Version control instructions

Git information is in a separate document.

The environment

Databases that you should need for development:

Even though our test run will run modifications to the test databases, don't check in these *.mdb and *.db files with your commit, to avoid bad merge-cases.

Solution structure

We use Fake and Paket. You have to run build.cmd on Windows (or sh ./build.sh on Mac/Linux) before opening the solutions.

The main source solution is SQLProvider.sln. The unit tests are located in another one, SQLProvider.Tests.sln, and when you open the solution, it will lock the bin\FSharp.Data.SqlProvider.dll, and after that you can't build the main solution.

  • To debug design-time features you "Attach to process" the main solution debugger to another instance of Visual Studio running the tests solution.
  • To debug runtime you attach it to e.g. fsi.exe and run the code in interactive.

Workarounds for "file in use" (issue #172)

  • Debugging execution: Have all the test-files closed in your test-project when you open it with VS. Then you can run tests from the tests from Tests Explorer window (and even debug them if you open the files to that instance of VS from src\SqlProvider).
  • Or you can open tests with some other editor than Visual Studio 2015

Referenced Files

  • Documentation is located at SQLProvider/docs/content/core and it's converted directly to *.html help files by the build-script.
  • src/ProvidedTypes.fsi, src/ProvidedTypes.fs and src/Code/ExpressionOptimizer.fs are coming from other repositoried restored by first build. Location is at packet.dependencies. Don't edit them manually.

Test cases

There are database specific test files as scripts in the test solution, /tests/SqlProvider.Tests/scripts/, but also one generic /tests/SqlProvider.Tests/QueryTests.fs which is running all the SQLite tests in the build script.

High level description of the provider

Context and design time

You have a source code like:

type sql = SqlDataProvider<...params...>
let dc = sql.GetDataContext()

What will first happen in the design-time, is that this will call createTypes of SqlDesignTime.fs and create (as lazily as possible) the database schema types (the shape of the database). These methods are added to the sql.datacontext and are stored to concurrent dictionaries. Visual Studio will do a lot of background processing so thread-safety is important here.

GetDataContext() will return a dynamic class called dataContext which will on design-time call class SqlDataContext in file SqlRuntime.DataContext.fs through interface ISqlDataContext. SqlDataContext uses ProviderBuilder to create database specific providers, fairly well documented ISqlProvider in file SqlRuntime.Common.fs.


The entity-items themselves are rows in the database data and they are modelled as dynamic sub-classes of SqlEntity, base-class in file SqlRuntime.Common.fs which can be basically think of as wrapper for Dictionary<string,obj> (a column name, and the value). SqlEntity is used for all-kind of result-data actually, so the data columns may not correspond to the actual data values. Mostly the results of the data are shaped as SqlQueryable<SqlEntity>, or SqlQueryable<'T> which is a SQLProvider's class for IQueryable<'T> items.

query {
    for cust in dbc.Main.Customers do
    where ("ALFKI" = cust.CustomerId)
    select cust
} |> Seq.toArray

This query is translated to a LINQ-expression-tree through Microsoft.FSharp.Linq.QueryFSharpBuilder. That will call IQueryable<'T>'s member Provider to execute two things for the LINQ-expression-tree: first CreateQuery and later Execute.

Parsing the LINQ-expression-tree

CreateQuery will hit our SqlQueryable<...>'s Provider (IQueryProvider) property. LINQ-expression-trees can be kind of recursive type structures, so we it will call CreateQuery for each linq-method. We get the expression-tree as parameter, and parse that with (multi-layer-) active patterns.

Our example the LINQ-expression tree is:

.Call System.Linq.Queryable.Where(
    '(.Lambda #Lambda1<System.Func`2[SqlEntity,Boolean]>))
.Lambda #Lambda1<System.Func`2[SqlEntity,Boolean]>(SqlEntity $cust) {
    .Call $cust.GetColumn("CustomerID") == "ALFKI"

so it would hit this in SqlRuntime.Linq.fs:

| MethodCall(None, (MethodWithName "Where" as meth), [ SourceWithQueryData source; OptionalQuote qual ]) ->

because the LINQ-expression-tree has ExpressionType.Call named "Where" with source of IWithSqlService (which is the SqlQueryable). What happens then is parsing of the Where-query. Where-queries are nested structures having known conditions (modelled with pattern Condition). If the conditions are having SqlColumnGets, a pattern that says that it's SqlEntity with method GetColumn, we know that it has to be part of SQL-clause.

We collect all the known patterns to IWithSqlServices field SqlExpression, being a type SqlExp, our non-complete known recursive model-tree of SQL clauses.

Execution of the query

Eventually there also comes the call executeQuery (or executeQueryScalar for SQL-queries that will return a single value like count), either by enumeration of our IQueryable or at the end of LINQ-expression-tree. That will call QueryExpressionTransformer.convertExpression. What happens there (in SqlRuntime.Linq.fs):

  • We create a projection-lambda. This is described in detail below.
  • We convert our SqlExp to real SQL-clause with QueryExpressionTransformer.convertExpression calling provider's GenerateQueryText-method. Each provider may have some differences in their SQL-syntax.
  • We gather the results as IEnumerable<SqlEntity> (or a single return value like count).
  • We execute the projection-lambda to the results.

In our example the whole cust object was selected. For security reasons we don't do SELECT * but we actually list the columns that are there at compile time.

The TupleIndex of IWithSqlService is a way to collect joined tables to match the sql-aliasses, here the [cust].

SELECT [cust].[Address] as 'Address', 
       [cust].[City] as 'City',
       [cust].[CompanyName] as 'CompanyName',
     [cust].[ContactName] as 'ContactName',
     [cust].[ContactTitle] as 'ContactTitle',
     [cust].[Country] as 'Country',
     [cust].[CustomerID] as 'CustomerID',
     [cust].[Fax] as 'Fax',
     [cust].[Phone] as 'Phone',
     [cust].[PostalCode] as 'PostalCode',
     [cust].[Region] as 'Region' 
FROM main.Customers as [cust] 
WHERE (([cust].[CustomerID]= @param1))

-- params @param1 - "ALFKI";


Now, if the select-clause would have been complex:

query {
    for emp in dc.Main.Employees do
    select (emp.BirthDate.DayOfYear + 3)
} |> Seq.toArray

We don't know the function of DayOfYear for each different SQL-providers (Oracle/MSSQL/Odbc/...), but we still want tihs code to work. The LINQ-expression-tree for this query is:

.Call System.Linq.Queryable.Select(
    '(.Lambda #Lambda1<System.Func`2[SqlEntity,Int32]>))
.Lambda #Lambda1<System.Func`2[SqlEntity,System.Int32]>(SqlEntity $emp) {
    (.Call $emp.GetColumn("BirthDate")).DayOfYear + 3

What happens now, is that in SqlRuntime.QueryExpression.fs we parse the whole LINQ-expression-tree, and find the parts that we do know to belong to SQL: the SqlEntity's emp.GetColumn("BirthDate"), and create a lambda-expression where this is replaced with a parameter:

fun empBirthDate -> empBirthDate.DayOfYear + 3

Now when we get the empBirthDate from the SQL result, we can execute this lambda for the parameter, in .NET-side, not SQL, and then we get the correct result. This is done with for e in results -> projector.DynamicInvoke(e) in SqlRuntime.Linq.fs.

Other things to know

  • SQLProvider also runs ExpressionOptimizer functions to simplify the LINQ-expression-trees
  • If you do IN-query (LINQ-Contains) to IEnumerable, it's as normal IN-query, but if the source collection is SqlQueryable, then the query is serialized as a nested query, where we have to check that the parameter names won't collide (i.e. param1 can be used only once).

    This documentation was written on 2017-04-11.

Fork me on GitHub