Wrong type promotion on lazy parquet dataframe

[Hoeze]

Polars version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of Polars.

Issue description

When reading a parquet dataframe lazily and selecting some transformed Float32 column, the data type gets incorrectly reported as being Float64. After collecting the result however, it's correctly reported as being Float32 again.

Reproducible example

import polars as pl
import numpy as np

df = pl.DataFrame({
    "x": np.array([1,1,1], dtype="int32"),
    "y": np.array([1,2,3], dtype="float32"),
}).write_parquet("test.parquet")

df = pl.read_parquet("test.parquet").lazy()

print(df.schema)
# {'x': Int32, 'y': Float32}

print(df.select(pl.col("y")).schema)
# {'y': Float32}

# THIS FAILS! Data type should be reported as Float32
print(df.select(-pl.col("y")).schema)
# {'literal': Float64}

# After collecting the dataset, the type is correct
print(df.select(-pl.col("y")).collect().schema)
# {'literal': Float32}

Expected behavior

print(df.schema)
# {'x': Int32, 'y': Float32}

print(df.select(-pl.col("y")).schema)
# {'literal': Float32}

Installed versions

---Version info---
Polars: 0.15.14
Index type: UInt32
Platform: Linux-6.1.0-1.el8.elrepo.x86_64-x86_64-with-glibc2.28
Python: 3.9.9 | packaged by conda-forge | (main, Dec 20 2021, 02:41:03) 
[GCC 9.4.0]
---Optional dependencies---
pyarrow: 9.0.0
pandas: 1.5.2
numpy: 1.23.2
fsspec: 2022.01.0
connectorx: <not installed>
xlsx2csv: <not installed>
matplotlib: 3.5.1

[stinodego]

This has nothing to do with Parquet. A more minimal example:

import polars as pl
from polars.testing import assert_frame_equal

lf = pl.LazyFrame({"a": [1.0, 2.0]}, schema={"a": pl.Float32})

result = lf.select(-pl.col("a"))

expected = pl.LazyFrame({"a": [-1.0, -2.0]}, schema={"a": pl.Float32})
assert_frame_equal(result, expected)
# AssertionError: LazyFrames are different (dtypes do not match)
# [left]:  {'a': Float64}
# [right]: {'a': Float32}

There are two issues here:

• We calculate the negation by doing pl.lit(0) - expr, which casts to the supertype of expr and Int32. We should implement negation directly to avoid this.
• The supertype in this case is Float64. So the schema is correct, but the result of the minus operation is wrong. This should be fixed.

I'll make a fix.

[Hoeze]

Thanks a lot, @stinodego!

[stinodego]

I opened a PR that will nicely fix the code example in this issue.

I opened a separate issue for the underlying cause that should be fixed in another way: #13804