Awesome

Comparison of Python pipeline packages: Airflow, Luigi, Gokart, Metaflow, Kedro, PipelineX

This article compares open-source Python packages for pipeline/workflow development: Airflow, Luigi, Gokart, Metaflow, Kedro, PipelineX.

In this article, terms of "pipeline", "workflow", and "DAG" are used almost interchangeably.

Summary

👍: good
👍👍: better

Package	Airflow	Luigi	Gokart	Metaflow	Kedro	PipelineX
Developer, Maintainer	Airbnb, Apache	Spotify	M3	Netflix	Quantum-Black (McKinsey)	Yusuke Minami
Wrapped packages			Luigi			Kedro, MLflow
Easiness/flexibility to define DAG			👍	👍	👍	👍👍
Modularity of DAG definition	👍👍				👍👍	👍👍
Unstructured data can be passed between tasks		👍👍	👍👍	👍👍	👍👍	👍👍
Built-in various data (file/database) existence check wrappers		👍👍	👍👍		👍👍	👍👍
Built-in various data (file/database) operation (read/write) wrappers			👍		👍👍	👍👍
Modularity, reusability, testability of data operation			👍		👍👍	👍👍
Automatic resuming option by detecting the intermediate data		👍👍	👍👍			👍👍
Force rerun of tasks by detecting parameter change			👍👍
Save parameters for experiments			👍👍			👍👍
Parallel execution	👍	👍	👍	👍	👍	👍
Distributed parallel execution with Celery	👍👍
Visualization of DAG	👍👍	👍	👍		👍	👍
Execution status monitoring in GUI	👍👍	👍	👍
Scheduling, Triggering in GUI	👍
Notification to Slack	👍		👍

Airflow

https://github.com/apache/airflow

Released in 2015 by Airbnb.

Airflow enables you to define your DAG (workflow) of tasks in Python code (an independent Python module).

(Optionally, unofficial plugins such as dag-factory enables you to define DAG in YAML.)

Pros:

Provides rich GUI with features including DAG visualization, execution progress monitoring, scheduling, and triggering.
Provides distributed computing option (using Celery).
DAG definition is modular; independent from processing functions.
Workflow can be nested using SubDagOperator.
Supports Slack notification.

Cons:

Not designed to pass data between dependent tasks without using a database. There is no good way to pass unstructured data (e.g. image, video, pickle, etc.) between dependent tasks in Airflow.
You need to write file access (read/write) code.
Does not support automatic pipeline resuming option using the intermediate data files or databases.

Luigi

https://github.com/spotify/luigi

Released in 2012 by Spotify.

Luigi enables you to define your pipeline by child classes of Task with 3 class methods (requires, output, run) in Python code.

Pros:

Support automatic pipeline resuming option using the intermediate data files in local or cloud (AWS, GCP, Azure) or databases as defined in Task.output method using Target class.
You can write code so any data can be passed between dependent tasks.
Provides GUI with features including DAG visualization, execution progress monitoring.

Cons:

You need to write file/database access (read/write) code.
Pipeline definition, task processing (Transform of ETL), and data access (Extract&Load of ETL) are tightly coupled and not modular. You need to modify the task classes to reuse in future projects.

Gokart

https://github.com/m3dev/gokart

Released in Dec 2018 by M3.

Gokart works on top of Luigi.

Pros:

In addition to Luigi's advantages:

Can split task processing (Transform of ETL) from pipeline definition using TaskInstanceParameter so you can easily reuse them in future projects.
Provides built-in file access (read/write) wrappers as FileProcessor classes for pickle, npz, gz, txt, csv, tsv, json, xml.
Saves parameters for each experiment to assure reproducibility. Viewer called thunderbolt can be used.
Reruns tasks upon parameter change based on hash string unique to the parameter set in each intermediate file name. This feature is useful for experimentation with various parameter sets.
Syntactic sugar for Luigi's requires class method using class decorator.
Supports Slack notification.

Cons:

Supported data formats for file access wrappers are limited. You need to write file/database access (read/write) code to use unsupported formats.

Metaflow

https://github.com/Netflix/metaflow

Released in Dec 2019 by Netflix.

Metaflow enables you to define your pipeline as a child class of FlowSpec that includes class methods with step decorators in Python code.

Pros:

Integration with AWS services (Especially AWS Batch).

Cons:

You need to write file/database access (read/write) code.
Pipeline definition, task processing (Transform of ETL), and data access (Extract&Load of ETL) are tightly coupled and not modular. You need to modify the task classes to reuse in future projects.
Does not support GUI.
Not much support for GCP & Azure.
Does not support automatic pipeline resuming option using the intermediate data files or databases.

Kedro

https://github.com/quantumblacklabs/kedro

Released in May 2019 by QuantumBlack, part of McKinsey & Company.

Kedro enables you to define pipelines using list of node functions with 3 arguments (func: task processing function, inputs: input data name (list or dict if multiple), outputs: output data name (list or dict if multiple)) in Python code (an independent Python module).

Pros:

Provides built-in file/database access (read/write) wrappers as DataSet classes for CSV, Pickle, YAML, JSON, Parquet, Excel, and text in local or cloud (S3 in AWS, GCS in GCP), as well as SQL, Spark, etc.
Any data format support can be added by users.
Pipeline definition, task processing (Transform of ETL), and data access (Extract&Load of ETL) are independent and modular. You can easily reuse in future projects.
Pipelines can be nested. (A pipeline can be used as a sub-pipeline of another pipeline. )
GUI (kedro-viz) provides DAG visualization feature.

Cons:

Does not support automatic pipeline resuming option using the intermediate data files or databases.
GUI (kedro-viz) does not provide execution progress monitoring feature.
Package dependencies which are not used in many cases (e.g. pyarrow) are included in the requirements.txt.

PipelineX:

https://github.com/Minyus/pipelinex

Released in Nov 2019 by a Kedro user (me).

PipelineX works on top of Kedro and MLflow.

PipelineX enables you to define your pipeline in YAML (an independent YAML file).

Pros:

In addition to Kedro's advantages:

Supports automatic pipeline resuming option using the intermediate data files or databases.
Optional syntactic sugar for Kedro Pipeline. (e.g. Sequential API similar to PyTorch (torch.nn.Sequential) and Keras (tf.keras.Sequential))
Optional syntactic sugar for Kedro DataSet catalog. (e.g. Use file name in the file path as the dataset instance name)
Backward-compatible to pure Kedro.
Integration with MLflow to save parameters, metrics, and other output artifacts such as models for each experiment.
Integration with common packages for Data Science: PyTorch, Ignite, pandas, OpenCV.
Additional DataSet including image set (a folder including images) useful for computer vision applications.
Lean project template compared with pure Kedro.

Cons:

GUI (kedro-viz) does not provide execution progress monitoring feature.
Package dependencies which are not used in many cases (e.g. pyarrow) are included in the requirements.txt of Kedro.
PipelineX is developed and maintained by an individual (me) at this moment.