[R] Reading a parquet with a Float16 column yields incorrect value

[rodigu]

Describe the bug, including details regarding any error messages, version, and platform.

i created a parquet file in polars where one of the columns has a Float16 type:

import polars as pl

df = pl.DataFrame(
    {"name": "john", "height": 1.78},
    schema={"name": pl.String, "height": pl.Float16},
).write_parquet("./data.parquet")

but reading it with arrow yields a different value:

library(arrow)

df <- read_parquet("./data.parquet")
print(df)

#   name height
# 1 john  16159

Component(s)

R

[raulcd]

In the example you are using pl.Float64, is this supposed to be Float16 instead? Also could you share versions of polars, and Arrow. Can you reproduce with PyArrow or is only R?

[rodigu]

In the example you are using pl.Float64, is this supposed to be Float16 instead? Also could you share versions of polars, and Arrow. Can you reproduce with PyArrow or is only R?

Yes, it was supposed to be Float16, sorry about that. I tested with other float types and those presented no issues.

Here are my versions:

Python 3.14.6
Polars 1.42.1

R 4.6.1
Arrow 24.0.0

PyArrow does not present the same issue when reading the same parquet file, only the arrow package with R.

[jmaspons]

Probably related to #46858

[rodigu]

Probably related to #46858

That's interesting.

How come it works on PyArrow but not R's arrow? It seems to me that that issue you referenced has to do with the C++ implementation itself. Wouldn't that mean PyArrow should have the same issue? Or does the PyArrow implementation draw from elsewhere?

[raulcd]

I don't think that one is related. Yes PyArrow and R Arrow are just bindings to the same C++ implementation. Maye @thisisnic knows if there are Float16 issues related to the R bindings?

[thisisnic]

I'm drafting a PR as we speak :) And it's to do with the conversion layer - while they both are based on the same C++ implementation, there are extra steps to convert from C++ to R/Python which have to be manually created. In the R codebase, it's in the R code's C++, some of which is here: https://github.com/apache/arrow/blob/main/r/src/array_to_vector.cpp

Or if you wanna see a PR where we're having to actively make decisions about the conversion, here's an example I was working on the other day: #50374

[thisisnic]

The root cause of the issue here is that the value you are returned is the raw value, but not the float value, so the fix here is to make sure it properly gets converted into a float instead of returning the raw bytes.

Or in other words, you're being returned the correct value, but the issue occurs when it's converted from C++ to R.