arrow-datafusion VS arrow-rs

Ballista is one possible Spark replacement, runs on top of Apache DataFusion https://arrow.apache.org/datafusion/

Have fun with in-memory columnar databases or SQL engines and see how fast they are (the ones that use CPU-efficient data structures and data element access, plus SIMD processing). For example Apache Arrow datafusion [1]

[1] https://github.com/apache/arrow-datafusion

This project looks cool: https://github.com/apache/arrow-datafusion

A while ago I posted a link to [Arc](https://news.ycombinator.com/item?id=26573930) a declarative method for defining repeatable data pipelines which execute against [Apache Spark](https://spark.apache.org/).

Today I would like to present a proof-of-concept implementation of the [Arc declarative ETL framework](https://arc.tripl.ai) against [Apache Datafusion](https://arrow.apache.org/datafusion/) which is an Ansi SQL (Postgres) execution engine based upon Apache Arrow and built with Rust.

The idea of providing a declarative 'configuration' language for defining data pipelines was planned from the beginning of the Arc project to allow changing execution engines without having to rewrite the base business logic (the part that is valuable to your business). Instead, by defining an abstraction layer, we can change the execution engine and run the same logic with different execution characteristics.

The benefit of the DataFusion over Apache Spark is a significant increase in speed and reduction in execution resource requirements. Even through a Docker-for-Mac inefficiency layer the same job completes in ~4 seconds with DataFusion vs ~24 seconds with Apache Spark (including JVM startup time). Without Docker-for-Mac layer end-to-end execution times of 0.5 second for the same example job (TPC-H) is possible. * the aim is not to start a benchmarking flamewar but to provide some indicative data *.

The purpose of this post is to gather feedback from the community whether you would use a tool like this, what features would be required for you to use it (MVP) or whether you would be interested in contributing to the project. I would also like to highlight the excellent work being done by the DataFusion/Arrow (and Apache) community for providing such amazing tools to us all as open source projects.

You should check out polars, datafusion, influxdb iox and databend, all written in native Rust and powered by the Apache Arrow format. Polars in particular is pretty dam fast and has bindings for Python.

A solution for either Polars or Datafusion (or something else?) would be fine. For both libraries, python packages exist, that contain the python bindings: https://github.com/pola-rs/polars/tree/master/py-polars https://github.com/apache/arrow-datafusion/tree/master/python

That matrix doesn't seem up to date. For example looking at the rust crate it does seem to support things like map, float16, and IPC. The changelog shows an impressive development pace.

Oh, and for anyone interested in pitching in on the Rust implementation, there's an issue logged here along with some discussion: https://github.com/apache/arrow-rs/issues/1323

There is more discussion about the decision here: https://github.com/apache/arrow-rs/issues/1120

Deeply nested parquet support would be nice. Even the official implementation lack this. https://github.com/apache/arrow-rs/issues/993

I'm still not sure how this differs from https://github.com/apache/arrow-rs. What does transmute even mean?

Cool! You may want to take a look at the Apache Arrow, rust project and datafusion: https://github.com/apache/arrow-rs and https://github.com/apache/arrow-datafusion

Congrats team and great work @elferherrera! Note that this backed by Polars and Arrow, and is as fast as it gets. :)