Read a Parquet File Using Pandas

Parquet is a columnar storage format that is optimized for distributed processing of large datasets. It is widely used in Big Data processing systems like Hadoop and Apache Spark. 

A partitioned parquet file is a parquet file that is partitioned into multiple smaller files based on the values of one or more columns. Partitioning can significantly improve query performance by allowing the processing system to read only the necessary files.

Concepts

• Parquet format: A columnar storage format that is optimized for distributed processing of large datasets.
• Partitioning: Dividing a dataset into smaller parts based on the values of one or more columns.
• Pandas DataFrame: A two-dimensional labeled data structure with columns of potentially different types.
• pyarrow: A Python package that provides a Python interface to the Arrow C++ library for working with columnar data.

Dataset link: [https://www.kaggle.com/datasets/pawankumargunjan/weather%5D

Example 1

Install the pyarrow package:

The pyarrow package provides a Python interface to the Arrow C++ library for working with columnar data. Install it using the following command

pip install pyarrow

Import the pandas or pyarrow packages:

import pyarrow.parquet as pa

Reading parquet data using pyarrow.parquet

table = pa.read_table('weather.2016.parquet')
table

Output:

pyarrow.Table
ForecastSiteCode: int64
ObservationTime: int64
ObservationDate: timestamp[ms]
WindDirection: int64
WindSpeed: int64
WindGust: double
Visibility: double
ScreenTemperature: double
Pressure: double
SignificantWeatherCode: int64
SiteName: string
Latitude: double
Longitude: double
Region: string
Country: string
----
ForecastSiteCode: [[3002,3005,3008,3017,3023,...,3882,3002,3005,3008,3017],[3023,3026,3031,3034,3037,...,3797,3866,3872,3876,3882]]
ObservationTime: [[0,0,0,0,0,...,12,13,13,13,13],[13,13,13,13,13,...,23,23,23,23,23]]
ObservationDate: [[2016-02-01 00:00:00.000,2016-02-01 00:00:00.000,2016-02-01 00:00:00.000,2016-02-01 00:00:00.000,2016-02-01 00:00:00.000,...,2016-03-12 00:00:00.000,2016-03-12 00:00:00.000,2016-03-12 00:00:00.000,2016-03-12 00:00:00.000,2016-03-12 00:00:00.000],[2016-03-12 00:00:00.000,2016-03-12 00:00:00.000,2016-03-12 00:00:00.000,2016-03-12 00:00:00.000,2016-03-12 00:00:00.000,...,2016-03-31 00:00:00.000,2016-03-31 00:00:00.000,2016-03-31 00:00:00.000,2016-03-31 00:00:00.000,2016-03-31 00:00:00.000]]
WindDirection: [[12,10,8,6,10,...,4,8,8,8,8],[9,9,10,8,8,...,1,0,1,1,1]]
WindSpeed: [[8,2,6,8,30,...,5,19,18,19,19],[25,26,24,23,23,...,5,10,2,3,2]]
WindGust: [[null,null,null,null,37,...,null,null,null,null,29],[36,41,37,34,37,...,null,null,null,null,null]]
Visibility: [[30000,35000,50000,40000,2600,...,4000,8000,3500,11000,28000],[4600,9000,30000,10000,2900,...,22000,null,50000,null,35000]]
ScreenTemperature: [[2.1,0.1,2.8,1.6,9.8,...,10,-99,7.4,8.1,9.2],[9.1,9.5,10.2,9.7,9.9,...,4.9,8.4,3.5,6.1,3.7]]
Pressure: [[997,997,997,996,991,...,1030,null,1019,1020,1019],[1019,1018,1020,1020,1021,...,1019,1018,1019,1019,1019]]
SignificantWeatherCode: [[8,7,-99,8,11,...,1,5,15,12,7],[15,12,12,12,15,...,0,-99,0,-99,0]]
...

Print the shape of the dataset

table.shape

Output:

(194697, 15)

Convert the pyarrow table dataset into a pandas dataframe.

df = table.to_pandas()
# Taking tanspose so the printing dataset will easy.
df.head().T

Output:

Example 2

Reading parquet data using pandas.read_parquet

import pandas as pd

df = pd.read_parquet('weather.2016.parquet')
df.head()

Output:

Example 3

Filtering the parquet data

import pandas as pd

df = pd.read_parquet('weather.2016.parquet', filters=[('Country', '=', 'ENGLAND')])
df.head()

Output

This will filter the parquet data by Country='ENGLAND' and print the first 5 rows of the DataFrame.

Aggregating the parquet data

grouped = df.groupby(['Country']).mean(numeric_only=True)
# Taking tanspose so the printing dataset will easy.
print(grouped.T)

Output

Country                      ENGLAND
ForecastSiteCode         3560.622936
ObservationTime            11.517152
WindDirection               8.534412
WindSpeed                   7.770786
WindGust                   36.035424
Visibility              22431.530727
ScreenTemperature           5.336209
Pressure                 1011.335307
SignificantWeatherCode     -3.614757
Latitude                   52.354470
Longitude                  -1.586393

Group the data by Country, calculate the mean for each group, and print the resulting data frame.

Example 4

Multiple filters

import pandas as pd
df = pd.read_parquet('weather.2016.parquet', 
                     filters=[('Country', '=', 'ENGLAND'),
                              ('WindSpeed','<', 7)])
# Taking tanspose so the printing dataset will easy.
df.head().T

Output:

This will filter the parquet data by Country='ENGLAND' and WindSpeed< 7 and print the first 5 rows of the DataFrame.

Conclusion

In conclusion, partitioning parquet files is a powerful way to optimize data storage and querying performance. By partitioning data based on one or more columns, you can easily filter, sort, and aggregate data within a subset of partitions, rather than having to scan the entire dataset.

In this article, we covered two methods for reading partitioned parquet files in Python: using pandas' read_parquet() function and using pyarrow's ParquetDataset class. We also provided several examples of how to read and filter partitioned parquet files using these methods with real-world weather data.

Overall, partitioning parquet files is an effective technique for optimizing data storage and retrieval. Whether you're dealing with big data or just trying to improve query performance, partitioning can help you get the most out of your data.