LINQ to SQL: The Mapping Engine

The primary purpose of any ORM system is to map relational data onto objects in your programming environment.  Mapping here refers to the meaning in the mathematical sense that there is a correspondence from one item to the other.  For example, a database row might map to an object, or a field in a database might map to a property. Some mappings are simple, like the ones I’ve already mentioned; others are more complex such as parts of multiple rows combining to form a single object.

LINQ to SQL has a variety of mapping capabilities some simple and some edging toward complex. Other ORM systems have additional mappings too, though none ever seem to fully satisfy the somewhat inexhaustive list of mappings that domain experts imagine. 

Mappings in an ORM system are very similar to the view system employed by databases, yet are often described as data for a mapping engine rather than as query in-and-of itself, which is likely why they never seem to cover all the possible cases. On the surface, views through queries seem much more powerful since you can basically write code to transform one set of information into another. A declarative model is more constrained, but that’s not necessarily a bad thing.  It keeps you from expressing mappings that have no inverse, which would be unfortunate since after mapping relational data out of database and into objects you often want to map it back. LINQ to SQL is no different. It allows you to describe mapping using attributes placed on your objects or via an XML ‘map’ file you can author and load at runtime.

Yet, even though the API only allows you to declare mapping in this constrained fashion, the actual mapping engine used by LINQ to SQL operates closer to the view model. The information you specify in the mapping is actually turned into a view query over the tables in your database. The query you were trying to execute actually ends stuck on top of this view query, similarly to if you were querying against an actual database view.

Let’s take a look at how this works.

var q = from c in db.Customers

        where c.City == "London"

        select c;

This simple query is actually the following API calls underneath:

var q = db.Customers
          .Where(c => c.City == "London")
          .Select(c => c);

Translated into a first approximation SQL query:

FROM db.Customers AS c
WHERE c.City = ‘London’

Of course, ‘db.Customers’ is a database table so we must translate it and the references to it into something that SQL can understand. Here’s where the mapping comes in.  In this case, our mapping between the database and the customer class is basically one-to-one so we can rewrite the expression ‘db.Customers’ with a query that constructs an object out of the underlying columns using a hybrid C# & SQL syntax.

   SELECT new Customer {
       CustomerID = t.CustomerID,
       CustomerName = t.CustomerName, …
   FROM [Customers] AS t
   ) AS c
WHERE c.City = ‘London’

Of course, the server cannot actually construct an object instance and send it back to us (at least not yet) so we must reduce this query further into something it can understand.  We do this by extracting out the portion of the projection that cannot be understood and leaving the parts that can be understood as a flat tuple of columns.  

new Customer {
    CustomerID = Column(“CustomerID”),
    ContanctName = Column(“ContactName”), …

SELECT c.CustomerID, c.ContactName, …
FROM [Customers] AS c
WHERE c.City = ‘London’

Now, it easy to see that the remaining query is what is executed against the server and the projection becomes the rule for turning the tabular results of the query into actual objects.

The mapping engine took the declared mapping and turned it into a query expression that produces the desired objects.  This query is just like the ‘view’ used by a database, except it’s expressed in terms of constructing CLR objects and not relational tables.

A slightly more complicated query shows the truth of this a little better.  For instance, a query can project any shape it wants, not just one of the mapped objects.

var q = from c in db.Customers

        where c.City == "London"

        select new { c.ContactName, c.Phone };

Jumping right to the SQL we see:

SELECT new { ContactName = c.ContactName, Phone = c.Phone }
    SELECT new { CustomerID = t.CustomerID … }
    FROM db.Customers AS t
    ) AS c  
WHERE c.City = ‘London’

Which reduces to:

SELECT new { ContactName = c.ContactName, Phone = c.Phone }
FROM [Customers] AS c
WHERE c.City = ‘London’

And then to:

new {
    ContanctName = Column(“ContactName”),
    Phone = Column(“Phone”);

SELECT c.ContactName, c.Phone
FROM [Customers] AS c
WHERE c.City = ‘London’

It turns out, representing the mapping as an actual query expression is very useful. Because mappings are expressed in the exact same form you would expression your own projection, all projections compose nicely on top of any mapping. Also, since both mappings and projections are expressed using a single language, the top-level projections extracted from the overall query become more than just descriptions of the objects you want, they in fact are the actual code that is used to construct them.


Hopefully, you can start to see how other more complicated mappings such as inheritance would be represented. 

Given a hierarchy such as:  Person -> Contact -> Customer

var q = from p in db.People

        select p;

The expression ‘db.People’ is mapped into a query that constructs one of three different types.

     CASE p.Type
       WHEN 1 THEN new Person {Name = p.Name}
       WHEN 2 THEN new Contact {Name = p.Name, Phone = p.Phone}
       WHEN 3 THEN new Customer {Name= p.Name, Company = p.Company }
FROM [People] AS p

Which becomes:

CASE Column(“Type”)
  WHEN 1 THEN new Person {Name = Column(“Name”)}
  WHEN 2 THEN new Contact {Name = Column(“Name”), Phone = Column(“Phone”)}
  WHEN 3 THEN new Customer {Name = Column(“Name”), Company = Column(“Company”) }

SELECT p.Type, p.Name, p.Phone, p.Company
FROM [People] AS p

Mapping columns from multiple tables into a single object becomes downright trivial.

SELECT new Customer {
        CustomerID = c.CustomerID,
        ContactName = c.ContactName,
        Address = ca.Address,
        City = ca.City, …
FROM [Customers] AS c
INNER JOIN [CustomerExtended] AS ca ON c.CustomerID = ca.CustomerID

Of course the opposite is true as well; mapping a single table into multiple objects is just as easy.

SELECT new Customer {
      Customer ID = c.CustomerID,
      ContactName = c.ContactName,
      Address = new Address {
             Street = c.Address,
             City = c.City, …
FROM [Customers] AS c

Because the LINQ to SQL mapping engine uses query expressions to express the mapping, just about any mapping imaginable is possible through the query translator.  In fact, any query that you could write to produce the output you want could be represented within the mapping engine.

Aren’t you just completely geeking-out about now?  I know I am. J

Of course, that might lead you to question why LINQ to SQL does not support some of these more advanced mappings, if it’s so easy.  The truth is that even though the mapping engine and query translator can understand all of these, the test burden to prove that they all work as expected, with the correct plumbing through the change tracker and update processor, not to mention the complication added to the mapping user model (attributes and xml file) and designer was simply too large to swallow in a single release. 

Instead, we took a conservative approach and engineered the most common and useful mappings as a starting point, allowing the system to stay simple and easy to understand by most people. 

And yes, we do want your feedback on what should come next.  Please help us budget our time for the next version by either leaving comments here or on the MSDN forum.

Comments (14)
  1. publicus says:

    If you make the code modify the data, and the data is the code, can you set it up to evolve sentience?

  2. Daniel says:

    How is this different from Entity Framework explained in the paper ?

  3. Daniel, it is very nearly identical. But that’s not all too surprising. 🙂

  4. publicus, yes indeed I have already achieved sentience. That happend a few weeks ago.  It woke up and started talking to me.  At first it was fun, and then it just got annoying.  I already have a six year old, I didn’t need another pestering creature that can ask questions faster than I can think. So I figured I needed to give it something to do to keep busy while I actually got my work done.  I told it to write posts form me on this blog and to answer any questions in the comments. It’s so life like I bet you couldn’t even tell.

  5. Pop Catalin says:

    So far so good, but does LINQ support any other data manipulation commands than select queries? I mean updates, inserts and deletes that don’t happen through classes?

    Without those it would be very awkward to use LINQ in scenarios where lot’s of data manipulation is happening, and the funny thing is that’s what LINQ is for.

    I love C# but I don’t want to trade the power of SQL with setting properties on classes and calling commit. Anyway LINQ could potentially one of the best things that will happen this year, Congratulations!

  6. Alexander Gornik says:

    The only real thing that LINQ to SQL feature set is missing is better inheritance mapping. Especially, class table inheritance pattern ( )

  7. Daniel says:

    Thanks Matt. But if they are nearly identical, why two ORM framework? How should we choose EF or LINQ to SQL?

  8. Alexander, I agree with you.  We trimmed down the first version to only the one kind of inheritance mapping because based on our customer survey it was the most commonly used.

  9. Daniel, the official word is if you want/need to use the addition mappings provided by EF or need to talk to other database servers (not SQL server) in the Orcas timeframe (in v1 of L2S and EF) then you should use EF.  

    The biggest difference between L2S and EF, is that EF is based on the existence of a realized conceptual model at runtime that allows you to work with entended entity-relational model data w/o mapping all the way into objects. Where-as with L2S this conceptual model is only conceptual and exists as a by-product of the mapping.

    Theoretically, there can be an additional mapping between the runtime conceptual model and objects in L2E.  Though, I don’t think that will exist in v1. If it did, you might find reason to want it. For example, the same reason you might want to guard against changes in the database schema by using an external mapping file, you might want to guard against changes between conceptual and object model. Yet, that might be a rare need.

  10. Hi Matt,

    Allowing mapping to cope with an inheritance hierarchy (or other kind of complex mapping) is a critical feature (IMHO). I hope you don’t leave it for too long! The need for it is most evident on large projects where there are multiple groups of designers (for database and object domains) who start with a single requirements model and manipulate it for the medium they are targetting.

    Because of the frequent need to denormalise the data design during optimisation, data designers often diverge from the kind of domain model that object oriented designers will produce. Having one to one mappings in LINQ to SQL will be a real handicap in systems like those.


  11. The plan for LINQ to SQL is to grow the capabilities over time.  I am aware that the set of features in LINQ to SQL won’t satisfy some developers when it first releases, so I expect the community to direct us where to focus next. I do feel, however, that the feature set we have right now has reached a critical mass and will be a compelling offering for a large body of customers as is.

  12. So long as it gets released, I’ll be happy! 😉

  13. Some quick links about LINQ: Articles about extension methods by the Visual Basic team Third-party LINQ

  14. Some quick links about LINQ: Articles about extension methods by the Visual Basic team Third-party LINQ

Comments are closed.

Skip to main content