taskproc 0.18.3

A lightweight task processing library written in pure Python

taskproc

A lightweight pipeline building/execution/management tool written in pure Python. Internally, it depends on DiskCache, cloudpickle networkx and concurrent.futures.

Why taskproc?

I need a pipeline-handling library that is thin and flexible as much as possible.

• Luigi is not flexible enough: The definition of the dependencies and the definition of the task computation is tightly coupled at luigi.Tasks, which is super cumbersome if one tries to edit the pipeline structure without changing the computation of each task.
• Airflow is too big and clumsy: It requires a message broker backend separately installed and run in background. It is also incompatible with non-pip package manager (such as Poetry).
• Most of the existing libraries tend to build their own ecosystems that unnecessarily forces the user to follow the specific way of handling pipelines.

taskproc aims to provide a language construct for defining computation by composition, ideally as simple as Python's built-in sytax of functions, with the support of automatic and configurable parallel execution and cache management.

Features

• Decomposing long and complex computation into tasks, i.e., smaller units of work with dependencies.
• Executing them in a distributed way, supporting multithreading/multiprocessing and local container/cluster-based dispatching.
• Automatically creating/discarding/reusing caches per task.

Nonfeatures

• Periodic scheduling
• Automatic retry
• External service integration (GCP, AWS, ...)
• Graphical user interface

Installation

pip install taskproc

Example

See here for a typical usage of taskproc.

Documentation

Basics

We define a task by class.

from taskproc import Task, Const, Cache

class Choose(Task):
    """ Compute the binomial coefficient. """

    def __init__(self, n: int, k: int):
        # Upstream tasks are prepared here.
        # Al attributes of a task with type `Task` are considered as the upstream tasks.
        if 0 < k < n:
            self.left = Choose(n - 1, k - 1)
            self.right = Choose(n - 1, k)
        elif k == 0 or k == n:
            # We can just pass a value without computation in place of a task.
            self.left = Const(0)
            self.right = Const(1)
        else:
            raise ValueError(f'{(n, k)}')

    def run_task(self) -> int:
        # Here we define the main computation of the task,
        # which is delayed until it is necessary.

        # The return values of the prerequisite tasks are accessible via `.get_result()`:
        return self.left.get_result() + self.right.get_result()

# Construct a concrete task with the class instantiation, which should be inside of `Cache`.
with Cache('./cache'):  # Specifies the cache directory
    task = Choose(6, 3)

# To run the task graph, use the `run_graph()` method.
ans, stats = task.run_graph()  # `ans` should be 6 Choose 3, which is 20.

# It greedily executes all the necessary tasks in the graph as parallel as possible
# and then produces the return value of the task on which we call `run_graph()`, as well as some execution stats.
# The return values of the intermediate tasks are cached on the specified location and reused on the fly whenever possible.

Commandline Interface

Task has a utility classmethod to run with commandline arguments, which is useful if all you need is to run a single task that does everything necessary. For example,

# taskfile.py

class Main(Task):
    def __init__(self):
        self.result = Choose(100, 50)
    
    def run_task(self):
        print(self.result.get_result())


if __name__ == '__main__':
    Main.cli()

Use --help option for more details.

Futures and Task Composition

To be more precise, any attributes of a task implementing the Future protocol are considered as upstream tasks. Tasks and constants are futures. One can pass a future into the initialization of another task to compose the computation.

class MyTask(Task):
    def __init__(self, upstream: Future[int], other_args: Any):
        ...

with Cache('./cache'):
    MyTask(
        upstream=UpstreamTaskProducingInt(),
        ...
    }).run_graph()

FutureList and FutureDict are for aggregating multiple Futures into one, allowing us to register multiple upstream tasks at once.

from taskproc import FutureList, FutureDict

class SummarizeScores(Task):

    def __init__(self, task_dict: dict[str, Future[float]]):
        self.score_list = FutureList([ScoringTask(i) for i in range(10)])
        self.score_dict = FutureDict(task_dict)

    def run_task(self) -> float:
        # `.get_result()` evaluates `FutureList[float]` and `FutureDict[str, float]` into
        # `list[float]` and `dict[str, float]`, respectively.
        return sum(self.score_dict.get_result().values()) / len(self.score_dict.get_result())

If the output is a sequence or a mapping, one can directly access its element by indexing. The result is also a Future.

class MultiOutputTask(Task):
    def run_task(self) -> dict[str, int]:
        return {'foo': 42, ...}

class DownstreamTask(Task):
    def __init__(self):
        self.dep = MultiOutputTask()['foo']

Input-output Specification

In general, tasks can be initialized with any JSON serializable objects which may or may not contain futures. Any non-jsonable objects can be also passed, as the output of a task.

SomeTask(1, 'foo', bar={'baz': TaskProducingNonJsonableObj(), 'other': [1, 2, 3]})

On the other hand, the output of a task, i.e., the return value of the .run_task() method, should be serializable with cloudpickle. The picked object is then compressed with gzip. The compression level can be changed as follows (defaults to 9).

class NoCompressionTask(Task):
    task_compress_level = 0
    ...

Deleting Cache

It is possible to selectively discard cache:

with Cache('./cache'):
    # After some modificaiton of `Choose(3, 3)`,
    # selectively discard the cache corresponding to the modification.
    Choose(3, 3).clear_task()

    # `ans` is recomputed tracing back to the computation of `Choose(3, 3)`.
    ans, _ = Choose(6, 3).run_graph()
    
    # Delete all the cache associated with `Choose`.
    Choose.clear_all_tasks()            

Data Directories

Use task.task_directory to get a fresh path dedicated to each task. The directory is automatically created and managed along with the cache:

class TrainModel(Task):
    def run_task(self) -> str:
        ...
        model_path = self.task_directory / 'model.bin'
        model.save(model_path)
        return model_path

Prefix Command

Tasks can be run with a prefix command, which is useful when working with a third-party job-scheduling or containerization command such as jbsub and docker.

class TaskWithJobScheduler(Task):
    task_prefix_command = 'jbsub -wait -queue x86_1h -cores 16+1 -mem 64g'
    ...

Execution Policy Configuration

One can control the task execution with concurrent.futures.Executor class:

from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor

class MyTask(Task):
    ...

with Cache('./cache'):
    # Limit the number of parallel workers
    MyTask().run_graph(executor=ProcessPoolExecutor(max_workers=2))
    
    # Thread-based parallelism
    MyTask().run_graph(executor=ThreadPoolExecutor())

One can also control the concurrency at a task level:

class TaskUsingGPU(Task):
    task_channel = 'gpu'
    ...

class AnotherTask(Task):
    task_channel = ['gpu', 'memory']
    ...

with Cache('./cache'):
    SomeDownstreamTask().run_graph(rate_limits={TaskUsingGPU.task_name: 1})  # Execute one at a time

Task Channels

Task prefixes and concurrency limits can be also specified via task channels, which is useful for controlling the computation resource of multiple tasks at once.

class TaskUsingGPU(Task):
    task_channel = 'gpu'
    ...

class AnotherTaskUsingGPU(Task):
    task_channel = ['gpu', 'memory']
    ...

with Cache('./cache'):
    # Channel-level concurrency control
    SomeDownstreamTask().run_graph(
        rate_limits={'gpu': 1, 'memory': 2},
        prefixes={'gpu': 'jbsub -wait -queue x86_1h -cores 16+1 -mem 64g'}
    ) 

Built-in properties/methods

Below is the list of the built-in properties/methods of Task. Do not override these attributes in the subclass.

Name	Owner	Type	Description
task_name	class	property	String id of the task class
task_id	instance	property	Integer id of the task, unique within the same task class
task_args	instance	property	The arguments of the task in JSON
task_directory	instance	property	Path to the data directory of the task
task_stdout	instance	property	Path to the task's stdout
task_stderr	instance	property	Path to the task's stderr
run_task	instance	method	Run the task
run_graph	instance	method	Run the task after necessary upstream tasks and save the results in the cache
get_result	instance	method	Get the result of the task (fails if the result is missing)
to_json	instance	method	Serialize itself as a JSON dictionary
clear_task	instance	method	Clear the cache of the task instance
clear_all_tasks	class	method	Clear the cache of the task class
cli	class	method	run_graph with command line arguments

TODO

• Potential bug
  • Current task argument serialization is not ideal since JSON is mapping two different values into the same text representation (e.g., tuple and list). Consider using consistency check x == json.loads(json.dumps(x)), or redesign the format.
• Better UX
  • Add the signature of task to the description to --kwargs of CLI.
  • Runtime validation of task arguments with type hint.
  • Add option to not cache result (need to address timestamp peeking and value passing).
• Simplify
  • Drop the support of ThreadPoolExecutor.
• Optional
  • Simple task graph visualizer.
  • Pydantic/dataclass support in task arguments (as an incompatible, but better-UX object with TypedDict).