An open source framework that bridges big data and AI.


Keywords
machine-learning-workflow, python
License
Apache-2.0
Install
pip install ai-flow-nightly==2023.8.31

Documentation

AIFlow

CI codecov

Introduction

AIFlow is an event-based workflow orchestration platform that allows users to programmatically author and schedule workflows with a mixture of streaming and batch tasks.

Most existing workflow orchestration platforms (e.g. Apache AirFlow, KubeFlow) schedule task executions based on the status changes of upstream task executions. While this approach works well for batch tasks that are guaranteed to end, it does not work well for streaming tasks which might run for an infinite amount of time without status changes. AIFlow is proposed to facilitate the orchestration of workflows involving streaming tasks.

For example, users might want to run a Flink streaming job continuously to assemable training data, and start a machine learning training job everytime the Flink job has processed all upstream data for the past hour. In order to schedule this workflow using non-event-based workflow orchestration platform, users need to schedule the training job periodically based on wallclock time. If there is traffic spike or upstream job failure, then the Flink job might not have processed the expected amount of upstream data by the time the TensorFlow job starts. The upstream job should either keep waiting, or fail fast, or process partial data, none of which is ideal. In comparison, AIFlow provides APIs for the Flink job to emit an event every time its event-based watermark increments by an hour, which triggers the execution of user-specified training job, without suffering the issues described above.

Learn more about AIFlow at https://ai-flow.readthedocs.io

Features

  1. Event-driven: AIFlow schedule workflow and jobs based on events. This is more efficient than status-driven scheduling and be able to schedule the workflows that contain stream jobs.

  2. Extensible: Users can easily define their own operators and executors to submit various types of tasks to different platforms.

  3. Exactly-once: AIFlow provides an event processing mechanism with exactly-once semantics, which means that your tasks will never be missed or repeated even if a failover occurs.

Articles on AIFlow

Contributing

We happily welcome contributions to AIFlow in any ways, whether reporting problems, drafting features, or contributing code changes. You can report problems to request features in the GitHub Issues. If you want to contribute code changes, please check out the contributing documentation.

Contact Us

For more information, we recommend you to join the AIFlow Community Group on the Google Groups to contact us: aiflow@googlegroups.com.

You can also join the group on the DingTalk. The number of the DingTalk group is 35876083, which group can also be joined by scanning the QR code below: