Work in progress! It contains a lot of tpyos, please let us know. There are comments at the bottom or you can submit a PR against master branch.

Please help with the documentation if you know Dynflow.


High level overview TODO

TODO to be refined

Dynflow (DYNamic workFLOW) is a workflow engine written in Ruby that allows to:

  • Keep track of the progress of running processes
  • Run the code asynchronously
  • Resume the process when something goes wrong, skip some steps when needed
  • Detect independent parts and run them concurrently
  • Compose simple actions into more complex scenarios
  • Extend the workflows from third-party libraries
  • Keep consistency between local transactional database and external services
  • Suspend the long-running steps, not blocking the thread pool
  • Cancel steps when possible
  • Extend the actions behavior with middlewares
  • Pick different adapters to provide: storage backend, transactions, or executor implementation

Dynflow has been developed to be able to support orchestration of services in the Katello and Foreman projects.


  • what problems does Dynflow solve?
  • maybe a little history


  • Action - building block of execution plans, a Ruby class inherited from Dynflow::Action, defines code to be run in each phase.
  • Phase - Each action has three phases: plan, run, finalize.
  • Input - A Hash of data coming to the action.
  • Output - A Hash of data that the action produces. It's persisted and can be used as input of other actions.
  • Execution plan - definition of the workflow: product of the plan phase,
  • Triggering an action - entering the plan phase, starting with the plan method of the action. The execution follows immediately.
  • Flow - definition of the run/finalize phase, holding the information about steps that can run concurrently/in sequence. Part of execution plan.
  • Executor - service that executes the run and finalize flows based on the execution plan. It can run in the same process as the plan phase or in different process (using the remote executor).
  • World - the universe where the Dynflow runs the code: it holds all needed configuration. Usually there's only one world per Dynflow process, besides configuration it also holds Persistence, Logger, Executor and all the other objects necessary for action executing. This concept allows us to avoid globally shared state. Also, the worlds can talk to each other, which is helpful for production and high-availability setups, having multiple worlds on different hosts handle the execution of the execution plans. If you're still confused and come from RoR world, think about it as similar thing that is Rails object for Ruby on Rails framework.


See the examples directory for the code in action. Running those files (except the example_helper.rb file) leads to the Dynflow runtime being initialized (including the web console where one can explore the features and experiment).


  • for async operations
  • for orchestrating system/ssh calls
  • for keeping consistency between local database and external systems
  • sub-tasks

How to use

World creation TODO

  • include executor description

Development vs production

  • In development execution runs in the same process, in production there is an executor process.


Action anatomy

digraph "Graphviz" { rankdir=LR Trigger -> Plan -> Run -> Finalize } Graphviz Trigger Trigger Plan Plan Trigger->Plan Run Run Plan->Run Finalize Finalize Run->Finalize

When action is triggered, Dynflow executes plan method on this action, which is responsible for building the execution plan. It builds the execution plan by calling plan_action and plan_self methods, effectively listing actions that should be run as a part of this execution plan. In other words, it compiles a list of actions on which method run will be called. Also it's responsible for giving these actions an order. A simple example of such plan action might look like this

# this would plan deletion of files passed as an input array
def plan(files)
  files.each do |filename|
    plan_action MyActions::File::Destroy, filename

Note that it does not have to be only other actions that are planned to run. In fact it's very common that the action plans itself, which means it will put its own run method call in the execution plan. In order to do that you can use plan_self. This could be used in MyActions::File::Destroy used in previous example

class MyActions::File::Destroy < Dynflow::Action
  def plan(filename)
    plan_self path: filename

  def run

In example above, it seems that plan_self is just shortcut to plan_action MyActions::File::Destroy, filename but it's not entirely true. Note that plan_action always triggers plan of a given action while plan_self plans only the run and finalize of Action, so by using plan_action we'd end up in endless loop.

Also note, that run method does not take any input. In fact, it can use input method that refers to arguments, that were used in plan_self.

Similar to the input mentioned above, the run produces output. After that some finalizing steps can be taken. Actions can use outputs of other actions as parts of their inputs establishing dependency. Action's state is serialized between each phase and survives machine/executor restarts.

As lightly touched in the previous paragraphs there are 3 phases: plan, run and finalize. Plan phase starts by triggering an action. run and finalize are only ran if you use plan_self in plan phase.

Input and Output

Both input and output are Hashes accessible by Action#input and Action#output methods. They need to be serializable to JSON so it should contain only combination of primitive Ruby types like: Hash, Array, String, Integer, etc.


One should avoid using directly

self.output = { key: data }

It might delete other data stored in the output (potentially by middleware and other parts of the action). Therefore it's preferred to use

output.update(key: data, another_key: another_data)
# or for one key
output[:key] = data


You may sometime find these input/output format definitions:

class AnAction < Dynflow::Action
  input_format do
    param :id, Integer
    param :name, String

  output_format do
    param :uuid, String

This might me quite handy especially in combination with subscriptions functionality.

The format follows apipie-params for more details. Validations of input/output could be performed against this description but it's not turned on by default. (It needs to be revisited and updated to be fully functional.)


Triggering the action means starting the plan phase, followed by immediate execution. Any action is triggered by calling:

world_instance.trigger(AnAction, *args)


In Foreman and Katello actions are usually triggered by ForemanTask.sync_task and ForemanTasks.async_task so following part is not that important if you are using ForemanTasks.

World#trigger method returns object of TriggerResult type. Which is Algebrick variant type where definition follows:

TriggerResult = Algebrick.type do
  # Returned by #trigger when planning fails.
  PlaningFailed   = type { fields! execution_plan_id: String, error: Exception }
  # Returned by #trigger when planning is successful but execution fails to start.
  ExecutionFailed = type { fields! execution_plan_id: String, error: Exception }
  # Returned by #delay when scheduling succeeded.
  Scheduled       = type { fields! execution_plan_id: String }
  # Returned by #trigger when planning is successful, #future will resolve after
  # ExecutionPlan is executed.
  Triggered       = type { fields! execution_plan_id: String, future: Future }

  variants PlaningFailed, ExecutionFailed, Triggered

If you do not know Algebrick you can think about these as Structs with types. You can see how it's used to distinguish all the possible results in ForemanTasks module.

def self.trigger_task(async, action, *args, &block)
  Match! async, true, false

  match trigger(action, *args, &block),
        # Raise if there is any error caused either by failed planning or
        # by faild start of execution.
        (on ::Dynflow::World::PlaningFailed.(error: ~any) |
                ::Dynflow::World::ExecutionFailed.(error: ~any) do |error|
          raise error
        # Succesfully triggered.
        (on ::Dynflow::World::Triggered.(
              execution_plan_id: ~any, future: ~any) do |id, finished|
          # block on the finished Future if this is called synchronously
          finished.wait if async == false
          return ForemanTasks::Task::DynflowTask.find_by_external_id!(id)


Scheduling an action means setting it up to be triggered at set time in future. Any action can be delayed by calling:

                     { start_at: + 360, start_before: + 400 },

This snippet of code would delay AnAction with arguments args to be executed in the time interval between start_at and start_before. Setting start_before to nil would delay execution of this action without the timeout limit.

When an action is delayed, an execution plan object is created with state set to scheduled, but it doesn't run the the plan phase yet, the planning happens when the start_at time comes. If the planning doesn't happen in time (e.g. after start_before), the execution plan is marked as failed (its state is set to stopped and result to error).

Since the args have to be saved, there must be a mechanism to safely serialize and deserialize them in order to make them survive being saved in a database. This is handled by a serializer. Different serializers can be set per action by overriding its delay method.

Planning of the delayed plans is handled by DelayedExecutor, an object which periodically checks for delayed execution plans and plans them. Scheduled execution plans don't do anything by themselves, they just wait to be picked up and planned by a DelayedExecutor. It means that if no DelayedExecutor is present, their planning will be delayed until a scheduler is spawned.

Plan phase

Planning always uses the thread triggering the action. Plan phase configures action's input for run phase. It starts by executing plan method of the action instance passing in arguments from World#trigger method

world_instance.trigger(AnAction, *args)
# executes following
an_action.plan(*args) # an_action is AnAction

plan method is inherited from Dynflow::Action and by default it plans itself if run method is present using first argument as input. It also calls finalize method if it is present.

class AnAction < Dynflow::Action
  def run
    output.update self.input

world_instance.trigger AnAction, data: 'nothing'

The above will just plan itself copying input to output in run phase.

In most cases the plan method is overridden to plan self with transformed arguments and/or to plan other actions. In the Rails application, the arguments of the plan phase are often the ActiveRecord objects, that are then used to produce the inputs for the actions.

Let's look at the argument transformation first:

class AnAction < Dynflow::Action
  def plan(any_array)
    # pick just numbers
    plan_self numbers: { |v| v.is_a? Number }

  def run
    # compute sum - simulating a time consuming operation
    output.update sum: input[:numbers].reduce(&:+)


It's considered a good practice to use just enough data for the input for the action to perform the job. That means not too much (such as using ActiveRecord's attributes), as it might have performance impact as well as causes issues when changing the attributes later.

On the other hand, the input should contain enough data to perform the job without the need for reaching to external sources. Therefore, instead of passing just the ActiveRecord id and loading the whole record again in run phase, just to use some attributes, it's better to use these attributes directly as input of the action.

Following theses rules should lead to the best results, both from readability and performance point of view.

Now let's see an example with action planning:

class SumNumbers < Dynflow::Action
  def plan(numbers)
    plan_self numbers: numbers

  def run
    output.update sum: input[:numbers].reduce(&:+)

class SumManyNumbers < Dynflow::Action
  def plan(numbers)
    # references to planned actions
    planned_sub_sum_actions = numbers.each_slice(10).map do |numbers|
      plan_action SumNumbers, numbers

    # prepare array of output references where each points to sum in the
    # output of particular action
    sub_sums = do |action|

    # plan one last action which will sum the sub_sums
    # it depends on all planned_sub_sum_actions because it uses theirs outputs
    plan_action SumNumbers, sub_sums

world_instance.trigger SumManyNumbers, (1..100).to_a

Above example will in parallel sum numbers by slices of 10 values: first action sums 1..10, second action sums 11..20, ..., tenth action sums 91..100. After all sub sums are computed one final action sums the sub sums into final sum.


This example is here to demonstrate the planning abilities. In reality this parallelization of compute intensive tasks does not have a positive effect on Dynflow running on MRI. The pool of workers may starve. It is not a big issue since Dynflow is mainly used to orchestrate external services.

TODO add link to detail explanation in How it works when available.

Action may access local DB in plan phase, see Database and Transactions.

Run phase

Actions has a run phase if there is run method implemented. (There may be actions just planning other actions.)

The run method implements the main piece of work done by this action converting input into output. Input is immutable in this phase. It's the right place for all the steps which are likely to fail. Action run phase are allowed to have side effects like: file operations, calls to other systems, etc. Local DB should not be accessed in this phase, see Database and Transactions

Finale phase

Main purpose of finalize phase is to be able access local DB after action finishes successfully, like: indexing based on new data, updating records as fully created, etc. Finalize phase does not modify input or output of the action. Action may access local DB in finalize phase and must be idempotent, see Database and Transactions.


As already mentioned, actions can use output of different actions as their input (or just parts). When they do it creates dependency between actions, which is automatically detected by Dynflow and the execution plan is built accordingly.

def plan
  first_action  = plan_action AnAction
  second_action = plan_action AnAction, first_action.output[:a_key_in_output]

second_action uses part of the first_action's output therefore it depends on the first_action.

If actions are planned without this dependency as follows

def plan
  first_action  = plan_action AnAction
  second_action = plan_action AnAction

then they are independent and they are executed concurrently.

There is also other mechanism how to describe dependencies between actions than just the one based on output usage. Dynflow user can specify the order between planned actions with DSL methods sequence and concurrence. Both methods are taking blocks and they specify how actions planned inside the block (or inner sequence and concurrence blocks) should be executed.

By default plan considers its space as inside concurrence. Which means

def plan
  first_action  = plan_action AnAction
  second_action = plan_action AnAction


def plan
  concurrence do
    first_action  = plan_action AnAction
    second_action = plan_action AnAction

You can establish same dependency between first_action and second_action without using output by using sequence

def plan
  sequence do
    first_action  = plan_action AnAction
    second_action = plan_action AnAction

As mentioned the sequence and concurrence methods can be nested and mixed with output usage to create more complex dependencies. Let see commented example:

def plan
  # Plans 3 actions of type AnAction to be executed in sequence
  # argument is the index in the sequence.
  actions_executed_sequentially = sequence do { |i| plan_action AnAction, i }

  # Depends on output of the last action in `actions_executed_sequentially`
  # so it's added to the above sequence to be executed as 4th.
  action1 = plan_action AnAction, actions_executed_sequentially.last.output

  # It's planned in default plan's concurrency scope so it's executed concurrently
  # with the other four actions.
  action2 = plan_action AnAction

The order than will be:

  • concurrently:
    • sequentially:
      1. *actions_executed_sequentially
      2. action1
    • action2

Let's see one more example:

def plan
  actions = sequence do do |i|
      concurrency do { plan_action AnAction, i }

Which results in order of execution:

  • sequentially:
    1. concurrently:
      • actions[0][0] argument: 0
      • actions[0][1] argument: 0
    2. concurrently:
      • actions[1][0] argument: 1
      • actions[1][1] argument: 1


It's on our todo-list to change that to be able to define acyclic-graph of dependencies between the actions. sequence and concurrence methods will then be deprecated and kept just for backward compatibility.


Internally dependencies are also modeled with objects representing Sequences and Concurrences, which makes it weaker than acyclic-graph so in some cases during the dependency resolution it might not lead into the most effective execution plan. Some actions will run in sequence even though they could be run concurrently. This limitation is likely to be removed in some of the further releases.

Database and Transactions

Dynflow was designed to help with orchestration of other services. The usual execution looks as follows, we use an ActiveRecord User as example of a resource.

  1. Trigger user creation, argument is an unsaved ActiveRecord user object
  2. Planning: The user is stored in local DB (in the Dynflow hosting application) within the plan phase. The record is marked as incomplete.
  3. Running: Operations needed for the user in external services with (e.g.) REST call. The phase finishes when the all the external calls succeeded successfully.
  4. Finalizing: The record in local DB is marked as done: ready to be used. Potentially, saving some data that were retrieved in the run phase back to the local database.

For that reason there are transactions around whole plan and finalize phase (all action's plan methods are in one transaction). If anything goes wrong in the plan phase any change made during planning to local DB is reverted. Same holds for finalizing, if anything goes wrong, all changes are reverted. Therefore all finalization methods has to be idempotent.

Internally Dynflow uses Sequel as its ORM, but users may choose what they need to access they data. There is an interface TransactionAdapters::Abstract where its implementations may provide transactions using different ORMs. The most common one probably being TransactionAdapters::ActiveRecord.

So in the above example 2. and 4. step would be wrapped in ActiveRecord transaction if TransactionAdapters::ActiveRecord is used.

Second outcome of the design is convention when actions should be accessing local Database:

  • allowed - in plan and finalize phases
  • disallowed - (or at least discouraged) in the run phase


TODO warning about AR pool configuration, needs to have sufficient size


Dynflow is designed to allow easy composition of small building blocks called Actions. Typically there are actions composing smaller pieces together and other actions doing actual steps of work as in following example:

class CreateInfrastructure < Dynflow::Action
  def plan
    sequence do
      concurrence do
        plan_action(CreateMachine, 'host1', 'db')
        plan_action(CreateMachine, 'host2', 'storage')
                  :db_machine      => 'host1',
                  :storage_machine => 'host2')

Action CreateInfrastructure does not have a run method defined, it only defines plan action where other actions composed together.


Even though composing actions is quite easy and allows to decompose business logic to small pieces it does not directly support extensions by plugins. For that there are subscriptions.

Actions can subscribe from a plugin, gem, any other library to already loaded Actions, doing so they extend the planning process with self.

Lets look at an example starting by definition of a core action

# This action can be extended without doing any
# other steps to support it.
class ACoreAppAction < Dynflow::Action
  def plan(arguments)
    plan_self(args: arguments)
    plan_action(AnotherCoreAppAction, arguments.first)

  def run
    puts "Running core action: #{input[:args]}"
    self.output.update success: true

followed by an action definition defined in a plugin/gem/etc.

class APluginAction < Dynflow::Action
  # plan this action whenever ACoreAppAction action is planned
  def self.subscribe

  def plan(arguments)
    # arguments are same as in ACoreAppAction#plan
    plan_self(args: arguments)

  def run
    puts "Running plugin action: #{input[:args]}"

Subscribed actions are planned with same arguments as action they are subscribing to which is called trigger. Their plan method is called right after planning of the triggering action finishes.

It's also possible to access target action and use its output which makes it dependent (running in sequence) on triggering action.

def plan(arguments)
  plan_self trigger_success: trigger.output[:success]

def run
  self.output.update 'trigger succeeded' if self.input[:trigger_success]

Subscription is designed for extension by plugins, it should not be used inside a single library/app-module. It would make the process definition hard to follow (all subscribed actions would need to be looked up).


Sometimes action represents tasks taken in different services, (e.g. repository synchronization in Pulp). Dynflow tries not to waste computer resources so it offers tools to free threads to work on other actions while waiting on external tasks or events.

Dynflow allows actions to suspend and be woken up on external events. Lets create a simulation of an external service before showing the example of suspending action.

class AnExternalService
  def start_synchronization(report_to) do
      sleep 1
      report_to << :done

The AnExternalService can be invoked to start_synchronization and it will report back a second later to action passed in argument report_to. It sends event :done back by << method.

Lets look at an action example.

class AnAction < Dynflow::Action

  def plan

  def run(event)
    case event
    when nil # first run
      suspend do |suspended_action|
        EXTERNAL_SERVICE.start_synchronization suspended_action
    when :done # external task is done
      output.update success: true
      # let the run phase finish normally
      raise 'unknown event'

Which is then executed as follows:

  1. AnAction is triggered
  2. It's planned.
  3. Its run phase begins.
  4. run method is invoked with no event (nil).
  5. Matches with case branch initiating the external synchronization.
  6. Action initializes the synchronization and pass in reference to suspended_action.
  7. Action is suspended, execution of the run method finishes immediately after suspend is called, its block parameter is evaluated right after suspending.
  8. Action is kept on memory to be woken up when events are received but it does not block any threads.
  9. Action receives :done event through suspend action reference.
  10. run method is executed again with :done event.
  11. Output is updated with success: true and actions finishes run phase.
  12. There is no finalize phase, action is done.

This event mechanism is quite flexible, it can be used for example to build a polling action abstraction which is a topic for next chapter.


Not all services support callbacks to be registered which would allow to wake up suspended actions only once at the end when the external task is finished. In that case we often need to poll the service to see if the task is still running or finished.

For that purpose there is Polling module in Dynflow. Any action can be turned into a polling one just by including the module.

class AnAction < Dynflow::Action
  include Dynflow::Action::Polling

There are 3 methods need to be always implemented:

  • done? - determines when the task is complete based on external task's data.
  • invoke_external_task - starts the external task.
  • poll_external_task - polls the external task status data and returns a status (JSON serializable data like: Hash, Array, String, etc.) which are stored in action's output.
  def done?
    external_task[:progress] == 1

  def invoke_external_task

  def poll_external_task
    data     = poll_data_with_rest_call
    progress = calculate_progress data # => a float in 0..1
    { progress: progress
      data:     data }

This action will do following in run phase:

  1. invoke_external_task on first run of the action
  2. suspends and then periodically:
    1. wakes up
    2. poll_external_task
    3. checks if done?:
      • true -> it concludes the run phase
      • false -> it schedules next polling

There are 2 other methods handling external task data which can optionally overridden:

  • external_task - reads the external task's stored data, by default it reads self.output[:task]
  • external_task= - writes the the external task's stored data, by default it writes to self.output[:task] = value

There are also other features implemented like:

  • Gradual prolongation of the polling interval.
  • Retries on a poll failing.

Please see the Polling module for more details.


Each Action phase can be in one of the following states:

  • Pending - Not yet executed.
  • Running - An action phase id being executed right now.
  • Success - Execution of an action phase finished successfully.
  • Error - There was an error during execution.
  • Suspended - Only run phase, when action sleeps waiting for events to be woken up.
  • Skipped - Failed actions can be marked as skipped allowing rest of the execution plan to finish successfully.
  • Skipping - Action is marked for skipping but execution plan was not yet resumed to mark it as Skipped.

Execution plan has following states:

  • Pending - Planning did not start yet.
  • Scheduled - Scheduled for later execution, not yet planned.
  • Planning - It's being planned.
  • Planned - It've been planned, running phase did not start yet.
  • Running - It's running, run and finalize phases of actions are executed.
  • Paused - It was paused when running. Happens on error or executor restart.
  • Stopped - Execution plan is completed.

Execution plan also has following results:

  • Success - Everything finished without error or skips.
  • Warning - When there are skipped steps.
  • Error - When one or more actions failed.
  • Pending - Execution plan still runs.

TODO how do I access such states as a programmer? TODO which Action phase states are "finish" and which requires user interaction?

Error handling

If an error is raised in plan phase, it is persisted in the Action object for later inspection and it bubbles up in World#trigger method which was used to trigger the action leading to this error. If you compare it to errors raised during run and finalize phase, there's one major difference: Those never bubble up in trigger because they are running in executor not in triggering Thread, they are persisted just in the Action object.

If there is an error in run phase, the execution pauses. You can inspect the error in console. The error may be intermittent or you may fix the problem manually. After that the execution plan can be resumed and it'll continue by rerunning the failed action and continuing with the rest of the actions. During fixing the problem you may also do the steps in the actions manually, in that case the failed action can be also marked as skipped. After resuming the skipped action is not executed and the execution plan continues with the rest.

If there is an error in finalize phase, whole finalize phase for all the actions is rollbacked and can be rerun when the problem is fixed by resuming.

If you encounter an error during run phase error! or usual raise can be used.

Rescue strategy TODO

Console TODO

  • where to access
  • screenshots

Testing TODO

Long-running actions

Dynflow was designed as an Orchestration tool, parallelization of heavy CPU computation tasks was not directly considered. Even with multiple executors single execution plan always runs on one executor, so without JRuby it wont scale well (MRI's GIL). However JRuby support should be added soon (TODO update when merged).

Another problem with long-running actions are blocked worker. Executor has only a limited pool of workers, if more of them become busy it may result in worsen performance.

Blocking actions for long time are also problematic.

Solutions are:

  • Using action suspending - suspending the action until a condition is met, freeing the worker.
  • Offloading computation - CPU heavy parts can be offloaded to different services notifying the suspended actions when the computation is done.

Multiple queues support

By default, a single queue and a pool of workers is used to process all the actions in the system. This can cause some actions to block execution of some higher-priority ones in the system.

To address this case, it's possible to define additional queues tied to additional pool of workers dedicated for it. This way, they can be processed more independently from the default queue.

To use the queue, one needs to register additional queues when defining the executor world:

config =
config.queues.add(:slow, :pool_size => 5)
world =

The action to use the queue just needs to override the queue method like this:

class MyAction < Dynflow::Action
  def queue

  def run
    sleep 60

In the current implementation, it's expected all the executors would have the same set of queues defined. In the future implementation it should be possible to have dedicated executors with just a subset of queues


Each action class has chain of middlewares which wrap phases of the action execution. It's very similar to rack middlewares. To create new middleware inherit from Dynflow::Middleware class. It has 5 methods which can be overridden: plan, run, finalize, plan_phase, finalize_phase. Where the default implementation for all the methods looks as following

def plan(*args)
  pass *args

When overriding user can insert code before and/or after the pass method which executes next middleware in the chain or the action itself which is at the end of the chain. Most usually the pass is always called somewhere in the overridden method. There may be some cases when it can be omitted, then it'll prevent all following middlewares and action from running.

Some implementation examples: KeepCurrentUser, Action::Progress::Calculate.

Each Action has a chain of middlewares defined. Middleware can be added by calling use in the action class.

class AnAction < Dynflow::Action
  use AMiddleware, after: AnotherMiddleware

Method use understands 3 option keys:

  • :before - makes this middleware to be ordered before a given middleware
  • :after - makes this middleware to be ordered after a given middleware
  • :replace - this middleware will replace given middleware

The :before and :after keys are used to build a graph from the middlewares which is then sorted down with topological sort to the chain of middleware execution.


  • when to use?

To use sub-plans, you must include the Dynflow::Action::WithSubPlans module and override the create_sub_plans method. Inside the create_sub_plans method, you use the trigger method to create sub-tasks that will be executed in no particular order during the run phase. The parent task will wait for the sub-tasks to finish without blocking a thread in a pool while waiting.

class MyAction < Actions::EntryAction
  include Dynflow::Action::WithSubPlans


  def create_sub_plans
      trigger(Actions::OtherAction, action_param1, action_opts),

Execution plan hooks

Dynflow allows actions to hook into the lifecycle of their execution plans. To use the hooks, the user has to define a method on the action and register it as a hook. Currently there are hook events for every execution plan's state which are executed when the execution plan transitions into the state. Additionally there are two more hook events, failure and success which are run when the execution plan transitions into the stopped state with error or success result.

Methods can be registered using execution_plan_hooks.use call, providing the method name as Symbol and optionally a :on => HOOK_EVENT parameter, where HOOK_EVENT can be one of the hook events or an array of them. In case the optional parameter is not provided, the method is executed on every state change. Similarly, for example inherited, hooks can be disabled by calling execution_plan_hooks.do_not_use, taking the same arguments. Hooks defined on an action are inherited when the action is sub-classed.

The hooks are executed for every action in the execution plan and the order of execution is not guaranteed.

class MyAction < Actions::EntryAction
  # Sets up a hook to call #state_change method when the execution plan changes
  #   its state
  execution_plan_hooks.use :state_change

  # Sets up a hook to call #success_notification method when the execution plan
  #   finishes successfully,
  execution_plan_hooks.use :success_notification, :on => :success

  # Disables running #state_change method when the execution plan starts or
  #   finishes planning
  execution_plan_hooks.do_not_use :state_change, :on => [:planning, :planned]

  def state_change(_execution_plan)
    # Do something on every state change

  def success_notification(_execution_plan)
    # Display a notification

How it works TODO


This part is based on the current work-in-progress on multi-executors support

Action states TODO

  • normal phases and Present phase
  • how to walk the execution plan

The world anatomy

The world represents the Dynflow's run-time and it acts as an external interface. It holds all the configuration and sub-components needed for the Dynflow to perform its job.

The Dynflow worlds is composed of the following sub-components:

  1. persistence - provides the durability functionality: see persistence
  2. coordinator - provides the coordination between worlds: see coordinator
  3. connector - provides messages passing between worlds: see connector
  4. executor - the run-time itself executing the execution plan. Not every worlds has to have the executor present (there might be pure client worlds: useful in production, see develpment vs. production.
  5. client dispatcher - responsible for communication between client requests and other worlds
  6. executor dispatcher - responsible for getting requests from other worlds and sending the responses
  7. delayed executor - responsible for planning and exectuion of scheduled tasks

The underlying technologies are hidden behind adapters abstraction, which allows choosing the right technology for the job, while keeping the rest of the Dynflow intact.

Client world vs. executor world

In the simplest case, the world handles both the client requests and the execution itself. This is useful for small all-in-one deployments and development.

In production, however, one might want to separate the client worlds (often running in the web-server process or client library) and the executor worlds (running as part of standalone service). This setup makes the execution more stable and high-available (in active-active mode).

There might be multiple client as well as executor worlds in the Dynflow infrastructure.

The executor world has still its own client dispatcher, so that it can act as a client for triggering other execution plans (useful in sub-plans feature).

Single-database model

Dynflow recognizes different expectations from the underlying technologies from the persistence (durability), coordinator (real-time synchronization) and connector (transport).

However, we also realize that for many use-cases, a single shared database is just enough infrastructure the user needs for the job. Forcing using different technologies would mean just useless overhead.

Therefore, it's possible to use a single shared SQL database to do all the work.


Something as simple as sqlite is enough for getting the Dynlfow up-and-running and getting something done.

For the best connector results, it's recommended to use PostgreSQL, as Dynflow can utilize the listen/notify feature for better response times.

Inner-world communication

1) the client prepares an execution plan, saves it into persistence and passes its id to the client dispatcher

2) the client dispatcher creates an IVar that represents the future value of the execution plan after it finishes execution. The client can use this value to wait for the execution plan to finish or hook other procedures on the finish-time.

3) the client dispatcher creates an envelope (see connector for more details), containing the request for execution, with receiver id set to AnyExecutor

4) the connector uses some scheduling algorithm to choose what executor to send the request to and replaces the AnyExecutor with it. It sends the envelope the the chosen executor.

5) the connector on the executor side receives the envelope and asks the executor dispatcher to handle it

6) the executor dispatcher acquires the lock on the execution plan and initiates the execution. In the mean-time, it lets the client know that the work was accepted (there might be additional logic on the client to handle a timeout after the work was not accepted)

7 - 9) the connector layer propagates the Accepted response back to client

10) the executor finishes execution

11 - 12) the connector layer propagates the Finished response

13) the client dispatcher resolves the original IVar, so that the client is able to find out about the finished execution.

The behavior is the same even in the "one world" scenario: in that case this world is participating both on the client and the executor side, connected through a direct in-memory connector.


The persistence making sure that the serialized states of the execution plans are persisted for recovery and status tracking. The execution plan data are stored in it, with the actual state.

Unlike coordinator, all the persisted data don't have to be available for all the worlds at the same time: every world needs just the data that it is actively working on. Also, all the data don't have to be fully synchronized between worlds (as long as the up-to-date data about relevant execution plans are available for the world).


Provides messages passing between worlds. The message has a form of envelope of the following structure:

  • sender id - the id of the world that produced the envelope
  • receiver id - the id of the world that should receive the message, or AnyExecutor, when load-balancing
  • request id - the client-unique id used for pairing the request - response at the client
  • message - the body of the message: the connector doesn't care about that as long as it's serializable


The connector is responsible for spreading the load across the executor worlds. It's determined by the AnyExecutor value at the request id field. The implementation available in the current version uses a simple round-robin algorithm for this purpose.


This component (especially important in a multi-executor setup): Makes sure no two executors are executing the same execution plan at the same time (a.k.a locking). It also provides the information about the worlds available in the system (the worlds register). Unlike the persistence, it's not required to persist the data (could be recalculated), but it needs to provide a globally shared state.

The main type of objects the coordinator works with is a record which consists of:

  • type - the type of the record
  • id - the id of the record (unique in scope of a type)
  • data - data in arbitrary format, based on the type

There is a special type of record called lock, that keeps the owner information as well. It's used for keeping information about what executor is actively working on what execution plan: the executor is not allowed to start executing the unless it has successfully acquired a lock for it.

Singleton Actions

Dynflow has a special module for actions of which there should be only one instance active at a time. This module provides a number of methods for managing the action's locks as well as a middleware for automatic locking.

It works in the following way. The middleware tries to acquire the lock for this action, which is owned by the execution plan. If another action already holds the lock, it fill fail and the execution plan will transition to stopped-error state. Having obtained the lock, the action goes through the planning as usually. In run phase, the middleware checks if the execution plan still owns the lock for the action. If the execution plan holds the lock or there is no lock at all and the action manages to acquire it again, the execution proceeds. If the lock is held by another execution plan, the current one fails. Unlocking can be either done manually from within the action or can be left to the execution plan. The execution plan unlocks all locks it holds whenever it transitions to paused or stopped state.

All that is needed to make an action a singleton one is including the module into it.

class ExampleSingletonAction < ::Dynflow::Action
  include ::Dynflow::Action::Singleton

Thread-pools TODO

  • how it works now
  • how it'll work
  • gotchas
    • worker pool sizing

Suspending -> events TODO

Use cases TODO

  • Embedded without a DB, like inside CLI tool for a complex installation
  • reserve resources in planning do not try to do ifs in run phase
  • Projects: katello, foreman, staypuft, fusor


Comments are temporally turned on here for faster feedback.