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.

pip install narwhals

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

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

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

2. use the subset of the Polars API supported by Narwhals

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)

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! 🪄

• 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.

Because they are so awesome.

Thanks to Olha Urdeichuk for the illustration!