Skip to content

Extending Narwhals, levels of support

List of supported libraries (and how to add yours!)

Currently, Narwhals has full API support for the following libraries:

Library 🔗 Link 🔗
️Polars 🐻‍❄️ github.com/pola-rs/polars
pandas 🐼 github.com/pandas-dev/pandas
cuDF github.com/rapidsai/cudf
Modin github.com/modin-project/modin
PyArrow ⇶ arrow.apache.org/docs/python

It also has lazy-only support for Dask, and interchange-only support for DuckDB and Ibis.

Levels

Narwhals comes with two levels of support ("full" and "interchange"), and we are working on defining a "lazy-only" level too.

Libraries for which we have full support can benefit from the whole Narwhals API.

For example:

import narwhals as nw
from narwhals.typing import FrameT


@nw.narwhalify
def func(df: FrameT) -> FrameT:
    return df.group_by("a").agg(
        b_mean=nw.col("b").mean(),
        b_std=nw.col("b").std(),
    )

will work for any of pandas, Polars, cuDF, Modin, and PyArrow.

However, sometimes you don't need to do complex operations on dataframes - all you need is to inspect the schema a bit before making other decisions, such as which columns to select or whether to convert to another library. For that purpose, we also provide "interchange" level of support. If a library implements the Dataframe Interchange Protocol, then a call such as

from typing import Any

import narwhals as nw
from narwhals.schema import Schema


def func(df: Any) -> Schema:
    df = nw.from_native(df, eager_or_interchange_only=True)
    return df.schema

is also supported, meaning that, in addition to the libraries mentioned above, you can also pass Ibis, DuckDB, Vaex, and any library which implements the protocol.

Interchange-only support

While libraries for which we have full support can benefit from the whole Narwhals API, libraries which have interchange only support can access the following methods after converting to Narwhals DataFrame:

  • .schema, hence column names via .schema.names() and column types via .schema.dtypes()
  • .columns
  • .to_pandas() and .to_arrow(), for converting to Pandas and Arrow, respectively.
  • .select(names) (Ibis and DuckDB), where names is a list of (string) column names. This is useful for selecting columns before converting to another library.

Extending Narwhals

If you want your own library to be recognised too, you're welcome open a PR (with tests)!. Alternatively, if you can't do that (for example, if you library is closed-source), see the next section for what else you can do.

We love open source, but we're not "open source absolutists". If you're unable to open source you library, then this is how you can make your library compatible with Narwhals.

Make sure that, in addition to the public Narwhals API, you also define:

  • DataFrame.__narwhals_dataframe__: return an object which implements public methods from Narwhals.DataFrame
  • DataFrame.__narwhals_namespace__: return an object which implements public top-level functions from narwhals (e.g. narwhals.col, narwhals.concat, ...)
  • DataFrame.__native_namespace__: return a native namespace object which must have a from_dict method
  • LazyFrame.__narwhals_lazyframe__: return an object which implements public methods from Narwhals.LazyFrame
  • LazyFrame.__narwhals_namespace__: return an object which implements public top-level functions from narwhals (e.g. narwhals.col, narwhals.concat, ...)
  • LazyFrame.__native_namespace__: return a native namespace object which must have a from_dict method
  • Series.__narwhals_series__: return an object which implements public methods from Narwhals.Series

If your library doesn't distinguish between lazy and eager, then it's OK for your dataframe object to implement both __narwhals_dataframe__ and __narwhals_lazyframe__. In fact, that's currently what narwhals._pandas_like.dataframe.PandasLikeDataFrame does. So, if you're stuck, take a look at the source code to see how it's done!

Note that this "extension" mechanism is still experimental. If anything is not clear, or doesn't work, please do raise an issue or contact us on Discord (see the link on the README).