Incorrect serialization of ints using Apache Arrow

Summary

When using Apache Arrow to serialize dataframes (the default since Streamlit 0.85), negative ints are handled incorrectly.

Steps to reproduce

Code snippet:

import pandas as pd
import streamlit as st

df = pd.DataFrame({
    "ints": [-1, 0, 1],
    "floats": [-1.0, 0.0, 1.0],
})
st.table(df)

Expected behavior:

Using the legacy serializer, the negative ints are displayed correctly.

streamlit run --global.dataFrameSerialization=legacy apps/neg.py

Actual behavior:

With Arrow serialization, the -1 becomes 18446744073709551615.

streamlit run --global.dataFrameSerialization=arrow apps/neg.py

Is this a regression?

Yes. The legacy serializer handled negative ints correctly. The arrow serializer became the default in Streamlit 0.85.

Debug info

• Streamlit version: Streamlit, version 0.88.0
• Python version: Python 3.9.7
• Using Conda? PipEnv? PyEnv? Pex? pip
• OS version: Debian GNU/Linux 11 (python:3.9 container)
• Browser version: Safari 14.1.2 on macOS

Additional information

None.

[asaini]

cc: @kantuni

[LukasMasuch]

Closing this since it seems to work fine with the current version. I did quick tests with the current version of streamlit and a few different browsers and wasn’t able to reproduce this issue.