Thursday, January 31, 2013

Data-binding among complex Expressions in C#

What could be more preferable than simply stating X = Y in one place then having the rest of the work transparently handled on your behalf?

Expression trees were one of the many language features introduced with C# 3.0 way back in 2007.  Integration with the compiler allows us to nonchalantly ask it to emit data that represents a line of code instead of going about its usual business of readying the code for execution only.  This leads to a convenient convergence of the benefits of reflection and those of compiler-checked code.  An example that is further described here is that we no longer must use a string literal to meta-identify a property; instead, we can use a lambda expression like o => o.SomeProperty.

It took me much of this month of January 2013 to apply expression trees in a way that goes far beyond that simple property example.  The result demonstrates not only the power of code as data, but also the applicability of the concept of data binding to coding in general instead of in the user-interface alone.

The project arose from the need to manage numerous interdependent properties on view-model and model classes.  A property must be updated whenever one of the properties it depends on changes.  Done on an individual bases, this would lead to a lot of event handlers.  My undeviating aim to keep application code as succinct and declarative as possible led me to seek an alternative.  What could be more preferable than simply stating X = Y in one place then having the rest of the work transparently handled on your behalf?  Nothing, of course.  So I create the following public interface and then implemented the guts to make it work:

/// <summary>
/// Keeps a target's properties updated with values determined by corresponding expressions. Properties referenced within a value expression are monitored such that property-change notifications cause the target properties to be updated.
/// </summary>
/// <typeparam name="TTarget">The type of object on which properties will be updated</typeparam>
public class PropertiesUpdater<TTarget> : IDisposable
{
    /// <summary>
    /// Constructs a PropertiesUpdater instance for properties of the specified target.
    /// </summary>
    /// <param name="target">The object on which properties will be updated</param>
    public PropertiesUpdater( TTarget target )
    {
        ...
    }

    /// <summary>
    /// Sets the specified property of target to the result of the specified value expression whenever the value of a property referenced within the value expression changes.
    /// </summary>
    /// <typeparam name="TProperty">The type of the property on target to be set</typeparam>
    /// <param name="targetPropertyExp">The property on target to be set</param>
    /// <param name="valueExp">The expression to be monitored for changes and evaluated as the target property's value</param>
    /// <param name="setInitialValue">Should the target property be immediately set to the evaluated value expression?</param>
    /// <returns>The instance on which this method was invoked, to allow multiple invocations to be chained together</returns>
    public PropertiesUpdater<TTarget> AddProperty<TProperty>(
        Expression<Func<TTarget, TProperty>> targetPropertyExp,
        Expression<Func<TProperty>> valueExp,
        bool setInitialValue = false )
    {
        ...
    }

    public PropertiesUpdater<TTarget> RemoveProperty<TProperty>(
        Expression<Func<TTarget, TProperty>> targetPropertyExp )
    {
         ...
    }

    public void Dispose()
    {
        ...
    }
}

I will present some usage examples to show the sort of advantage that PropertiesUpdater provides.  The restaurant POS system I am now working on includes a Table domain object that represents the sort of table you sit at.  Table’s OpenOrder property is bound in the user-interface and so its private setter raises a property-changed event for notifying the UI of changes.  Table ’s constructor contains the following code for determining OpenOrder’s value:

new PropertiesUpdater<Table>( this )
    .AddProperty(
        t => t.OpenOrder,
        () => DineInParts.Select( p => p.Owner ).SingleOrDefault( o => o != null && o.IsOpen )
    );

PropertiesUpdater recognizes most LINQ-to-Objects methods and can interpret the lambdas passed to them.  After calling AddProperty in the above example, each of the following causes the target Table.OpenOrder property to be updated on this: (1) changes to the Table.DineInParts collection property’s value on this, (2) adds/removes of the collection instance’s elements, and (3) changes to values of the Owner and IsOpen properties referenced via collection elements.  Under the covers, PropertiesUpdater relies on the referenced objects implementing INotifyPropertyChanged and the referenced collections implementing INotifyCollectionChanged.

Next I will present two usages taken from the constructor of the order detail screen’s view-model in the restaurant POS system-

// Base the phone number and name on their sources in the domain model.
// Initialize these without causing the contact to be overwritten.
new PropertiesUpdater<OrderDetailsScreenPM>( this )
    .AddProperty(
        pm => pm.PhoneNumber,
        () => domainItem.Contact is VerifiedCustomer ? ((VerifiedCustomer)domainItem.Contact).PhoneNumber : null,
        setInitialValue: true
    ).AddProperty(
        pm => pm.Name,
        () => domainItem.Contact != null ? domainItem.Contact.Name : null,
        setInitialValue: true
    );

// Store a contact in the domain model that has the user-provided phone number and name.
new PropertiesUpdater<Order>( domainItem )
    .AddProperty(
        dm => dm.Contact,
        () => PhoneNumber != null
            // If a phone# is provided then lookup a verified contact; if not found then create one.
            ? (DataAccess.Instance.FindCustomer( PhoneNumber ) ?? new VerifiedCustomer { PhoneNumber = PhoneNumber, Name = Name })
            // If only a name is provided then create an unverified contact.
            : Name != null ? new Customer { Name = Name } : null
    );

The first thing you may notice about this example is that the “one line” permitted for the value expression is actually spanning multiple lines.  The “one line” limit just means the expression must not contain a semicolon.  The expression can in fact be however complex is needed.

In the value expressions for the first instantiated PropertiesUpdater, notice the presence of property paths that navigate across more than one property.  These are domainItem.Contact.PhoneNumber and domainItem.Contact.Name.  What if Contact's value changes?  The former value of Contact is no longer of interest (in particular, PhoneNumber's and Name’s values on the former contact), therefore the PropertiesUpdater stops observing the former contact.  This is a similar situation as when an item is removed from a collection in which case the removed element is no longer observed.  In the opposing case, PropertiesUpdater will begin observing a newly introduced value.

The Wikipedia article on Reactive programming provides a canonical example that is in fact PropertyUpdater’s primary purpose.  So, I decided to call the namespace for holding PropertiesUpdater and its supporting classes “Qnomad.Reactive”, at least until someone who knows better tells me that’s an inappropriate classification.  As for Microsoft’s Reactive Extensions (Rx), I didn’t see its connection to what PropertiesUpdater provides until I came across an interesting article that fills in the gap: Data binding in code using Reactive Extensions.  When I first saw the title, I worried my month-long work on PropertiesUpdater may have been for naught.  The single usage example that the article builds up to, however, falls far short of what PropertiesUpdater can accomplish.  The explained functionality reminds me of the basic building block with which PropertiesUpdater observes referenced objects, which resembles Josh Smith’s PropertyObserver.  Using Rx for this purpose, however, doesn’t seem necessary or even useful.  This leaves me wondering how “reactive” Rx really is; although, I suppose it may just be reactive in a different sense.  As with other CS concepts that become buzzwords, “reactive” is likely to go through a period of overuse before its actual meaning gets whittled down in the minds of the masses.  Instead of waiting for that to happen, though, I would love to hear some clarification sooner.

Realizing the significant time expense of implementing PropertiesUpdater about a week into it, I checked if something similar already exists.  I came across the WhenAny method of the ReactiveUI project, but it pales in comparison to what I ended up accomplishing with PropertiesUpdater.  It wasn’t until after completing PropertiesUpdater and describing it in the above paragraphs that I finally found three projects — Bindable LINQ, Obtics, and Continuous LINQ — that appear to accomplish a similar feat.  My initial impression, however, is that PropertiesUpdater can more easily accomplish the same with its mere two public methods and comparatively tiny implementation.  Anyway, I’ll further explore these two alternatives later, if there is occasion.

Wednesday, December 5, 2012

Collection accessors language feature

Collections are independent types from the domain objects on which they are used to define properties, yet they are usually closely tied to internal business logic of the single domain object that "owns" them.

What do you do in each of the following scenarios:
  1. You have business logic that determines whether or not an item can be added to the collection.
  2. You have some logic that needs to run before an item is added to the collection and another piece of logic that need run after the item was added.
These are the possible approaches that I can think of for tackling these scenarios:
  1. Put that logic in the view-model. The logic will need to be repeated everywhere the collection is modified.
  2. Expose the collection as IEnumerable then add Add/Remove methods directly onto the domain object where you put the check logic. This works but the public collection property can always be cast to ICollection or derivative and directly modified, thus violating the intended encapsulation.
  3. Create a custom collection.
But I do not feel satisfied with any of these options.  I always wished to see a language feature like the add/remove block that was introduced in C# for events but as being available for public collection properties as well.

    public collection IList Details
    {
       add { if (isMaxReached) throw new Exception("max reached"); details.Add(value); }
       remove { details.Remove(value); }
    }

All modifications to the collection would pass through these accessors.

Wednesday, April 13, 2011

Reusable Tree Operations for the .NET Framework

An earlier article of mine describes how a common ITreeNode<T> interface can facilitate reusable tree operations thanks to the introduction of extension methods. That article discusses reusable tree operations with respect to domain-models. Trees, however, are a ubiquitous data structure and also occur throughout the .NET Framework, as discussed in [Eberhardt 2010]. Ideally, the .NET Framework would include an ITreeNode<T> interface that's consistently implemented within the Framework, but this is not so. [Eberhardt 2010] describes a way to provide tree operations to those Framework types by means of the adapter pattern and code generation. His approach, however, has drawbacks. Wrapping a node instance every time it is navigated is likely to impact performance, and relying on code generation complicates maintenance. Upon looking into the matter for myself, I came up with an approach that suffers neither of these drawbacks.

Before continuing onto my solution, I want to note what tree operations, specifically, should be available. Leveraging the functionality already provided by LINQ to Objects reduces this list considerably. For example, instead of numerous methods that deal with ancestors, such as IsAncestorOf and FindAncestorOfType<T>, a single Ancestors method can be provided with LINQ to Objects taking care of the rest.
//item.IsAncestorOf( newParent )
newParent.Ancestors().Contains( item )

//tvi = this.FindAncestorOfType<TreeViewItem>();
tvi = this.Ancestors().OfType<TreeViewItem>().First();
LINQ to XML sets a good example by borrowing its "axis methods" — that query ancestors, siblings, descendants — from XPath. [Eberhardt 2010] does an excellent job of describing these.

Now I will show my solution for adapting .NET Framework tree types in order to make common tree operations available on them. The code that I will be discussing is available as part of Qnomad's CoreHelpers library. To create my solution, I applied a variation of the adapter pattern. Instead of instantiating a wrapper for every tree node, there's just one wrapper per adapted tree type, e.g. DependencyObject. The library also includes a singleton of this sort for the ITreeNode<T> type discussed in my earlier article. The tree operations, then, use the singletons to navigate the hierarchy. The various pieces of this puzzle are outlined below.
  • internal static class TreeOperations: Contains an implementation for each tree operation. Each method signature is of the form:
    internal static IEnumerable<T> Operation<T>( T startNode, ITreeNodeAdapter<T> adapter )
  • internal interface ITreeNodeAdapter<T>
    {
    bool HasSubItems( T treeNode );
    IEnumerable<T> GetSubItems( T treeNode );
    T GetParent( T treeNode );
    }
  • Next come the implementations of ITreeNodeAdapter<T>. The library contains these two:
    • internal class WpfTreeNodeAdapter : ITreeNodeAdapter<DependencyObject>
    • internal class ITreeNodeAdapterImpl<T> : ITreeNodeAdapter<ITreeNode<T>>
You've undoubtedly noticed that everything so far has been declared as internal. The public methods are all on the corresponding "Extensions" classes. The library's WpfExtentions class, for example, contains this method:
  public static IEnumerable<DependencyObject> Ancestors( this DependencyObject startNode )
{
return TreeOperations.Ancestors( startNode, WpfTreeNodeAdapter.Instance );
}
Well, that pretty much wraps it up. So now you see a better way for exposing common tree operations on .NET Framework tree types.

Monday, March 21, 2011

Advantages of the Domain Model

Prior to arriving at my previous consulting assignment, the company had already completed their first WPF project. The architecture they used on that earlier project included a rather unusual, in my opinion, hybrid object/relational in-memory representation of business objects. Some of the business objects were wrapped in View Models, while the View was directly referencing ADO.NET data rows of other business objects. I initially kept an open mind, programming their new app to that earlier architecture. But I eventually realized that it provided me no way to abstract away from the underlying relational structures. I then spent the weekend refactoring the new app, and building its Data Access layer. Having broken away from the status quo established by their first WPF project, however, my efforts were not greeted warmly. I had to spend a subsequent weekend producing the below “Advantages of the Domain Model”, as well as the Microsoft Embraces the Domain Model article I posted earlier, for convincing them to go with a purely object-oriented in-memory representation of business objects.

Advantages of the Domain Model


Confining object/relational mapping logic to a clearly demarcated data access layer (DAL) within your application, facilitates the creation of a single, coherent, object-oriented Domain Model to represent your application’s business objects. Adherence to this practice leads to more robust, scalable, and maintainable software.

Separation of concerns


View
View Model
Domain Model
Data Access
Data Services
Separation of concerns is a fundamental principle within computer science. As applied to an application’s overall architecture, it provides for organizing a complex software system into separate, but interdependent, parts. The new app’s codebase is easier to understand and maintain than the old app, thanks to the separated Data Access, Domain Model, and View Model layers. The responsibilities of all three layers are lumped into one layer within the old app.
  • Conceptual model. The Domain Model provides a concise representation of the domain of interest. This can aid in communication with experts in the problem domain who are not necessarily developers.

  • Division of labor and organization of code. Need to retrieve or persist data? Then call into the DataAccess singleton. Need to execute some business logic when a condition’s selected field changes? Then go to the Domain Model and put the logic into the Condition.SelectedField setter. Need to tell the View how to display itself differently? Then go to the View Model.

  • Code Reuse. Both the old and new apps contain object-relational mapping logic. The difference is that the new app’s Data Access layer, by way of generalized functions, eliminates the needless repetition found within the old app’s numerous Load and Save methods.

  • Consistent interface to in-memory data. Hiding the relational model specifics, as is accomplished by the new app’s Data Access layer, provides a consistent representation of business objects. The View layer in the old app, on the other hand, contains bindings that reference both .NET properties as well as database fields.

  • Maintainability. Because of the separation of responsibility, the new app can be modified more easily when business needs change. Changing a database field’s name, for example, requires a limited maintenance effort in the new app. Since data access is coded concisely in the Data Access layer, it involves altering just one string.

Increased Productivity


Coding directly to ADO.NET objects, as the old app does, means that you cannot easily navigate the relationships between business objects. The new app’s Data Access layer, on the other hand, shields upper layers from database specifics. This allows for more natural modeling of data, and simpler handling of complex relationships between entities.
  • The business objects are strongly typed, so we benefit from compile-time type-checking.

  • Business objects’ members expose themselves through Intellisense, eliminating typos and problems with remembering property names.

  • We can navigate the object-oriented Domain Model by using the OO-dot syntax, such as condition.SelectedField.Type.AvailableOperators. That is, we just choose an object instance to be our point-of-entry, then navigate from there. The ADO.NET objects in the old app, on the other hand, don’t allow you to easily navigate the data model.

Thursday, January 20, 2011

WPF Dialog Service via Binding

View Model logic sometimes involves showing a modal dialog box. The dilemma has always been: how do you show the dialog while abiding by the rule that the View Model not reference the View? A common solution has become to create a dialog service class, then extract its interface for the View Model to consume. In my opinion, using the interface is only slightly better than using the actual class. It still looks to me like the View Model is referencing the View. Why attempt such a charade when the View Model's primary means of communication with the View is perfectly adequate?

Although the View Model must not reference View classes, the View Model can still talk to the View. If there wasn't such a capability, then displaying the View Model's properties to the user would be impossible. The communication method I am referring to, of course, is data binding. Data binding is already the primary means of communication between the View and the View Model. We can use it for showing a dialog box as well.

Showing a dialog box using the solution I have created involves two easy steps:
  1. Add the MessageBoxHelper.Message attached property into the View.
  2. Within the View Model, set the property specified in the first step when you want the dialog box to be shown.
The View Model's property mentioned in these steps is of type DialogPM. This class contains no View-specific logic. It's purpose, rather, is to transport parameters for the MessageBox.Show(... method call from the View Model to the View and, if that method's result is important, back from the View to the View Model.

Download the Qnomad CoreHelpers library to see the complete example.

Sunday, January 2, 2011

Microsoft Embraces the Domain Model

Soon after I introduced a Domain Model into the application I was creating for my employer, I found myself at a team meeting trying to explain to fellow programmers, and my manager at the time in particular, why such a change is beneficial. I started by describing how, by the nature of data flowing from the relational database to the object-oriented runtime environment and back, we have an inherent need to translate between two different data models. The manager stopped me in my tracks when she said that I sounded “too academic”. Although left speechless at the moment, in retrospect (and at the risk of sounding like a smart ass) I should have responded, “the law of gravity may sound academic when described by a physicist, but that alters neither its truth nor practical consequence.” Just as the laws of physics describe our physical world, the world our applications live in can be described in terms of data models. The problem stemming from the presence of two data models, instead of just one, is called the Object-Relational Impedance Mismatch.

The Object-Relational Impedance Mismatch is an “academic” (to quote my former manager) sounding term for a practical problem facing software application development.
…work directly against the relational structure of database tables.

This is all well and good for simple databases and simple applications. However, as a database gets richer and an application more sophisticated, we see increasing divergence between the way an application needs to look at the data—its “conceptual” or “object” model—and the way information is structured in the database—the “storage” or “relational” model.” A manufacturer's ordering system, for example, may store its order information in multiple related tables, yet the application programmer really wants to work with a single, conceptual “order” entity without having to do a complex JOIN in every order-related query. (This is a case of what’s called the “object-relational impedance mismatch,” if you’ve heard that term.)
The above quote is taken from an article in the MSDN Library titled Microsoft Data Development Technologies: Past, Present, and Future by Kraig Brockschmidt and Lorenz Prem, Program Managers at Microsoft.

Microsoft released the ADO.NET Entity Framework in August 2008 as part of the .NET Framework version 3.5 SP1, and then an improved version in the .NET Framework 4 released in April 2010. What is the ADO.NET Entity Framework? From its overview:
The ADO.NET Entity Framework enables developers to create data access applications by programming against a conceptual application model instead of programming directly against a relational storage schema. The goal is to decrease the amount of code and maintenance required for data-oriented applications.
The “conceptual application model” is commonly known as the Domain Model.

With the Entity Framework finally here, it is clear that Microsoft has at long last embraced the Domain Model pattern. This is significant given that, in the 12 year hiatus since Microsoft's last data-access paradigm shift when it released ADO in October 1996, they’ve been careful to find the new emblem of application architecture for years to come. In general, Microsoft is the most immune to “fads”. They tend to do things their own way, and avoid industry standards up until the standards become so engrained that the obvious choice becomes to embrace them or get left behind. Microsoft’s embrace of the Domain Model is evidence, in my eyes, that this has become an engrained industry standard.

For those of us using the .NET Framework to build applications, Microsoft’s commitment to the Domain Model goes beyond confirming an industry standard. Having Visual Studio 2010 installed on our machines, while relying on it so much for development, we also have access to the extensive tooling support Microsoft has built around the pattern. From the MSDN library: “conceptual and relational representations can be easily created using designers in Visual Studio, and the Entity Framework designer in Visual Studio will create a default mapping without any effort on your part.” It’s a safe bet that this tooling support will be continually improved with future Visual Studio releases.

Wednesday, September 16, 2009

Listening to HTML Events from BHO

Creating a plug-in for Internet Explorer was an interesting proposition. After years of .NET development, the first point I investigated was how to write my add-in using C#. I was surprised to find no official support from Microsoft for .NET development of IE add-ins. Thanks to COM Interop, however, it's still possible. More so, information and sample code abound for making it happen.

A number of Code Project articles have been written about how to setup your add-in project using COM Interop. But I wanted to get straight to writing C# code and seeing its effects within IE. In other words, I wanted an existing library that did the COM Interop stuff for me, and which I could simply reference from my project. Two such solutions exist: one commercial and one open source. The commercial one is Add-in Express for IE. I was always suspicious of their tool since a free trial isn't available, so I never shelled out the $149 (USD). The free open-source alternative known as SpicIE, however, is good enough.

My plug-in needs to be kept in sync with whatever text is currently selected within the browser window. Listening to HTML events raised on the browser's current page is necessary in order to do this. The SpicIE Contrib companion website includes a bare bones sample showing how to do this. This sample, however, hooks up the event handlers only after the entire page has loaded. I was looking for something a bit more interactive than that. In other words, I want to hook up the event handlers ideally as soon as the user is able to interact with page elements. This turned out to be a non-trivial task.

In response to the above problem, I developed a reusable class called HtmlPageEventManager whose purpose is to subscribe to a given list of HTML events for each new webpage that's opened. A goal is to attach handlers as soon as the user can begin interacting with page elements, even before the page load is complete. Using this class is simple -- just call the constructor:
var evts = new List<HtmlEvent>() {
HtmlEvent.onclick, HtmlEvent.ondblclick, HtmlEvent.onkeydown,
HtmlEvent.onselectstart, HtmlEvent.onselectionchange,
HtmlEvent.onfocusin
};
new HtmlPageEventManager( this, evts, this.HtmlEventHandler );
You can download my solution here to try it out. And feel free to use it in your own projects if you find it useful.

I encountered some issues while creating HtmlPageEventManager.
  • As mentioned above, I want to add the HTML event handlers as soon as possible, i.e. even before the page load is complete. Doing so in the NavigateComplete handler seemed to work great, until I tested opening a hyperlink in a new tab or window. The new tab or window often times didn't have the HTML event handlers functioning correctly. I could see from debugging that IHTMLDocument3.attachEvent was indeed called on the correct document or element instance and returned 'true' indicating handlers were attached, but the handlers were thereafter never invoked. It seemed that the document/element wasn't "ready" to have events attached, although that didn't concur with the 'true' return value.
    Via the debugger, I discovered that from within OnNavigateComplete, I can detect the problem situation by looking at the IWebBrowser2.Busy property, and adjust my strategy concerning when to add the handlers. So now the registrations occur in OnNavigateComplete only when browser.Busy is true. And if they haven't been registered by the time OnDocumentComplete is reached, then they get registered there instead.

  • When you "Refresh" a webpage, the event registrations were always being lost. That is, they weren't being reconnected on the new page elements. I found out that the NavigateComplete and DocumentComplete handles aren’t called on a "Refresh". When I tested it earlier, it seemed like they were called. But what I observed must have been those two handlers called for frames within the main document – not the main document itself. I overcame this issue by subscribing to the BHO's DownloadComplete event, where I again call my "RegisterEventHandlers" helper method that attaches the HTML event handlers. I incorporated the idea from this code of a "normalPageLoad" member variable to conditionally call RegisterEventHandlers.