Flowex

Railway Flow-Based Programming.

The library is not supported anymore, see the ALF project.

Flowex is a set of abstractions built on top Elixir GenStage which allows writing program with Flow-Based Programming paradigm.

I would say it is a mix of FBP and so-called Railway Oriented Programming (ROP) approach.

Flowex DSL allows you to easily create "pipelines" of Elixir GenStages.

Dedicated to my lovely girlfriend Chryścina.

Resources

• Railway Flow-Based Programming with Flowex - post
• Flowex: Flow-Based Programming with Elixir GenStage - presentation
• Flow-based programming with Elixir - presentation
• Flow-Based REST API with Flowex and Plug - post
• Multi language FBP with Flowex - presentation
• Multi-language Flowex components - post
• Flow-Based REST API with Flowex and Plug - post

Contents

• Installation
• A simple example to get the idea
• More complex example for understanding interface
• Flowex magic!
• Run the pipeline
• How it works
• Error handling
• Pipeline and pipe options
• Synchronous and asynchronous calls
• Bottlenecks
• Module pipes
• Data available in pipes
• Starting strategies
• Debugging with Flowex.Sync.Pipeline
• Contributing

Installation

Just add flowex as dependency to the mix.exs file.

A simple example to get the idea

Let's consider a simple program which receives a number as an input, then adds one, then multiplies the result by two and finally subtracts 3.

defmodule Functions do
  def add_one(number), do: number + 1
  def mult_by_two(number), do: number * 2
  def minus_three(number), do: number - 3
end

defmodule MainModule do
  def run(number) do
    number
    |> Functions.add_one
    |> Functions.mult_by_two
    |> Functions.minus_three
  end
end

So the program is a pipeline of functions with the same interface. The functions are very simple in the example.

In the real world they can be something like validate_http_request, get_user_from_db, update_db_from_request and render_response. Furthermore, each of the function can potentially fail. But for getting the idea let's stick the simplest example.

FBP defines applications as networks of "black box" processes, which exchange data across predefined connections by message passing.

To satisfy the FBP approach we need to place each of the function into a separate process. So the number will be passed from 'add_one' process to 'mult_by_two' and then 'minus_three' process which returns the final result.

That, in short, is the idea of Flowex!

More complex example for understanding interface

Let's define a more strict interface for our function. So each of the function will receive a predefined struct as a first argument and will return a map:

def add_one(%{number: number}, opts) do
  %{number: number + 1, a: opts.a}
end

The function receives a structure with number field and the options map with field a and returns map with new number. The second argument is a set of options and will be described later. Let's rewrite the whole Functions module in the following way:

defmodule Functions do
  defstruct number: nil, a: nil, b: nil, c: nil

  def add_one(%{number: number}, %{a: a}) do
    %{number: number + 1, a: a}
  end

  def mult_by_two(%{number: number}, %{b: b}) do
    %{number: number * 2, b: b}
  end

  def minus_three(%{number: number}, %{c: c}) do
    %{number: number - 3, c: c}
  end
end

The module defines three functions with the similar interface. We also defined as struct %Functions{} which defines a data-structure being passed to the functions.

The main module may look like:

defmodule MainModule do
  def run(number) do
    opts = %{a: 1, b: 2, c: 3}
    %Functions{number: number}
    |> Functions.add_one(opts)
    |> Functions.mult_by_two(opts)
    |> Functions.minus_three(opts)
  end
end

Flowex magic!

Let's add a few lines at the beginning.

defmodule FunPipeline do
  use Flowex.Pipeline

  pipe :add_one
  pipe :mult_by_two
  pipe :minus_three

  defstruct number: nil, a: nil, b: nil, c: nil

  def add_one(%{number: number}, %{a: a}) do
    %{number: number + 1, a: a}
  end

  # mult_by_two and minus_three definitions skipped
end

We also renamed the module to FunPipeline because we are going to create "Flowex pipeline". Flowex.Pipeline extend our module, so we have:

• pipe macro to define which function evaluation should be placed into separate GenStage;
• error_pipe macro to define function which will be called if error occurs;
• start, supervised_start and stop functions to create and destroy pipelines;
• call function to run pipeline computations synchronously.
• cast function to run pipeline computations asynchronously.
• overridable init function which, by default, accepts opts and return them

Let's start a pipeline:

opts = %{a: 1, b: 2, c: 3}

pipeline = FunPipeline.start(opts)

#returns
%Flowex.Pipeline{in_name: :"Flowex.Producer_#Reference<0.0.7.504>",
 module: FunPipeline, out_name: :"Flowex.Consumer_#Reference<0.0.7.521>",
 sup_pid: #PID<0.136.0>}

What happened:

• Three GenStages have been started - one for each of the function in pipeline. Each of GenStages is :producer_consumer;
• One additional GenStage for error processing has been started (it is also :producer_consumer);
• 'producer' and 'consumer' GenStages for input and output have been added;
• All the components have been placed under Supervisor.

The next picture shows what the 'pipeline' is.

The start function returns a %Flowex.Pipeline{} struct with the following fields:

• module - the name of the module
• in_name - unique name of 'producer';
• out_name - unique name of 'consumer';
• sup_name - unique name of the pipeline supervisor

Note, we have passed options to start function. This options will be passed to each function of the pipeline as a second argument. There is supervised_start function which allows to place pipeline's under external supervisor. See details in Starting strategies section.

Run the pipeline

One can run calculations in pipeline synchronously and asynchronously:

• call function to run pipeline computations synchronously.
• cast function to run pipeline computations asynchronously.

FunPipeline.call/2 function receive a %Flowex.Pipeline{} struct as a first argument and must receive a %FunPipeline{} struct as a second one. The call function returns a %FunPipeline{} struct.

FunPipeline.call(pipeline, %FunPipeline{number: 2})
# returns
%FunPipeline{a: 1, b: 2, c: 3, number: 3}

As expected, pipeline returned %FunPipeline{} struct with number: 3. a, b and c were set from options.

If you don't care about the result, you should use cast/2 function to run and forget.

FunPipeline.cast(pipeline, %FunPipeline{number: 2})
# returns
:ok

Run via client

Another way is using Flowex.Client module which implements GenServer behavior. The Flowex.Client.start\1 function receives pipeline struct as an argument. Then you can use call/2 function or cast/2. See example below:

{:ok, client_pid} = Flowex.Client.start(pipeline)

Flowex.Client.call(client_pid, %FunPipeline{number: 2})
# returns
%FunPipeline{a: 1, b: 2, c: 3, number: 3}

#or
Flowex.Client.cast(client_pid, %FunPipeline{number: 2})
# returns
:ok

How it works

The following figure demonstrates the way data follows: Note: error_pipe is not on the picture in order to save place.

The things happen when you call Flowex.Client.call (synchronous):

• self process makes synchronous call to the client gen_server with %FunPipeline{number: 2} struct;
• the client makes synchronous call 'FunPipeline.call(pipeline, %FunPipeline{number: 2})';
• the struct is wrapped into %Flowex.IP{} struct and begins its asynchronous journey from one GenStage to another;
• when the consumer receives the Information Packet (IP), it sends it back to the client which sends it back to the caller process.

The things happen when you cast pipeline (asynchronous):

• self process makes cast call to the client and immediately receives :ok
• the client makes cast to pipeline;
• the struct is wrapped into %Flowex.IP{} struct and begins its asynchronous journey from one GenStage to another;
• consumer does not send data back, because this is cast

Error handling

What happens when error occurs in some pipe?

The pipeline behavior is like Either monad. If everything ok, each 'pipe' function will be called one by one and result data will skip the 'error_pipe'. But if error happens, for example, in the first pipe, the :mult_by_two and :minus_three functions will not be called. IP will bypass to the 'error_pipe'. If you don't specify 'error_pipe' flowex will add the default one:

def handle_error(error, _struct, _opts) do
  raise error
end

which just raises an exception.

To specify the 'error' function use error_pipe macro:

defmodule FunPipeline do
  use Flowex.Pipeline
  # ...
  error_pipe :if_error

  def if_error(error, struct, opts) do
    # error is %Flowex.PipeError{} structure
    # with :message, :pipe, and :struct fields
    %{number: :oops}
  end
  #...
end

You can specify only one error_pipe! Note: The 'error_pipe' function accepts three arguments. The first argument is a %Flowex.PipeError{} structure which has the following fields:

• :message - error message;
• :pipe - is {module, function, opts} tuple containing info about the pipe where error occured;
• :struct - the input of the pipe.

Pipeline and pipe options

In addition to specifying options when starting pipeline one can pass component's options to the pipe macro. And remember about pipeline's init function which can add or override options. The flow is the following: The options passed to start function are available in pipeline init function. The function can merge additional options. Then opts passed to pipe macro are merged. So there are three levels that options pass before appearing in component:

• pipeline start function;
• pipeline init function;
• pipe opts.

Let's consider an example:

defmodule InitOptsFunPipeline do
  use Flowex.Pipeline

  defstruct [:from_start, :from_init, :from_opts]
  pipe :component, opts: %{from_opts: 3}

  def init(opts) do
    # opts passed to start function is available here
    Map.put(opts, :from_init, 2)
   end

  def component(_data, opts) do
    # here all the options is available
    opts
  end
end

Suppose we've started the pipeline with options %{from_start: 1}. init function adds :from_init option. Then :from_opts are merged.

The test below illustrates what is going on:

describe "function pipeline" do
  let :pipeline, do: InitOptsFunPipeline.start(%{from_start: 1})
  let :result, do: InitOptsFunPipeline.call(pipeline(), %InitOptsFunPipeline{})

  it "returns values from different init functions" do
    expect(result())
    |> to(eq %InitOptsFunPipeline{from_start: 1, from_init: 2, from_opts: 3})
  end
end

Synchronous and asynchronous calls

Note, that call function on pipeline module or Flowex.Client is synchronous. While communication inside the pipeline is asynchronous: One might think that there is no way to effectively use the pipeline via call/2 method.

That's not true!

In order to send a large number of IP's and process them in parallel one can use several clients connected to the pipeline:

Bottlenecks

Each component of pipeline takes a some to finish IP processing. One component does simple work, another can process data for a long time. So if several clients continuously push data they will stack before the slowest component. And data processing speed will be limited by that component.

Flowex has a solution! One can define a number of execution processes for each component.

defmodule FunPipeline do
  use Flowex.Pipeline

  pipe :add_one, count: 1
  pipe :mult_by_two, count: 3
  pipe :minus_three, count: 2
  error_pipe :if_error, count: 2

  # ...
end

And the pipeline will look like on the figure below:

Module pipes

One can create reusable 'pipe' - module which implements init and call functions. Each module must define a struct it works with. Only fields defined it the stuct will be passed to call function.

defmodule ModulePipeline do
  use Flowex.Pipeline

  defstruct [:number, :a, :b, :c]

  pipe AddOne, count: 1
  pipe MultByTwo, count: 3
  pipe MinusThree, count: 2
  error_pipe IfError, count: 2
end

#pipes

defmodule AddOne do
  defstruct [:number]

  def init(opts) do
    %{opts | a: :add_one}
  end

  def call(%{number: number}, %{a: a}) do
    %{number: number + 1, a: a}
  end
end

defmodule MultByTwo do
  defstruct [:number]

  def init(opts) do
    %{opts | b: :mult_by_two}
  end

  def call(%{number: number}, %{b: b}) do
    %{number: number * 2, b: b}
  end
end

defmodule MinusThree do
  defstruct [:number]

  def init(opts) do
    %{opts | c: :minus_three}
  end

  def call(%{number: number}, %{c: c}) do
    %{number: number - 3, c: c}
  end
end

defmodule IfError do
  defstruct [:number]

  def init(opts), do: opts

  def call(error, %{number: _number}, _opts) do
    %{number: error}
  end
end

Of course, one can combine module and functional 'pipes'!

Data available in pipes

If your pipeline consists of function pipes only, each function will receive pipeline struct as an input. The situation is a little more complex with module pipes. Each module defines its own struct and data will be cast to that struct. Map returned from the call function will be merged to the previos data. Let's consider an example:

defmodule DataAvailable do
  use Flowex.Pipeline

  defstruct [:top, :c1, :foo]

  pipe Component1
  pipe :component2
  pipe Component3

  def component2(%__MODULE__{top: top}, _opts) do
    %{top: top + 2, c3: 2}
  end
end

defmodule Component1 do
  defstruct [:top, :c1]
  def init(opts), do: opts

  def call(%__MODULE__{c1: c1, top: top}, _opts) do
    %{top: top + c1, bar: :baz}
  end
end

defmodule Component3 do
  defstruct [:c3, :top]
  def init(opts), do: opts

  def call(%__MODULE__{c3: c3, top: top}, _opts) do
    %{top: top + c3, c3: top - c3, foo: :set_foo}
  end
end

And suppose we passed %DataAvailable{top: 100, c1: 1} to DataAvailable.call function.

Data in IP before calling first pipe is %{c1: 1, foo: nil, top: 100}. Before entering the first pipe the data will be cast to %Component1{c1: 1, top: 100}. The returned value of first pipe is merged to IP data, so the data is %{bar: :baz, c1: 1, foo: nil, top: 101}.

Function component2 receives %DataAvailable{c1: 1, foo: nil, top: 101} structure and returned value %{c3: 2, top: 103} is merged with previous data, so IP data is %{bar: :baz, c1: 1, c3: 2, foo: nil, top: 103}

Last component receives %Component3{c3: 2, top: 103}, returns %{c3: 101, foo: :set_foo, top: 105} and data is %{bar: :baz, c1: 1, c3: 101, foo: :set_foo, top: 105}. Before returning data from pipeline they are casted to DataAvailable structure, so final result is %DataAvailable{c1: 1, foo: :set_foo, top: 105}}

Starting strategies

Using start/1 function one can start pipelines in any process. Pipelines will be alive while the process is alive. The supervised_start function accepts supervisor pid as the first argument and opts as the second argument. And starts pipeline's supervisor under predefined supervisor process.

In general there are three ways to start pipelines in your project:

1. Start pipelines in arbitrary supervised process:

   defmodule PipelineGenServer do
   use GenServer
   
   def init(_opts) do
   pipeline_one = PipelineOne.start
   pipeline_two = PipelineTwo.start
   
   {:ok, %{pipeline_one: pipeline_one, pipeline_two: pipeline_two}}
   end
   end

   You can also store pipeline structure in Agent or Application environment.

2. Start one pipeline per application. In that case pipeline supervisor will be the main supervisor in the application:

   defmodule OnePipelinePerApp do
   use Application
   
   def start(_type, _opts) do
   pipeline = PipelineOne.start
   Application.put_env(:start_flowex, :pipeline, pipeline)
   {:ok, pipeline.sup_pid}
   end
   end

3. Start several pipelines inside one application using supervised_start function. In that case pipeline supervisors will be placed under application supervisor:

   defmodule TwoPipelinesPerApp do
   use Application
   
   def start(_type, _opts) do
   {:ok, supervisor_pid} = Supervisor.start_link([], strategy: :one_for_one, name: :multi_flowex_sup)
   
   pipeline_one = PipelineOne.supervised_start(supervisor_pid)
   pipeline_two = PipelineTwo.supervised_start(supervisor_pid)
   
   Application.put_env(:start_flowex, :pipeline_one, pipeline_one)
   Application.put_env(:start_flowex, :pipeline_two, pipeline_two)
   
   {:ok,supervisor_pid}
   end
   end

You can find the examples in 'Start-Flowex' project

Debugging with Flowex.Sync.Pipeline

If you are faced with some error that is hard to debug or an error that causes GenServers to crash, you may find the Flowex.Sync.Pipeline module useful. Adding one Sync word will completely change the behavior.

defmodule FunPipeline do
  use Flowex.Sync.Pipeline
  # The same code as before
  # ...
end  

Interface remains the same but all the code will be evaluated in one simple GenServer. So all you pipes will be evaluated synchronously in separate process. Use this option only for debug purposes.

Contributing

Contributions are welcome and appreciated!

Request a new feature by creating an issue.

Create a pull request with new features or fixes.

Flowex is tested using ESpec. So run:

mix espec