Roger Alsing Weblog

 THIS IS AN OLD POST FROM MY OLD BLOG (2006) 

AOP is gaining momentum in .NET and there are starting to pop up quite a few AOP frameworks.
A comparison of (some of) those can be found here:
List of existing approaches

As you can see there are a few different approaches to accomplish AOP in .net
the ones that I know are:

• ContextBoundObject - Inherit ContextBound
• Source transformation - Extra compile step
• IL instrumentation - Extra compile step
• .NET profiling API - Modify the Jitted IL
• Runtime Subclass proxies - Reflection emit subclasses
• Runtime Interface proxies - Reflection emit proxies with interfaces

These approaches can be divided into two categories:

• Inheritance based (Runtime proxies)
• Code mangling (all the others)

(Ok ContextBound might be a separate category but since its too limited Ill just ignore it ;-) )

So what is the difference between these two approaches?

Code mangling can be very powerful and can modify your code in ways that is not supported by inheritance.
e.g. you could add interception to non virtual members, you could intercept field access and you can change base class of other classes.

Runtime proxies can only intercept ctors, virtual members or interface members.
So runtime proxies are in most cases considered to be the less of the two.
But here is my take on it:

If we take a look at good’ol OO, we have constructs like private, sealed/final, virtual, override etc.
Those constructs all have a purpose, to let the author of a class specify things like:

“It’s all OK to alter the behaviour of this method and things should work all OK if you do because its virtual”

Or:

“You may not alter this method because nasty things might happen”
(e.g. sealed or private etc)

No one has ever questioned those constructs in OOP.
So why should different rules apply to AOP?

the goal is the same as in inheritance, to alter or extend behaviour / functionality of a class.

So in my opinion inheritance based AOP is the way to go because its 100% OO compatible.
It doesn’t break the above rules.

While you might be able to do cool hacks with code mangling, you are also opening up for a whole lot of problems.
e.g., it requires much more of a consumer of a virtual method than a consumer of a non virtual method.

Why?

Because when you consume a non virtual method you know exactly what might happen, you know what exceptions that might occur and you know what the expected behaviour is.
When consuming a virtual method, you don’t know what might happen because you might call the method on your base class or on a subclass and the subclass could have been provided from someone else.
so when you consume the virtual method you might have to add some extra exception handling or make your code a bit more generic because you can’t be certain that things behave exactly as when you call the base implementation.

So if AOP would allow you to break the fundamental OO rules, you could alter the behaviour of a non virtual method and introduce behaviours that is not expected by the consumer.
and I think that is VERY bad.
If the author of the consuming code wrote his code knowing that it was a non virtual method he called, he did not prepare his code for anything else, and the AOP’ed variant of the provider method might screw things up big time.

And since the entire framework is designed w/o AOP in mind, I bet there is plenty of code in both your applications and in the framework itself that simply expects those non virtual methods to behave as they did when the code was written.

So in short, I don’t think AOP should be allowed to break the fundamental OO rules because most of the code you use is designed for those rules.
(However I think its a completely different story with things like AspectJ where the language itself is designed for AOP and those who use it know that every kind of method might be intercepted.)
The next thing is debugging.

Code mangling can accomplish lots of weird stuff like changing the base class of a type or introduce fields.
That might seem like a cool feature but since we all write our code in .NET languages that is NOT designed for AOP, e.g. C# or VB.NET
How are we supposed to debug such code?

The code executing is not the same code you have in your source files.
For me, that is just madness.

In runtime proxies we do not have those problems because the proxies are just dumb objects redirecting your calls to your interceptors, mixins and base methods.
so you will never ever need to debug the proxies themselves since they just redirect the calls.
and your base class and interceptors are still normal .NET code which you can debug just fine.
However there is a scenario where I do agree it is nice to be able to intercept non virtuals.
Debug aspects.
e.g. tracing, code coverage, interaction tests and various mock aspects.
Such aspects is a special case for me since they do pretty much the same as your IDE debugger.
in the debugger you can step into a private method, you can view all private variables etc.
Debug aspects does the same but in a different way.
Well that’s it :-)

And as stated in the beginning, this is just my take on it.
I’m not a fanatic so I could convert if someone just provides me some valid arguments why breaking OO rules is good in AOP :-)

//Roger

 THIS IS AN OLD POST FROM MY OLD BLOG (2006)

 Most O/R Mappers today support some sort of Query Objects.
Query Objects are often some sort of semi fluent combination of objects and methods where you can build a query.
 
Query objects also seem to be the preferred way to build queries in code because they are “Typed”.
Being “Typed” comes with two major benefits, you get intellisense (to some extent) and you get compiler warnings if you do something way too funky with your query.
 
When Mats Helander started to build NPersist, he went for a DSL approach instead of QO’s and invented NPath which object query language similar to SQL in syntax.
 
So why go for an untyped DSL instead of typed QO?
 
Well, one thing that QO’s does not do is to verify your logic, you need unit tests for that.
And once you’ve got unit tests, the benefit of being strictly typed is reduced because with unit tests you can clearly see if your untyped code does work or not.
 
QO’s still got the benefit of intellisense,
But let’s look at the readability of two queries:
 
(the queries come from : http://www.mygenerationsoftware.com/phpbb2/viewtopic.php?t=107)
 
SQL:

SELECT * FROM Employees
WHERE LastName LIKE 'A%' AND LastName IS NOT NULL OR
( City LIKE 'Ind%' OR HireDate BETWEEN '1/1/95' AND '4/4/04' )

[edit: added npath sample]

NPath:

SELECT * FROM Employee
WHERE LastName LIKE 'A%' AND LastName != NULL OR
( City LIKE 'Ind%' OR HireDate BETWEEN #1/1/95# AND #4/4/04# )

NOTE: npath specifies classes and properties , not tables and fields
See: *** Missing Page *** : Old blog is gone
[/edit]
 

dOOdas ORM:

Dim emps As Employees = New Employees
emps.Where.LastName.Value = "A%"
emps.Where.LastName.Operator = WhereParameter.Operand.Like
Dim wp As WhereParameter = emps.Where.TearOff.LastName
wp.Operator = WhereParameter.Operand.IsNotNull
emps.Query.AddConjunction WhereParameter.Conj.Or
emps.Query.OpenParenthesis()
emps.Where.City.Conjuction = WhereParameter.Conj.And
emps.Where.City.Value = "Ind%"
emps.Where.City.Operator = WhereParameter.Operand.Like
emps.Where.HireDate.Conjuction = WhereParameter.Conj.Or
emps.Where.HireDate.Operator = WhereParameter.Operand.Between
emps.Where.HireDate.BetweenBeginValue = "1/1/95"
emps.Where.HireDate.BetweenEndValue = "4/4/04"
emps.Query.CloseParenthesis()

Are QO’s really sane?
If the whole purpose of an O/R Mapper is to reduce code and enable RAD, shouldn’t it be easier to query your DB?
I was a LLBL gen pro user a few years ago and loved most parts of it, but I just couldn’t stand the QO’s.
Just making a query with a few conditions and some sorting would generate queries similar to the one above.
Even if Frans has excellent documentation and the classes got intellisense and all, I still had to spend way too much time
just to make a simple query that should take a few seconds to write.
(I have to add that I haven’t looked at LLBL Gen since and Frans might have altered the API or added some DSL, I’m just stating what I experienced with it about 2 years ago.
my arguments are also targeted against QO’s in general, not only the ones in LLBL Gen)

Later on I got in contact with Mats and saw NPath for the first time, and I loved it, NPath is the #1 reason I teamed up with Mats and started co-develop NPersist.
With NPath I can write queries that are similar to SQL but instead of targeting your DB schema you target your domain classes and properties, and instead of creating joins you just traverse property paths.

So even if NPath is untyped I still find it way more productive than any QO’s and I get a smaller code base which is easy to read and since I do unit tests I can verify that my untyped queries does work as intended.
It would be very interesting to hear other opinions on this topic and why QO’s still is the most common way to make queries in O/R Mappers?
Is it because of the intellisense?
or because most people don’t do unit tests?
or just because it looks cool with some huge object graph?
or because most O/R vendors don’t know how to make a decent parser?

Fire away! :-)

[edit]
Just added this sample incase anyone is interested in NPath and in mem querying:
http://web.archive.org/web/20070103130842/http://blogs.wdevs.com/phirephly/archive/2006/02/03/12096.aspx
[/edit]

THIS IS AN OLD POST FROM MY OLD BLOG (2006)

There are quite a few AOP frameworks for .NET out there today.
But all (?) of them suffer from the same problem, its hard or impossible to debug them.

I myself would pick debugging over AOP any day, so forcing our users to debug via coredbg or such tool was not an option when we started to build NAspect.

NAspect supports an intact callstack when stepping from your consumer code into your interceptors and finally into your real provider method.
Here is a few screenshots:

Place a breakpoint in your consumer code:

Single step into the method, notice how we ended up in our first interceptor instead of the provider method:

Single step into the proceed method, now we arrive at the real provider method.
Note that the call stack is intact and shows all steps from your consumer to your interceptor to your provier method.

This is as far as I know not possible with any other AOP framework for .NET today.

Another problem area when dealing with AOP is to know what interceptors that actually got attached to your method.
You can ofcourse step through the entire call flow and check the callstack, but that is pretty messy if you got a few hundred methods with a few interceptors each.

NAspect support VS.NET 2005 DebuggerVisualizers.
Instead of trying to figure out if your config file works or single step through every method to see the interceptors, we can simply enter debug mode and visualize the AOP’ed instance:

On the left side of the visualizer you get a list of all adviced methods in your type,
On the right side you get a list of interceptors for the currently selected method.

The next step will be to build a debugger visualizer that shows an UML graph of the proxied instance, showing what mixins and baseclass it relates to.

//Roger

Console.WriteLine (”Hello World”);