Jeroen van Zundert

I am not so convinced yet on the change from `exprs` to `columns`, because it documents the output (i.e. the dataframe returned has those columns), but not the inputs anymore....

Can reproduce on master: ```python >>> pl.from_dict({"num": [13888061535333455402]}, schema={"num": pl.UInt64}).item() 13888061535333455402 >>> pl.from_dicts([{"num": 13888061535333455402}], schema={"num": pl.UInt64}).item() 13888061535333455872 ``` Seems to happen in the Rust layer when reading the sequence of...

I don't know all the implementation details, but since the ordering is not being preserved when casting to Arrow format already, the IO part is really one way to manifest...

Given you point to PyArrow source code as the potential issue, and your stack trace points to Pandas, is this a Polars bug at all? For example, does this also...

Three thoughts I have: 1. It is, per your example, easy enough to code using the existing expression API if you use a simple constant, so the argument for adding...

On the naming: ok, maybe Pandas has just confused my brain, `mad` would be better than `aad`. I think overall this would be useful and worth implementing, but leaving to...

The script posted above has been useful to me, but it does ignore pins and bounds, on purpose, whereas I was looking for something where I could selectively not upgrade...

If you specify all your (direct) dependencies in `pyproject.toml` using pins (i.e. `==`), then the lock file can only change if someone has run `uv sync --upgrade` as, by default,...