October 30th, 2019

Introducing Orleans 3.0

Reuben Bond
Principal Software Development Engineer

This is a guest post from the Orleans team. Orleans is a cross-platform framework for building distributed applications with .NET. For more information, see https://github.com/dotnet/orleans.

We are excited to announce the Orleans 3.0 release. A great number of improvements and fixes went in, as well as several new features, since Orleans 2.0. These changes were driven by the experience of many people running Orleans-based applications in production in a wide range of scenarios and environments, and by the ingenuity and passion of the global Orleans community that always strives to make the codebase better, faster, and more flexible. A BIG Thank You to all who contributed to this release in various ways!

Major changes since Orleans 2.0

Orleans 2.0 was released a little over 18 months ago and since then Orleans has made significant strides. Some of the headline changes since 2.0 are:

  • Distributed ACID transactions — multiple grains can join a transaction regardless of where their state is stored
  • A new scheduler, which alone increased performance by over 30% in some cases
  • A new code generator based on Roslyn code analysis
  • Rewritten cluster membership for improved recovery speed
  • Co-hosting support

As well as many, many other improvements and fixes.

Since the days of working on Orleans 2.0, the team established a virtuous cycle of implementing or integrating certain features, such as generic host, named options, in close collaboration with the .NET team before those features are ready to be part of the .NET Core releases, contributing feedback and improvements “upstream”, and in later releases switching to their final implementations shipped with .NET releases. During development of Orleans 3.0, this cycle continued, with Bedrock code used by Orleans 3.0.0-beta1 before it finally shipped as part of .NET 3.0. Similarly, support for TLS on TCP socket connections was implemented as part of Orleans 3.0, and is intended to become part of a future release of .NET Core. We view this ongoing collaboration as our contribution to the larger .NET ecosystem, in the true spirit of Open Source.

Networking layer replacement with ASP.NET Bedrock

Support for securing communication with TLS has been a major ask for some time, both from the community as well as from internal partners. With the 3.0 release we are introducing TLS support, available via the Microsoft.Orleans.Connections.Security package. For more information, see the TransportLayerSecurity sample.

Implementing TLS support was a major undertaking due to how the networking layer in previous versions of Orleans was implemented: it could not be easily adapted to use SslStream, which is the most common method for implementing TLS. With TLS as our driving force, we embarked upon a journey to rewrite Orleans’ networking layer.

Orleans 3.0 replaces its entire networking layer with one built on top of Project Bedrock, an initiative from the ASP.NET team. The goal of Bedrock is to help developers to build fast and robust network clients and servers.

The ASP.NET team and the Orleans team worked together to design abstractions which support both network clients and servers, are transport-agnostic, and can be customized using middleware. These abstractions allow us to change the network transport via configuration, without modifying internal, Orleans-specific networking code. Orleans’ TLS support is implemented as a Bedrock middleware and our intention is for this to be made generic so that it can be shared with others in the .NET ecosystem.

Although the impetus for this undertaking was to enable TLS support, we see an approximately 30% improvement in throughput on average in our nightly load tests.

The networking layer rewrite also involved replacing our custom buffer pooling with reliance on MemoryPool<byte> and in making this change, serialization now takes more advantage of Span<T>. Some code paths which previously relied on blocking via dedicated threads calling BlockingCollection<T> are now using Channel<T> to pass messages asynchronously. This results in fewer dedicated threads, moving the work to the .NET thread pool instead.

The core wire protocol for Orleans has remained fixed since its initial release. With Orleans 3.0, we have added support for progressively upgrading the network protocol via protocol negotiation. The protocol negotiation support added in Orleans 3.0 enables future enhancements, such as customizing the core serializer, while maintaining backwards compatibility. One benefit of the new networking protocol is support for full-duplex silo-to-silo connections rather than the simplex connection pairs established between silos previously. The protocol version can be configured via ConnectionOptions.ProtocolVersion.

Co-hosting via the Generic Host

Co-hosting Orleans with other frameworks, such as ASP.NET Core, in the same process is now easier than before thanks to the .NET Generic Host.

Here is an example of adding Orleans alongside ASP.NET Core to a host using UseOrleans:

var host = new HostBuilder()
  .ConfigureWebHostDefaults(webBuilder =>
  {
    // Configure ASP.NET Core
    webBuilder.UseStartup<Startup>();
  })
  .UseOrleans(siloBuilder =>
  {
    // Configure Orleans
    siloBuilder.UseLocalHostClustering();
  })
  .ConfigureLogging(logging =>
  {
    /* Configure cross-cutting concerns such as logging */
  })
  .ConfigureServices(services =>
  {
    /* Configure shared services */
  })
  .UseConsoleLifetime()
  .Build();

// Start the host and wait for it to stop.
await host.RunAsync();

Using the generic host builder, Orleans will share a service provider with other hosted services. This grants these services access to Orleans. For example, a developer can inject IClusterClient or IGrainFactory into an ASP.NET Core MVC controller and call grains directly from their MVC application.

This functionality can be used to simplify your deployment topology or to add additional functionality to an existing application. Some teams internally use co-hosting to add Kubernetes liveness and readiness probes to their Orleans silos using the ASP.NET Core Health Checks.

Reliability improvements

Clusters now recover more quickly from failures thanks to extended gossiping. In previous versions of Orleans, silos would send membership gossip messages to other silos, instructing them to update membership information. Gossip messages now include versioned, immutable snapshots of cluster membership. This improves convergence time after a silo joins or leaves the cluster (for example during upgrade, scaling, or after a failure) and alleviates contention on the shared membership store, allowing for quicker cluster transitions. Failure detection has also been improved, with more diagnostic messages and refinements to ensure faster, more accurate detection. Failure detection involves silos in a cluster collaboratively monitoring each other, with each silo sending periodic health probes to a subset of other silos. Silos and clients also now proactively disconnect from silos which have been declared defunct and they will deny connections to such silos.

Messaging errors are now handled more consistently, resulting in prompt errors being propagated back to the caller. This helps developers to discover errors more quickly. For example, when a message cannot be fully serialized or deserialized, a detailed exception will be propagated back to the original caller.

Improved extensibility

Streams can now have custom data adaptors, allowing them to ingest data in any format. This gives developers greater control over how stream items are represented in storage. It also gives the stream provider control over how data is written, allowing steams to integrate with legacy systems and/or non-Orleans services.

Grain extensions allow additional behavior to be added to a grain at runtime by attaching a new component with its own communication interface. For example, Orleans transactions use grain extensions to add transaction lifecycle methods, such as Prepare, Commit, and Abort, to a grain transparently to the user. Grain extensions are now also available for Grain Services and System Targets.

Custom transactional state can now declare what roles it is able to fulfil in a transaction. For example, a transactional state implementation which writes transaction lifecycle events to a Service Bus queue cannot fulfil the duties of the transaction manager since it is write-only.

The predefined placement strategies are publicly accessible now, so that any placement director can be replaced during configuration time.

Join the effort

Now that Orleans 3.0 is out the door we are turning our attention to future releases — and we have some exciting plans! Come and join our warm, welcoming community on GitHub and Gitter and help us to make these plans a reality.

Orleans Team

Category
.NET

Author

Reuben Bond
Principal Software Development Engineer

Reuben Bond is a Principal Software Development Engineer on the Orleans team.

7 comments

Discussion is closed. Login to edit/delete existing comments.

  • Robson Previato

    Can you suggest a start path for learning more about Distributed Application?

  • Muhammad Miftah

    Now that 3.0 has been finalised, please update the documentation at http://dotnet.github.io/orleans/Documentation/index.html.
    It still references v1.0 API in some places. I tried learning as a newbie, but it's annoying getting stung trying to follow the documentation only to have to realise I need to go trawling through Github or the object explorer in VS to find the v2.0 API or most recent equivalent of the API mentioned in the docs because it refers to...

    Read more
  • Jonathan LEI

    I haven’t tried Orleans, but it looks very similar to Dapr from the description alone. What’s the difference between the two? Thanks.

    • Jack Bond

      Orleans? Dapr? Service Fabric?

      Could definitely use some guidance here.

      • William Howard

        Don’t forget docker and Kubernetes 🙂

        Probably need one called JAMSS (Just another micro-service system)

      • Arshia Ghafoori

        Orleans is not the same as any of these. It’s an actor model framework most comparable to Akka, but the “virtual” actor model makes it easier to use.

      • Jonathan LEI

        To be fair Docker is not really an alternative here lol. Dapr runs on Docker. The level of abstraction is different.