I think that's the only parsing of strings/ binary to number we should support.

We probably should remove the cast from binary to integer support. This was never intended to work.

Personally, I think this is a useful capability to have. My use case stems from the fact that Delta Tables do not have unsigned ints, so oftentimes storing them as a binary type is a valid alternative.

FWIW: Databricks does support casting from binary to numerical types source.

It just seems to ambigous to me. Endianness, encoding etc all must be assumed. Maybe we can support it, but as a specific method, not as a generic cast.

I agree that it is ambiguous, but I think the only thing that is ambiguous is in endianness. I understand not wanting to introduce arguments to cast function. Maybe a separate function similar to decode could work in this case? This would allow for arbitrary endianness.

There is also no ambiguity for single byte sizes, so this could potentially be part of cast.

At the least, removing current behavior for Binary -> Numerical would be appreciated, since it is not I expected, and leads to silent errors.

I vote for removing the binary -> numerical cast and possibly add integer<->byte reinterpret function(s) in the future.

Will this issue and related PR fix the ability to cast pl.Binary series to pl.Array or pl.List? Please redirect me if it was asked somewhere else before.

I'm trying to import data from postgres table with bytea columns, which need to be interpreted as fixed-size float32 arrays. Using read_database_uri with connectorx engine, I get bytea columns as pl.Binary series, however it looks like it's not possible to cast them into any other usable data type:

import polars as pl
import numpy as np

# This data is actually coming from postgres bytea column via connectorx
a = np.random.rand(512).astype(np.float32).tobytes()
b = np.random.rand(512).astype(np.float32).tobytes()
df = pl.DataFrame(data=[{"data": a}, {"data": b}], schema={"data": pl.Binary})

print(df["data"].cast(pl.Array(pl.Float32, 512)))  # Not supported
print(df["data"].cast(pl.List(pl.Float32)))  # Not supported
print(df["data"].to_numpy().view("<f4"))  # Cannot change data-type for object array

Not yet. For now it is just numerical types. That being said, it can be extended to lists/arrays. My plan for now is:

• Merging currently open MR
• Adding support for lists/arrays
• Adding tobytes which is an inversion of from_buffer, following numpy convention