[ARROW-6791] Memory Leak

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 0.14.0, 0.14.1
• Component/s: Python
• Environment: Ubuntu 18.04, 32GB ram, conda-forge installation
• External issue URL: https://github.com/apache/arrow/issues/23126
• Assignee: Unassigned
• Reporter: George Prichard
• Created: 04/Oct/19 10:31
• Updated: 11/Jan/23 07:49
• Resolved: 07/Oct/19 02:50

Description

Memory leak with large string columns crashes the program. This only seems to affect 0.14.x  - it works fine for me in 0.13.0. It might be related to earlier similar issues? e.g. https://github.com/apache/arrow/issues/2624

Below is a reprex which works in earlier versions, but crashes on read (writing is fine) in this one. The real-life version of the data is full of URLs as the strings. 

Weirdly it crashes my 32GB Ubuntu 18.04, but runs (if very slowly for the read) on my 16GB Macbook. 

Thanks so much for the excellent tools! 

 

 

import pandas as pd

n_rows = int(1e6)
n_cols = 10
col_length = 100

df = pd.DataFrame()

for i in range(n_cols):
    df[f'col_{i}'] = pd.util.testing.rands_array(col_length, n_rows)

print('Generated df', df.shape)
filename = 'tmp.parquet'

print('Writing parquet')
df.to_parquet(filename)

print('Reading parquet')
pd.read_parquet(filename)