beavers 0.13.0

Python stream processing

Beavers

Documentation / Installation / Repository / PyPI

Beavers is a python library for stream processing, optimized for analytics.

It is used at Tradewell Technologies, to calculate analytics and serve model predictions, for both realtime and batch jobs.

Key Features

• Works in real time (eg: reading from Kafka) and replay mode (eg: reading from Parquet files).
• Optimized for analytics, using micro-batches (instead of processing records one by one).
• Similar to incremental, it updates nodes in a dag incrementally.
• Taking inspiration from kafka streams, there are two types of nodes in the dag:
  • Stream: ephemeral micro-batches of events (cleared after every cycle).
  • State: durable state derived from streams.
• Clear separation between the business logic and the IO. So the same dag can be used in real time mode, replay mode or can be easily tested.
• Functional interface: no inheritance or decorator required.
• Support for complicated joins, not just "linear" data flow.

Limitations

• No concurrency support. To speed up calculation use libraries like pandas, pyarrow or polars.
• No async code. To speed up IO use kafka driver native thread or parquet IO thread pool.
• No support for persistent state. Instead of saving state, replay historic data from kafka to prime stateful nodes.

Talks

• Unified batch and stream processing in python | PyData Global 2023