (typically by sending objectsWithFetchSpecification). An existing enterprise object is unaffected by changes to the editing context's fetchTimestamp.

When an EODatabase records a snapshot, it now also records a timestamp for that snapshot. The timestamp is an NSTimeInterval (relative to the reference date as documented for the NSDate class). Methods that return a snapshot from EODatabase or EODatabaseContext now have variants that take an extra argument to specify a minimal timestamp for the snapshot. If the recorded snapshot's timestamp is earlier than the requested time, null/nil is returned. If the snapshot is recent enough for the request, the snapshot is returned as usual.

Related API Changes

The following tables summarize new API added to support the snapshot timestamp feature.

Handling Missing Faults

In previous versions of EOF, if a fault fired but no corresponding database row could be found (for example because of a referential integrity problem or because the row was deleted without EOF's knowledge), the delegate method databaseContextFailedToFetchObject/databaseContext:failedToFetchObject:globalID: was called. If the delegate didn't fix the problem, an exception was raised immediately. In 4.5, The delegate is invoked, but the exception is delayed or avoided, and an empty enterprise object is returned. If the application later tries to save an object graph that requires the missing fault, the exception is raised during saveChanges. If the object is never needed, no exception is raised.

Related API Changes

The following tables summarize new API added to support the better handling of missing faults.

Automatic Database Reconnection

In EOF 4.5, a concrete adaptor can now implement methods that cause EOF to automatically attempt to reconnect to a database server when a connection is unexpectedly dropped. This behavior handles the problem of transient communication failures. By default reconnection is attempted by all of the adaptors that ship with EOF 4.5.

EOF now sends isDroppedConnectionException to the adaptor if an exception is raised during fetching or saving. If the adaptor returns true/YES, then it attempts to reconnect to the database and retry the operation. (The adaptor context must also implement handleDroppedConnetion to clean up the state of the context and its channels before the reconnection is attempted.) If the reconnection attempt fails, the exception from the failure is raised as usual.

The delegate method reconnectionDictionaryForAdaptor can be used to provide a new connection dictionary for the reconnection attempt. If the delegate is not implemented, the adaptor uses its existing connection dictionary when reconnecting to the server.

You can completely override the database reconnection behavior with the delegate method databaseContextShouldHandleDatabaseException (Java) or databaseContext: shouldHandleDatabaseException (Objective-C). For more information, see Table 5-20 .

Related API Changes

The following tables summarize new API added to support the database reconnection feature.

Setting Access Layer Delegates

EOAdaptor, EOAdaptorContext, and EODatabaseContext now implement the static method setDefaultDelegate to simplify the setting of delegates for new instances of those classes (there are corresponding class methods in Objective-C). By default, the default delegate for those classes is null/nil. However, setDefaultDelegate can be used to establish a default delegate for a particular class that will be assigned to any new instance of that class when the instance is initialized.

Related API Changes

The following tables summarize new API added to support the database reconnection feature.

Key Value Coding Changes

EOF 4.5 introduces two improvements to key-value coding: The addition of key binding objects and the enforcement of lowercase key names. The follow sections describe each.

Key Bindings

New key-value coding primitives have been introduced. The new primitives are based on bindings, which associate a class/key pair to a mechanism for accessing the key. There are two types of bindings, get bindings and set bindings. They are represented by the new class EOKeyBinding.

To access an object's data with key-value coding, clients typically use the EOKeyValueCoding and EOKeyValueCodingAdditions methods valueForKey, takeValueForKey, and so on. However, in EOF 4.5, clients can optimize access by obtaining and caching bindings for particular class/key combinations and applying those bindings directly. Classes can override the EOKeyValueCoding methods (such as valueForKey) directly, but overriding keyValueBindingForKey to return an appropriately initialized binder object provides the maximum performance benefit.

Clients caching bindings are responsible for checking that the class of the object to which a binding is applied matches the target class of the binding. Using a binding obtained from one class on an object of another results in undefined behavior.

Enforcing Lowercase Key Names

EOF expects keys to begin with a lowercase letter. It now logs a warning if that restriction is violated. For backwards compatibility with previous releases which did not strictly check capitalization, you can use the EOKeyValueCoding.KeyBinding static method suppressCapitalizedKeyWarning to suppress the warning for capitalized keys (EOKeyBinding class method in Objective-C). However, note that this method is deprecated and will be removed in a future release.

Related API Changes

The following new classes have been added to support key value bindings:

• EOKeyValueCoding.KeyBinding (Java) and EOKeyBinding (Objective-C)
• EOKeyValueCoding.KeyGetBinding (Java) and EOKeyGetBinding (Objective-C)
• EOKeyValueCoding.KeySetBinding (Java) and EOKeySetBinding (Objective-C)

For more information, see the class specification for EOKeyValueCoding.KeyBinding/EOKeyBinding.

In addition, the EOKeyValueCoding interface/informal protocol has been enhanced to create and return key bindings. In Java, the new methods are defined in a new interface called EOKeyValueCoding.KeyBindingCreation. In Objective-C, the new methods are defined in EOKeyValueCoding.h. For more information, see the EOKeyValueCoding.KeyBindingCreation interface specification (Java) or the EOKeyBindingCreation informal protocol specification (Objective-C).

Recursive Reader and Writer Locks

A new class, EOMultiReaderLock, provides EOF with recursive reader and writer locks. The locks are recursive; a single thread can request a lock many times, but a lock is actually taken only on the first request. Likewise, when a thread indicates it's finished with a lock, it takes an equal number of unlock calls to return the lock.

There is no limit on the number of reader locks that can be taken by a process. However, there can only be one writer lock at a time, and a writer lock is not issued until all reader locks are returned. (Reader locks aren't issued to new threads when there is a thread waiting for a writer lock, but threads that already have a reader lock can increment their lock count.)

Thread safety is maintained by using mutex locks (binary semaphores), which ensures that no more than one critical section of the class can be processed at a time. The queueing order of requests for writer locks is not managed by the class; the underlying implementation of mutex signaling manages the queue order.

EOMultiReaderLock correctly handles promotion of a read lock to a write lock, and the extension of a reader lock to the current writer. This prevents a thread from deadlocking on itself when requesting a combination of lock types.

EOMultiReaderLocks are slightly more time-expensive than NSRecursiveLocks because the recursion count has to be stored per-thread, causing each request for a reader lock to incur a hash. Writer locks are even more expensive because EOMultiReaderLock must poll the hashtable until all reader locks have been returned before the writer lock can be taken.

Related API Changes

A new class, EOMultiReaderLock, has been added to the Framework to support the new multi reader and writer lock feature. For more information, see the corresponding class specification.

LDAP Adaptor Example

EOF 4.5 comes with a new sample adaptor: the LDAP adaptor. It provides read, modify, and delete access and limited support for inserting. Additionally it provides a simple way to verify a user's password on the Web with an LDAP server without requiring a model.

LDAP Client Libraries

The LDAP adaptor is ready to use when you install WebObjects. The source code is provided so that you can use a different client library or enhance the adaptor yourself.

The LDAP adaptor consists of two pieces: The adaptor framework and an LDAP client library. Apple ships a client library based on the public University of Michigan LDAP Client. You can build and install this library as a framework, or you can substitute your own client library, such as the Netscape LDAP library. (The latter has not been tested, but both libraries conform to RFC 1823, the LDAP client API).

Creating Models

You create a model for an LDAP server with EOModeler the way you create models for other adaptors. Simply choose New from the Model menu, and choose LDAP as the adaptor for the new model.

Logging In

The LDAP adaptor defines the following connection keys, which are represented in the login panel:

Server Name

The server's address. Either an IP address (e.g. 17.205.15.205) or a name (e.g. ldap.bigfoot.com). If you are inside a corporate firewall, it might not allow LDAP traffic to pass through. Check with your administrator if you have trouble accessing LDAP servers on the open internet.

User Name

If your LDAP server requires you to login, enter a username. This version of the LDAP adaptor supports simple authentication only.

Password

If your LDAP server requires a password, enter it here.

Search Base

Entries on an LDAP server (roughly equivalent to rows in a database) are organized in trees. The search base indicates what branch of the tree to begin searching in. Common designs are to have the top level of the tree have country or organization branches, so good search bases have the form "c=US" or "o=Apple Computer". An approximate database equivalent is to say the search base is ANDed with any qualifiers you pass in.

Schema Base

This value is used only during reverse engineering the LDAP server. The Adaptor looks at this location for the subschema entry. Note that during reverse engineering, the scope is temporarily overridden to Base because the subschema does not have any children.

Value Separator

The character used to separate multiple values for a property.

Port

LDAP servers are almost always on port 389, but secure servers might use a different port.

Timeout

Maximum time to wait during LDAP operations before raising an exception. Timeouts come from the client library and are raised as EOGeneralAdaptorExceptions.

Scope

LDAP entries belong to an object class. The Adaptor maps each class to an entity. Like classes in object-oriented programming, LDAP classes have an inheritance hierarchy. Depending on the scope you set, results include results from the that object class only (Base), its immediate children (One Level), or the entire subtree (Subtree). The more you include, the larger the result set is for a given search. Object class hierarchies and entry hierarchies work together to give order to the LDAP server's data, and the Adaptor provides control over how ordered or chaotic the data appears.

If Reverse Engineering Fails

The LDAP specification does not require LDAP servers to provide a means of learning the server's layout; providing a schema to clients is optional. For this reason, the LDAP adaptor might not be able to reverse-engineer your LDAP server to generate an EOModel automatically. If this is the case, you might need to create your model by hand.

Generally speaking, Netscape Directory Servers automatically create and provide schemas to clients. So, if you have a Netscape Directory server, you should be fine.

However, if the adaptor can't find subschema information, it creates a default model with two entities: Person and NoSchemaDataAvailable. If the server is ldap.bigfoot.com (a well known server) and the scope is set to Subtree, this model works for finding names. The NoSchemaDataAvailable entity is designed to be an error message and should be deleted from the model. (EOAdaptors aren't allowed to return null/nil when reverse engineering a database, so the best alternative is to return a minimal model.)

Adding Entries to the Server

The LDAP adaptor provides limited support for adding entries to the server. If you plan to use the adaptor to insert entries, keep the following points in mind:

• The LDAP adaptor doesn't provide automatic primary key generation. so set the primary key (the distinguished name, or "dn") as a class property and set an appropriate value yourself before saving a new entry.
• Consider adding an objectclass attribute to the entity so your can add the entry to multiple classes at once.
• You must enforce object rules in custom code. The adaptor detects object class violations once you attempt to save changes, but the violations cause the adaptor to raise an exception.

Performing Authentication

One of the common applications for LDAP is to verify a user's password on the Web. The authentication is generally done by the LDAP server, not by retrieving the user's password from an LDAP entry. So, in essence, all that is needed is a mutable connection dictionary from a model and a request to validate the connection. The LDAP Adaptor provides a simple way to do this without a model:

java.lang.Throwable athenticateUser()

or in Objective-C,

+ (NSException *)authenticateUserString:(NSString *)userString
    password:(NSString *)password
    withServer:(NSString *)server
    matchOnAttribute:(NSString *)searchAtt
    searchBase:(NSString *)base
    searchScope:(NSString *)scope;

The method is pretty easy to use and the header file contains specifics on its use. Here is a sample code fragment from a WebObjects application that has one page, "Main" with three instance variables: userName1, password1, and isLoggedIn. This method is tied to the web page's Submit button.

- authenticateUser {
    NSException *exception;
    exception = [LDAPAdaptor authenticateUserString:userName1
            password:password1
            withServer:@"bigbird.apple.com"
            matchOnAttribute:@"cn"
            searchBase:@"o=apple computer" searchScope:@"Subtree"];ì
    if (exception) {
        NSLog(@"Auth failed.");
        isLoggedIn = NO;
    } else {
        NSLog(@"Auth successful.");
        isLoggedIn = YES;
    }
    return self;
}

Or, in Java (isLoggedIn is a string instead of a BOOL in this example):

public WOComponent authenticateUser() {
    java.lang.Throwable ex = null;
    ex = LDAPAdaptor.authenticateUser (
        userName,
        pswField,
        "bigbird.apple.com",
        "cn",
        "o=Apple Computer",
        "Subtree");
    if (ex == null) {
          isLoggedIn = "You are now logged in.";
    } else {
        isLoggedIn = ex.getMessage();
    }
    return this;
}

Miscellaneous API Enhancements

This section summarizes other miscellaneous methods added in EOF 4.5 not covered in the other sections.

Deprecated API

Nested transactions are no longer supported. EOF never actually used nested transactions. Furthermore, the concrete adaptors were not guaranteed to support them, especially since the SQL/92 standard doesn't allow nested transactions. New features in EOF 4.5 make nested transactions impossible to support.

Consequently, the following methods are deprecated.

For backwards compatibility, the Sybase Adaptor still allows you to attempt to begin a nested transaction, but the implementation ignores the nesting.