Session – WPF: Extensible BitmapEffects, Pixel Shaders, and WPF Graphics Futures

PDC 2008, Day #4, Session #4, 1 hr 15 mins

David Teitlebaum
Program Manager
WPF Team

My final session at PDC 2008 was a talk about the improvements in WPF graphics that are available in .NET Framework 3.5 SP1.  David also touched briefly some possible future features (i.e. that would appear in .NET Framework 4.0).

David’s main topic was to walk through the details of the new Shader Effects model, which replaces the old Bitmap Effects feature.

What are Bitmap Effects?

These are effects that are applied to an individual UI element, like a button, to create some desired visual effect.  This includes things like drop shadow, bevels and blur effects.


The BitmapEffect object was introduced in Framework 3.0 (the first WPF release).  But there were some problems with it, that led to now replacing it with Shader Effects in 3.5SP1.

Problems with BitmapEffect:

  • They were rendered in software
  • Blur operations were very slow
  • There were various limitations, including no ClearType support, no anisotripic filtering, etc.

New Shader Effects

Basic characteristics in the new Shader Effects include:

  • GPU accelerated
  • Have implemented hardware acceleration of the most popular bitmap effects
    • But did not implement outer glow
  • Can author custom hardware-accelerated bitmap effects using HLSL
  • There is a software-only fallback pipeline that is actually faster than the old Bitmap Effects
  • New Shader Effects run on most video cards
    • Require PixelShader 2.0, which is about 5 years old

How Do You Do Shader Effects?

Here’s an outline of how you use the new Shader Effect model:

  • Derive a custom class from the new ShaderEffect class (which derives from Effect)
  • You write your actual pixel shader code in HLSL, which is used for doing custom hardware-accelerated stuff using Direct3D
    • Language is C-like
    • Compiled to byte-code, consumed by video driver, runs on GPU
  • Some more details about HLSL, as used in WPF
    • DirectX 10 supports HLSL 4.0
    • WPF currently only supports Pixelshader 2.0

So what do pixel shaders really do?  They basically take in a texture (bitmap) as input, do some processing on each point, and return a revised texture as an output.

Basically, you have a main function that accepts the coordinates of the current single pixel to be mapped.  Your code then accesses the original input texture through a register, so it just uses the input parameter (X/Y coordinate) to index into the source texture.  It then does some processing on the pixel in question and returns a color value.  This resultant color value just represents—the resulting RGB color at the specified coordinate.

The final step is to create, in managed code, a class that derives from ShaderEffect and hook it up to the pixel shader code (e.g. file) that you wrote.  You can then apply your shader to any WPF UIElement using XAML.  (By setting the Effect property).

Direct3D Interop

David’s next topic was to talk a bit about interop’ing with Direct3D.  This just means that your WPF application can easily host Direct3D content by using a new class called D3DImage.

This was pretty cool.  David demoed displaying a Direct3D wireframe in the background (WPF 3D subsystem can’t do wireframes), with WPF GUI elements in the foreground, overlaying the background image.

The basic idea is that you create a Direct3D device in unmanaged code and then hook it to a new instance of a WPF D3DImage element, which you include in your visual hierarchy.

WPF Futures

Finally, David touched very briefly on some possible future features.  These are things that may show up in WPF 4.0 (.NET Framework 4.0), or possibly beyond that.

Some of the features likely included in WPF 4.0 include:

  • Increased graphical richness  (e.g. Pixelshader 3.0)
  • Offloading more work to the GPU
  • Better rendering quality
    • Integrate DirectWrite for text clarity
    • Layout rounding

And some of the possible post-4.0 features include:

  • Better exploitation of hardware
  • Vertex shaders
  • Shader groups
  • Shaders in WPF 3D
  • 3D improvements
  • Better media extensibility


You can get at David’s PDC08 slide deck for this talk here:

And you can find full video from the session at:

Session – Windows 7: Unlocking the GPU with Direct3D

PDC 2008, Day #4, Session #3, 1 hr 15 mins

Allison Klein

I jumped off the Azure track (starting to be a bit repetitive) and next went to a session focused on Direct3D.

Despite the title, this session really had nothing to do with Windows 7, other than the fact that it talked a  lot about Direct3D 11, which will be included in Windows 7 and available for Windows Vista.

Direct3D 10

Direct3D 10 is the currently shipping version, and supported by most (all?) modern video cards, as well as integrated graphics chips.  I’m not entirely sure, but I think that Direct3D 10 shipped out-of-the-box with Windows Vista.  It is also available for Windows XP.

Allison spent about half of the talk going through things that are different in Direct3D 10, as compared with Direct3D 9.

I’m not inclined to rehash all of the details.  (I’ll include a link to Allison’s slide deck when it is available).

The main takeaway was that it’s very much worth programming to the v10 API, as opposed to the v9 API.  Some of the reasons for this include:

  • Much more consistent behavior, across devices
  • Cleaner API
  • Elimination of large CAPS (device capability) matrix, for a more consistent experience across devices
  • Built-in driver that allows D3D10 to talk to D3D9 hardware
  • Addition of WARP 10 software rasterizer, to support devices that don’t support WDDM directly.  This is actually quite a bit faster than earlier software-only implementations

Direct3D 11

In the second half of her talk, Allison talked about the advances coming in Direct3D 11.  She mentioned that D3D11 will ship with Windows 7 and also be available for Windows Vista.

Again, the details are probably more appropriate for a game developer.  (See the slide deck).  But the high level points are:

  • Direct3D 11 is a strict superset of 10—there are no changes to existing 10 features
  • Better support for character authoring, for denser meshes and more detailed characters
  • Addition of tessellation to the rendering pipeline, for better performance and quality
  • Much more scalable multi-threading support.
    • Much more flexibility in what can be distributed across threads
  • Dynamically linking in custom shaders
  • Introduction of object-oriented features (interfaces/classes) to HLSL
  • New block compression
  • Direct3D11 will be available in the Nov 2008 SDK


Finally, Allison touched briefly on some future directions that the Direct3D is thinking about.

The main topic that she talked about here was in potentially using the GPU to perform highly parallel general purpose compute intensive tasks.  The developer would use HLSL to write a “compute shader”, which would then get sent to the GPU to do the work.  As an example, she talked about using this mechanism for post-processing of an image.

Session – Services Symposium: Enterprise Grade Cloud Applications

PDC 2008, Day #4, Session #2, 1 hr 30 mins

Eugenio Pace

My second session on Thursday was a continuation of the cloud services symposium from the first session.  There was a third part to the symposium, which I did not attend.

The presenter for this session, Eugenio, was not nearly as good a presenter as Gianpaolo from the previous session.  So it was a bit less dynamic, and harder to stay interested.

The session basically consisted of a single demo, which illustrated some of the possible solutions to the Identity, Monitoring, and Integration challenges mentioned in the previous session.


Eugenio pointed out the problems involved in authentication/authorization.  You don’t want to require the enterprise users to have a unique username/password combination for each service that they use.  And pushing out the enterprise (e.g. Active Directory) credential information to the third party service is not secure and creates a management nightmare.

The proposed solution is to use a central (federated) identity system to do the authentication and authorization.  This is the purpose of the Azure Access Control service.


The next part of the demo showed how Azure supports remote management, on the part of IT staff at an individual customer site, of their instance of your application.  The basic things that you can do remotely include:

  • Active real-time monitoring of application health
  • Trigger administrative actions, based on the current state

The end result (and goal) is that you have the same scope of control over your application as you’d have if it were on premises.

Application Integration

Finally, Eugenio did some demos related to “process integration”—allowing your service to be called from a legacy service or system.  This demo actually woke everyone up, because Eugenio brought an archaic green-screen AS400 system up in an emulator and proceeded to have it talk to his Azure service.


The conclusions were recommendations to both IT organizations and ISVs:

  • Enterprise IT organization
    • Don’t settle for sub-optimal solutions
    • Tap into the benefits of Software+Services
  • ISV
    • Don’t give them an excuse to reject your solution
    • Make use of better tools, frameworks, and services

Session – Services Symposium: Expanding Applications to the Cloud

PDC 2008, Day #4, Session #1, 1 hr 30 mins

Gianpaolo Carraro

As the last day of PDC starts, I’m down to four sessions to go.  I’ll continue doing a quick blog post on each session, where I share my notes, as well as some miscellaneous thoughts.

The Idea of a Symposium

Gianpaolo started out by explaining that they were finishing PDC by doing a pair of symposiums, each a series of three different sessions.  One symposium focused on parallel computing and the other on cloud-based services.  This particular session was the first in the set of three that addressed cloud services.

The idea of a symposium, explained Gianpaolo, is to take all of the various individual technologies and try to sort of fit the puzzle pieces together, providing a basic context.

The goal was also present some of the experience that Microsoft has gained in early usage of the Azure platform over the past 6-12 months.  He said that he himself has spent the last 6-12 months using the new Services, so he had some thoughts to share.

This first session in the symposium focused on taking existing business applications and expanding them to “the cloud”.  When should an ISV do this?  Why?  How?

Build vs. Buy and On-Premises vs. Cloud

Gianpaolo presented a nice matrix showing the two basic independent decisions that you face when looking for software to fulfill a need.

  • Build vs. Buy – Can I buy a packaged off-the-shelf product that does what I need?  Or are my needs specialized enough that I need to build my own stuff?
  • On-Premises vs. Cloud – Should I run this software on my own servers?  Or host everything up in “the cloud”?

There are, of course, tradeoffs on both sides of each decision.  These have been discussed ad infinitum elsewhere, but the basic tradeoffs are:

  • Build vs. Buy – Features vs. Cost
  • On-Premises vs. Cloud – Control vs. Economy of Scale

Here’s the graph that Gianpaolo presented, showing six different classes of software, based on how you answer these questions.  Note that on the On-Premises vs. Cloud scale, there is a middle column that represents taking applications that you essentially control and moving them to co-located servers.

This is a nice way to look at things.  It shows that, for each individual software function, it can live anywhere on this graph.  In fact, Gianpaolo’s main point is that you can deploy different pieces of your solution at different spots on the graph.

So the idea is that while you might start off on-premises, you can push your solution out to either a co-located hosting server or to the cloud in general.  This is true of both packaged apps as well as custom-developed software.


The main challenge in moving things out of the enterprise is dealing with the various issues that show up now when your data needs to cross the corporate/internet boundary.

There are several separate types of challenges that show up:

  • Identify challenges – as you move across various boundaries, how does the software know who you are and what you’re allowed to access?
  • Monitoring and Management challenges – how do you know if your application is healthy, if it’s running out in the cloud?
  • Application Integration challenge – how do various applications communicate with each other, across the various boundaries?

Solutions to the Identity Problem

Gianpaolo proposed the following possible solutions to this problem of identity moving across the different boundaries:

  • Federated ID
  • Claim-based access control
  • Geneva identity system, or Cardspace

The basic idea was that Microsoft has various assets that can help with this problem.

Solutions to the Monitoring and Management Problem

Next, the possible solutions to the monitoring and management problem included:

  • Programmatic access to a “Health” model
  • Various management APIs
  • Firewall-friendly protocols
  • Powershell support

Solutions to the Application Integration Problem

Finally, some of the proposed solutions to the application integration problem included:

  • ServiceBus
  • Oslo
  • Azure storage
  • Sync framework

The ISV Perspective

The above issues were all from an IT perspective.  But you can look at the same landscape from the perspective of an independent software vendor, trying to sell solutions to the enterprise.

To start with, there are two fundamentally different ways that the ISV can make use of “the cloud”:

  • As a service mechanism, for delivering your services via the cloud
    • You make your application’s basic services available over the internet, no matter where it is hosted
    • This is mostly a customer choice, based on where they want to deploy
  • As a platform
    • Treating the cloud as a platform, where your app runs
    • Benefits are the economy of scale
    • Mostly an ISV choice
    • E.g. you could use Azure without your customer even being aware of it

When delivering your software as a service, you need to consider things like:

  • Is the feature set available via cloud sufficient?
  • Firewall issues
  • Need a management interface for your customers

Some Patterns

Gianpaolo presented some miscellaneous design considerations and patterns that might apply to applications deployed in the cloud.


  • Design for average load, handling the ‘peak’ as an exception
  • I.e. only go to the cloud for scalability when you need to

Worker / Queue / Blob Pattern

Let’s say that you have a task like encoding and publishing of video.  You can push the data out to the cloud, where the encoding work happens.  (Raw data places in a “blob” in cloud storage).  You then add an entry to a queue, indicating that there is work to be done, and a separate worker process eventually does the encoding work.

This is a nice pattern for supporting flexible scaling—both the queues and the worker processes could be scaled out separately.

CAP: Pick 2 out of 3

  • Consistency
  • Availability
  • Tolerance to network Partitioning

Eventual Consistency (ACID – BASE)

The idea here is that we are all used to the ACID characteristics listed below.  We need to guarantee that the data is consistent and correct—which means that performance likely will suffer.  As an example, we have a process submit data synchronously because we need to guarantee that the data gets to its destination.

But Gianpaolo talked about the idea of “eventual consistency”.  For most applications, while it’s important for your data to be correct and consistent, it’s not necessarily for it to be consistent right now.  This leads to a model that he referred to as BASE, with the characteristics listed below.

  • ACID
    • Atomicity
    • Consistency
    • Isolation
    • Durability
  • BASE
    • Basically Available
    • Soft state
    • Eventually consistent

Fundamental Lesson

Basically the main takeaway is:

  • Put the software components in the place that makes the most sense, given their use

Session – Building Mesh-Enabled Web Applications Using the Live Framework

PDC 2008, Day #3, Session #5, 1 hr 15 mins

Arash Ghanaie-Sichanie

Throughout the conference, I bounced back and forth between Azure and Live Mesh sessions.  I was trying to make sense of the difference between them and understand when you might use one vs. the other.

I understand that Azure Services is the underlying platform that Live Services, and Live Mesh, are built on.  But at the outset, it still wasn’t quite clear what class of applications Live Services were targeted at.  Which apps would want to use Live Services and which would need to drop down and use Azure?

I think that after Arash’s talk, I have a working stab at answering this question.  This is sort of my current understanding, that I’ll update as things become more clear.

Question: When should I use Live Services / Live Mesh ?


  • Create a Live Mesh application (Mesh-enabled) if your customers are currently, or will become, customers.
  • Otherwise, use Azure cloud-based services outside of Live, or the other services built on top of Azure:
    • Azure storage services
    • Sync Framework
    • Service Bus
    • SQL Data Services

Unless I’m missing something, you can’t make use of features of the Live Operating Environment, either as a web application or a local desktop client, unless your application is registered with, and your user has added your application to his account.

The one possible loophole that I can see is that you might just have your app always authorize, for all users, using a central account.  That account could be pre-created and pre-configured.  Your application would then use the same account for all users, enabling some of the synchronization and other options.  But even this approach might not work—it’s possible that any device where local Mesh data is to be stored needs to be registered with that Live account and so your users wouldn’t normally have the authorization to join their devices to your central mesh.

Three Takeaways

Arash listed three main takeaways from this talk:

  • Live Services add value to different types of applications
  • Mesh-enabled web apps extend web sites to the desktop
  • The Live Framework is a standards-based API, available to all types of apps

How It All Works

The basic idea is to start with a “Mesh-enabled” web site.  This is a web site that delivers content from the user’s Mesh, including things like contacts and files.  Additionally, the web application could store all of its data in the Mesh, rather than on the web server where it is hosted.

Once a web application is Mesh-enabled, you have the ability to run it on various devices.  You basically create a local client, targeted at a particular platform, and have it work with data through the Live Framework API.  It typically ends up working with a cached local copy of the data, and the data is automatically synchronized to the cloud and then to other devices that are running the same application.

This basically implements the Mesh vision that Ray Ozzie presented first at Mix 2008 in Las Vegas and then again at PDC 2008 in Los Angeles.  The idea is that we move a user’s data into the cloud and then the data follows them around, no matter what device they’re currently working on.  The Mesh knows about a user’s:

  • Devices
  • Data, including special data like Contacts
  • Applications
  • Social graph  (friends)

The User’s Perspective

The user must do a few things to get your application or web site pointing at his data.  As a developer, you send him a link that lets him go sign up for a account and then register your application in his Mesh.  Registration, from the user’s perspective, is a way for him to authorize your application to access his Mesh-based data.

Again, it’s required that the user has, or get, a account.  That’s sort of the whole idea—we’re talking about developing applications that run on the Live platform.

Run Anywhere

There is still work to be done, on the part of the developer, to be able to run the application on various devices, but the basic choices are:

  • Locally, on a client PC or Mac
  • In a web browser, anywhere
  • From the Live Desktop itself, which is hosted in a browser

(In the future, with Silverlight adding support for running Silverlight-sandboxed apps outside of the browser, we can imagine that as a fourth option for Mesh-enabled applications).

The Developer’s Perspective

In addition to building the mesh application, the developer must also register his application, using the Azure Services Portal.  Under the covers, the application is just an Azure-based service.  And so you can leverage the Azure standard goodies, like running your service in a test mode and then deploying/publishing it.

One other feature that is made available to Mesh application authors is the ability to view basic Analytics data for your application.  Because the underlying Mesh service is aware of your application, wherever it is running, it can collect data about usage.    The data is “anonymized”, so you can’t see data about individual users, but can view general metrics.

Arash talked in some detail about the underlying security model, showing how a token is granted to the user/application.


Mesh obviously seems a good fit for some applications.  You can basically run on a platform that gives you a lot of services, as well as access to useful data that may already exist in a particular user’s Mesh environment.  There is also some potential for cross-application sharing of data, once common data models are agreed upon.

But the choice to develop on the Mesh platform implies a decision to sign your users up as part of the Mesh ecosystem.  While the programming APIs are entirely open, using basic HTTP/REST protocols, the platform itself is owned/hosted/run exclusively by Microsoft.

Not all of your users will want to go through the hassle of setting up a Live account in order to use your application.  What makes it a little worse is that the process for them is far from seamless.  It would be easier if you could hide the Live branding and automate some of the configuration.  But the user must actually log into and authorize your application.  This is a pretty high barrier, in terms of usability, for some users.

This also means that your app is betting on the success of the platform itself.  If the platform doesn’t become widely adopted, few users will live (pardon the pun) in that environment.  And the value of hosting your application in that ecosystem becomes less clear.

It remains to be seen where Live as a platform will go.  The tools and the programming models are rich and compelling.  But whether the platform will live up to Ozzie’s vision is still unclear.

Session – Offline-Enabled Data Services

PDC 2008, Day #3, Session #4, 1 hr 15 mins

Pablo Castro

I attended a second session with Pablo Castro (the previous one was the session on Azure Tables).  This session focused on a future capability in ADO.NET Data Services that would allow taking data services “offline” and then occasionally synching them with the online data.

Background – Astoria

ADO.NET Data Services was recently released as part of the .NET Framework 3.5 SP1 release.  It was formerly known as project “Astoria”.

The idea of ADO.NET Data Services is to allow creating a data service that lives on the web.  Your data source can be anything—a SQL Server database, third-party database, or some local data store.  You wrap your data source using the Entity Data Model (EDM) and ADO.NET Data Services Runtime.  Your data is now available over the web using standard HTTP protocols.

Once you have a data service that is exposed on the web, you can access it from any client.  Because the service is exposed using HTTP/REST protocols, you can access your data using simple URIs.  By using URIs, you are able to create/read/update/delete your data (POST/GET/PUT/DELETE in REST terms).

Because we can access the data using HTTP, our client is not limited to a .NET application, but can be any software that knows about the web and HTTP.  So we can consume our data from Javascript, Ruby, or whatever.  And the client could be a web-based application or a thick client somewhere that has access to the web.

If your client is a .NET client, you can use the ADO.NET Data Services classes in the .NET Framework to access your data, rather than having to build up the underlying URIs.  You can also use LINQ.

So that’s the basic idea of using ADO.NET Data Services and EDM to create a data service.  For more info, see:

The Data Synchronization Landscape

Many of the technologies that made an appearance at PDC 2008 make heavy use of data synchronization.  Data synchronization is available to Azure cloud services or to Live Mesh applications.

The underlying engine that handles data synchronization is the Microsoft Sync Framework.  The Sync Framework is a synchronization platform that allows creating sync providers, sitting on top of data that needs to be synchronized, as well as sync consumers—clients that consume that data.

The basic idea with sync is that you have multiple copies of your data in different physical locations and local clients that make use of that data.  Your client would work with its own local copy of the data and then the Sync Framework would ensure that the data is synched up with all of the other copies of the data.

What’s New

This session talked about an effort to add support in ADO.NET Data Services for offline copies of your Astoria-served data, using the Sync Framework to do data synchronization.

Here are the basic pieces (I’m too lazy to draw a picture).  This is just one possible scenario, where you want to have an application that runs locally and makes use of a locally cached copy of your data, which exists in a database somewhere:

  • Data mainly “lives” in a SQL Server database.  Assume that the database itself is not exposed to the web
  • You’ve created a data service using ADO.NET Data Services and EDM that exposes your SQL Server Data to the web using a basic REST-based protocol.  You can now do standard Create/Read/Update/Delete operations through this interface
  • You might have a web application running somewhere that consumes this data.  E.g. A Web 2.0 site built using Silverlight 2, that allows viewing/modifying the data.  Note that the web server does not have a copy of your data, but goes directly to the data service to read/write its data.
  • Now you create a thick client that also wants to read/write your data.  E.g. A WPF application.  To start with, you assume that you have a live internet connection and you configure the application to read/write data directly from/to your data service

At this point, you have something that you could build today, with the tools in the .NET Framework 3.5 SP1.  You have your data out “in the cloud” and you’ve provided both rich and thin clients that can access the data.

Note: If you were smart, you would have reused lots of code between the thin (Silverlight 2) and thick (WPF) clients.  Doing this gives your users the most consistent GUI between online and offline versions.

Now comes the new stuff.  Let’s say that you have cases when you want your thick WPC client to be able to work even though the Internet connection is not present.  Reasons for doing this include:

  • You’re working on a laptop, somewhere where you don’t have an Internet connection (e.g. airplane)
  • You want the application to be more reliable—i.e. app is still usable even if the connection disappears from time to time
  • You’d like the application to be slightly better performing.  As it stands, the performance depends on network bandwidth.  (The “lunchtime slowdown” phenomenon).

Enter Astoria Offline.  This is the set of extensions to Astoria that Pablo described, which is currently not available, but planned to be in Alpha by the end of the year.

With Astoria Offline, the idea is that you get a local cache of your data on the client PC where you’re running your thick client.  Then what happens is the following:

  • Your thick (WPF) application works directly with the offline copy of the data
  • Performance is improved
  • Much more reliable—the data is always there
  • You initiate synchronization (or set it up from time to time) to synch data back to the online copy

This synchronization is accomplished using the new Astoria Offline components.  When you do synchronize, the synchronization is two-ways, which means that you update both copies with any changes that have occurred since you last synched:

  • All data created locally is copied up to the online store
  • Data created online is copied down
  • Changes are reconciled—two-way
  • Deletions are reconciled—two-way

Pablo did a basic demo of this scenario and it worked just as advertised.  He showed that the client worked with the local copy of the data and that everything synched properly.  He also showed off some early tooling in Visual Studio that will automate much of the configuration that is required for all of this to work.

Interestingly, it looked like in Pablo’s example, the local copy of the data was stored in SQL Server Express.  This was a good match, because the “in the cloud” data was stored in SQL Server.

How Did They Do It?

Jump back to the description of the Microsoft Sync Framework.  Astoria Offline is using the sync framework to do all of the underlying synchronization.  They’ve written a sync provider that knows about the entity data model and interfaces between EDM and the sync framework.

Extensibility Points

I’m a little fuzzier on this area, but I think I have a general sense of what can be done.

Note that the Sync Framework itself is extensible—you can write your own sync providers, providing synchronized access to any data store that you care to support.  Once you do this, you get 2-way (or more) synchronization between your islands of custom data.

But if I understood Pablo correctly, it sounds like you could do this a bit differently with Astoria Offline in place.  It seems like you could pump your custom data from the Entity Framework, by building a custom data source so that the EDM can see your data.  (EntityFrameworkSyncProvider fits in here somewhere).  I’m guessing that once you serve up your data in a relational manner to the EDM, you can then synch it using the Astoria Offline mechanisms.  Fantastic stuff!

Going Beyond Two Nodes

One could imagine going beyond just an online data source and an offline copy.  You could easily imagine topologies that had many different copies of the data, in various places, all being synched up from time to time.

Other Stuff

Pablo talked about some of the other issues that you need to think about.  Conflict detection and resolution is a big one.  What if two clients both update the same piece of data at the same time?  Classic synchronization issue.

The basic things to know about conflicts, in Astoria Offline, are:

  • Sync Framework provides a rich model for detecting/resolving conflicts, under the covers
  • Astoria Offline framework will detect conflicts
  • The application provides “resolution handlers” to dictate how to resolve the conflict
    • Could be locally—e.g. ask the user what to do
    • Or online—automatic policies

Pablo also talked briefly about the idea of Incremental Synchronization.  The idea is that you might want to synch things a little bit at a time, in a batch-type environment.

There was a lot more stuff here, and a lot to learn.  Much of the concepts just bubble up from the Sync Framework.


Astoria Offline is potentially game-changing.  In my opinion, of the new technologies presented at PDC 2008, Astoria Offline is the one most likely to change the landscape.  In the past, vendors have generally had to pick between working with live data or local data.  Now they can do both.

In the past, the online vs. offline data choice was driven by whether you needed to share the data across multiple users.  So the only apps that went with offline data were the ones that didn’t need to share their data.  What’s interesting about Astoria Offline is that these apps/scenarios  can now use this solution to leave their data essentially local, but make the data more mobile, across devices.  Imagine an application that just stores local data that only it consumes.  But now if you want to run that app on multiple machines, you have to manually copy the data—or move it to a share seen by both devices.  With Astoria Offline, you can set up a sync to an online location that each device synchs to, as needed.  So you can just move from device to device and your data will just follow you.  So you can imagine that this makes it much easier to move apps out to mobile devices.

This vision is very similar to what Live Mesh and Live Services promise.  But the difference is that here you don’t need to subscribe to your app and its data living in the MS Live space.  Your data can be in whatever format you like, and nobody needs to sign up with MS Live.

When Can I Get It?

Pablo mentioned a basic timeline:

  • Early Alpha by the end of the year
  • CTPs later, i.e. next year


In addition to the links I listed above, you might also check out:

Session – Windows Azure Tables: Programming Cloud Table Storage

PDC 2008, Day #3, Session #3, 1 hr 15 mins

Pablo Castro, Niranjan Nilakantan

Pablo and Niranjan did a session that went into some more detail on how the Azure Table objects can be used store data in the cloud.


This talk dealt with the “Scalable Storage” part of the new Azure Services platform.  Scalable Storage is a mechanism by which applications can store data “in the cloud” in a highly scalable manner.

Data Types

There are three fundamental data types available to applications using Azure Storage Services:

  • Blobs
  • Tables
  • Queues

This session focused mainly on Tables.  Specifically, Niranjan and Pablo addressed the different ways that an application might access the storage service programmatically.


Tables are a “massively scalable” data type for cloud-based storage.  They are able to store billions of rows, are highly available, and “durable”.  The Azure platform takes care of scaling out the data automatically to multiple servers, if necessary.  (With some hints on the part of the developer).

Programming Model

Azure Storage Services are accessed through the ADO.NET Data Services (Astoria).  Using ADO.NET Data Sercices, there are basically two ways for an application to access the service.

  • .NET API   (System.Data.Services.Client)
  • REST interface   (using HTTP URIs directly)

Data Model

It’s important to note that Azure knows nothing about your data model.  It does not store data in a relational database or access it via a relational model.  Rather, you specify a Table that you’d like to store data in, along with a simple query expression for the data that you’d like to retrieve.

A Table represents a single Entity and is composed of a collection of rows.  Each row is uniquely defined by a Row Key, which the developer specifies.  Additionally, the developer specifies a Partition Key, which is used by Azure in knowing how to split the data across multiple servers.

Beyond the Record Key and Partition Key, the developer can add any other properties that she likes, up to a total of 255 properties.  While the Record and Partition Keys must be string data types, the other properties support other data types.


Azure storage services are meant to be automatically scalable, meaning that the data will be automatically spread across multiple servers, as needed.

In order to know how to split up the data, Azure uses a developer-specified Partition Key, which is one of the properties of each record.  (Think “field” or “column”).

The developer should pick a partition key that makes sense for his application.  It’s important to remember two things:

  • Querying for all data having a single value for a partition key is cheap
  • Querying for data having multiple partition key values is more expensive

For example, if your application often retrieves data by date and shows data typically for a single day, then it would make sense to have a CurrentData property in your data entity and to make that property the Partition Key.

The way to think of this is that each possible unique value for a Partition Key represent a “bucket” that will contain one or more records.  If you pick a key that results in only one record per bucket, that would be inefficient.  But if you pick a key that results in a set of records in the bucket that you are likely to ask for together, this will be efficient.

Accessing the Data Programmatically

Pablo demonstrated creating the classes required to access data stored in an Azure storage service.

He started by creating a class representing the data entity to be stored in a single table.  He selected and defined properties for the Partition and Record key, as well as other properties to store any other desired data in.

Pablo also recommended that you create a single class to act as an entry point into the system.  This class then acts as a service entry point for all of the data operations that your client application would like to perform.

He also demonstrated using LINQ to run queries against the Azure storage service.  LINQ automatically created to corresponding URI to retrieve, create, update, or delete the data.


Pablo and Niranjan also touched on a few other issues that most applications will deal with:

  • Dealing with concurrent updates  (uses Etag and if-match)
  • Pagination  (using continuation tokens)
  • Using Azure Queues for pseudo-transactional deletion of data


Pablo and Niranjan demonstrated that it was quite straightforward to access Azure storage services from a .NET application.  It’s also the case that non-.NET stacks could make use of the same services using a simple REST protocol.

It was also helpful to see how Pablo used ADO.NET Data Services to construct a service layer on top of the Azure storage services.  This seems to make consuming the data pretty straightforward.

(I still might have this a little confused—it’s possible that Astoria was just being used to wrap Azure services, rather than exposing the data in an Astoria-based service to the client.  I need to look at the examples in a little more detail to figure this out).

Original Materials

You can find the video of the session at: