Croupier

Croupier is a smart task definition and execution library, which can be used for dataflow programming.

What does it mean

You use Croupier to define tasks. Tasks have:

• An id

• Zero or more input files or k/v store keys

• Zero or more output files or k/v store keys

• A Proc that consumes the inputs and returns a string

• After the Proc returns data which is saved to the output(s) unless the task has the no_save flag set to true, in which case it's expected to have already saved it.

  Note: the return value for procs depends on several factors, see below. Note: A reference to a k/v key is of the form kv://mykey

And here is the fun part:

Croupier will examine the inputs and outputs for your tasks and use them to build a dependency graph. This expresses the connections between your tasks and the files on disk, and between tasks, and will use that information to decide what to run.

So, suppose you have task1 consuming input.txt producing fileA and task2 that has fileA as input and outputs fileB. That means your tasks look something like this:

  graph LR;
      id1(["📁 input.txt"])-->idt1["⚙️ task1"]-->id2(["📁 fileA"]);
      id2-->idt2["⚙️ task2"]-->id3(["📁 fileB"]);

Croupier guarantees the following:

• If task1 has never run before, it will run and create fileA
• If task1 has run before and input.txt has not changed, it will not run.
• If task1 has run before and input.txt has changed, it will run
• If task1 runs, task2 will run and create fileB
• task1 will run before task2

That's a very long way to say: Croupier will run whatever needs running, based on the content of the dependency files and the dependencies between tasks. In this example it may look silly because it's simple, but it should work even for thousands of tasks and dependencies.

The state between runs is kept in .croupier so if you delete that file all tasks will run.

Further documentation at the doc pages

Notes

Notes about proc return types

• Procs in Tasks without outputs can return nil or a string, it will be ignored.

• Procs with one output and no_save==false should return a string which will be saved to that output.

  If no_save==true then the returned value is ignored.

• Procs with multiple outputs and no_save==false should return an Array(String) which will be saved to those outputs.

  If no_save==true then the returned value is ignored.

No target conflicts

If there are two or more tasks with the same output they will be merged into the first task created. The resulting task will:

• Depend on the combination of all dependencies of all merged tasks
• Run the procs of all merged tasks in order of creation

Tasks without output

A task with no output will be registered under its id and is not expected to create any output files. Other than that, it's just a regular task.

Tasks with multiple outputs

If a task expects the TaskManager to create multiple files, it should return an array of strings.

The Task Set Is Fixed Before Running

Tasks are created at setup time, before the first run_tasks (or auto_run) call. Creating a task while a run is in progress raises UsageError — mid-run registration would mutate the task registries while the run reads them.

If the task set depends on things only known at runtime (files in a folder, entries in a database), build the graph when those things change:

• Batch mode: collect the files, create the tasks, call run_tasks. Unchanged tasks skip via early cutoff, so re-running with the same set is cheap.
• Auto mode: stop with auto_stop, rebuild the task graph, and call auto_run again. (TaskManager.cleanup resets everything if you want to build from scratch.)

Tasks can still discover dependencies at runtime: add_input grows an existing task's inputs from inside a proc, and is safe during parallel runs. The addition is deferred until the run ends, so it takes effect on the next run. task.inputs << input does the same.

TaskManager.tasks and Task#inputs are read-only views. Change them only through Task.new, TaskManager.remove_task and add_input.

Installation

1. Add the dependency to your shard.yml:

   dependencies:
     croupier:
       github: ralsina/croupier
   

2. Run shards install

Usage

This is the example described above, in actual code:

require "croupier"

Croupier::Task.new(
  output: "fileA",
  inputs: ["input.txt"],
) {
  puts "task1 running"
  File.read("input.txt").downcase
}

Croupier::Task.new(
  output: "fileB",
  inputs: ["fileA"],
) do
  puts "task2 running"
  File.read("fileA").upcase
end

Croupier::TaskManager.run_tasks

If we create a input.txt file with some text in it and run this program, it will print task1 running and task2 running and produce fileA with that same text in lowercase, and fileB with the text in uppercase.

The second time we run it, it will do nothing because all tasks dependencies are unchanged.

If we modify index.txt or fileA then one or both tasks will run, as needed.

Auto Mode

Besides run_tasks, there is another way to run your tasks, auto_run. It will run tasks as needed, when their input files change. This allows for some sorts of "continuous build" which is useful for things like web development.

You start the auto mode with TaskManager.auto_run and stop it with TaskManager.auto_stop. It runs in a separate fiber so your main fiber needs to do something else and yield. For details on that, see Crystal's docs.

This feature is still under development and may change, but here is an example of how it works, taken from the specs:

# We create a proc that has a visible side effect
x = 0
counter = TaskProc.new { x += 1; x.to_s }
# This task depends on a file called "i" and produces "t1"
Task.new(output: "t1", inputs: ["i"], proc: counter)
# Launch in auto mode
TaskManager.auto_run

# We have to yield and/or do stuff in the main fiber
# so the auto_run fibers can run
Fiber.yield

# Trigger a build by creating the dependency
File.open("i", "w") << "foo"
Fiber.yield

# Stop the auto_run
TaskManager.auto_stop

# It should only have ran once
x.should eq 1
File.exists?("t1").should eq true

When Tasks Fail

When a task's proc raises, the task fails with a Croupier::TaskFailure that keeps the original exception available as its #cause.

A failed run surfaces as Croupier::RunFailure:

• Without keep_going, the run aborts on the first failure and its state is not saved. The RunFailure carries that single failure.
• With keep_going: true, the run completes everything it can and saves its state, then raises RunFailure at the end. Its #errors array carries every failure, each with the original exception as its #cause.

begin
  TaskManager.run_tasks(keep_going: true)
rescue failure : Croupier::RunFailure
  failure.errors.each { |e| Log.error { e.message } }
  exit 1
end

Tasks that can't run because an input is missing (neither a task output, an existing file, nor a kv:// key) raise Croupier::UnknownInputsError. In auto mode these are expected (inputs appear incrementally) and retried with backoff. Failed tasks leave their inputs marked as modified, so the next run retries them.

Development

Let's try to keep test coverage good :-)

• To run tests: make test or crystal spec
• To check coverage: make coverage
• To run mutation testing: make mutation

Other than that, anything is fair game. In the TODO.md file there is a section for things that were considered and decided to be a bad idea, but that is conditional and can change when presented with a good argument.

Contributing

1. Fork it (https://github.com/ralsina/croupier/fork)
2. Create your feature branch (git checkout -b my-new-feature)
3. Commit your changes (git commit -am 'Add some feature')
4. Push to the branch (git push origin my-new-feature)
5. Create a new Pull Request

Contributors

• Roberto Alsina - creator and maintainer