arrow2 VS db-benchmark

One of the interesting components of Polars that I've been watching is the use of the Apache Arrow memory format, which is a standard layout for data in memory that enables processing (querying, iterating, calculating, etc) in a language agnostic way, in particular without having to copy/convert it into the local object format first. This enables cross-language data access by mmaping or transferring a single buffer, with zero [de]serialization overhead.

For some history, there's has been a bit of contention between the official arrow-rs implementation and the arrow2 implementation created by the polars team which includes some extra features that they find important. I think the current status is that everyone agrees that having two crates that implement the same standard is not ideal, and they are working to port any necessary features to the arrow-rs crate and plan on eventually switching to it and deprecating arrow2. But that's not easy.

https://github.com/apache/arrow-rs/issues/1176

https://github.com/jorgecarleitao/arrow2/pull/1476

https://github.com/jorgecarleitao/arrow2 https://github.com/apache/arrow-datafusion https://github.com/apache/arrow-ballista https://github.com/pola-rs/polars https://github.com/duckdb/duckdb

Currently datafusion and polars aren't directly operable iirc because they use different underlying arrows implementations, but there seems to be work being done on that here https://github.com/jorgecarleitao/arrow2/issues/1429

[arrow2](https://github.com/jorgecarleitao/arrow2) and [parquet2](https://github.com/jorgecarleitao/parquet2) are great foundational libraries for and DataFrame libs in Rust.

[1] https://github.com/jorgecarleitao/arrow2

In lockstep with arrow2's 0.13 release, we have published polars 0.23.0.

This is so freaking life-saving! - we have been using StreamingIterator and FallibleStreamingIterator in libraries (arrow2 and parquet2) and the existing landscape is quite confusing for new users!

arrow2 has support for mssql via ODBC (which microsoft has first class support to). Here are the integration tests we have (both read and write) against mssql specifically.

Real-world performance is complicated since data science covers a lot of use cases.

If you're just reading a small CSV to do analysis on it, then there will be no human-perceptible difference between Polars and Pandas. If you're reading a larger CSV with 100k rows, there still won't be much of a perceptible difference.

Per this (old) benchmark, there are differences once you get into 500MB+ territory: https://h2oai.github.io/db-benchmark/

I do think it was important for duckdb to put out a new version of the results as the earlier version of that benchmark [1] went dormant with a very old version of duckdb with very bad performance, especially against polars.

[1] https://h2oai.github.io/db-benchmark/

https://news.ycombinator.com/item?id=33270638 :

> Apache Ballista and Polars do Apache Arrow and SIMD.

> The Polars homepage links to the "Database-like ops benchmark" of {Polars, data.table, DataFrames.jl, ClickHouse, cuDF, spark, (py)datatable, dplyr, pandas, dask, Arrow, DuckDB, Modin,} but not yet PostgresML? https://h2oai.github.io/db-benchmark/ *

LLM -> Vector database: https://en.wikipedia.org/wiki/Vector_database

/? inurl:awesome site:github.com "vector database"

I agree with your conclusion but want to add that switching from Julia may not make sense either.

According to these benchmarks: https://h2oai.github.io/db-benchmark/, DF.jl is the fastest library for some things, data.table for others, polars for others. Which is fastest depends on the query and whether it takes advantage of the features/properties of each.

For what it's worth, data.table is my favourite to use and I believe it has the nicest ergonomics of the three I spoke about.

Same. Numba does wonders for me in most scenarios. Yesterday I've discovered pola-rs and looks like I will add it to the stack. It's API is similar to pandas. Have a look at the benchmarks of cuDF, spark, dask, pandas compared to it: Benchmarks

The syntax has similarities with dplyr in terms of the way you chain operations, and it’s around an order of magnitude faster than pandas and dplyr (there’s a nice benchmark here). It’s also more memory-efficient and can handle larger-than-memory datasets via streaming if needed.

If interested in benchmarks comparing different dataframe implementations, here is one:

https://h2oai.github.io/db-benchmark/

Bad examples. Both numpy and pandas are notoriously un-optimized packages, losing handily to pretty much all their competitors (R, Julia, kdb+, vaex, polars). See https://h2oai.github.io/db-benchmark/ for a partial comparison.