narwhals 0.6.3

Extremely lightweight compatibility layer between pandas, Polars, cuDF, and Modin

Narwhals

Extremely lightweight compatibility layer between Polars, pandas, modin, and cuDF (and possibly more?).

Seamlessly support all, without depending on any!

• ✅ Just use a subset of the Polars API, no need to learn anything new
• ✅ No dependencies (not even Polars), keep your library lightweight
• ✅ Separate lazy and eager APIs
• ✅ Use Polars Expressions

Note: this is work-in-progress, and a bit of an experiment, don't take it too seriously.

Installation

pip install narwhals

Or just vendor it, it's only a bunch of pure-Python files.

Usage

There are three steps to writing dataframe-agnostic code using Narwhals:

1. use narwhals.LazyFrame or narwhals.DataFrame to wrap a pandas or Polars DataFrame/LazyFrame in a Narwhals class

2. use the subset of the Polars API supported by Narwhals. Just like in Polars, some methods (e.g. to_numpy) are only available for DataFrame, not LazyFrame

3. use narwhals.to_native to return an object to the user in its original dataframe flavour. For example:

   • if you started with pandas, you'll get pandas back
   • if you started with Polars, you'll get Polars back
   • if you started with Modin, you'll get Modin back (and compute will be distributed)
   • if you started with cuDF, you'll get cuDF back (and compute will happen on GPU)

Example

Here's an example of a dataframe agnostic function:

from typing import Any
import pandas as pd
import polars as pl

import narwhals as nw


def my_agnostic_function(
    suppliers_native,
    parts_native,
):
    suppliers = nw.LazyFrame(suppliers_native)
    parts = nw.LazyFrame(parts_native)

    result = (
        suppliers.join(parts, left_on="city", right_on="city")
        .filter(nw.col("weight") > 10)
        .group_by("s")
        .agg(
            weight_mean=nw.col("weight").mean(),
            weight_max=nw.col("weight").max(),
        )
    )
    return nw.to_native(result)

You can pass in a pandas or Polars dataframe, the output will be the same! Let's try it out:

suppliers = {
    "s": ["S1", "S2", "S3", "S4", "S5"],
    "sname": ["Smith", "Jones", "Blake", "Clark", "Adams"],
    "status": [20, 10, 30, 20, 30],
    "city": ["London", "Paris", "Paris", "London", "Athens"],
}
parts = {
    "p": ["P1", "P2", "P3", "P4", "P5", "P6"],
    "pname": ["Nut", "Bolt", "Screw", "Screw", "Cam", "Cog"],
    "color": ["Red", "Green", "Blue", "Red", "Blue", "Red"],
    "weight": [12.0, 17.0, 17.0, 14.0, 12.0, 19.0],
    "city": ["London", "Paris", "Oslo", "London", "Paris", "London"],
}

print("pandas output:")
print(
    my_agnostic_function(
        pd.DataFrame(suppliers),
        pd.DataFrame(parts),
    )
)
print("\nPolars output:")
print(
    my_agnostic_function(
        pl.LazyFrame(suppliers),
        pl.LazyFrame(parts),
    ).collect()
)

pandas output:
    s  weight_mean  weight_max
0  S1         15.0        19.0
1  S2         14.5        17.0
2  S3         14.5        17.0
3  S4         15.0        19.0

Polars output:
shape: (4, 3)
┌─────┬─────────────┬────────────┐
│ s   ┆ weight_mean ┆ weight_max │
│ --- ┆ ---         ┆ ---        │
│ str ┆ f64         ┆ f64        │
╞═════╪═════════════╪════════════╡
│ S2  ┆ 14.5        ┆ 17.0       │
│ S3  ┆ 14.5        ┆ 17.0       │
│ S4  ┆ 15.0        ┆ 19.0       │
│ S1  ┆ 15.0        ┆ 19.0       │
└─────┴─────────────┴────────────┘

Magic! 🪄

Scope

• Do you maintain a dataframe-consuming library?
• Is there a Polars function which you'd like Narwhals to have, which would make your job easier?

If, I'd love to hear from you!

Note: You might suspect that this is a secret ploy to infiltrate the Polars API everywhere. Indeed, you may suspect that.

Why "Narwhals"?

Because they are so awesome.

Thanks to Olha Urdeichuk for the illustration!