The Dual Schema Problem

A few months ago, Ted Neward wrote a great article about the history of the Object Relational Impedance Mismatch problem and how LINQ is addressing it in a new way. Basically, LINQ is introducing new language abstractions and complementary libraries to enable queries as a first class concept within the language. However, I don’t believe that O/R Impedance Mismatch is the whole problem. More specifically, it’s a follow-on problem to what I would call the Dual Schema problem.

In a nutshell, the Dual Schema problem is that you have to design and implement two separate versions of your persistent entities. There’s the in memory version, typically written in an OO language like C# or Java. Then there’s the on disk version, typically written in SQL. Regardless of the difficulties translating between the two versions (i.e. the aforementioned impedance mismatch), you have to first deal with the complexity of keeping the two versions in sync. While LINQ does a great job eliminating much of the friction translating between on disk and in memory formats, it could go much farther by eliminating the need for translation in the first place.

A variety of solutions to the Dual Schema problem have evolved, primarily outside the hallowed halls of enterprise vendors (i.e. MS and others like us). One such solution is Ruby on Rails. In a Rails environment, I simply declare the existence of a given persistent entity:

class Person < ActiveRecord::Base
end

The ActiveRecord base class (a standard part of Rails) will dynamically create methods and attributes on the Person object at runtime, based on the schema of the People table in the database. (Rails is smart enough to understand English plurals, hence the automatic connection of Person and People.) So technically there are still two schemas, but the in-memory version is automatically derived of the on-disk version.

(Note, DLinq provides a conceptually similar tool – SqlMetal – that can generate the static types from a given database schema. However, as static types they have to be defined at compile time. So while SqlMetal reduces the effort to keep schemas in sync, it doesn’t eliminate it the way Rails does.)

By slaving the object schema to the database schema, Rails essentially solves the Dual Schema problem. The problem with the Rails approach is that defining a database schema requires a significant amount of skill and effort. Defining classes is typically trivial in comparison.The fact Rails allows you to implement a persistent entity with almost no code doesn’t help you much if you have to write and maintain a ton of SQL code to define your database schema.

I believe the Rails model is actually backwards. It would be much better for the developer if they could define their persistent entity in code and slave the database schema to the object model instead of the other way around.

Of course, this approach isn’t exactly news. In his article, Ted writes of the rise and fall of OO database management systems, which were supposed to solve the Dual Schema and Impedance Mismatch problems. I’m certainly not suggesting a return to the heyday of OODBMS. However, one of the reasons Ted points out OODBMS failed was because big companies were already wedded to RDBMS. But those big companies are the short head. As you move down the long tail of software, relational database as the primary storage paradigm makes less and less sense. For the vast majority of applications, relational databases are overkill.

Ted’s other point about OODBMS is that loose coupling between the data store and the in memory representation is a feature, not a flaw. He’s totally right. But can’t we advance the state of the art in database typing to the level of modern day OO languages? How about eliminating anachronisms like fixed length strings? What if we derive the database schema from the object model – Rails in reverse if you will – but is still loosely coupled enough to allow for schema evolution?

An example of this code-centric model for data storage is Consus. It’s written by Konstantin Knizhnik, who has written a bunch of open source, object-oriented and object-relational databases across a wide variety of languages and execution environments, including CLR. Consus is actually written in Java, but he provides version compiled for .NET using Visual J#. Consus lets you to define your data either as tables or objects. So you can do this:

Statement st = db.createStatement();
st.executeUpdate("create table Person (name string, address string, salary bigint)");
st.executeUpdate("insert into Person values ('John Smith', '1 Guildhall St.', 75000)");
ResultSet rs = st.executeQuery(
    "select name, address, salary from Person where salary > 100000");

Or you can do this:

class Person {
    String name;
    String address;
    long salary;
    Person(String aName, long aSalary, String aAddress) {
        name = aName;
        salary = aSalary;
        address = aAddress;
    }
};

Person p = new Person("John Smith", 75000, "1 Guildhall St.");
ConsusStatement st = db.createStatement();
stmt.insert(p);
ConsusResultSet cursor = (ConsusResultSet)st.executeQuery(
    "select from Person where salary > 100000");

Consus also handles OO concepts like derivation and containment. Of course, the embedded queries are ugly, but you could imagine DLinq style support for Consus. In fact, one of the primary issues with Consus is that it supports both object and tuple style queries. When you explicitly request tables (i.e. “select name, address salary from Person”), you’ve got a tuple style query. When you don’t (i.e. “select from Person”) you’ve got an object style query. Of course, the issues with tuple style queries are well documented in Ted’s article and is exactly the problem that LINQ is designed to solve.

(Konstantin, if you’re reading this, drop me a line and I’ll look into getting you hooked up with the LINQ folks if you’re interested in adding LINQ support to Consus.NET.)

The tradeoff between the Rails approach and the Consus approach is one of performance. I have a ton of respect for Konstantin and the work he’s done on Consus and other OO and OR databases available from his site. However, I sure the combined developer forces at major database vendors like Microsoft (and other DB companies) means SQL Server (and the like) will out perform Consus by a significant margin, especially on large scale databases. So if execution performance is your primary criteria, the Ruby on Rails approach is better (leaving aside discussion of the Ruby runtime itself). However, in the long run execution performance is much less important than developer productivity. So I believe that  for all the current interest in Rails, I think a Consus-style model will become dominant.

Identity Woman

I’ve been to Mashup Camp, ETech and MIX in the past month and a half. So has Kaliya, otherwise known as Identity Woman. She calls MS out for a lack of power bars @ MIX06 as well as the lack of an official wiki for the event. She’s right on both counts. I thought some of the logistics of ETech were suspect, but they sure did assume that every single attendee would need power and laid out power strips accordingly.

I kept seeing Kaliya at these events, but I didn’t get to meet her until she came to the SPARK @ MIX session Monday afternoon. She’s one of the organizers of the Internet Identity Workshop. Sounds like a cool event, but it’s right before my daughter Rileyanne’s first birthday and you know Jules would NEVER forgive me for missing that.

CLR Everywhere

Big news from the Game Developers Conference this week is the XNA Framework. From the press release:

The XNA Framework contains a custom implementation of the Microsoft® .NET Framework and new game-development-specific libraries designed to help game developers more easily create cross-platform games on Windows® and Xbox 360 using the highly productive C# programming language. Using the XNA Framework, game developers will benefit from the ability to re-use code and game assets in developing multiplatform titles, without sacrificing performance or flexibility.

So now I can the Xbox to the list of personal technology I can program, joining my phone, my TV, my server and my laptop all with C# or VB (VB plug brought to you by Brian and Erik who think I should to switch). I wonder if XNA Framework will support WPF?

BTW, I still want an easy way to program my house and my car, but I’m sure those are coming.

Where else should the CLR live?

Overheard at MIX

“We would be happy if Firefox included WPF/e in their distro” and “There will be a version of WPF/e for Linux”

Overheard at SPARK

“Web 3.0 is Web 2.0 without the duct tape”