float and decimal have different round behavior (round up; half even)

[Julian-J-S]

Description

pl.DataFrame(
    {"d": ["0.5", "1.5"], "f": [0.5, 1.5]},
    schema_overrides={"d": pl.Decimal},
).with_columns(
    pl.all().round(decimals=0).name.suffix("_0"),
)


shape: (2, 4)
┌──────────────┬─────┬──────────────┬─────┐
│ d            ┆ f   ┆ d_0          ┆ f_0 │
│ ---          ┆ --- ┆ ---          ┆ --- │
│ decimal[*,1] ┆ f64 ┆ decimal[*,1] ┆ f64 │
╞══════════════╪═════╪══════════════╪═════╡
│ 0.5          ┆ 0.5 ┆ 0.0          ┆ 1.0 │
│ 1.5          ┆ 1.5 ┆ 2.0          ┆ 2.0 │
└──────────────┴─────┴──────────────┴─────┘

• 0.5 -> 0.0: decimal
• 0.5 -> 1.0: float (note: both 0.5 and 1.5 have EXACT representation as float)

Request:

• same/consistent round behavior across types
• (BONUS): configurable round behavior ⭐

[syedtaz]

Hi.

The polars-ops/series/ops/round module uses the default floating point round() function (which rounds up) and implements banker's rounding for decimals. Rust's stdlib includes round_ties_even() since 1.77 -- switching the functions around seems to work.

>>> import polars as pl
>>> pl.DataFrame({"d": ["0.5", "1.5"], "f": [0.5, 1.5]}, schema_overrides={"d": pl.Decimal}).with_columns(pl.all().round(decimals=0).name.suffix("_0")
... )
shape: (2, 4)
┌──────────────┬─────┬──────────────┬─────┐
│ d            ┆ f   ┆ d_0          ┆ f_0 │
│ ---          ┆ --- ┆ ---          ┆ --- │
│ decimal[*,1] ┆ f64 ┆ decimal[*,1] ┆ f64 │
╞══════════════╪═════╪══════════════╪═════╡
│ 0.5          ┆ 0.5 ┆ 0.0          ┆ 0.0 │
│ 1.5          ┆ 1.5 ┆ 2.0          ┆ 2.0 │
└──────────────┴─────┴──────────────┴─────┘

I am happy to work on this feature properly and implement the configurable rounding behavior if this enhancement is going to be accepted. I am not sure if there is some code out there that depends on the rounding-up behavior -- the docs for polars.Series.round at least don't seem to commit to anything.