DacFx Public Model Tutorial



Recently there has been an increasing interest in extending the capabilities of SSDT and DacFx. The walkthrough guides for creating new build and deployment contributors and database unit test conditions are a useful start in exploring the tools, but they only scratch the surface of what’s possible. They also don’t really show the best practices for developers when building and debugging extensions such as deployment contributors. In this article we’ll fix that by covering the key concepts behind DacFx extensions, solve real customer issues and highlight best practices. All of the code in this tutorial is available at https://github.com/Microsoft/DACExtensions.

What is the public model?

The key to most extensibility is the public model API. Dacpacs and SSDT projects both model a database’s schema. The public model API lets you access that model programmatically. You can load, query and manipulate the schema to do whatever you’d like. Most scenarios will rely on some level of querying the model and examining the objects that describe the database.

The public model API is loosely typed: the TSqlModel contains loosely typed TSqlObjects that represent all the elements in your schema. Each object will have some Properties that describe its state, and Relationships to other objects in the model. Whether the object you’re looking at is a Table, View, Index or anything else, they’re all represented by the one TSqlObject class.

Of course if everything is a TSqlObject, how can you tell Tables and Views apart? How can you even know what properties and relationships a Table has? That’s where the strongly-typed metadata classes come in. The majority of classes in the model API are actually metadata classes – you’ll see Table, View, etc. Each class has a number of fields that list the Properties and Relationships for that type of object. To lookup Tables in the model you pass in the Table.TypeClass to GetObjects, and then only tables are returned. To get the Columns for a table, you ask for relationships and pass in the Columns relationship class. If all this seems complicated, the code examples should make it clearer. The important thing to note is that you’ll pass in these metadata descriptions whenever you query the model.

Scenario: using the public model API

OK, let’s get started with using the public model! We’ll show the basics of loading, reading, adding to and saving the model. All the code in this example is in the SampleConsoleAppModelEndToEnd.cs sample, and can be run by specifying “RunEndToEnd” when running the SampleConsoleApp application.

Loading a model

Loading a model is really simple – either point to the location of an existing Dacpac, or create an empty model and add scripts to it.


All the database options that you can specify in an SSDT project can be defined using TSqlModelOptions when creating a new model.

The samples show how to copy options from an existing model to a new model if you need to do this.

When adding objects to the model there are certain properties such as how they treat ANSI nulls and QuotedIdentifiers that can be defined.

Reading a Table, its properties and relationships

Reading top level types such as Tables and Views is easy, as shown below. Top level types are any type that could be defined independently in TSQL – Tables, Views, but also things like a Primary Key Constraint since this can be specified in an ALTER TABLE statement.


To examine something like a specific Column for a table, you need to first look up the relevant table and then get referenced columns. Note how the Table.Columns metadata relationship is used to find columns for the table.


The example below shows how the Column.Length metadata property is used to get the length of a column.

If you know the return type for a given property you can use generics to cast to that type. In the example below “Length” is cast to an int. Properties usually have simple return types, such as int, bool, string. Some int properties actually map to Enum values – for example DataCompressionOption.CompressionLevel maps to the CompressionLevel enumeration – and you can cast directly to that Enum type when getting the property. Note that if the property is not found on that object, a default value for that may be returned instead.


Finally, a very small number of types in the model have actual “Metadata” properties. These are useful when a type can actually represent conceptually similar things, where each has different properties. A Column can be a regular column, a computed column or a ColumnSet, and what properties are relevant for the column will vary depending on the ColumnType.


DacQueryScopes can be quite important. It specifies what kind of objects you want to search for. Depending on the scope you pass in, different types of objects can be returned:

What are you looking for?Correct query scope
The objects you defined in this dacpacUserDefined, All
Built in types (for example SQL data types like nvarchar)BuiltIn, Default, All
Referenced objects added using composite projects in SSDT (“Same Database” references)SameDatabase , All
System objects from master.dacpacSystem, All

You may notice that “different database” references aren’t on this list. That’s because they’re not really useful for anything other than validating the model, and you can never have a TSqlObject that describes them. The only time you’ll get to see any information about them is when querying what types of things an object references, and there’s a special call with an external query scope that’ll include some information about them.


GetReferenced is only one of several methods to traverse relationships in the model, depending on the type of relationship. See the example below showing how a Table relates to an index and a column differently:

To simplify this a little in the public model, we added GetChildren and GetParent methods. In this case you shouldn’t need to understand which object has a reference to the other, or what the relationship is. It will just return all the objects that are logical children of a Table:


Some relationships have properties associated with them. For instance the relationship between a table constraint and the columns that it refers to has an Ascending property. These properties are queryable using ModelRelationshipInstance.GetProperty<T>.


Saving a dacpac

The public model supports building dacpacs and even updating the model inside an existing dacpac. Unfortunately the API does not fully support everything that an SSDT project supports. This may change in the future, but for now the feature support is as follows:

Refactor logYes
Deployment contributorsYes
Pre / Post deployment scriptNo
CLR objectsNo
XML Schema CollectionNo


Scenario: filtering developer schemas

A real example raised in the forums was how to filter out objects for specific schemas. For example a user may have a “dev” or “test” schema that is populated with some data used during testing. However these should never be deployed to the production environment. The question is, how can you achieve this without using separate projects for the “dev” and “test” schema elements? Two general solutions come to mind here, each with different benefits and drawbacks. We’ll outline both approaches and show the key code required to solve this problem. For full code examples we recommend going to the samples solution and debugging into the sample application and unit tests. That’s really the best way to learn what’s going on here.

Our sample data

Here’s the sample data we’ll use for this scenario. It’s very simple – just a few schemas, tables and views we want to work with. Our goal is to start with a dacpac that includes all of these schema objects, and ensure that what’s deployed to a database only includes objects in the “prod” schema.

Solution 1: Filtering the model and creating a new dacpac

The first solution assumes that whenever you build your project, you’d like to output two dacpacs: a “production” dacpac that doesn’t contain the “dev” or “test” schemas, and a “dev” dacpac that contains all objects. “production” would be used when deploying to a production database, while the “dev” dacpac is used during development.

Let’s look at the key steps required to do this. All the code in this example is in the SampleConsoleAppModelFilterExample.cs sample, and can be run by specifying “FilterModel” when running the SampleConsoleApp application. There are also unit tests for this in the “SampleTestsTestFiltering.cs” file.

Filtering the model and building a new dacpac

Let’s create a simple “IFilter” interface that takes in a set of TSqlObjects and performs some action. We’ll write a schema filter and apply it to all the objects in our model, then save it to a dacpac. The basic process is as follows:

And the filter works by examining the first part of the TSqlObject.Name property. ObjectIdentifiers describe the name. The internal part of the name always starts with the schema. Even the name describes a reference to an external object (for example to master DB or a different database) the external parts of the name are in a separate property. This makes it easy to write a schema-based filter. Here’s a simplified version (again look at the sample files for a fully fleshed out example):

Finally, there’s the CreateFilteredModel method that reads all objects from the current model and copies only objects that pass the filter into a new model:


The schema name comparison currently uses a simple string comparison. Ideally it would compare based on the SQL Database Collation for the model by using SqlString objects for comparison. This is the kind of feature we may add in future releases, but you could also write this yourself fairly easily.

Updating the model in the existing dacpac

The API also supports updating the model inside and existing dacpac. This might be useful if you have other resources such as pre and post deployment scripts inside a dacpac. The public API doesn’t have support for including these when building a dacpac yet, so the best solution would be to copy the dacpac file and then update the model inside it. There’s a unit test in TestFiltering.cs that shows how this is done. The API call is really simple:

Deploying the filtered Dacpac

Deploying a Dacpac is really simple using the DacServices API. DacServices supports publishing Dacpacs, creating Dacpacs from a database, and a number of other useful features. To actually deploy our filtered Dacpac to production (or in this example, to localdb) we’d just do as follows:

Solution 2: Filtering at deployment time

So filtering objects in a dacpac is one option, but what if you want to avoid the need to create a new dacpac? Isn’t there a way to just change things when you’re actually deploying the dacpac? That’s exactly what we’ll show you next by creating a custom Deployment Plan Modifier contributor that runs during the deployment pipeline. These are covered in a separate walkthrough but this example will show you how to specify the contributors to run at deployment time rather than when building a project.

As usual the full code for this example is in the samples. To see how this example works look at the SamplesTestsTestFiltering.cs unit test class. The “TestFilterPlanWhenPublishing” unit test runs this end to end. In this case a unit test was used since it avoided the need to install the sample to the extensions directory before running the sample code (see Best Practices for more information).

Creating a filtering deployment contributor class

A basic contributor class just requires an Export attribute and to extend the DeploymentPlanModifier class. Here’s a “Hello World” contributor and how to add it to the deployment:

Note that this doesn’t cover actual installation of the contributor DLL – that’s covered under the Best Practices section later in the document.

Reading and filtering the deployment plan steps

During deployment a number of different objects are available to a contributor. In this case the Deployment Plan is the most interesting thing. It describes each step in the deployment, and contributors can add new steps and remove or replace existing steps. For this example, what we need is to block any CreateElementSteps that mention the schemas to be filtered. Understanding what step you need to examine might not be immediately obvious – in this case you could probably guess, but sometimes the best thing to do would be writing a dummy contributor that steps through a plan and then debugging a deployment, or writing the step type and contents to a file. That lets you understand the precise types to work with.

Here’s the code that actually filters out steps. We’re reusing the filter code from the 1st scenario since the logic is all the same. The only difference is that each step has only 1 object, so we’ll apply our filter and if no objects are left afterwards, we know that the step should be removed.


We’ve skipped a number of steps here, most importantly how the filter is actually initialized. It’s fairly simple code and if you debug through the example you’ll see exactly how this works

While writing this example, we added an “Initialize” method to the IFilter interface. This doesn’t look right on an interface, so in a real-world example we’d probably change this to be an abstract “Filter” class with an empty default implementation of the Initialize method, or simply use a Factory pattern for creating the filters instead.

Configuring the contributor at deployment time, and running the deployment

Now that we’ve written the sample contributor, let’s see how it would be used during deployment:

And that’s that! Now you have the ability to filter by schema when deploying a dacpac.

Building on this example

Follow up scenarios you could try for yourself:

ScenarioHint (how to do it)
Extract a dacpac from a database and filter out some objects. For example filter out all Users and Logins so that later you could replace them with new ones.Use the DacServices API to extract the dacpac, then run the ModelFilterer on it with a new “FilterObjectType” filter
Implement a more relaxed “Block on Table Loss” function instead of the current “Block on possible Data Loss”.

This is another real-world example, a team wanted to allow columns to be dropped, but wanted to block the deployment if tables were removed.

Write a DeploymentPlanModifier contributor that looked at the ModelComparisonResult in the deployment context, and block if there are any tables in the list of elements to be dropped. If there are, block deployment by publishing an error message with severity “Error”.
Implement a more relaxed “Drop Objects not in source” option that doesn’t drop elements in a “reserved” schema.

Another real world example.

Note that you will need a recent release of DacFx to make this work as there was a bug in the previous version.

In addition to filtering CreateElementSteps, you would also filter DropElementSteps and AlterElementSteps if they relate to the reserved schema.

Learnings and best practices

Testing deployment contributors

The walkthrough guides discuss how to install a contributor so that your Visual Studio projects can make use of them. That’s great in a way, but it’s really not what you want to use during testing. It’s too cumbersome to copy the DLL each time you run it, and if you actually open Visual Studio to test it, you’ll need to shut it down every time you want to change your contributor code.

The best way to test contributors is to write unit tests and reference your contributor DLL and the DacFx DLLs. To be picked up during deployment the contributor code must be in a DLL file (not and executable), and that must either be under the standard DacFx extensions directory on your machine or else be in the same directory as the “Microsoft.Data.Tools.Schema.Sql.dll” file. If you are writing unit tests, the 2nd option has one really powerful benefit. Unit tests usually copy all referenced DLLs to the same location and that means that if your unit test references the DacFx DLLs and your contributor DLL, you can easily run tests without needing to copy the contributor code into the extensions directory. When the test is run, both will be in the same location and hence the DacFx extension manager will find your contributor.

The deployment plan filtering example uses this approach and it makes it really easy to make changes to the contributor and verify that everything works.


Extending DacFx can help solve common issues that your team runs into. Extensions can be really powerful– the APIs are intended to let you do everything our tools can do internally. We’re not quite there yet, but we’ll be updating the current APIs and adding new ones in the future, so stay tuned!

Hopefully after reading this tutorial you’ll take a chance to think about an issue you’ve had that SSDT/DacFx doesn’t solve for you right now, and if you could solve it yourself. If you’d like to share your solution with others, think about publishing it online or adding it to the samples project https://dacsamples.codeplex.com/.

Kevin Cunnane

Follow Kevin   


  • Avatar
    Damon New

    Kevin – How would you go about “Solution 1: Filtering the model and creating a new dacpac”, when you need to filter a dacpac which contains data? Applying dacpac.UpdateModel(filteredModel, metadata) throws the exception: Microsoft.SqlServer.Dac.DacServicesException: ‘Cannot update the model as the package contains data. This is not supported since it could cause errors during deployment’.

    My goal is to produce a production-only .dacpac from a similar database schema (which contains dev, test, and production objects), so the dacpac can be delivered to a customer for import into an on-prem SQL Server. Seems that the only way to exclude non-production objects is by filtering at deployment time, but you place this responsibility on the client. If they don’t filter properly, they end up getting all your dev/test objects.

Leave a comment