For people using Polars, how well does the surrounding Python visualization and statistics ecosystem work without installing pandas or PyArrow? Some libraries still import pandas directly, while others convert data with to_pandas() internally, which can pull in PyArrow as well. In many cases, it seems possible to work around this by passing NumPy arrays or constructing a pandas DataFrame from individual NumPy columns, especially for numeric data without missing values. What has your experience been, and which libraries support Polars directly without these dependencies?
5 Answers
There’s no single answer because pandas is deeply embedded in many statistics packages. You can often avoid it with NumPy inputs, but that may mean losing dataframe labels, nullable types, or convenient formula interfaces. My practical approach is to use Polars for the heavy data work, convert only the smaller analysis or plotting subset when needed, and choose newer libraries with dataframe-agnostic support whenever possible.
A growing number of visualization tools use Narwhals, so they can accept Polars directly without requiring pandas or PyArrow. Plotly, Altair, Bokeh, HoloViews, and related tools are good options, depending on the type of chart and backend. Seaborn is still more tightly coupled to pandas, so replacing it may be necessary if avoiding pandas is a priority.
PyArrow is a fairly large dependency, but it also provides useful nullable types and a strong interchange format. If the only goal is occasional conversion to pandas, its size can feel disproportionate. Still, when a package already depends on Arrow or needs robust support for missing values and mixed data types, using PyArrow is often simpler than maintaining custom NumPy conversion code.
The ecosystem is in transition. Some projects use Narwhals for dataframe compatibility, while others prefer Arrow interfaces or the data interchange protocol. There are also efforts to use the Arrow C data interface through capsules, which can avoid a direct PyArrow dependency, but support is inconsistent and the ecosystem is not yet as mature as the pandas-based path. For now, pandas and PyArrow are common practical dependencies, but they are not necessarily permanent requirements.
I wouldn’t treat them as universally mandatory. Libraries are gradually adding more direct Polars, Narwhals, or capsule-based support, so it is worth checking each package individually before adding a conversion step.
In statistical and psychology workflows, I’ve found that converting to pandas is usually the least frustrating option. I keep Polars for cleaning, joining, reshaping, and general data preparation, then convert to pandas when I reach a statistics or visualization library. It’s not ideal, but it tends to avoid compatibility issues and saves time.
Altair works well for many static plots, while Plotly and Bokeh are useful when interactivity matters. HoloViews is also worth considering because it can sit above different plotting backends.