[ARROW-7556] [Python][Parquet] Performance regression in pyarrow-0.15.1 vs pyarrow-0.12.1 when reading a "partitioned parquet table" ?

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 0.15.1
• Component/s: Benchmarking, Python
• Labels: dataset-parquet-read parquet
• External issue URL: https://github.com/apache/arrow/issues/23815
• Assignee: Unassigned
• Reporter: Julien MASSIOT
• Created: 12/Jan/20 22:44
• Updated: 11/Jan/23 07:54

Description

Hi,

I am currently running a small test with pyarrow to load 2 "partitioned" parquet tables.

The performance seems to be 2 times less with pyarrow-0.15.1 than with pyarrow-0.12.1.

In my test:

• the 'parquet' tables I am loading are called 'reports' & 'memory'
• they have been generated through pandas.to_parquet by specifying the partitions columns
• they are both partitioned by 2 columns 'p_type' and 'p_start'
• it is small tables:
  • reports
    • 90 partitions (1 parquet file / partition)
    • total size: 6.2MB
  • memory
    • 105 partitions (1 parquet file / partition)
    • total size: 9.1MB

 

Here is the code of my simple test that tries to read them (I'm using a filter on the p_start partition):

 

// code placeholder

import os
import sys
import time
import pyarrow
from pyarrow.parquet import ParquetDataset
def load_dataframe(data_dir, table, start_date, end_date):
    return ParquetDataset(os.path.join(data_dir, table),
                          filters=[('p_start', '>=', start_date),
                                   ('p_start', '<=', end_date)
                                   ]).read().to_pandas()

print(f'pyarrow version;{pyarrow.__version__}')

data_dir = sys.argv[1]

for i in range(1, 10):
    start = time.time()
    start_date = '201912230000'
    end_date = '202001080000'
    load_dataframe(sys.argv[1], 'reports', start_date, end_date)
    load_dataframe(sys.argv[1], 'memory', start_date, end_date)
    print(f'loaded;in;{time.time()-start}')

 

 

Here are the results:

• with pyarrow-0.12.1

$ python -m cProfile -o load_pyarrow_0.12.1.cprof load_df_from_pyarrow.py parquet/

pyarrow version;0.12.1

loaded;in;0.5566098690032959
loaded;in;0.32605648040771484
loaded;in;0.28951501846313477
loaded;in;0.29279112815856934
loaded;in;0.3474299907684326
loaded;in;0.4075736999511719
loaded;in;0.425199031829834
loaded;in;0.34653329849243164
loaded;in;0.300839900970459

(~350ms to load the 2 tables)

 

 

• with pyarrow-0.15.1

 

$ python -m cProfile -o load_pyarrow_0.15.1.cprof load_df_from_pyarrow.py parquet/

pyarrow version;0.15.1
loaded;in;1.1126022338867188
loaded;in;0.8931224346160889
loaded;in;1.3298325538635254
loaded;in;0.8584625720977783
loaded;in;0.9232609272003174
loaded;in;1.0619215965270996
loaded;in;0.8619768619537354
loaded;in;0.8686420917510986
loaded;in;1.1183602809906006

(>800ms to load the 2 tables)

 

 

Is there a performance regression here ?

Am I missing something ?

 

In attachment, you can find the 2 .cprof files.

load_pyarrow_0.12.1.cprof

load_pyarrow_0.15.1.cprof