ml-pipeline-engine 2.2.0

Фреймворк для работы с пайплайном ML моделей

ML Pipeline Engine

Un-spaghetti and speed-up your data processing pipelines without usage of any complex DAG framework

Key benefits

• Fast in-memory pipeline runtime. Suitable for online transaction processing
• No Graph DSL needed. Just use Pyhton type annotations to declare node's dependencies
• Strong topological analysis based pipeline execution engine with MVCC support
• Synchronous and asyncronous processor nodes supported
• Asyncronous pipeline interface enabled by default
• Simple but powerful control flow operators - no need to declare branches or node groups, just declare right dependencies
• Built-in pipeline visualizer
• Support for pipeline lifecycle events

Table of Contents

• Usage
• Development
  • Environment setup

Usage

"""
main.py
"""

import asyncio
import time

from ml_pipeline_engine.chart import PipelineChart
from ml_pipeline_engine.dag_builders.annotation import build_dag
from ml_pipeline_engine.dag_builders.annotation.marks import Input
from ml_pipeline_engine.node import ProcessorBase
from ml_pipeline_engine.parallelism import threads_pool_registry


# 1. Setup thread pool

threads_pool_registry.auto_init()


# 2. Define nodes and their dependencies


class InvertNumber(ProcessorBase):
    def process(self, num: float) -> float:
        return -num


class AsyncAddConst(ProcessorBase):
    async def process(self, num: Input(InvertNumber), const: float = 0.2) -> float:
        await asyncio.sleep(2)

        return num + const


class DoubleNumber(ProcessorBase):
    def process(self, num: Input(InvertNumber)) -> float:
        time.sleep(2)

        return num * 2


class AddNumbers(ProcessorBase):
    def process(self, num1: Input(AsyncAddConst), num2: Input(DoubleNumber)) -> float:
        return num1 + num2


# 3. Define pipeline

pipeline = PipelineChart(
    "example_pipeline",
    build_dag(input_node=InvertNumber, output_node=AddNumbers),
)

# 4. Run it


async def main():
    start = time.time()

    result = await pipeline.run(input_kwargs=dict(num=3.0))

    end = time.time()

    assert result.error is None
    assert result.value == -8.8

    # Execution engine used concurrency, basing on graph topology analysis,
    # so AsyncAddConst and DoubleNumber nodes were ran in parallel
    assert end - start < 2.1


if __name__ == "__main__":
    asyncio.run(main())

See additional usage examples in docs: docs/examples/.

Development

Environment setup

Clone the project

git clone https://github.com/tochka-public/ml-pipeline-engine.git

Go to the project directory

cd ml-pipeline-engine

Use Python>=3.9 and the package manager poetry to install ml-pipeline-engine dependencies

poetry install --no-root

For further contribution, use pre-commit hooks to maintain consistent code format

pre-commit install -f --hook-type pre-commit --hook-type pre-push

Run tests

python -m pytest tests