One useful feature in WPF 4.0 is the ability to databind to dynamic (runtime generated) properties using the DynamicObject as a base class or implementing the IDynamicMetaObjectProvider interface. I am currently working on a WPF application and this ability to bind to runtime generated properties would have been very useful in a previous story we implemented two weeks ago.

The story concerned merging two object graphs and then visualizing what properties that were changed/conflicted in the UI (for example with a different color).

In order to not add “XXX_HasMergeChange” for every property in the presentation model we solved this by using a value converter and some WPF binding magic that some might call a HACK. The solution was only partial as it only worked in the Grid and not on everything else.

If we had WPF 4.0 we could have solved this like this:

In the above style trigger the data trigger is binding to a property that doesn’t exist on the presentation model. How does WPF then get the value for this property? By calling the TryGetMember method:

WPF will call the TryGetMember and that will check if the property ends with “_HasMergeChange”, if that is the case it will try to look up the property in the MergeChanges dictionary.

The above is just a simple proof of concept, if I would go forward with this I would have to figure out a more generic way to define the style and data trigger to be able to reuse the xaml style markup for example, but that shouldn’t be a big problem. I also tested property change notifications using the INotifyPropertyChanged interface and they work for dynamic properties as well.

To learn more about the new features in WPF 4.0 read ScottGu’s recent post.

Among the Linq extensions methods that came with .NET 3.5 is one called Except. This method takes two lists (first and second)

The MSDN docs say:

This method returns those elements in first that do not appear in second. It does not also return those elements in second that do not appear in first.

This appears to be a lie. Review the code below and guess the output:

The User object has overridden the Equals and GetHashCode methods which the Except method use to determine equality. Since there is only one object in list2 that is also “equal” to an object in list1 I would expect (granted the MSDN docs are correct) that list3 would contain three users with the id zero.

The actual result? list3 will only contain ONE user object (with id zero). When I debug I see that Equals is called to compare objects in list1 with each other.

Using reflector I can see why, Except is implemented using the internal class System.Linq.Set:

It starts by adding all items from list2 into the set class then for each item from list1 that can be added to the set it yield returns. This filters all items from list1 that are equal to an item in list2 BUT it also filters all items in list1 that are equal any other item in list1!

Maybe not such a common usage scenario, and I don’t recommend overriding Equals and GetHashCode in this manner. Anyway frustrated by this because I burnt an hour on debugging before figured out it that what was to blame. 

NHibernate has pretty good support for batching, something that can significantly increase performance when inserting or updating large number of objects.

Example:

In the above example you can see that the order lines are created in one statement. In a recent mail conversation with Patrik Löwendahl he asked for assistance in getting batching to work. The first thing to check is what id generator you are using, you cannot use native (sql identity) id generator and expect batching to work for inserts. The reason for this that for identity inserts NHibernate issues a "select SCOPE_IDENTITY()" statement after each insert statement to fetch the generated ID. If you want to use batching for inserts you need to use the guid or hilo id generator.

Another issue i came across was that batching does not work as you would hope for associations. For example if you want to save a thousands orders and each order has five order lines, this would result in six thousands calls if batching was disabled and two thousands calls with batching enabled. As you see in the screenshot above, batching is only done on the order lines and not for everything.

You can optimize further by using the stateless session. However inserting entities using nhibernate’s stateless session ignores associations. But by looping through all orders and calling session.Insert(order), and then doing a nested loop to do the same for all order lines you can insert all orders and order lines in just two calls to the database.

The problem Patrik had was very weird and confusing. To verify that batching was actually happening he used SQL Profiler, while I used NHProfiler. The weird thing is that they show a very different picture. NHProfiler shows order lines as being issued in one command while SQL Profiler shows them as separate RPC calls.

+

This result left me very confused. The NHProfiler result clearly indicates that batching is being used but SQL Profiler shows the same results as when batching is off. However when batching is enabled the performance is significantly better, what is going on here?? After some Googling on SQL Profiler and batching I found this comment on stackoverflow:

On MS SQL Server, SQL Profiler shows each insert statement seems to be on it's own. After reviewing your comment, I viewed a TCP Dump of the conversation and do see that it is batching multiple commands together. SQL Profiler shows each insert as a "RPC Completed" event which was confusing me. Thanks for your help.

I appears that batching IS being done but not like I thought it would be (for example a Batch Starting command in SQL Profiler). The difference, when batching is turned on, is that all the statements are sent to the database in one go without waiting to listen for a response. That explains the SQL Profiler result, however I still find the NHProfiler result puzzling as it indicates that the order lines are created using a single call to sp_executesql.

Ayende care to explain? :)

Some links on NHibernate batching:

• Bulk Data Operations With NHibernate’s Stateless Sessions
• NHibernate Perf Tricks

There is a new release of NHibernate available, download it now. It contains a host of great new features, like support for Dependency Injection for entities using an inversion of control container of your choosing. There is also a new ANTLR based HQL parser that has allowed for some HQL improvements, like the with clause.  The new ANTLR based HQL parser is also central to the forthcoming LINQ support and is the result of some great work by  Steve Strong and Fabio.

This release also includes the long sought after support for executable bulk queries. This is a feature that the java (Hibernate) version has had for some time and is now fully ported to NHibernate.

For a complete list of new features: link

I have been doing more and more talks lately. I am not a natural speaker I usually need to practice a few times before in order to talk more fluently. But practice has made me more comfortable with it and I feel that I am getting better at it. Two days ago I held a long (3.5 hour) talk on ASP.NET MVC which was both my longest talk and my most successful, at least considering the positive response a got, which was very encouraging.

The talk was mostly a long code demo. One thing that can often kill code demos are that they can slow down the tempo of a presentation when there is too much typing of unimportant text/code. Like creating a new class, constructor, etc and then later getting to the really important part of a particular function.

In this MVC code demo I tried to have as much prepared as possible. I started with a standard MVC template project, but had hidden (excluded from the project) controllers and views which I included as the code demo progressed. These controllers/views included some existing functionality which I then expanded upon. That way I did not need to type class definitions defining simple controller actions and views before getting to the interesting bits.

I did the same with many other parts of the presentation, for example when explaining how to unit test controller actions I already had an almost empty test method already written, and only needed to show how to unit test the controller actions and how to assert on the result.

I also had code snippets in the toolbox for some of the tricky parts of the code demo that I could use if something did not work or I felt that it took to long to write. Never let a problem in the code demo completely halt the presentation, have a backup plan or just move a long if you cannot fix the problem on the first 2 tries.

I feel that I still have much to work on when it comes to presentation technique. I often talk a little to fast, need to focus on keeping calm and talking in a slow and articulate manner.  It doesn’t matter how nice your powerpoint or code demo is if the audience can’t hear what you are saying!

Next up is trying keynote, nice to see how it compares to powerpoint.

I listened to the panel discussion on the pros and cons of stored procedures from the currently ongoing TechEd09 today. It was not what I hoped for, the panel consisted almost exclusively of pro stored procedure people with the exception of Jeffrey Palermo who for an NHibernate guy appeared very pro stored procedure.

I was hoping for a more balanced debate. The arguments were to much focused on the real vs. perceived benefits of stored procedures in terms of performance, database coupling, vendor coupling, security etc.

The really big issues I have personally with stored procedures (sprocs from now on) were never fully brought up. Especially when you compare sprocs and a manually coded DAL (which I find is the most common) with NHibernate.

Code duplication
In my experience systems which rely heavily on sprocs also show a large amount of code duplication, for example duplication of SQL queries in the same sprocs in the case of dynamic queries that filter on different columns based on input. I have seen sprocs that feature the same basic query duplicated 12 times with small variation in the where/order clause. Also duplication between different sprocs can usually be very high. And the final point is the sad fact that sprocs usually contain some business logic, logic that sometimes also exist in the application itself.

Productivity & Lines of code
This topic was also not really touched upon. Data access layers which use sprocs often feature massively more amount of code to deal with calling the sprocs and mapping them to entities. The amount of TSQL you need to write for the basic CRUD sprocs is also a huge time waster and possible maintenance nightmare.

Some of these issues could be argued that it is just incompetent programmers/DBAs and sprocs are not to blame. Maybe it is not fair to compare sprocs with an ORM like NHibernate. But I think you can compare having to write and maintain sprocs compared to letting NHibernate generate adhoc SQL. Sure I think sprocs still have their usage in specific and relatively rare scenarios but the panel discussion too often concluded on the wishy washy "it depends". Of course it depends, context is everything (as Scott Bellware always says), but that does not mean that one method shouldn't be the preferred "best practice" choice.

Sorry for the rant. Kind of frustrated with a current legacy system (which uses sprocs) :)

The picture to the right shows a hand that illustrates how the cross product of two vectors (a and b) generate a vector that is perpendicular to both a and b. Now try to imagine four vectors that are all perpendicular to each other. This is kind of tricky, mainly because it is impossible for four 3-dimensional vectors to all be perpendicular to each other.

But it is not impossible if the vectors are 4-dimensional, however that creates another problem: it is (at least for me) not possible to mentally picture a 4-dimensional space. Luckily I don’t have to the math works anyway :) The reason I post this is that I am porting an old 4D Julia Raytracer from C++ to C# and was just struck by this magic function:

/// <summary>
/// Calculates a quaternion that is perpendicular to three other quaternions. 
/// Quaternions are handled as vectors in 4d space
/// </summary>
public static Quaternion Cross4D(Quaternion q1,Quaternion q2,Quaternion q3)
{
  double b1c4=q2.r*q3.k-q2.k*q3.r;
  double b1c2=q2.r*q3.i-q2.i*q3.r;
  double b1c3=q2.r*q3.j-q2.j*q3.r;
  double b2c3=q2.i*q3.j-q2.j*q3.i;
  double b2c4=q2.i*q3.k-q2.k*q3.i;
  double b3c4=q2.j*q3.k-q2.k*q3.j;

  var r = -q1.i*b3c4+q1.j*b2c4-q1.k*b2c3;
  var i =  q1.r*b3c4-q1.j*b1c4+q1.k*b1c3;
  var j = -q1.r*b2c4+q1.i*b1c4-q1.k*b1c2;
  var k =  q1.r*b2c3-q1.i*b1c3+q1.j*b1c2;

  return new Quaternion(r, i, j, k);
}

This is a function that calculates a cross product for a 4-dimensional vector (the Quaternion class is used as a 4-dimensional vector in imaginary space). The reason for the naming of variables in the calculation relates to how the cross product formula is derived (as the determinant of a matrix). Anyway, I just found it funny that no matter how hard I try I cannot picture what this function actually generates. This is probably nothing new for mathematicians or physicist who I guess daily has to fight against the limitations of the human mind.

But the math works, I can position the camera in 4D space and render pictures of the 4-dimensional Julia Set :)

On a side note this app was MUCH easier to parallelize (using Parallel.For from the Parallel Extensions Library) than GenArt. Because the algorithm works like a raytracer the outer ray casting loop is easily implemented using Parallel.For which instantly gave a 4x performance increase on my quad core CPU.

I took some time to upload some of the old animations to youtube, they were rendered many years ago using the C++ version.

Here is an animation of a camera move around the Julia set, the camera is moving in the second (i) and fourth (k) imaginary dimension.

 

Here is another one, which I really like, the Julia constant is moving in a small circle in the first and second dimension. It gives me a strange impression of something organic and fluid. When I and my friend presented this rendering technique the first slide had this animation in a loop :)

You can see an interesting artifact of the rendering algorithm in the video above. The Julia Set actually hits the camera plane. The camera has a near plane where we start traversing the rays and a far plane where we stop, what happens is that the middle expands beyond the near plane, creating a flat surface.

Here are some other rendered animations that I uploaded to youtube:

• 4D Julia Fractal - 4D Camera Move 2
• 4D Julia Fractal - Constant Walk
• 4D Julia Fractal - Rotation of standard Julia