20 releases (breaking)

0.15.0 Jun 22, 2024
0.13.0 Apr 14, 2024
0.12.0 Feb 29, 2024
0.9.0 Nov 18, 2023
0.2.0 Feb 15, 2023

#138 in Rust patterns

Download history 3419/week @ 2024-03-27 4265/week @ 2024-04-03 5088/week @ 2024-04-10 4595/week @ 2024-04-17 4323/week @ 2024-04-24 6219/week @ 2024-05-01 7510/week @ 2024-05-08 7976/week @ 2024-05-15 8821/week @ 2024-05-22 10439/week @ 2024-05-29 9586/week @ 2024-06-05 11820/week @ 2024-06-12 13958/week @ 2024-06-19 15409/week @ 2024-06-26 13290/week @ 2024-07-03 13176/week @ 2024-07-10

57,814 downloads per month
Used in 7 crates

MIT license

24KB
339 lines

1. Shared library plugins for Polars

Documentation for this functionality may also be found in the Polars User Guide. This is new functionality and should be preferred over 2. as this will circumvent the GIL and will be the way we want to support extending polars.

Parallelism and optimizations are managed by the default polars runtime. That runtime will call into the plugin function. The plugin functions are compiled separately.

We can therefore keep polars more lean and maybe add support for a polars-distance, polars-geo, polars-ml, etc. Those can then have specialized expressions and don't have to worry as much for code bloat as they can be optionally installed.

The idea is that you define an expression in another Rust crate with a proc_macro polars_expr.

The macro may have one of the following attributes:

  • output_type -> to define the output type of that expression
  • output_type_func -> to define a function that computes the output type based on input types.
  • output_type_func_with_kwargs -> to define a function that computes the output type based on input types and keyword args.

Here is an example of a String conversion expression that converts any string to pig latin:

fn pig_latin_str(value: &str, capitalize: bool, output: &mut String) {
    if let Some(first_char) = value.chars().next() {
        if capitalize {
            for c in value.chars().skip(1).map(|char| char.to_uppercase()) {
                write!(output, "{c}").unwrap()
            }
            write!(output, "AY").unwrap()
        } else {
            let offset = first_char.len_utf8();
            write!(output, "{}{}ay", &value[offset..], first_char).unwrap()
        }
    }
}

#[derive(Deserialize)]
struct PigLatinKwargs {
    capitalize: bool,
}

#[polars_expr(output_type=String)]
fn pig_latinnify(inputs: &[Series], kwargs: PigLatinKwargs) -> PolarsResult<Series> {
    let ca = inputs[0].str()?;
    let out: StringChunked =
        ca.apply_to_buffer(|value, output| pig_latin_str(value, kwargs.capitalize, output));
    Ok(out.into_series())
}

This can then be exposed on the Python side:

import polars as pl
from polars.type_aliases import IntoExpr
from polars.utils.udfs import _get_shared_lib_location

from expression_lib.utils import parse_into_expr

lib = _get_shared_lib_location(__file__)


def pig_latinnify(expr: IntoExpr, capitalize: bool = False) -> pl.Expr:
    expr = parse_into_expr(expr)
    return expr.register_plugin(
        lib=lib,
        symbol="pig_latinnify",
        is_elementwise=True,
        kwargs={"capitalize": capitalize},
    )

Compile/ship and then it is ready to use:

import polars as pl
from expression_lib import language

df = pl.DataFrame({
    "names": ["Richard", "Alice", "Bob"],
})


out = df.with_columns(
   pig_latin = language.pig_latinnify("names")
)

Alternatively, you can register a custom namespace, which enables you to write:

out = df.with_columns(
   pig_latin = pl.col("names").language.pig_latinnify()
)

See the full example in example/derive_expression: https://github.com/pola-rs/pyo3-polars/tree/main/example/derive_expression

2. Pyo3 extensions for Polars

See the example directory for a concrete example. Here we send a polars DataFrame to rust and then compute a jaccard similarity in parallel using rayon and rust hash sets.

Run example

$ cd example && make install $ venv/bin/python run.py

This will output:

shape: (2, 2)
┌───────────┬───────────────┐
│ list_a    ┆ list_b        │
│ ------           │
│ list[i64] ┆ list[i64]     │
╞═══════════╪═══════════════╡
│ [1, 2, 3][1, 2, ... 8] │
│ [5, 5][5, 1, 1]     │
└───────────┴───────────────┘
shape: (2, 1)
┌─────────┐
│ jaccard │
│ ---     │
│ f64     │
╞═════════╡
│ 0.75    │
│ 0.5     │
└─────────┘

Compile for release

$ make install-release

What to expect

This crate offers a PySeries and a PyDataFrame which are simple wrapper around Series and DataFrame. The advantage of these wrappers is that they can be converted to and from python as they implement FromPyObject and IntoPy.

Dependencies

~16–28MB
~440K SLoC