Float32 accuracy issue about polars HOT 4 CLOSED

Maybe consider opening an issue at IEEE_754? ;)

from polars.

@Macfly to expound a bit on Ritchie's answer, numbers are stored in the computer in binary format with a limited number of bits, and so we cannot represent all numbers exactly. If you take a look at https://www.omnicalculator.com/other/floating-point and select Number to Floating Point and enter 799.32, you'll see that the number stored as a 32-bit float actually represents the number 799.32000732421875. If you set the number of bits to 64, you get a whole lot more precision, but it's still not exact: 799.32000000000005.

from polars.

I understand the floating accuracy but why with Numpy and Pandas, it is displayed as 799.32 with float32.
Is there something special with polars that always show the exact floating representation?

from polars.

numpy does the same thing actually

>>> pl.Series([799.32], dtype=pl.Float32).item()
799.3200073242188
>>> np.float32(799.32).item()
799.3200073242188
>>> np.float32(799.32).astype(np.float64)
799.3200073242188

when you do Series([799.32], dtype=pl.Float32)[0], you are actually performing these operations:

>>> np.float64(799.32).astype(np.float32).astype(np.float64).item()
799.3200073242188

you start with a 64-bit Python float, which gets casted to a 32-bit float by polars, and then back to a 64-bit Python float when you do Series[0]

from polars.