modak 0.3.4

A simple, opinionated task manager

modak

modak is a simple-to-use, opinionated task queue system with dependency management, resource allocation, and isolation control. Tasks are run respecting topological dependencies, resource limits, and optional isolation.

This library only has two classes, Tasks, which are an abstract class with a single method to override, run(self) -> None, and a TaskQueue which manages the execution order. Additionally, modak comes with a task monitor TUI which can be invoked with the modak shell command.

The TaskQueue has been written in Rust to get past issues with parallelism and the GIL. Instead of using a thread pool or even a multiprocessing pool, the tasks are serialized into bytes and passed to the Rust-side manager, which handles dispatching and execution. Each task is then run as a separate subprocess spawned in a Rust thread. This means the only way to share state between tasks is by writing to an output file and having a task depend on that file.

By default, modak scripts will create a state file called .modak in the current working directory. This can be changed by setting it in the TaskQueue's initialization method. The modak CLI also supports an optional argument to point to the location of the state file.

Features

• Topological task scheduling
• Persistent state and log files
• Resource-aware execution
• Isolated task handling
• Skipping of previously completed tasks

Installation

pip install modak

Or with uv:

pip install modak

FAQ

Q: What do you mean by "opinionated"?

A: The library is meant to do one thing (and hopefully do it well): run tasks and write output files. Some users might want more flexibility, like writing to a database or having a target that isn't written to at all, but that is not a goal of this library. If you need this level of control, try airflow or luigi.

Q: Why make another task manager?

A: luigi is nice, but I've been annoyed by the poor type hints for task parameters. It's also very confusing for first-time users, and has a lot of features that I don't really think people use unless they are working with products like Spotify. I built modak with research pipelines in mind, so I wanted something that was so simple to use, you don't have to think too hard about what you're doing and can focus on the data instead. I haven't used airflow much, but it also seems like a tool intended for enterprise. My goal here is simplicity and a minimal learning curve. There are only two classes. luigi has the added annoyance of running a web server to visualize the state of the DAG, which is very tricky to use on a remote server if you don't have the proper permissions.

Q: Isn't Rust a bit overkill?

A: Rust isn't as scary as it sounds. I don't actually care much about memory safety (although I'll take it for free), I like the development experience.

Q: Any sharp corners?

A: In development, I've found that libraries that do something when imported need to be handled with care. Such libraries should be imported inside the run method of the task. This is because the task gets serialized and sent to the __main__ module, but the imports from your code are run before serialization. An example of this is the loguru library, which sets up the global logger on import. If loguru is only imported outside the task, the logger instance will have no sink added because these lines will not be run when the task is deserialized. This will not effect most code, it's just something to be aware of.

Examples

A simple chain of tasks

from modak import Task, TaskQueue

class PrintTask(Task):
    def run(self):
        self.logger.info(f"Running {self.name}")

t1 = PrintTask(name="task1")
t2 = PrintTask(name="task2", inputs=[t1])
t3 = PrintTask(name="task3", inputs=[t2])

queue = TaskQueue('example1')
queue.run([t3])

Fan-in, fan-out

from pathlib import Path
from modak import Task, TaskQueue

class DummyTask(Task):
    def run(self):
        self.logger.info(f"Running {self.name}")
        for output in self.outputs:
            output.write_text(f"Output of {self.name}")

# Leaf tasks
a = DummyTask(name="A", outputs=[Path("a.out")])
b = DummyTask(name="B", outputs=[Path("b.out")])
c = DummyTask(name="C", outputs=[Path("c.out")])

# Fan-in: D depends on A, B, C
d = DummyTask(name="D", inputs=[a, b, c], outputs=[Path("d.out")])

# Fan-out: E and F both depend on D
e = DummyTask(name="E", inputs=[d], outputs=[Path("e.out")])
f = DummyTask(name="F", inputs=[d], outputs=[Path("f.out")])

queue = TaskQueue('example2')
queue.run([e, f])

A complex workflow

from pathlib import Path
from modak import Task, TaskQueue

class SimTask(Task):
    def run(self):
        self.logger.info(f"{self.name} starting with {self.resources}")
        for out in self.outputs:
            out.write_text(f"Generated by {self.name}")

# Raw data preprocessing
pre_a = SimTask(name="PreA", outputs=[Path("a.pre")], resources={"cpu": 1})
pre_b = SimTask(name="PreB", outputs=[Path("b.pre")], resources={"cpu": 1})
pre_c = SimTask(name="PreC", outputs=[Path("c.pre")], resources={"cpu": 1})

# Feature extraction (can run in parallel)
feat1 = SimTask(name="Feature1", inputs=[pre_a], outputs=[Path("a.feat")], resources={"cpu": 2})
feat2 = SimTask(name="Feature2", inputs=[pre_b], outputs=[Path("b.feat")], resources={"cpu": 2})
feat3 = SimTask(name="Feature3", inputs=[pre_c], outputs=[Path("c.feat")], resources={"cpu": 2})

# Aggregation step
aggregate = SimTask(
    name="Aggregate",
    inputs=[feat1, feat2, feat3],
    outputs=[Path("agg.out")],
    resources={"cpu": 3}
)

# Final model training (expensive, must be isolated)
train = SimTask(
    name="TrainModel",
    inputs=[aggregate],
    outputs=[Path("model.bin")],
    isolated=True,
    resources={"cpu": 3, "gpu": 1}
)

# Side analysis and visualization can run independently
viz = SimTask(name="Visualization", inputs=[feat1, feat2], outputs=[Path("viz.png")], resources={"cpu": 1})
stats = SimTask(name="Stats", inputs=[feat3], outputs=[Path("stats.txt")], resources={"cpu": 1})

queue = TaskQueue(
    'example3',
    workers=4,
    resources={"cpu": 4, "gpu": 1}
)

queue.run([train, viz, stats])

Future Plans

I'll probably make small improvements to the TUI and add features as I find the need. Contributions are welcome, just open an issue or pull request on GitHub and I'll try to respond as soon as I can.