Skip to content
A .NET dataflow and etl framework built upon Microsoft TPL Dataflow library
C#
Branch: master
Clone or download
Pull request Compare This branch is even with gridsum:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.nuget
.vscode
Gridsum.DataflowEx.Demo
Gridsum.DataflowEx.Test
Gridsum.DataflowEx.TestHelper
Gridsum.DataflowEx
images
.gitignore
Gridsum.DataflowEx.sln
License.txt
README.md
global.json

README.md

Welcome to DataflowEx

Gridsum.DataflowEx is a high-level dataflow framework redesigned on top of Microsoft TPL Dataflow library with Object-Oriented Programming in mind. It does not replace TPL Dataflow but provides reusability, abstraction and management over underlying dataflow blocks to make your life easier. You can get compiled binaries on nuget.org.

Update: Gridsum.DataflowEx v2.0 has been released! This new major version fully supports .NET standard 2.0, which means you are able to utilize the library in a wide range of platforms including .Net Core, Xamarin and UWP.

Here is a list of DataflowEx cool features:

  • Block chain/graph encapsulation as a reusable component
  • Inheritance and polymorphism for dataflows and their hehaviors
  • Sophisticated dataflow lifecycle management (e.g. Failure propagation, dynamic flow expansion)
  • Advanced dataflow chaining/linking
  • Extensive logging and built-in dataflow health monitor
  • Dataflow event counting and aggregation
  • Auto completion support for circular dataflow graph
  • Dataflow-friendly sql bulk inserter
  • Dataflow-friendly ETL dimension lookup component
  • Compatibility and integration with TPL Dataflow

Interested? O.K. Let's start the DataflowEx tour.

1. Prerequisites

If you are not familiar with TPL Dataflow yet, please take your time to watch two videos:

Beginner (15min): http://channel9.msdn.com/posts/TPL-Dataflow-Tour

Advanced (63min): https://channel9.msdn.com/Shows/Going+Deep/Stephen-Toub-Inside-TPL-Dataflow

2. Background

The very first question you may ask: what's wrong with TPL Dataflow? Nothing. The library from Microsoft library looks simply great. However, in the tough real world there are some obstacles when we apply RAW TPL Dataflow. Let's look at an example:

var splitter = new TransformBlock<string, KeyValuePair<string, int>>(
    input =>
        {
            var splitted = input.Split('=');
            return new KeyValuePair<string, int>(splitted[0], int.Parse(splitted[1]));
        });

var dict = new Dictionary<string, int>();
var aggregater = new ActionBlock<KeyValuePair<string, int>>(
    pair =>
        {
            int oldValue;
            dict[pair.Key] = dict.TryGetValue(pair.Key, out oldValue) ? oldValue + pair.Value : pair.Value;
        });

splitter.LinkTo(aggregater, new DataflowLinkOptions() { PropagateCompletion = true});

splitter.Post("a=1");
splitter.Post("b=2");
splitter.Post("a=5");

splitter.Complete();
aggregater.Completion.Wait();
Console.WriteLine("sum(a) = {0}", dict["a"]); //prints sum(a) = 6

A wonderful Dataflow demo, right? A splitter block who cuts kv pair strings connects to an aggregator block who sums value for every key. So far so good if we need this dataflow only once. But what if I need the same dataflow graph somewhere else in my application? Or, expose the same functionality as reusable components in a library?

Things are getting complicated. Obviously Copy&Paste is not an acceptable choice. What about put everything about the dataflow construction in a static method? Hmmm, this is a step forward to reuse the code but, what should be the return value of the static method as the handle of the graph? Returning the head block helps posting new data to the pipeline but then we miss the tail block which we need to wait completion on. Not to mention the dict variable which contains our state/data... Last but not least, static method is an anti-pattern for testing as you can hardly change any behavior of underlying blocks.

Clearly we need a class representing the graph and being the handle of all the stakeholders. Object oriented design is a perfect fit here to solve all problems mentioned above. That is why we gave birth to Gridsum.DataflowEx.

3. Introduction to DataflowEx

Code tells a lot. Let's migrate the above example to DataflowEx and see what it looks like:

using Gridsum.DataflowEx;
using System.Threading.Tasks.Dataflow;

public class AggregatorFlow : Dataflow<string>
{
    //Blocks
    private TransformBlock<string, KeyValuePair<string, int>> _splitter; 
    private ActionBlock<KeyValuePair<string, int>> _aggregater;

    //Data
    private Dictionary<string, int> _dict;

    public AggregatorFlow() : base(DataflowOptions.Default)
    {
        _splitter = new TransformBlock<string, KeyValuePair<string, int>>(s => this.Split(s));
        _dict = new Dictionary<string, int>();
        _aggregater = new ActionBlock<KeyValuePair<string, int>>(p => this.Aggregate(p));

        //Block linking
        _splitter.LinkTo(_aggregater, new DataflowLinkOptions() { PropagateCompletion = true });

        /* IMPORTANT */
        RegisterChild(_splitter);
        RegisterChild(_aggregater);
    }

    protected virtual void Aggregate(KeyValuePair<string, int> pair)
    {
        int oldValue;
        _dict[pair.Key] = this._dict.TryGetValue(pair.Key, out oldValue) ? oldValue + pair.Value : pair.Value;
    }

    protected virtual KeyValuePair<string, int> Split(string input)
    {
        string[] splitted = input.Split('=');
        return new KeyValuePair<string, int>(splitted[0], int.Parse(splitted[1]));
    }

    public override ITargetBlock<string> InputBlock { get { return _splitter; } }

    public IDictionary<string, int> Result { get { return _dict; } }
}

Though there seems to be more code, it is quite clear. We have a class AggregatorFlow representing our flow finally, which inherits from Dataflow<TIn> with type parameter string. This means the AggregatorFlow class reprensents a dataflow graph itself and accepts strings as input.

In this form, dataflow blocks and data become class members. Block behaviors become class methods (which allows the outside to override!). We also implemented the abstract InputBlock property of Dataflow<TIn> and exposes our internal data as an extra Result property.

There are two important calls to RegisterChild() in the constructor. We will come back to this later.

Now let's come to the consumer side of the AggregatorFlow class:

var aggregatorFlow = new AggregatorFlow();
aggregatorFlow.InputBlock.Post("a=1");
aggregatorFlow.InputBlock.Post("b=2");
aggregatorFlow.InputBlock.Post("a=5");
aggregatorFlow.InputBlock.Complete();
await aggregatorFlow.CompletionTask;
Console.WriteLine("sum(a) = {0}", aggregatorFlow.Result["a"]); //prints sum(a) = 6

You see that we now operate on a single instance of AggregatorFlow, without knowing the implementation detail of it. This gives you the possibility to encapsulate complex dataflow logic in your Dataflow class and pass on your consumers a clean high-level view of the graph.

Note: The Dataflow class exposes a CompletionTask property (just like IDataflowBlock.Completion) to represent the life of the whole dataflow. The whole dataflow won't complete till every single child block in the flow completes . Here in the example we await on the task to make sure the dataflow completes. More on this topic below.

By the way, Dataflow<TIn> provides some helper methods to boost productivity. To achieve the same effect:

var aggregatorFlow = new AggregatorFlow();
await aggregatorFlow.ProcessAsync(new[] { "a=1", "b=2", "a=5" }, completeFlowOnFinish:true);
Console.WriteLine("sum(a) = {0}", aggregatorFlow.Result["a"]); //prints sum(a) = 6

It is now that easy with ProcessAsync as DataflowEx handles the tedious Post-and-complete boilerplate code for you :)

This is the basic idea of DataflowEx which empowers you with a fully functional handle of your dataflow graph. Find more in the following topics.

4. Understanding DataflowEx design

Just like IDataflowBlock is the fundamental piece in TPL Dataflow, IDataflow is the counterpart in DataflowEx library. Take a look at the IDataflow design:

public interface IDataflow
{
    IEnumerable<IDataflowBlock> Blocks { get; }
    Task CompletionTask { get; }
    void Fault(Exception exception);
    string Name { get; }    
}

public interface IDataflow<in TIn> : IDataflow
{
    ITargetBlock<TIn> InputBlock { get; }
}

public interface IOutputDataflow<out TOut> : IDataflow
{
    ISourceBlock<TOut> OutputBlock { get; }
    void LinkTo(IDataflow<TOut> other);
}

public interface IDataflow<in TIn, out TOut> : IDataflow<TIn>, IOutputDataflow<TOut>
{
}

IDataflow looks like IDataflowBlock, doesn't it? Well, remember IDataflow now represents a dataflow graph which may contain one or more low-level blocks. Since a graph may have inputs and outputs, those strongly-typed generic IDataflow interfaces are designed.

Note: If you see IOutputDataflow<TOut>.LinkTo(IDataflow<TOut> other), congratulations to you as you find out the API supports (and encourages) graph level data linking.

So on top of IDataflow there is an implementation called Dataflow, which should be the base class for all DataflowEx flows. Besides acting as the handle of the graph, it has many useful functionalities built-in. Let's explore them one by one.

4.1 Lifecycle management

A key role of the Dataflow base class is to monitor the health of its children and provides a single completion state to the outside, namely the CompletionTask property.

So first things first, we need a way to tell the dataflow who is its child. That is done through the Dataflow.RegisterChild method (you have seen it in the last example). Dataflow class will now keep the reference of the child in its internal data structure and the lifecycle of child will start to affect its parent.

Note: RegisterChild() method is not restricted to be called inside dataflow constructor. In fact, it can be called by outside invokers and used wherever necessary. Dataflow class uses a smart mechanism here to ensure dynamically registered child will affect the Dataflow's CompletionTask, even if you acquire the CompletionTask reference beforehand. This feature empowers the scenario that your dataflow is changing its shape at runtime. Just don't forget to call RegisterChild() when new child is created on demand.

There are 2 kinds of child that you can register, a dataflow block or a sub dataflow. The latter means Dataflow nesting is supported! Feel free to build different levels of dataflow components to provide even better modularity and encapsulation. Let's look at an example:

using System.Threading.Tasks.Dataflow;
using Gridsum.DataflowEx;

public class LineAggregatorFlow : Dataflow<string>
{
    private Dataflow<string, string> _lineProcessor;
    private AggregatorFlow _aggregator;
    public LineAggregatorFlow() : base(DataflowOptions.Default)
    {
        _lineProcessor = new TransformManyBlock<string, string>(line => line.Split(' ')).ToDataflow();
        _aggregator = new AggregatorFlow();
        _lineProcessor.LinkTo(_aggregator);
        RegisterChild(_lineProcessor);
        RegisterChild(_aggregator);
    }

    public override ITargetBlock<string> InputBlock { get { return _lineProcessor.InputBlock; } }
    public int this[string key] { get { return _aggregator.Result[key]; } }
}

//consumer here
var lineAggregator = new LineAggregatorFlow();
await lineAggregator.ProcessAsync(new[] { "a=1 b=2 a=5", "c=6 b=8" });
Console.WriteLine("sum(a) = {0}", lineAggregator["a"]); //prints sum(a) = 6

The example builds a LineAggregatorFlow on top of the existing AggregatorFlow to provide further functionalities. You get the idea how existing modules can be reused and seamlessly integrated to build a clearly-designed sophisticated dataflow graph.

Tip: Notice that instead of creating a Dataflow class for _lineProcessor, IPropagatorBlock<TIn, TOut>.ToDataflow() is used to avoid class creation as we just want a trivial wrapper over the delegate. This extension method is defined in DataflowUtils.cs where there are more helpers to convert from blocks or delegates to Dataflows.

Back to the topic of lifecycle, when a child is registered (no matter it is a block or sub flow), how will it affect the parent? The following rules answer the question:

  • The parent comes to its completion only when all of its children completes.
  • The parent fails if any of its children fails.
  • When one of the children fails, the parent will notify and wait other children to shutdown, and then comes to Faulted state itself.

So, in this form, the parent takes care of each child to guarantee nothing is wrong. Whenever something bad happens, the parent dataflow takes a fast-fail approach and propagates the error to other running children so they comes to a grinding halt.

Notice: It is common need for Dataflow instances to gracefully stop, no matter it is a natural completion or exceptional shutdown (E.g. A file writer may want to flush buffered data to disk). To meet this need Dataflow class provides a virtual method CleanUp() which will be called after all children of the dataflow is finished or faulted. You can override this method to provide your customized tear-down operation for your dataflow. In case of exceptional shutdown, the exception object will be assigned to the input parameter of CleanUp(); in case of a happy ending, the exception parameter value is null. But bear in mind that, when there is an error, this method doesn't change the fact that the dataflow is down. It is just your last chance to do some work before the original exception gets thrown eventually.

This is the general idea about lifecyle management in DataflowEx. A parent keeps his child under umbrella and never leaves any baby behind.

4.2 Graph construction

Normally you construct your graph in the constructor of your own dataflow class which inherits from Dataflow. There are typically 3 steps to construct a dataflow graph:

  1. Create dataflow blocks or sub-flow instances
  2. Connect flows and blocks to shape a network
  3. Register flows and blocks as children

As you can see, previous examples all follows the same pattern.

The 2nd step is worth digging into here. If you are dealing with raw blocks, ISourceBlock.LinkTo is your friend. And probably you want to set DataflowLinkOptions.PropagateCompletion to true if you want completion to be passed down automatically. This is traditional TPL Dataflow linking, as demonstrated by class AggregatorFlow.

Tip: When programming TPL Dataflow, how many times do you find your blocks never complete? And how many times do you find out the reason to be forgetting to set ? :)

But the real connecting power resides in the dataflow level connecting. DataflowEx put some effort here to provide a number of utilites and best practices to help graph construction: Dataflow classes have rich linking APIs. So you don't bother call block-level linking any more.

The first to mention is IOutputDataflow.LinkTo (as demonstrated in LineAggregatorFlow), counterpart of the low level ISourceBlock.LinkTo. As its name implies, it connects the output block to the input block of the given parameter, and propagates completion by default. DataflowEx encourages completion propagation.

There is one more secret about IOutputDataflow<TOut>.LinkTo (default implementation in Dataflow<TIn, TOut>): it supports one target dataflow being linked-to multiple times and guarantees the target dataflow receives a completion signal only when ALL upstream dataflows complete. Notice this behavior is different from block level linking with PropagateCompletion set to true, which means the target block receives a completion signal when any of the upstream blocks completes. We think our choice is what you need in most cases.

Does Dataflow class allow an orphan child that links to no one and is not linked to? Yes. Your graph need not be a fully-connected one if you wish. You can have 'islands'. But you take care of by yourself how these islands receives a completion signal. Please always ensure that completion signal will be correctly propagated along the dataflow chain when your job is done. If a child never gets a completion signal, the parent's CompletionTask will never come to an end.

Note: Dataflow class doesn't require children to be connected. The dataflow network/linking is constructed at your wish. So if a children will not get completion signal automatically through linking, you need to manually complete it on some condition, or you can use RegisterDependency(). The orphan child will complete when all of its dependencies complete. RegisterDependency() has nothing to do with data. It only handles completion.

Tip: To tell you the truth, Dataflow.LinkTo() also uses RegisterDependency() internally to achieve the 'WhenAll' behavior.

All right. Time for a demo to show the graph construction and complex completion propagation.

public class ComplexIntFlow : Dataflow<int>
{
    private ITargetBlock<int> _headBlock;
    public ComplexIntFlow() : base(DataflowOptions.Default)
    {
        Dataflow<int, int> node2 = DataflowUtils.FromDelegate<int, int>(i => i);
        Dataflow<int, int> node3 = DataflowUtils.FromDelegate<int, int>(i => i * -1)

        Dataflow<int, int> node1 = DataflowUtils.FromDelegate<int, int>(
            i => {
                    if (i % 2 == 0) { node2.Post(i); }
                    else { node3.Post(i); }
                    return 999;
                });
        
        Dataflow<int> printer = DataflowUtils.FromDelegate<int>(Console.WriteLine);

        node1.Name = "node1";
        node2.Name = "node2";
        node3.Name = "node3";
        printer.Name = "printer";

        node1.LinkTo(printer);
        node2.LinkTo(printer);
        node3.LinkTo(printer);

        //Completion propagation: node1 ---> node2
        node2.RegisterDependency(node1);
        //Completion propagation: node1 + node2 ---> node3
        node3.RegisterDependency(node1);
        node3.RegisterDependency(node2);

        this.RegisterChild(node1);
        this.RegisterChild(node2);
        this.RegisterChild(node3);
        this.RegisterChild(printer, t => { 
            if (t.Status == TaskStatus.RanToCompletion) 
                Console.WriteLine("Printer done!");
        });

        this._headBlock = node1.InputBlock;
    }

    public override ITargetBlock<int> InputBlock { get { return this._headBlock; } }
}

//Consumer
var intFlow = new ComplexIntFlow();
await intFlow.ProcessAsync(new[] { 1, 2, 3});

Code tells :) In this example (1) Node1, Node2 and Node3 all flow to the printer node. (2) Node2's life cycle depends on Node1. (3) Node3 depends on Node1 and Node2. So when the graph comes to it completion, the order will be Node1 -> Node2 -> Node3 -> Printer. This is powered by linking and dependency registration built-in in DataflowEx.

4.3 Logging is your friend

To help you better understand how DataflowEx works and sometimes diagnose your application, DataflowEx provides extensive logging where we think necessary. So you get insights of the blocks/dataflows managed by DataflowEx without tedious debugging. Please just check the log, which is always our first advice.

Technically Common.Logging is used as the underlying logging framework so that it can easily be integrated to your own logging system with an adapter, no matter you are using Log4Net, NLog or Enterprise Library Logging. For detailed documentation on Common.Logging, please checkout this link.

I'll use NLog as an example to print out logging of ComplexIntFlow. In the app.config, configure a NLogLoggerFactoryAdapter to route common.logging messages to NLog.

<configuration>
  <configSections>
    <sectionGroup name="common">
      <section name="logging" type="Common.Logging.ConfigurationSectionHandler, Common.Logging" requirePermission="false" />
    </sectionGroup>
  </configSections>
  
  <common>
    <logging>
      <factoryAdapter type="Common.Logging.NLog.NLogLoggerFactoryAdapter, Common.Logging.NLog20">
        <arg key="configType" value="FILE" />
        <arg key="configFile" value="~/NLog.config" />
      </factoryAdapter>
    </logging>
  </common>
</configuration>

Meanwhile, set up NLog in you NLog.config:

<nlog xmlns="http://www.nlog-project.org/schemas/NLog.xsd"
      xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">

  <variable name="logFormat" value="${date:format=yy/MM/dd HH\:mm\:ss} [${logger}].[${level}] ${message} ${exception:format=tostring} "/>

  <targets>
    <target xsi:type="Console" name="console" layout="${logFormat}"/>
    <target xsi:type="File" name ="file" fileName="Gridsum.DataflowEx.Demo.log" layout="${logFormat}" keepFileOpen="true"/>
  </targets>

  <rules>
    <logger name ="Gridsum.DataflowEx*" minlevel="Trace" writeTo="console,file"></logger>
  </rules>
</nlog>

Then the logging infomation will be written to both the console and a log file.

Note: Notice that "Gridsum.DataflowEx*" is used as the logger name in the rules section. This represents the logging categories used by DataflowEx. In fact, most logging of DataflowEx is recorded under category "Gridsum.DataflowEx" and a few uses subcategories with the prefix (e.g. "Gridsum.DataflowEx.Databases"). So an extra wildcard ‘*’ includes every message from the library.

O.K. Let's take a look of the log file after executing the ComplexIntFlow demo:

14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[printer] now has 2 dependencies.  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[printer] now has 3 dependencies.  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[node3] now has 2 dependencies.  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1] Successfully pulled and posted 3 Int32s to [ComplexIntFlow1]->[node1].  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1] Finished reading from enumerable and posting to the dataflow.  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1] Telling myself there is no more input and wait for children completion  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[node1] completed  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[node2] completed  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[node3] All of my dependencies are done. Completing myself.  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[node3] completed  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[printer] All of my dependencies are done. Completing myself.  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1]->[printer] completed  
14/09/28 11:18:33 [Gridsum.DataflowEx].[Info] [ComplexIntFlow1] completed  

Nice and clear, isn't it? The output also proves that dataflow dependency works correctly introduced in the last topic.

Logging provides more than lifecycle information. It includes block/dataflow buffer information as well, which is extremely helpful when checking dataflow health, finding bottlenecks and diagnosing deadlock problems. If you need flow-level buffer monitor, set DataflowOptions.FlowMonitorEnabled to true and then pass the DataflowOptions object to the constructor of dataflow classes. If you need block-level buffer information, set DataflowOptions.BlockMonitorEnabled to true. With everything set up, the Dataflow base class will start an asynchronous loop to check and log states of all its children with an given interval (the default interval is 10 seconds).

To demonstrate buffer monitoring logging, let's create a SlowFlow on purpose:

public class SlowFlow : Dataflow<string>
{
    private Dataflow<string, char> _splitter;
    private Dataflow<char> _printer;

    public SlowFlow(DataflowOptions dataflowOptions)
        : base(dataflowOptions)
    {
        _splitter = new TransformManyBlock<string, char>(new Func<string, IEnumerable<char>>(this.SlowSplit), 
            dataflowOptions.ToExecutionBlockOption())
            .ToDataflow(dataflowOptions, "SlowSplitter");

        _printer = new ActionBlock<char>(c => Console.WriteLine(c),
            dataflowOptions.ToExecutionBlockOption())
            .ToDataflow(dataflowOptions, "Printer");

        RegisterChild(_splitter);
        RegisterChild(_printer);

        _splitter.LinkTo(_printer);
    }

    private IEnumerable<char> SlowSplit(string s)
    {
        foreach (var c in s)
        {
            Thread.Sleep(1000); //slow down
            yield return c;
        }
    }

    public override ITargetBlock<string> InputBlock { get { return _splitter.InputBlock; } }
}

//consumer
var slowFlow = new SlowFlow( new DataflowOptions
            {
                FlowMonitorEnabled = true, 
                MonitorInterval = TimeSpan.FromSeconds(2),
                PerformanceMonitorMode = DataflowOptions.PerformanceLogMode.Verbose
            });

await slowFlow.ProcessAsync(new[]
                                {
                                    "abcd", 
                                    "abc", 
                                    "ab", 
                                    "a"
                                });

The slow flow class has a slow splitter block if you see the Thread.Sleep(). Then we properly set some options when constructing a SlowFlow instance:

  1. FlowMonitorEnabled tells the flow to log its buffer status periodically. Since the dataflow option is normally passed on to its children. Child flows will also log their buffer status.
  2. MonitorInterval sets the time interval between two adjacent buffer status calls.
  3. PerformanceMonitorMode sets whether or not to output logging information if a flow has 0 items buffered. Setting it to Verbose will log empty flows while setting to Succint will not.

Let's have a look at the log after executing the above code:

14/09/29 11:38:00 [Gridsum.DataflowEx].[Info] [SlowFlow1] Telling myself there is no more input and wait for children completion  
14/09/29 11:38:02 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[SlowSplitter] has 3 todo items (in:3, out:0) at this moment.  
14/09/29 11:38:02 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1] has 3 todo items (in:3, out:0) at this moment.  
14/09/29 11:38:02 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[Printer] has 0 todo items (in:0, out:0) at this moment.  
14/09/29 11:38:04 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[Printer] has 0 todo items (in:0, out:0) at this moment.  
14/09/29 11:38:04 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1] has 3 todo items (in:3, out:0) at this moment.  
14/09/29 11:38:04 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[SlowSplitter] has 3 todo items (in:3, out:0) at this moment.  
14/09/29 11:38:06 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[Printer] has 0 todo items (in:0, out:0) at this moment.  
14/09/29 11:38:06 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1] has 2 todo items (in:2, out:0) at this moment.  
14/09/29 11:38:06 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[SlowSplitter] has 2 todo items (in:2, out:0) at this moment.  
14/09/29 11:38:08 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[SlowSplitter] has 1 todo items (in:1, out:0) at this moment.  
14/09/29 11:38:08 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1] has 1 todo items (in:1, out:0) at this moment.  
14/09/29 11:38:08 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[Printer] has 0 todo items (in:0, out:0) at this moment.  
14/09/29 11:38:10 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[Printer] has 0 todo items (in:0, out:0) at this moment.  
14/09/29 11:38:10 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1] has 0 todo items (in:0, out:0) at this moment.  
14/09/29 11:38:10 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[SlowSplitter] has 0 todo items (in:0, out:0) at this moment.  
14/09/29 11:38:10 [Gridsum.DataflowEx].[Info] [SlowFlow1]->[SlowSplitter] completed  
14/09/29 11:38:10 [Gridsum.DataflowEx].[Info] [SlowFlow1]->[Printer] completed  
14/09/29 11:38:10 [Gridsum.DataflowEx].[Info] [SlowFlow1] completed  

Now we easily find out the bottle neck of the flow to be [SlowFlow1]->[SlowSplitter], which has more todo items than other modules.

What if you have a flow that contains many blocks? Fortunately there is another property on DataflowOptions you can set, BlockMonitorEnabled, to disclose block level buffer status. After setting it to true, you get logs like:

14/09/29 11:53:55 [Gridsum.DataflowEx.PerfMon].[Debug] [SlowFlow1]->[SlowSplitter]->(TransformManyBlock<String, Char>) has 3 todo items (in:3, out:0) at this moment.   

The evil block has nowhere to escape :)

Note: DataflowEx provides a default DataflowOptions instance: DataflowOptions.Default. It has FlowMonitorEnabled set to true, BlockMonitorEnabled set to false, MonitorInterval as 10 seconds and PerformanceMonitorMode being Succint.

4.4 Error handling

If we look back to the very first Dataflow implementation, the AggregatorFlow class, there is a couple of conditions required for the input, in a silent way. For example, in the Split method:

protected virtual KeyValuePair<string, int> Split(string input)
{
    string[] splitted = input.Split('=');
    return new KeyValuePair<string, int>(splitted[0], int.Parse(splitted[1]));
}

If the input does not obey the "key={int}" format, an exception will be thrown. So, what is next? What will happen to the Dataflow graph?

DataflowEx takes a fast-fail approach on exception handling just like TPL Dataflow. When an exception is thrown, the low-level block ends to the Faulted state first. Then the Dataflow instance who is the parent of the failing block gets notified. It will immediately propagate the fatal error: notify its other children to shutdown immediately. After all its children is done/completed, the parent Dataflow also comes to its completion, with the original exception wrapped in the CompletionTask whose status is also Faulted.

In the case of Dataflow nesting, the exception will be raised all the way from leaf to root, according to the child registration tree. One thing to remember is that, DataflowEx never ever swallows an unhandled exception. The library propagates it, store it in the CompletionTask and finally throw it where the top level Dataflow's CompletionTask is awaited.

Thus for AggregatorFlow, if we pass in some invalid input (e.g. "a=badstring"), an unhandled exception is thrown on Main method (Exception will also be logged by DataflowEx internally):

Unhandled Exception: System.AggregateException: One or more errors occurred. ---> System.AggregateException: One or more
 errors occurred. ---> System.FormatException: Input string was not in a correct format.
   at System.Number.StringToNumber(String str, NumberStyles options, NumberBuffer& number, NumberFormatInfo info, Boolea
n parseDecimal)
   at System.Number.ParseInt32(String s, NumberStyles style, NumberFormatInfo info)
   at Gridsum.DataflowEx.Demo.AggregatorFlow.Split(String input) in c:\Users\karld_000\Documents\SourceTree\DataflowEx\G
ridsum.DataflowEx.Demo\AggregatorFlow.cs:line 41
   at System.Threading.Tasks.Dataflow.TransformBlock`2.ProcessMessage(Func`2 transform, KeyValuePair`2 messageWithId)
   at System.Threading.Tasks.Dataflow.TransformBlock`2.<>c__DisplayClass10.<.ctor>b__3(KeyValuePair`2 messageWithId)
   at System.Threading.Tasks.Dataflow.Internal.TargetCore`1.ProcessMessagesLoopCore()
--- End of stack trace from previous location where exception was thrown ---
   at System.Runtime.CompilerServices.TaskAwaiter.ThrowForNonSuccess(Task task)
   at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
   at Gridsum.DataflowEx.Exceptions.TaskEx.<AwaitableWhenAll>d__3`1.MoveNext() in c:\Users\karld_000\Documents\SourceTre
e\DataflowEx\Gridsum.DataflowEx\Exceptions\TaskEx.cs:line 59
--- End of stack trace from previous location where exception was thrown ---
   at System.Runtime.CompilerServices.TaskAwaiter.ThrowForNonSuccess(Task task)
   at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
   at Gridsum.DataflowEx.Exceptions.TaskEx.<AwaitableWhenAll>d__3`1.MoveNext() in c:\Users\karld_000\Documents\SourceTre
e\DataflowEx\Gridsum.DataflowEx\Exceptions\TaskEx.cs:line 59
   --- End of inner exception stack trace ---
   at Gridsum.DataflowEx.Exceptions.TaskEx.<AwaitableWhenAll>d__3`1.MoveNext() in c:\Users\karld_000\Documents\SourceTre
e\DataflowEx\Gridsum.DataflowEx\Exceptions\TaskEx.cs:line 71
--- End of stack trace from previous location where exception was thrown ---
   at System.Runtime.CompilerServices.TaskAwaiter.ThrowForNonSuccess(Task task)
   at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
   at Gridsum.DataflowEx.Dataflow.<GetCompletionTask>d__1c.MoveNext() in c:\Users\karld_000\Documents\SourceTree\Dataflo
wEx\Gridsum.DataflowEx\Dataflow.cs:line 345
--- End of stack trace from previous location where exception was thrown ---
   at System.Runtime.CompilerServices.TaskAwaiter.ThrowForNonSuccess(Task task)
   at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
   at Gridsum.DataflowEx.Demo.Program.<CalcAsync>d__6.MoveNext() in c:\Users\karld_000\Documents\SourceTree\DataflowEx\G
ridsum.DataflowEx.Demo\Program.cs:line 66
   --- End of inner exception stack trace ---
   at System.Threading.Tasks.Task.ThrowIfExceptional(Boolean includeTaskCanceledExceptions)
   at System.Threading.Tasks.Task.Wait(Int32 millisecondsTimeout, CancellationToken cancellationToken)
   at Gridsum.DataflowEx.Demo.Program.Main(String[] args) in c:\Users\karld_000\Documents\SourceTree\DataflowEx\Gridsum.
DataflowEx.Demo\Program.cs:line 50

Hence, if you think the exception is expected or tolerable, please catch it in the very beginning in the delegate passed to the low level block. Otherwise, the domino effect will spread and fail the whole dataflow graph.

5. DataflowEx In Depth

5.1 Advanced Linking

LinkTo() is the most powerful mechanism TPL Dataflow provides to help intuitive and efficient dataflow graph construction. There is also an overload of LinkTo() that accepts a predicate as the filter for conditional linking.

DataflowEx also provides similar LinkTo() at higher level (i.e. the Dataflow level), as you already see in previous demos. It uses block level LinkTo under the hood but you just easily operate on Dataflow nodes no matter how complex these node are internally. As completion propagation is concerned, DataflowEx's LinkTo() ensures completion propagation by default and goes beyond that: if A->C and B->C, Dataflow C will complete after both A and B complete. This feature has been mentioned in the previous Graph construction section but I want to re-emphasize here.

Anything more on linking? Yes. Let's consider how many times you have a dataflow generating a serious of base class objects and wants to redirect different child types to different specific components that only handles the child type? LinkSubTypeTo() is born for this:

//public void LinkSubTypeTo<TTarget>(IDataflow<TTarget> other) where TTarget : TOut

Dataflow<TIn, TOut> flow1;
Dataflow<TOutSubType1> flow2;
Dataflow<TOutSubType2> flow3;

flow1.LinkSubTypeTo(flow2);
flow1.LinkSubTypeTo(flow3);

LinkSubTypeTo() make our life so much easier :)

Internally LinkSubTypeTo() uses a even more powerful linking mechanism called TransformAndLink() which allows you to indicate a transform function on output objects as well as a filtering predicate. The transform function is a perfect bridge between output flow and downstream flows. It can be as simple as a type cast (LinkSubTypeTo), or a complex mapping function that converts, for example, a weakly-typed json object to your strongly typed domain objects.

Note: Previously to implement TransformAndLink(), DataflowEx used an extra TransformBlock to achieve this. But starting from 1.0.9.6, we avoid the overhead of TransformBlock and push transform function down to the LinkPropagator level (a no-buffer block wrapper TPL dataflow uses to achieve link filtering), which brings better performance.

Let's look at a demo for TransformAndLink():

public static async Task TransformAndLinkDemo()
{
    var d1 = new BufferBlock<int>().ToDataflow();
    var d2 = new ActionBlock<string>(s => Console.WriteLine(s)).ToDataflow();
    var d3 = new ActionBlock<string>(s => Console.WriteLine(s)).ToDataflow();

    d1.TransformAndLink(d2, _ => "Odd: "+ _, _ => _ % 2 == 1);
    d1.TransformAndLink(d3, _ => "Even: " + _, _ => _ % 2 == 0);

    for (int i = 0; i < 10; i++)
    {
        d1.Post(i);
    }

    await d1.SignalAndWaitForCompletionAsync();
    await d2.CompletionTask;
    await d3.CompletionTask;
}

By using TransformAndLink, we direct odd numbers to d1 and even numbers to d2, and at the same time converts integers to meaningful strings to be printed out:

Even: 0
Even: 2
Even: 4
Even: 6
Even: 8
Odd: 1
Odd: 3
Odd: 5
Odd: 7
Odd: 9

The final linking helper in DataflowEx to introduce is LinkLeftTo(). When we use conditional linking extensively in our application, there is a common trap that if one output matches NONE of output linking predicates, it will remain forever in the upstream dataflow, and thus our dataflow graph never completes. To address the issue, DataflowEx trackes all predicates the dataflow used previously and allows you to redirect LEFT objects (i.e. output objects that match no predicates) to a given target:

Demo time:

public static async Task LinkLeftToDemo()
{
    var d1 = new BufferBlock<int>().ToDataflow(name: "IntGenerator");
    var d2 = new ActionBlock<int>(s => Console.WriteLine(s)).ToDataflow();
    var d3 = new ActionBlock<int>(s => Console.WriteLine(s)).ToDataflow();
    var d4 = new ActionBlock<int>(s => Console.WriteLine(s + "[Left]")).ToDataflow();

    d1.LinkTo(d2, _ => _ % 3 == 0);
    d1.LinkTo(d3, _ => _ % 3 == 1);
    d1.LinkLeftTo(d4); // same as d1.LinkTo(d4, _ => _ % 3 == 2);

    for (int i = 0; i < 10; i++)
    {
        d1.Post(i);
    }

    await d1.SignalAndWaitForCompletionAsync();
    await d2.CompletionTask;
    await d3.CompletionTask;
}
//output:
//1
//2[Left]
//5[Left]
//8[Left]
//0
//3
//4
//7
//6
//9

Tip: In this demo, LinkLeftTo is based on the predicates previously used in LinkTo(). But to clarify, predicates used in TransformAndLink() will also be taken into account when DataflowEx calculates what stands for 'Left'.

Besides LinkLeftTo(), you could also use the handy LinkLeftToNull() if you just want to silently ignore those unmatched objects (will be linked to DataflowBlock.NullTarget()). Or, you can use LinkLeftToError() if it is a fatal error if there is an output object that matches none predicate conditions. This is a quite useful diagnosing tip to make sure your graph doesn't have a linking leak.

Note: When using LinkLeftToNull() there is an garbage recorder which counts different types of the ignored output. Override Dataflow.OnOutputToNull() to customize the behavior. Regarding recorders and the statistics you can get from recorder, we will dig into that later.

If we slightly modify the above demo to use LinkLeftToError():

public static async Task LinkLeftToDemo()
{
    var d1 = new BufferBlock<int>().ToDataflow(name: "IntGenerator");
    var d2 = new ActionBlock<int>(s => Console.WriteLine(s)).ToDataflow();
    var d3 = new ActionBlock<int>(s => Console.WriteLine(s)).ToDataflow();
    var d4 = new ActionBlock<int>(s => Console.WriteLine(s + "[Left]")).ToDataflow();

    d1.LinkTo(d2, _ => _ % 3 == 0);
    d1.LinkTo(d3, _ => _ % 3 == 1);
    d1.LinkLeftToError();//d1.LinkLeftTo(d4); 

    for (int i = 0; i < 10; i++)
    {
        d1.Post(i);
    }

    await d1.SignalAndWaitForCompletionAsync();
    await d2.CompletionTask;
    await d3.CompletionTask;
}

Exception will be thrown and d1 fails as expected:

14/10/22 12:28:39 [Gridsum.DataflowEx].[Error] [IntGenerator] This is my error destination. Data should not arrive here: 2

Unhandled Exception: 14/10/22 12:28:39 [Gridsum.DataflowEx].[Error] [IntGenerator] Exception occur. Shutting down my chi
ldren... System.IO.InvalidDataException: An object came to error region of [IntGenerator]: 2
   at Gridsum.DataflowEx.Dataflow`2.<LinkLeftToError>b__5c(TOut survivor) in c:\Users\karld_000\Documents\SourceTree\Dat
aflowEx\Gridsum.DataflowEx\Dataflow.cs:line 846
   at System.Threading.Tasks.Dataflow.ActionBlock`1.ProcessMessageWithTask(Func`2 action, KeyValuePair`2 messageWithId)
--- End of stack trace from previous location where exception was thrown ---
   at System.Runtime.CompilerServices.TaskAwaiter.ThrowForNonSuccess(Task task)
   at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
   at Gridsum.DataflowEx.Exceptions.TaskEx.<AwaitableWhenAll>d__3`1.MoveNext() in c:\Users\karld_000\Documents\SourceTre
e\DataflowEx\Gridsum.DataflowEx\Exceptions\TaskEx.cs:line 59
...

To sum up, a) DataflowEx provides its own versatile LinkTo() system and it is preferred over raw TPL Dataflow LinkTo(). b) although a Dataflow instance may be a complex graph itself, it is treated a fundamental node in the data network when connecting to its siblings.

5.2 Cyclic graph and ring completion detection

Your dataflow graph becomes really, really complicated when there is a circult after linking components properly according to your application logic. Consider a real world example, a web crawler, which has an http request maker component and a link analysis component: they consume messages from and provide resources to each other.

A cyclic graph, or a ring, is well supported by the linking mechanism of TPL Dataflow, as far as data propagation is concerned. Messages do flow fluently along the linking arrows, even if there is a ring. But when it comes to graph completion, things become tricky: A depends on B's output, so probably A shouldn't end until B does. Well, B also depends on A... So now, the graph never has a chance to complete? Hmmm...

You may want to shutdown one of the ring components manually and expect the completion propagates along the ring. But chances are that other components are still busy processing and producing new messages. When these new messages flow to the component that are already shut down, they cannot continue their trip and will stay in the buffer of working blocks. So some components never complete because of thier buffer status. So it is a dead lock.

Clearly, a subtle mechanism is needed to automatically shutdown one proper component at the right time. The right time is, when there is nothing left in the ring and every node is idle; 'Automatcially' means you could hardly tell when to start the ring-completion domino safely so a smart background scheduler is needed to pull the trigger at 'the right time'; 'Proper' means we should choose properly which node is the first to complete in the ring and its completion can lead to the full completion of the ring.

It is difficult that you implement this mechanism yourself. Luckily DataflowEx comes with a solution out of the box: RegisterChildRing(). It tells the Dataflow that some children is forming a ring and should enable the automatic smart auto-complete mechanism for the ring. Take a look at the method signature:

/// <summary>
/// Register a continuous ring completion check after the preTask ends.
/// When the checker thinks the ring is O.K. to shut down, it will pull the trigger 
/// of heartbeat node in the ring. Then nodes in the ring will shut down one by one
/// </summary>
/// <param name="preTask">The task after which ring check begins</param>
/// <param name="ringNodes">Child dataflows that forms a ring</param>
protected async void RegisterChildRing(Task preTask, params IRingNode[] ringNodes)

The preTask parameter gives you a way to delay the child ring monitoring loop until a pre-condition is satisfied. The ring nodes parameter array, as its name implies, is the child components that forms a ring. Notice that the required type of a ring node is IRingNode which inherits from IDataflow.

Note: IRingNode only add one property IsBusy to IDataflow. As TPL Dataflow doesn't expose too much information of a running block, this interface helps the ring monitor to tell whether the dataflow/block is working (e.g. executing an item transform in the TransformBlock). So a typical implementation is to set IsBusy to true in the first line of the transform method and set it back to false in the last line. We know this is a bit tedious but it's the best we can do at this moment without Microsoft TPL Dataflow improvement.

O.K. Demo time again. Let's consider we have a heater and a more powerful cooler that changes the air temperature. At last, when the temperature drops below 0 degree the flow should quit.

public class CircularFlow : Dataflow<int>
{
    private Dataflow<int, int> _buffer;

    public CircularFlow(DataflowOptions dataflowOptions) : base(dataflowOptions)
    {
        //a no-op node to demonstrate the usage of preTask param in RegisterChildRing
        _buffer = new BufferBlock<int>().ToDataflow(name: "NoOpBuffer"); 

        var heater = new TransformManyDataflow<int, int>(async i =>
                {
                    await Task.Delay(200); 
                    Console.WriteLine("Heated to {0}", i + 1);
                    return new [] {i + 1};
                }, dataflowOptions);
        heater.Name = "Heater";

        var cooler = new TransformManyDataflow<int, int>(async i =>
                {
                    await Task.Delay(200);
                    int cooled = i - 2;
                    Console.WriteLine("Cooled to {0}", cooled);

                    if (cooled < 0) //time to stop
                    {
                        return Enumerable.Empty<int>(); 
                    }

                    return new [] {cooled};
                }, dataflowOptions);
        cooler.Name = "Cooler";

        var heartbeat = new HeartbeatNode<int>(dataflowOptions) {Name = "HeartBeat"};
        
        _buffer.LinkTo(heater);

        //circular
        heater.LinkTo(cooler);
        cooler.LinkTo(heartbeat);
        heartbeat.LinkTo(heater);

        RegisterChildren(_buffer, heater, cooler, heartbeat);
        
        //ring registration
        RegisterChildRing(_buffer.CompletionTask, heater, cooler, heartbeat);
    }

    public override ITargetBlock<int> InputBlock { get { return _buffer.InputBlock; } }
}

public static async Task CircularFlowAutoComplete()
{
    var f = new CircularFlow(DataflowOptions.Default);
    f.Post(20);
    await f.SignalAndWaitForCompletionAsync();
}

As you can see, there is a circular dependency in the example graph which theoratically never ends. But RegisterChildRing() does some magic to ensure the flow will end after jobs are done.

Note: RegisterChildRing() requires a heartbeat node in the ring to function correctly. You can use the HeartbeatNode class to initialize an instance like the example does. This heartbeat node, together with IsBusy property mentioned above, helps the ring monitor to correctly count the status of underlying blocks in case the flow is still running (internally ring monitor use a 2-round ring scan to ensure correctness of the ring status). The heartbeat node is also the first node to complete in the completion chain.

The output of the above code is listed below:

14/10/27 15:55:34 [Gridsum.DataflowEx].[Info] [Heater] now has 2 dependencies.
14/10/27 15:55:34 [Gridsum.DataflowEx].[Info] [CircularFlow1] A ring is set up: (Heater=>Cooler=>HeartBeat)
14/10/27 15:55:34 [Gridsum.DataflowEx].[Info] [CircularFlow1] Telling myself there is no more input and wait for childre
n completion
14/10/27 15:55:34 [Gridsum.DataflowEx].[Info] [CircularFlow1]->[NoOpBuffer] completed
14/10/27 15:55:34 [Gridsum.DataflowEx].[Info] [CircularFlow1] Ring pretask done. Starting check loop for (Heater=>Cooler
=>HeartBeat)
Heated to 11
Cooled to 9
Heated to 10
Cooled to 8
Heated to 9
Cooled to 7
Heated to 8
Cooled to 6
Heated to 7
Cooled to 5
Heated to 6
Cooled to 4
Heated to 5
Cooled to 3
Heated to 4
Cooled to 2
Heated to 3
Cooled to 1
Heated to 2
Cooled to 0
Heated to 1
Cooled to -1
14/10/27 15:55:44 [Gridsum.DataflowEx].[Debug] [CircularFlow1] 1st level empty ring check passed for (Heater=>Cooler=>He
artBeat)
14/10/27 15:55:54 [Gridsum.DataflowEx].[Debug] [CircularFlow1] 2nd level empty ring check passed for (Heater=>Cooler=>He
artBeat)
14/10/27 15:56:04 [Gridsum.DataflowEx].[Info] [CircularFlow1] Ring completion detected! Completion triggered on heartbea
t node: (Heater=>Cooler=>HeartBeat)
14/10/27 15:56:04 [Gridsum.DataflowEx].[Info] [CircularFlow1]->[HeartBeat] completed
14/10/27 15:56:04 [Gridsum.DataflowEx].[Info] [CircularFlow1]->[Heater] All of my dependencies are done. Completing myse
lf.
14/10/27 15:56:04 [Gridsum.DataflowEx].[Info] [CircularFlow1]->[Heater] completed
14/10/27 15:56:04 [Gridsum.DataflowEx].[Info] [CircularFlow1]->[Cooler] completed
14/10/27 15:56:04 [Gridsum.DataflowEx].[Info] [CircularFlow1] completed

DataflowEx provides extensive logging on cyclic flow events. Please don't forget to turn them on to help debugging.

Note: In this demo, the completion condition is relatively simple (temperature < 0) so yes it could be easily replaced by manual completion. But there are complex real world situations that are really hard to find an ending condition, especially when the count of output items of the transform-many function in some of the ring nodes is indeterministic. A spider is one of the complex examples. The DataflowEx built-in DbDataJoiner class is another one. These scenarios are where ring completion detection feature really shines.

5.3 Introducing StatisticsRecorder

Logging is our friend. It provides messages with detail to help diagnosing our applications. But sometimes we need overviews rather than detailed logging. This requirement is particularly important in a data flow because it might process billions of items per day and you don't want to drawn in the sea of details. For example, for a log parser program, you probably need an aggregater to sum up counts of error/warning messages in each category, rather than simply dumping everything into a single large log file.

That is why StatisticsRecorder is introduced, to be the built-in event counter/aggregator in DataflowEx, where you get overviews of your interest on your running dataflow. You can choose to access and dump statistics from the class at the end of your program, or at any other checkpoints.

StatisticsRecorder supports two kinds of recording: Type recording and event recording. Type recording is simply used to count the number of a particular type while event recording is an general-purpose extension feature where you can inject custom 'events' into the statistics recorder.

Note: StatisticsRecorder uses DataflowEvent class to store event information, which is composed of two strings: Level1 and Level2. Level1 is the parent node and Level2 is the child node of the hierarchy. StatisticsRecorder will aggregate event count on both attributes and supports queries like count(L1 = level1) and count(L1 = level1 && L2 = level2).

To record a type once, simply call StatisticsRecorder.RecordType(Type). To record an event, call RecordEvent(DataflowEvent) or its overloads. Both methods are thread-safe.

Built on top of RecordType() and RecordEvent(), StatisticsRecorder provides a Record(object) method that accepts any object:

/// <summary>
/// Records a processed object in dataflow pipeline
/// </summary>
/// <param name="instance">The object passing dataflow</param>
public virtual void Record(object instance)
{
    this.RecordType(instance.GetType());

    var eventProvider = instance as IEventProvider;
    if (eventProvider != null)
    {
        this.RecordEvent(eventProvider.GetEvent());
    }
}

In most cases you simply use Record() to monitor the object stream in your flow. It records the object type and extracts valuable event information wherever applicable. If the default implementation needs some adjustment in your scenario, don't forget the method is virtual.

Note: As shown, objects that carries event information should implement IEventProvider which allows Record() to get corresponding event information from the object. This is quite useful if you wish to include more information than just the type of the object in the statistics.

Let's look at a demo which processes person information stream. In this demo StatisticsRecorder not only keeps track of people count processed but also treats old person as special event.

public class Person : IEventProvider
{
    public string Name { get; set; }
    public int Age { get; set; }

    public DataflowEvent GetEvent()
    {
        if (Age > 70)
        {
            return new DataflowEvent("OldPerson");
        }
        else
        {
            //returning empty so it will not be recorded as an event
            return DataflowEvent.Empty; 
        }
    }
}

public class PeopleFlow : Dataflow<string, Person>
{
    private TransformBlock<string, Person> m_converter;
    private TransformBlock<Person, Person> m_recorder;
    private StatisticsRecorder m_peopleRecorder;
    
    public PeopleFlow(DataflowOptions dataflowOptions)
        : base(dataflowOptions)
    {
        m_peopleRecorder = new StatisticsRecorder(this) { Name = "PeopleRecorder"};

        m_converter = new TransformBlock<string, Person>(s => JsonConvert.DeserializeObject<Person>(s));
        m_recorder = new TransformBlock<Person, Person>(
            p =>
                {
                    //record every person
                    m_peopleRecorder.Record(p);
                    return p;
                });

        m_converter.LinkTo(m_recorder, new DataflowLinkOptions() { PropagateCompletion = true});

        RegisterChild(m_converter);
        RegisterChild(m_recorder);
    }

    public override ITargetBlock<string> InputBlock { get { return m_converter; } }
    public override ISourceBlock<Person> OutputBlock { get { return m_recorder; } }
    public StatisticsRecorder PeopleRecorder { get { return m_peopleRecorder; } }
}

public static async Task RecorderDemo()
{
    var f = new PeopleFlow(DataflowOptions.Default);
    var sayHello = new ActionBlock<Person>(p => Console.WriteLine("Hello, I am {0}, {1}", p.Name, p.Age)).ToDataflow(name: "sayHello");
    f.LinkTo(sayHello, p => p.Age > 0);
    f.LinkLeftToNull(); //object flowing here will be recorded by GarbageRecorder
    
    f.Post("{Name: 'aaron', Age: 20}");
    f.Post("{Name: 'bob', Age: 30}");
    f.Post("{Name: 'carmen', Age: 80}");
    f.Post("{Name: 'neo', Age: -1}");
    await f.SignalAndWaitForCompletionAsync();
    await sayHello.CompletionTask;

	Console.WriteLine("Total people count: " + f.PeopleRecorder[typeof(Person)]);
    Console.WriteLine(f.PeopleRecorder.DumpStatistics());
    Console.WriteLine(f.GarbageRecorder.DumpStatistics());
}

Tip: Although StatisticsRecorder allows you to access its data by indexers programatically, DumpStatistics() is the most convenient way to print out beautiful formatted overview gathered by StatisticsRecorder.

And its output:

Hello, I am aaron, 20
Hello, I am bob, 30
Hello, I am carmen, 80
Total people count: 4
[[PeopleFlow1]-PeopleRecorder] Entities: Person(4)
[[PeopleFlow1]-PeopleRecorder] Events: OldPerson(1)

[[PeopleFlow1]-GarbageRecorder] Entities: Person(1)
[[PeopleFlow1]-GarbageRecorder] Events:

As expected, people recorder captures the old person event as well as the total person object count. In addition, the GarbageRecorder is a handy statistics recorder embedded in the Dataflow class to monitor objects flowing to null target when using LinkLeftToNull(). It starts and functions implicitly.

To sum up, StatisticsRecorder is the aggregation engine in DataflowEx for reporting/monitoring purpose. Feel free to extend it and enrich your dataflow statistics.

6. Built-in Components

Here is another big reason why many projects inside Gridsum use DataflowEx: its powerful built-in components. Provided as generic reusable Dataflow classes, you get the their power out of the box. Data bulk insertion, data branching,

6.1 Bulk insertion support

The very 1st feature/component in DataflowEx we strongly recommend is the DbBulkInserter which enables bulk insertion to SQL Server. Though there are many ORM solutions out there but when it comes to insertion, they are just not designed to use the most efficient way: bulk insert. Using SqlBulkCopy internally, DbBulkInserter is born to solve the problem in a speedy way. And it nicely fits in dataflow style programming.

Inheriting from Dataflow, DbBulkInserter is a generic class and accepts a generic parameter T, which is the type of your domain objects. Just connect your output flows to DbBulkInserter. Bulk insertion magic will then happen internally in DbBulkInserter. The only extra thing to do is to set up column mapping from your domain object properties to database table columns, taking advantage of C# attributes.

In the last demo, we have a PeopleFlow that outputs Person objects. Let's dump those objects to SQL Server!

//Please note the added attributes to the person class
public class Person : IEventProvider
{
    [DBColumnMapping("LocalDbTarget", "NameCol", "N/A", ColumnMappingOption.Mandatory)]
    public string Name { get; set; }

    [DBColumnMapping("LocalDbTarget", "AgeCol", -1, ColumnMappingOption.Optional)]
    public int? Age { get; set; }

    public DataflowEvent GetEvent()
    {
        if (Age > 70)
        {
            return new DataflowEvent("OldPerson");
        }
        else
        {
            //returning empty so it will not be recorded as an event
            return DataflowEvent.Empty; 
        }
    }
}

public static async Task BulkInserterDemo()
{
    string connStr;

    //initialize table
    using (var conn = LocalDB.GetLocalDB("People"))
    {
        var cmd = new SqlCommand(@"
        IF OBJECT_id('dbo.People', 'U') IS NOT NULL
            DROP TABLE dbo.People;
        
        CREATE TABLE dbo.People
        (
            Id INT IDENTITY(1,1) NOT NULL,
            NameCol nvarchar(50) NOT NULL,
            AgeCol INT           NOT NULL
        )
        ", conn);
        cmd.ExecuteNonQuery();
        connStr = conn.ConnectionString;
    }

    var f = new PeopleFlow(DataflowOptions.Default);
    var dbInserter = new DbBulkInserter<Person>(connStr, "dbo.People", DataflowOptions.Default, "LocalDbTarget");
    f.LinkTo(dbInserter);

    f.Post("{Name: 'aaron', Age: 20}");
    f.Post("{Name: 'bob', Age: 30}");
    f.Post("{Age: 80}"); //Name will be default value: "N/A"
    f.Post("{Name: 'neo' }"); // Age will be default value: -1
    await f.SignalAndWaitForCompletionAsync();
    await dbInserter.CompletionTask;
}

Bulk insertion in .Net (namely SqlBulkCopy) used to be a very complex and heavyweight component to use. Now with DbBulkInserter, it's that easy! DbBulkInserter uses SqlBulkCopy internally and auto-generates the underlying property accessors to be used by SqlBulkCopy.

Let's put some words on the parameters of DBColumnMapping attribute constructor.

  • The 1st parameter is a string of your choice called 'destLabel' that defines a db target (i.e. your destination table). Mappings under same dest label form a group which defines a specific object relational mapper. Using different dest labels enables a type to be mapped to multiple table schemas: simply tag the attributes with multiple DBColumnMapping.
  • The 2nd parameter is the column name of your destination table. That's it.
  • The 3rd parameter is the default value to output instead if the property's value happen to be null. This works for reference types and nullable value types.
  • The last parameter of DBColumnMapping attribute constructor is a ColumnMappingOption, where you can indicate the mapping to be mandatory or optional. In the case of 'Optional', DataflowEx will simply ignore the given mapping (and warn you in the log) when a corresponding DB column is not found for the mapping. In the case of 'Mandatory', exception will be thrown.

Tip: The 'Optional' option could be quite useful when you want to maintain only one set of mappings (i.e. mappings with same destlabel) to match multiple tables that share a great part of their columns in common but has some minor schema difference.

Now take a glimpse of what's been put in the table, just as expected (Notice the default values are also taking effect):

ScreenShot

Notice: One thing that you may complain about attribute tagging is, db column mappings are pre-defined along with the domain object types. Chances are that your object types are defined in a dependent library which is completely out of your control. Yes, this was a problem. Luckily, starting from DataflowEx 1.1.0 TypeAccessorConfig.RegisterMapping() is added to enable manual mapping construction and mapping overwriting at runtime. Check out the API's unit test for how to use it.

If you want more insights of how DbBulkInserter works, please check the log where you get the internals of DbBulkInserter, especially how type properties are mapped to columns of a database table.

14/11/13 17:55:08 [Gridsum.DataflowEx.Databases.TypeAccessor<Person>].[Debug] Populated column offset for DBColumnMapping: [ColName:NameCol, ColOffset:1, DefaultValue:N/A] on property node: Person->Name by table dbo.People  
14/11/13 17:55:08 [Gridsum.DataflowEx.Databases.TypeAccessor<Person>].[Debug] Populated column offset for DBColumnMapping: [ColName:AgeCol, ColOffset:2, DefaultValue:-1] on property node: Person->Age by table dbo.People  
14/11/13 17:55:08 [Gridsum.DataflowEx].[Info] [PeopleFlow1] Telling myself there is no more input and wait for children completion  
14/11/13 17:55:08 [Gridsum.DataflowEx].[Info] [PeopleFlow1] completed  
14/11/13 17:55:08 [Gridsum.DataflowEx.Databases].[Debug] [DbBulkInserter<Person>1] starts bulk-inserting 4 Person to db table dbo.People  
14/11/13 17:55:08 [Gridsum.DataflowEx.Databases].[Info] [DbBulkInserter<Person>1] bulk-inserted 4 Person to db table dbo.People  
14/11/13 17:55:08 [Gridsum.DataflowEx].[Info] [DbBulkInserter<Person>1] completed  

So, be sure to check the log to diagnose issues!

One demo is not enough to show the real power of DbBulkInserter, which supports recursive property expansion for complex custom type. Look at an example of deep mapping searching:

public class Order
{
    [DBColumnMapping("OrderTarget", "Date", null)]
    public DateTime OrderDate { get; set; }

    [DBColumnMapping("OrderTarget", "Value", 0.0f)]
    public float? OrderValue { get; set; }

    public Person Customer { get; set; }
}

public class Person : IEventProvider
{
    [DBColumnMapping("PersonTarget", "NameCol", "N/A", ColumnMappingOption.Mandatory)]
    [DBColumnMapping("OrderTarget", "CustomerName", "Unknown Customer")]
    public string Name { get; set; }

    [DBColumnMapping("PersonTarget", "AgeCol", -1, ColumnMappingOption.Optional)]
    [DBColumnMapping("OrderTarget", "CustomerAge", -1)]
    public int? Age { get; set; }

    public DataflowEvent GetEvent()
    {
        // omit some code here
    }
}

public static async Task BulkInserterDemo2()
{
    string connStr;

    //initialize table
    using (var conn = LocalDB.GetLocalDB("BulkInserterDemo"))
    {
        var cmd = new SqlCommand(@"
        IF OBJECT_id('dbo.Orders', 'U') IS NOT NULL
            DROP TABLE dbo.Orders;
        
        CREATE TABLE dbo.Orders
        (
            Id INT IDENTITY(1,1) NOT NULL,
            Date DATETIME NOT NULL,
            Value FLOAT NOT NULL,
            CustomerName NVARCHAR(50) NOT NULL,
            CustomerAge INT NOT NULL
        )
        ", conn);
        cmd.ExecuteNonQuery();
        connStr = conn.ConnectionString;
    }

    var dbInserter = new DbBulkInserter<Order>(connStr, "dbo.Orders", DataflowOptions.Default, "OrderTarget");

    dbInserter.Post(new Order {OrderDate = DateTime.Now, OrderValue = 15, Customer = new Person() {Name = "Aaron", Age = 38}}); 
    dbInserter.Post(new Order {OrderDate = DateTime.Now, OrderValue = 25, Customer = new Person() {Name = "Bob", Age = 30}}); 
    dbInserter.Post(new Order {OrderDate = DateTime.Now, OrderValue = 35, Customer = new Person() {Age = 48}}); 
    dbInserter.Post(new Order {OrderDate = DateTime.Now, OrderValue = 45, Customer = new Person() {Name = "Neo"}});

    await dbInserter.SignalAndWaitForCompletionAsync();
}

Whose result is:

ScreenShot

In this case, type Order is the root type and it has a property whose type is Person. DbBulkInserter expands the property in Order class and grabs some mapping down from the Person class.

Tip: Make sure the destination label is the same for different levels of properties. In this demo, Person class added two mappings for label "OrderTarget", same as the label used in Order class.

One final point: DbBulkInserter is very careful about 'null's when generating deep property accessors like A.B.C. Since A.B could be null, DbBulkInserter generates something like 'A.B == null ? D : (A.B.C ?? D)' rather than A.B.C ?? D (D is the default value defined on C's DBColumnMapping) to avoid NullReferenceException. This affects the performance, of course, especially when your property tree is tall. So DataflowEx gives you an attribute, [DBColumnPath], to allow you to turn off the null check in the IL of the generated property accessor. Simply tag it on A.B with option DBColumnPathOptions.DoNotGenerateNullCheck and the null check will be stripped out: only A.B.C ?? D is generated. But remember you take the risk to guarantee A.B is not null (otherwise NullReferenceException will be thrown at runtime).

In the last demo, if you are sure each of the Order objects has a non-null Customer propety, try tagging DBColumnPath like this to get better performance:

public class Order
{
    [DBColumnMapping("OrderTarget", "Date", null)]
    public DateTime OrderDate { get; set; }

    [DBColumnMapping("OrderTarget", "Value", 0.0f)]
    public float? OrderValue { get; set; }

    [DBColumnPath(DBColumnPathOptions.DoNotGenerateNullCheck)]
    public Person Customer { get; set; }
}

Tip: Another option you may use when constructing a DBColumnPath attribute is DBColumnPathOptions.DoNotExpand, which disables the expansion of a certain property. This means this property path will be totally ignored when generating a mapping from your object model to the database table.

6.2 DataBroadcaster

When beginners touch Microsoft TPL Dataflow, one thing they complain is that an item can only travel to one of the many destinations. This is due to the design principle of TPL Dataflow but admittedly yes, there are scenarios this feature could be quite useful. That's why DataflowEx brings DataBroadcaster to make your life easier.

DataBroadcaster acts simply like a copy machine. When it is linked to multiple targets, whenever an item flows in, it passes the reference of the same item to multiple targets. Optionally you can indicate a clone function to DataBroadcaster if you want to copy the item before handing to targets.

Code talks:

public static async Task BroadcasterDemo()
{
    var broadcaster = new DataBroadcaster<string>();

    var printer1 = new ActionBlock<string>(s => Console.WriteLine("Printer1: {0}", s)).ToDataflow();
    var printer2 = new ActionBlock<string>(s => Console.WriteLine("Printer2: {0}", s)).ToDataflow();
    var printer3 = new ActionBlock<string>(s => Console.WriteLine("Printer3: {0}", s)).ToDataflow();

    broadcaster.LinkTo(printer1);
    broadcaster.LinkTo(printer2);
    broadcaster.LinkTo(printer3);

    broadcaster.Post("first message");
    broadcaster.Post("second message");
    broadcaster.Post("third message");

    await broadcaster.SignalAndWaitForCompletionAsync();
    await printer1.CompletionTask;
    await printer2.CompletionTask;
    await printer3.CompletionTask;
}

Logs and outputs will be:

14/11/19 11:23:32 [Gridsum.DataflowEx].[Info] [DataBroadcaster<String>1] now links to its 1th target (TargetDataflow<String>1)
14/11/19 11:23:32 [Gridsum.DataflowEx].[Info] [DataBroadcaster<String>1] now links to its 2th target (TargetDataflow<String>2)
14/11/19 11:23:32 [Gridsum.DataflowEx].[Info] [DataBroadcaster<String>1] now links to its 3th target (TargetDataflow<String>3)
14/11/19 11:23:32 [Gridsum.DataflowEx].[Info] [DataBroadcaster<String>1] Telling myself there is no more input and wait for children completion
Printer2: first message
Printer1: first message
Printer1: second message
Printer1: third message
Printer3: first message
Printer3: second message
Printer3: third message
Printer2: second message
Printer2: third message

Tip: TPL Dataflow also has a BroadcastBlock but it provides a buffer for storing at most one element at time, overwriting each message with the next as it arrives, which is very different from DataflowEx's DataBroadcaster that guarantees no data loss. IMHO, in most cases DataBroadcaster rather than BroadcastBlock is what you want.

6.3 DataDispatcher

DataDispatcher also falls into the 'one source multi targets' category. But it is different from DataBroadcaster in several aspects:

  1. It ensures one input item only goes to one target.
  2. Target dataflow nodes are created dynamically on demand, depending on the input items and the dispatch function.
  3. It is an abstract class that you should inherit from and add your logic

The second point is the core design of DataDispatcher. In many cases you can just use multiple conditional LinkTo() to dispatch object to where they should go, but DataDispatcher allows you to create targets on-the-fly according to the input.

Consider you are writing a file logging system which will allow logs to be distributed to different text files according to their logging level. DataDispatcher is the perfect place to start:

public class MyLog
{
    public LogLevel Level { get; set; }
    public string Message { get; set; }
}

/// <summary>
/// Logger that accepts logs and dispatch them to appropriate dynamically created log writer
/// </summary>
public class MyLogger : DataDispatcher<MyLog, LogLevel>
{
    public MyLogger() : base(log => log.Level)
    {
    }
    
    /// <summary>
    /// This function will only be called once for each distinct dispatchKey (the first time)
    /// </summary>
    protected override Dataflow<MyLog> CreateChildFlow(LogLevel dispatchKey)
    {
        //dynamically create a log writer by the dispatchKey (i.e. the log level)
        var writer = new LogWriter(string.Format(@".\MyLogger-{0}.log", dispatchKey));

        //no need to call RegisterChild(writer) here as DataDispatcher will call automatically
        return writer;
    }
}

/// <summary>
/// Log writer node for a single destination file
/// </summary>
internal class LogWriter : Dataflow<MyLog>
{
    private readonly ActionBlock<MyLog> m_writerBlock;
    private readonly StreamWriter m_writer;

    public LogWriter(string fileName) : base(DataflowOptions.Default)
    {
        this.m_writer = new StreamWriter(new FileStream(fileName, FileMode.Append));

        m_writerBlock = new ActionBlock<MyLog>(log => m_writer.WriteLine("[{0}] {1}", log.Level, log.Message));

        RegisterChild(m_writerBlock);
    }

    public override ITargetBlock<MyLog> InputBlock { get { return m_writerBlock; } }

    protected override void CleanUp(Exception e)
    {
        base.CleanUp(e);
        m_writer.Flush();
    }
}

Guess what? This sophisticated logging framwork is implemented in about 5 minutes :), with the help of DataDispatcher.

Cannot wait to invoke the wonderful 'logging library':

public static async Task MyLoggerDemo()
{
    var mylogger = new MyLogger();

    mylogger.Post(new MyLog { Level = LogLevel.Error, Message = "I am Error!" });
    mylogger.Post(new MyLog { Level = LogLevel.Warn, Message = "I am Warn!" });
    mylogger.Post(new MyLog { Level = LogLevel.Error, Message = "I am Error2!" });
    mylogger.Post(new MyLog { Level = LogLevel.Warn, Message = "I am Warn2!" });
    mylogger.Post(new MyLog { Level = LogLevel.Info, Message = "I am Info!" });

    await mylogger.SignalAndWaitForCompletionAsync();
}

Worked like a magic:

ScreenShot

There are many real world scenarios where DataDispatcher really shines. We hope you can find your own case. But apart from the abstract DataDispatcher, it's worth mentioning DataflowEx comes with an implementation of DataDispatcher called MultiDbBulkInserter, which combines the power of DataDispatcher and DbBulkInserter to support dumping data to multi-tenant databases that share the same schema. We will not prepare a separate demo here but if you are interested please check out the test code for how to use the class (see method TestMultiDbBulkInserter()):

https://github.com/gridsum/DataflowEx/blob/master/Gridsum.DataflowEx.Test/DatabaseTests/BulkInserterTest.cs

6.4 DbDataJoiner

If you have experience with SSIS, or ETL, a 'LookUp Component' should sound famaliar to you. The component normally searches a column of a dimension table and returns the relative dimension key. It will also insert to the dimension table if the lookup fails. Normally there is also a cache inside the component to improve performance.

Now instead of involving a heavy-weight enterprise ETL system, a fully-functional lookup component is available in a lightweight form, provided by DataflowEx out of the box as the DbDataJoiner class. As its name implies, the class needs to join the remote table in the database and pull out the keys matched. And yes, don't worry, DbDataJoiner also has a cache built-in.

To demonstrate the usage of DbDataJoiner, look at the following demo:

//extend the Order class in the last demo
public class OrderEx : Order
{
    public OrderEx()
    {
        this.OrderDate = DateTime.Now;
    }

    // This is the field we want to populate
    [DBColumnMapping("LookupDemo", "ProductKey")]
    public long? ProductKey { get; set; }

    public Product Product { get; set; }
}

//This is the 'Dimension' class for OrderEx
public class Product
{
    public Product(string category, string name)
    {
        this.Category = category;
        this.Name = name;
    }

    [DBColumnMapping("LookupDemo", "Category")]
    public string Category { get; private set; }

    [DBColumnMapping("LookupDemo", "ProductName")]
    public string Name { get; private set; }

    [DBColumnMapping("LookupDemo", "ProductFullName")]
    public string FullName { get { return string.Format("{0}-{1}", Category, Name); } }
}

public class ProductLookupFlow : DbDataJoiner<OrderEx, string>
{
    public ProductLookupFlow(TargetTable dimTableTarget, int batchSize = 8192, int cacheSize = 1048576)
        : base(
        //Tells DbDataJoiner the join condition is 'OrderEx.Product.FullName = TABLE.ProductFullName'
        o => o.Product.FullName, 
        dimTableTarget, DataflowOptions.Default, batchSize, cacheSize)
    {
    }

    protected override void OnSuccessfulLookup(OrderEx input, IDimRow<string> rowInDimTable)
    {
		//what we want is just the key column of the dimension row
        input.ProductKey = rowInDimTable.AutoIncrementKey;
    }
}

Simple code for a complex job, right? Three things to pay attention here:

  1. Appropriately tag DBColumnMapping for columns of the dimension table. (Don't mess them up with fact table mappings. They are independent.)
  2. In constructor, tell DbDataJoiner which property of your domain object to join with a column in the dimension table.
  3. Override OnSuccessfulLookup() to control the behavior when the dimension is looked up. Normally you just want the key column as the demo shows.

Now let's see how to use ProductLookupFlow and make our example complete. We will inherit from the Order class in previous demo BulkInserterDemo2 but this time extends it and mapps it to a fact table: dbo.FactOrders, which contains a 'ProductKey' column and this is the column we want to populate using ProductLookupFlow.

public static async Task ETLLookupDemo()
{
    string connStr;

    //initialize table
    using (var conn = LocalDB.GetLocalDB("ETLLookupDemo"))
    {
        //We will create a dimension table and a fact table
        //We also populate the dimension table with some pre-defined rows
        var cmd = new SqlCommand(@"
        IF OBJECT_id('dbo.Product', 'U') IS NOT NULL
            DROP TABLE dbo.Product;
        
        CREATE TABLE dbo.Product
        (
            ProductKey INT IDENTITY(1,1) NOT NULL,
            Category nvarchar(50) NOT NULL,
            ProductName nvarchar(50) NOT NULL,
            ProductFullName nvarchar(100) NOT NULL                    
        )

        INSERT INTO dbo.Product VALUES ('Books', 'The Great Gatsby', 'Books-The Great Gatsby');
        INSERT INTO dbo.Product VALUES ('Games', 'Call of Duty', 'Games-Call of Duty');

        IF OBJECT_id('dbo.FactOrders', 'U') IS NOT NULL
            DROP TABLE dbo.FactOrders;
        
        CREATE TABLE dbo.FactOrders
        (
            Id INT IDENTITY(1,1) NOT NULL,
            Date DATETIME NOT NULL,
            Value FLOAT NOT NULL,
            ProductKey INT NULL,
        )
        ", conn);
        cmd.ExecuteNonQuery();
        connStr = conn.ConnectionString;
    }

    var lookupNode = new ProductLookupFlow(new TargetTable("LookupDemo", connStr, "dbo.Product"));

    lookupNode.Post(new OrderEx { OrderValue = 10, Product = new Product("Books", "The Call of the Wild") });
    lookupNode.Post(new OrderEx { OrderValue = 20, Product = new Product("Games", "Call of Duty") });
    lookupNode.Post(new OrderEx { OrderValue = 30, Product = new Product("Games", "Call of Duty") });
    lookupNode.Post(new OrderEx { OrderValue = 20, Product = new Product("Books", "The Call of the Wild") });
    lookupNode.Post(new OrderEx { OrderValue = 20, Product = new Product("Books", "The Great Gatsby") });

    var factInserter = new DbBulkInserter<OrderEx>(connStr, "dbo.FactOrders", DataflowOptions.Default, "OrderTarget");
    lookupNode.LinkTo(factInserter);

    await lookupNode.SignalAndWaitForCompletionAsync();
    await factInserter.CompletionTask;
}

After running the demo, ProductLookupFlow did two things for us. First, it populated the dimension table if a match row is not found:

ScreenShot

Secondly, of course, it did a good job setting the ProductKey column for the fact table:

ScreenShot

Tip: More about how DbDataJoiner works internally: basically it has a cache to do client-side look up. If that fails, it will send distinct dimension rows as a batch to a temp dimension table which will then be merged to dimension table as an server side lookup. Row matching information is returned by an output clause of the MERGE statement. Find more detail in the source code of DbDataJoiner.

We build DbDataJoiner not only because we need the functionality in production but also we want to show the potential of DataflowEx framework as well: it is the most complex Dataflow implementation in DataflowEx. Feel free to use it or extend it to meet your need. If you wish, pull requests are always welcome.

7. DataflowEx Best Practices

7.1 Building your own Dataflow<TIn, TOut>

DataflowEx shows its real power with real-world custom Dataflow implementations. In above demos, we already demonstrate some common patterns but here is a bit more helpful tips.

First reminder: never ever forget to register each of the children. If you create a dataflow graph without registering all the nodes as children, the CompletionTask of your dataflow is simply incorrent: your dataflow will complete before some inner components do. (e.g. let's say the child you forget to register is a DbBulkInserter then probably you will lose some data when your application quits. )

Second tip is about a design choice: since Dataflow support both raw block and sub Dataflow as children, which one should we use? How to properly handle linking when I have a mixed kind of children (i.e. have some blocks as well as some sub dataflow)? Well, the best practice we recommend here is:

  1. If your dataflow only involves low-level raw TPL blocks, stay with blocks as your children and use traditional TPL Dataflow linking. But just don't forget to set PropagateCompletion to true in the DataflowBlockOptions parameter.
  2. If your dataflow involves raw blocks as well as sub dataflows, escalate all blocks to sub dataflows (you may use ToDataflow() extension method) and link them together using flow level LinkTo() API in DataflowEx.
  3. When multiple blocks points to the same destination and the target should complete when all predecessors finish, alway escalate to flow level and use DataflowEx's LinkTo().

Last but not least, bear in mind the basic 3 steps to build your own Dataflow: create, register and link.

7.2 What you should know about DataflowOptions

One thing that we touched but haven't yet explored is the option class of DataflowEx: DataflowOptions. We simply used DataflowOptions.Default. This is OK for demo purpose but you may want to adjust some options of this class seriously in the tough real world.

Let's look at the default property values in DataflowOptions.Default:

private static DataflowOptions s_defaultOptions = new DataflowOptions()
{
    BlockMonitorEnabled = false,
    FlowMonitorEnabled = true,
    PerformanceMonitorMode = PerformanceLogMode.Succinct,
    MonitorInterval = DefaultInterval,
    RecommendedCapacity = 100000,
    RecommendedParallelismIfMultiThreaded = Environment.ProcessorCount
};

Basically the settings fall into two categories: monitoring/logging control and block construction hint. The former category controls the dataflow logging loop behavior (e.g. logging interval, logging mode, etc. Check property comments for more explanation) while the latter category gives hints to the options that raw blocks should adopt (e.g. the buffer size and parallelism for multi-threaded blocks).

As a dataflow consumer, you should know what the properties mean and set them properly. A common case, for example, is that you want to set block capacity to a larger value for your flow to fully utilize your hardware. The same applies to RecommendedParallelismIfMultiThreaded which helps you to precisely allocate CPU resources for parallel processing nodes in your dataflow.

As a dataflow builder, or library provider, you should always respect the DataflowOptions passed in. That is to say, when constructing raw blocks in your dataflow, please extract relative properties from DataflowOptions and put that into corresponding field in the DataflowBlockOptions instance, which will then be passed to the block constructor. In addition, for Dataflow inheritance and child dataflows you normally pass the same DataflowOptions object down to their constructor. These are the DataflowOptions rules DataflowEx components obey and so should your Dataflow implementations.

Tip: DataflowEx provides ToExecutionBlockOption() method and ToGroupingBlockOption() method to help you migrate settings from the given flow option to a new block option instance, on which you could make further modification.

7.3 Performance Considerations

One of the most asked questions about DataflowEx is: does the framework affect TPL Dataflow performance?

The short answer is: Relax. No, it doesn't.

The longer answer: DataflowEx and particularly the Dataflow class is a thin wrapper over underlying blocks from the perspective of performance. Yes, it maintains some internal states and does some periodic status checking per dataflow but these effects are trivial from a performance perspective considering the typical amount of blocks in a Dataflow application. The real hot path is the data processsing part and data travelling between the nodes but this is where DataflowEx has clearly no side effects because the underlying mechanism is just TPL Dataflow.

To sum up, DataflowEx introduces very little overhead. When the pipeline is created and starts running, it is just raw TPL Dataflow stuff.

Tip: Using DataflowEx doesn't cause performance problems. But you still need to properly design your dataflow at the level of TPL Dataflow blocks. One simple design rule is to 'avoid too many blocks'. If one simple task could be implemented within a block, don't split it into two linked blocks: this hurts performance. Bear in mind that every single TPL block comes with a cost like buffering, threading and additional data travelling. Balancing between proper task granularity and component modularity/reusability is a key part in Dataflow/DataflowEx performance tuning. Again, don't have too many dataflows or blocks.

8. Have a try now!

Thanks for reading such a long tutorial. Congratulations that you've reached the end.

Tip: You can get full C# code of all demos used in this tutorial in the Gridsum.DataflowEx.Demo project. It will be a good start to clone the repo and play with the existing examples.

DataflowEx is being used at Gridsum internally to process hundreds of millions of messages per day. We wish you try the library in your application and enjoy it just like we do. For any issue or feedback please start a thread on github issue forum.

You can’t perform that action at this time.