Interview: Virtual Actors with Microsoft Orleans

Introduction

In recent years, the ever increasing demand for computing resources has rendered traditional single-threaded programming inadequate for most modern applications. Faced by heavy performance and scalability challenges, many developers are forced to turn to concurrent and distributed programming.

While multithreaded programming has been in use for many years, those who have used it will know that building a performant shared memory system free of race conditions can be very challenging to get right.

It is possible to avoid the complications of shared memory systems, and indeed multithreading, by using a message passing system. An actor model is a framework where processing is done by a large number of single-threaded actors, which communicate together by sending asynchronous messages.

As it turns out, Microsoft have their own actor model, and it’s called Orleans. Sergey Bykov (SB), Principal Software Development Engineer Lead at Microsoft, and project lead of the Orleans project, has very kindly agreed to answer my (DD) questions about Orleans.

Orleans Overview

DD: What is Microsoft Orleans?

SB: The home page of our docs says the following.

“Orleans is a framework that provides a straightforward approach to building distributed high-scale computing applications, without the need to learn and apply complex concurrency or other scaling patterns. It was created by Microsoft Research and designed for use in the cloud.

“Orleans has been used extensively in Microsoft Azure by several Microsoft product groups, most notably by 343 Industries as a platform for all of Halo 4 and Halo 5 cloud services, as well as by a growing number of other companies.”

SB: In other words, Orleans provides a programming model (backed by the Orleans runtime) for building distributed scalable applications almost as easily as single machine apps. The goal of the project from the beginning was to democratize cloud development by making a broad range of developers with little to no distributed systems expertise productive and successful in building scalable distributed systems in general, and cloud services in particular.

The introduction explains that Orleans is built around a distributed actor model, and the key innovation there is the notion of Virtual Actors. Detailed description is in our publication.

DD: Out of curiosity, why the name, ‘Orleans’?

SB: It was a general rule within Microsoft that codenames should be chosen from geographical names like names of cities because those aren’t trademark-able. Over time, the codename of Orleans accrued enough brand recognition that we decided to stick with it when we went open source.

DD: Tell us a little about the history of Microsoft Orleans.

SB: Orleans started in 2009 as a research project within a new Microsoft Research lab that eventually was named eXtreme Computing Group (XCG, it was later merged with MSR’s Redmond lab). The goal for the project was to try to create something that would qualitatively simplify creating software for the cloud. The two major challenges we focused on were 1) the complexity of building distributed systems that has traditionally been the domain of a relatively small population of expert developers; and 2) the pattern of major re-architectures required from nearly every successful web property as they experienced exponential growth of their user base.

We took on building a framework with a programming mode that would make mainstream single-machine developers productive in the cloud and would help build systems and services that could easily scale through several orders of magnitude of growing scale. While focusing on mainstream developers, we wanted Orleans to be as appealing to expert developers, by reducing the amount of low level ceremony they have to deal with. As we went through several early prototypes and iterations, we learned quite a bit from building first Orleans applications, and even more so when we started collaborating with internal product groups. The programming model has evolved, and we arrived to what we ended up naming the “Virtual Actor Model”.

Using Orleans

DD: How does Microsoft Orleans compare with other actor models?

SB: The Actor Model is quite old, and there are many various implementations of it. There’s a much smaller number of available Distributed Actor Model solutions. The most popular ones are Erlang/OTP and its JVM “younger sibling” Akka. Erlang and Akka organically grew from being single process actor libraries into the multi-machine scenarios by gradually adding remoting and distribution features. They brought the fault tolerance model of hierarchical supervision trees that are easy within a single process, acceptable for small-scale fixed topologies, but are difficult to manage at cloud scale, especially for developers with limited distributed systems experience.

The Virtual Actor Model of Orleans removed a lot of coordination and fault tolerance complexity from developers’ shoulders by providing an intuitive notion of actors that don’t need to be created, destroyed or looked up. The “Virtual” qualifier comes from the analogy with virtual memory. Actors in Orleans live “eternal” life, always available for a call to process, and the Orleans runtime is responsible for instantiating their physical “incarnations” in memory on an as needed basis, and for removing idle ones to free up resources. The Orleans runtime also transparently handles failures of servers by keeping track of instantiated actors and recreating them when needed on a different server in case of a failure. As a result, the developer writes much less code (we’ve received anecdotal reports of 3-5 times reduction of code, up to 10 times in some cases) and much simpler code, free from data races and complex distributed coordination logic.

The effort of Orleans to ‘democratize’ distributed programming and to raise developer productivity received an endorsement of sort from the inventor of Actor Model, Carl Hewitt. In his recent publication Actor Model of Computation for Scalable Robust Information Systems he wrote that: “Orleans is an important step in furthering a goal of the Actor Model that application programmers need not be so concerned with low-level system details.” Obviously, that made the Orleans team very proud.

DD: In Microsoft Orleans, virtual actors are also known as grains. They run within host processes called silos. Why were these names devised?

SB: Early on we had the intuition that we’d end up with a novel programming model. In hindsight, that was prescient. The “grains” term is distinct from the already overloaded term actor, where it’s hard to tell upfront if somebody is talking about single machine concurrency or about a distributed case. In the end, “grain” is a shorthand for “Orleans actor” or “virtual actor”. When we needed to name the runtime containers for grains, we naturally went down the agricultural path with “silos”. Just imagine the confusion if called them “containers”.

DD: Who is using Microsoft Orleans, and how well does it support their systems’ scalability?

SB: Orleans has been used in production inside Microsoft since 2011. It is enjoing a growing adoption outside Microsoft after we publicly released a binary preview, and then open-sourced it. We see a wide range of systems built with Orleans: online gaming, finance, collaboration solutions, fraud detection, social network analysis, and more. One of the hottest areas is IoT. There we see Orleans-based systems that manage devices like thermostats and even, I’m not joking, mousetraps. One of the fascinating projects is the green power storage facility in Hawaii. We showed some scalability numbers in our paper.

DD: Is Microsoft Orleans meant only to be used in the cloud?

SB: The advent of the cloud brought the challenges of building reliable scalable distributed systems into the spotlight. Orleans as a project focused on solving those fundamental challenges. As a result, Orleans is equally applicable in any cloud and on premises. We have customers running Orleans in AWS and some interested in GCP, but also those that use it in private datacenters and on corporate IT infrastructures. Our first target was naturally Azure, and we built providers and extensions for it first. But Orleans was designed with extensibility in mind, and it is fairly easy to make it run pretty much anywhere.

Development and Support

DD: What is the Microsoft Orleans team currently working on, and is there a roadmap for future development?

SB: Our current focus is on making Orleans run on .NET Core, support for geo-distribution, improvements to streaming, application lifecycle and the upgrade and versioning process. Even though the project moved out of Microsoft Research to the product group, we have an ongoing collaboration with Research, which gives us a healthy pipeline of new ideas and advanced prototypes. Support for geo-disribution is one example. We also have support for indexing of actors, ACID multi-actor transactions, and reactive computations at various stages of readiness. Orleans is one of the most popular Microsoft open source projects, right next to .NET Core and Roslyn. We continue to work on it and recently substantially increased our investments.

DD: What resources are available for developers building their systems upon Microsoft Orleans?

SB: We keep hearing that our documentation is very good compared to other open source projects, but we keep improving it (and samples) as people point to topics that aren’t clear or can be explained better. The community around the project is our biggest “brain trust” and the best source of support for new people. It’s an amazing group of experienced and passionate engineers around the globe that come to our GitHub repo and Gitter chat not only because they use Orleans for their projects and contribute to it, but also because they enjoy hanging out with this very welcoming and encouraging community that always tries to help, even with topics not directly related to Orleans.

Custom Loggers in Akka .NET

Akka .NET supports a flexible logging mechanism that can adapt with various logging providers, as we have seen in my earlier article on logging with Akka .NET. Aside from the default logger that writes to the console, you can plug in various loggers of your own choosing (e.g. NLog), set them up in configuration, and work with them using a common interface.

Sometimes, you may have specific logging requirements that are not covered by any of the existing logging plugins for Akka .NET. In such cases, you would need to write your own custom logger. Unfortunately, the Akka .NET Logging documentation does not explain how to do this at the time of writing this article.

This article is intended to fill this gap, explaining how to write a custom logger for Akka .NET, but also touching upon various topics such as reading custom configuration, actor lifecycle hooks, and cleanup of resources used by the ActorSystem. The source code for this article is available at the Gigi Labs BitBucket repository.

In a Nutshell

Akka .NET Logging is basically a port of Akka logging. Any logger is an actor that receives the following messages: Debug, Info, Warning, Error and InitializeLogger. These are all standard Akka .NET messages, and we will see how to use them in a minute. At the time of writing this article, I’ve come across the following loggers that one can refer to in order to see how custom loggers are built:

Main Program

For the sake of this article, we can go along with the following simple program:

            using (var actorSystem = ActorSystem.Create("MyActorSystem"))
            {
                var logger = Logging.GetLogger(actorSystem, actorSystem, null);
                logger.Info("ActorSystem created!");

                Console.WriteLine("Press ENTER to exit...");
                Console.ReadLine();
            }

In my earlier Akka .NET logging article, we had done logging only from within actors. The above code shows how you can use the same configured logger directly from the ActorSystem.

We are now going to need a little configuration.

<?xml version="1.0" encoding="utf-8" ?>
<configuration>

  <configSections>

<section name="akka" type="Akka.Configuration.Hocon.AkkaConfigurationSection, Akka" />
  </configSections>

  <startup>
    <supportedRuntime version="v4.0" sku=".NETFramework,Version=v4.6.1" />
  </startup>

  <akka>
    <hocon>
      <![CDATA[ akka { loglevel = DEBUG loggers = ["AkkaNetCustomLogger.Loggers.ConsoleLogger, AkkaNetCustomLogger"] actor { debug { receive = on # log any received message autoreceive = on # log automatically received messages, e.g. PoisonPill lifecycle = on # log actor lifecycle changes event-stream = on # log subscription changes for Akka.NET event stream unhandled = on # log unhandled messages sent to actors } } } ]]>
    </hocon>
  </akka>

</configuration>

Here, we’re turning on all internal Akka .NET logging so that we can automatically get some logging output.

The important thing here is the loggers configuration, where we’re specifying the custom logger that we want Akka .NET to use. This is exactly how we had set up Akka .NET to use NLog in my earlier article, but this time we’re going to use a class called ConsoleLogger (which we have yet to create).

Writing the ConsoleLogger

As I mentioned earlier, any custom logger needs to handle five messages: Debug, Info, Warning, Error, and InitializeLogger. That’s the first thing we’ll set up our ConsoleLogger to do.

    public class ConsoleLogger : ReceiveActor
    {
        public ConsoleLogger()
        {
            Receive<Debug>(e => this.Log(LogLevel.DebugLevel, e.ToString()));
            Receive<Info>(e => this.Log(LogLevel.InfoLevel, e.ToString()));
            Receive<Warning>(e => this.Log(LogLevel.WarningLevel, e.ToString()));
            Receive<Error>(e => this.Log(LogLevel.ErrorLevel, e.ToString()));
            Receive<InitializeLogger>(_ => this.Init(Sender));
        }

        // ...
    }

The actual logging messages will use a common Log() helper method, since they differ only in log level (the operation of writing to the log destination is the same for all). Note that when we convert each of these classes to string, that includes the log level, so we don’t need to write it separately. In the case of the ConsoleLogger, we are passing in the LogLevel merely so we can use a different colour for each level.

The special message we haven’t covered yet is InitializeLogger. When the ActorSystem creates the logger actor, the internal event bus needs to know whether the logger is ready to start accepting messages. It does this by sending the logger actor an InitializeLogger message, and expects a LoggerInitialized message in return:

        private void Init(IActorRef sender)
        {
            using (var consoleColour = new ScopedConsoleColour(ConsoleColor.Green))
                Console.WriteLine("Init");

            sender.Tell(new LoggerInitialized());
        }

Aside from sending back the required message, I’m also logging the init operation itself, so that we can later observe the sequence of events. I am using my trusty ScopedConsoleColour class to change the colour, and then reset it back after the message has been written.

If you don’t send the LoggerInitialized message back, the actor system reports a timeout from initialising the logger, and basically you get no logging. (Well, ironically, the timeout itself is logged… presumably using the DefaultLogger as a fallback.)

akkanet-logging-notinitialized

Now we can implement our Log() helper method:

        private void Log(LogLevel level, string message)
        {
            ConsoleColor colour = ConsoleColor.Gray;

            switch (level)
            {
                case LogLevel.DebugLevel:
                    colour = ConsoleColor.Gray;
                    break;
                case LogLevel.InfoLevel:
                    colour = ConsoleColor.White;
                    break;
                case LogLevel.WarningLevel:
                    colour = ConsoleColor.Yellow;
                    break;
                case LogLevel.ErrorLevel:
                    colour = ConsoleColor.Red;
                    break;
                default: // shouldn't happen
                    goto case LogLevel.InfoLevel;
            }

            using (var consoleColour = new ScopedConsoleColour(colour))
                Console.WriteLine(message);
        }

Here we’re doing nothing but using a different colour per level, and writing the message to the console (which is pretty much what StandardOutLogger does). Remember, the level is already part of the message, so we don’t need to format it into the output message. (And if you’re outraged at the use of goto above, I suggest you read about goto case in C#.)

Actor Life Cycle Hooks

If your custom logger depends on external resources (which is most likely the case) such as the filesystem or a database, you will want to initialise those resources when the logger actor is created, and clean them up when it is destroyed. That work typically goes into actor life cycle hooks, i.e. overridable methods that allow you to run arbitrary code when an actor starts, stops, or restarts.

We don’t need to do this for ConsoleLogger, so we will simply log the start and stop operation instead. However, we will use these hooks more realistically when we implement the FileLogger.

        protected override void PreStart()
        {
            base.PreStart();

            using (var consoleColour = new ScopedConsoleColour(ConsoleColor.Green))
                Console.WriteLine("PreStart");
        }

        protected override void PostStop()
        {
            using (var consoleColour = new ScopedConsoleColour(ConsoleColor.Green))
                Console.WriteLine("PostStop");

            base.PostStop();
        }

We can now run this and see the logging in action:

akkanet-logging-console

Now, for something interesting, put a breakpoint inside PostStop(), and press ENTER to cause the program to continue and terminate. One would expect PostStop() to run as the ActorSystem is shutting down. But in fact, it doesn’t.

Next, go back to the main program, and add a second Console.ReadLine() at the end:

            using (var actorSystem = ActorSystem.Create("MyActorSystem"))
            {
                var logger = Logging.GetLogger(actorSystem, actorSystem, null);
                logger.Info("ActorSystem created!");

                Console.WriteLine("Press ENTER to exit...");
                Console.ReadLine();
            }

            Console.ReadLine();

Run it again, and when you press ENTER, the breakpoint is hit and the PostStop event is written to the console while waiting for the second ENTER:

akkanet-logging-poststop

When we disposed the ActorSystem earlier, the program terminated before the ActorSystem had the chance to do its cleanup work. It appears that when the ActorSystem shuts down, it doesn’t clean resources right away; most likely it is done using asynchronous messaging just like in the rest of Akka .NET. For this reason, in your program’s stopping code, you might want to wait a little bit between destroying the ActorSystem and actually letting the application terminate, in order to let it gracefully free its resources.

Writing the FileLogger

We will now write a second custom logger, this time one that writes to file.

First, change the HOCON configuration to use the new logger.

          loggers = ["AkkaNetCustomLogger.Loggers.FileLogger, AkkaNetCustomLogger"]

Next, let’s write the FileLogger. As with the ConsoleLogger, we need to handle the same five messages:

    public class FileLogger : ReceiveActor
    {
        private StreamWriter writer;

        public FileLogger()
        {
            ReceiveAsync<Debug>(async e => await this.LogAsync(e.ToString()));
            ReceiveAsync<Info>(async e => await this.LogAsync(e.ToString()));
            ReceiveAsync<Warning>(async e => await this.LogAsync(e.ToString()));
            ReceiveAsync<Error>(async e => await this.LogAsync(e.ToString()));
            Receive<InitializeLogger>(_ => Sender.Tell(new LoggerInitialized()));
        }

        // ...
    }

The logger keeps a reference to a StreamWriter, which wraps the file we will be writing to.

The LogAsync() method simply dumps the messages into that StreamWriter (remember, streams and toilets must always be flushed):

        private async Task LogAsync(string message)
        {
            await this.writer.WriteLineAsync(message);
            await this.writer.FlushAsync();
        }

We can open the file itself either during the InitializeLogger handler or in PreStart(). Let’s use a fixed filename for now:

        protected override void PreStart()
        {
            base.PreStart();

            string filePath = "log.txt";
            var fileStream = File.OpenWrite(filePath);
            this.writer = new StreamWriter(fileStream);
        }

We can then do the cleanup in PostStop():

        protected override void PostStop()
        {
            // dispose the StreamWriter, and implicitly the
            // underlying FileStream with it
            this.writer.Dispose();

            base.PostStop();
        }

We only need to Dispose() our StreamWriter; doing that will automatically also close the underlying FileStream.

Now, while this is enough to log to file, there is a problem. We can’t actually use another program to read the log file while the program is running:

akkanet-logging-cantread

We can fix this by changing the way we open the file.

        protected override void PreStart()
        {
            base.PreStart();

            string filePath = "log.txt";
            var fileStream = File.Open(filePath, FileMode.Append, FileAccess.Write, FileShare.Read);
            this.writer = new StreamWriter(fileStream);
        }

You’ll see that the log messages are now written to file:

akkanet-logging-fileoutput

However, there is another problem. If you press ENTER to close the program, the following happens:

akkanet-logging-race

Upon further inspection, it seems that logging messages are coming in around the same time that PostStop() is running, causing a race condition on the underlying resource. I’ve opened a bug report for this, but until this is sorted, you can flush synchronously as a workaround:

        private async Task LogAsync(string message)
        {
            await this.writer.WriteLineAsync(message);
            this.writer.Flush();
        }

So, if there is this problem, how do existing logging adapters that have a file-based component (e.g. NLog) do their cleanup? Well, I’ve checked a few, and it seems they don’t.

Loading Custom Configuration

We’ve managed to write a file logger, but we’re using a fixed filename. How can we make it configurable?

It turns out we can just add an arbitrary setting anywhere in the HOCON configuration, and read it from inside the actor. So, let’s add this:

        akka
        {
          loglevel = DEBUG
          loggers = ["AkkaNetCustomLogger.Loggers.FileLogger, AkkaNetCustomLogger"]
          logfilepath = "logfile.txt"

We can get to the setting we want using the configuration system in Akka .NET:

        protected override void PreStart()
        {
            base.PreStart();

            string filePath = "log.txt";

            filePath = Context.System.Settings.Config
                .GetString("akka.logfilepath", filePath);

            var fileStream = File.Open(filePath, FileMode.Append, FileAccess.Write, FileShare.Read);
            this.writer = new StreamWriter(fileStream);
        }

Basically we’re reading the “akka.logfilepath” key from the HOCON config. We’re also passing in filePath as a default in case the setting is not found.

Running Multiple Loggers

So far we’ve been using either one logger or the other. But if you notice, the loggers configuration in HOCON is actually an array. Thus there is nothing stopping us from using multiple loggers at once:

        akka
        {
          loglevel = DEBUG
          loggers = ["AkkaNetCustomLogger.Loggers.FileLogger, AkkaNetCustomLogger",
                     "AkkaNetCustomLogger.Loggers.ConsoleLogger, AkkaNetCustomLogger"]
          logfilepath = "logfile.txt"

Yes, it works:

akkanet-logging-multiple