Polars

Default stance

When no data processing library has been specified, use Polars. Always use the lazy API: build the whole query, then execute it with one .collect(), so Polars can optimize the plan before any data is read. Eager execution skips that optimization, so the same logic does more work.

Data already in memory is no exception. When a query starts from a DataFrame — a function argument, pl.from_pandas(pdf), pl.DataFrame(...), or an earlier collect() — call .lazy() before the first operation and .collect() at the end. .lazy() is free (it wraps the data, it does not copy it) and it is what turns a step-by-step eager pipeline into one optimized plan. Do this unless the user asks for eager execution.

# start from a file
pl.scan_csv("data.csv").filter(...).collect()