Sarvesh Kumar

How to Convert CSV to Parquet Files?

Convert CSV to Parquet Files

Apache Parquet

Apache Parquet is a columnar data storage format, which provides a way to store tabular data column wise. Columns of same date-time are stored together as rows in Parquet format, so as to offer better storage, compression and data retrieval.

What is Row Oriented Storage Format?

In row oriented storage, data is stored row wise on to the disk.

data is stored row wise on to the disk

For example above table has three columns of different data types (Integer, String and Double). This table will be saved on disk row wise in row oriented storage format as following-

This table will be saved on disk row wise

Columnar Storage Format

In columnar storage format above table will be stored column wise.

Columnar Storage Format

As you can see in this format all the IDs are together and so are names and salaries. A Query selecting Name column will require less I/O time as all the values are adjacent unlike in row oriented format.

Actual implementation of columnar format for Apache Parquet is defined here.

Using Apache Parquet

Using Parquet format has two advantages

  1. Reduced storage
  2. Query performance

Depending on your business use case, Apache Parquet is a good option if you have to provide partial search features i.e. not querying all the columns, and you are not worried about file write time.

Apache Parquet format is supported in all Hadoop based frameworks. Queries selecting few columns from a big set of columns, run faster because disk I/O is much improved because of homogeneous data stored together.

To use Apache spark we need to convert existing data into parquet format. In this article we will learn to convert CSV files to parquet format and then retrieve them back.

CSV to Parquet

We will convert csv files to parquet format using Apache Spark. For Introduction to Spark you can refer to Spark documentation.

Below is pyspark code to convert csv to parquet. You can edit the names and types of columns as per your input.csv

from pyspark import SparkContext
from pyspark.sql import SQLContext
from pyspark.sql.types import *

if __name__ == "__main__":
    sc = SparkContext(appName="CSV2Parquet")
    sqlContext = SQLContext(sc)
    schema = StructType([
            StructField("col1", IntegerType(), True),
            StructField("col2", IntegerType(), True),
            StructField("col3", StringType(), True),
            StructField("col4", StringType(), True),
            StructField("col5", StringType(), True),
            StructField("col6", DoubleType(), True)])
    rdd = sc.textFile("/home/sarvesh/Desktop/input.csv").map(lambda line: line.split(","))
    df = sqlContext.createDataFrame(rdd, schema)


Above code will create parquet files in input-parquet directory. Files will be in binary format so you will not able to read them. You can check the size of the directory and compare it with size of CSV compressed file.

For a 8 MB csv, when compressed, it generated a 636kb parquet file.

The other way: Parquet to CSV

You can retrieve csv files back from parquet files.

from pyspark import SparkContext
from pyspark.sql import SQLContext
from pyspark.sql.types import *

if __name__ == "__main__":
    sc = SparkContext(appName="Parquet2CSV")
    sqlContext = SQLContext(sc)

    readdf ='/home/sarvesh/Desktop/submissions-parquet') row: str(row[0]) + "," + str(row[1]) + ","+ str(row[2]) + ","+ str(row[3])+ ","+
                              str(row[4])+","+ str(row[5])).saveAsTextFile("/home/sarvesh/Desktop/parquet-to-csv.csv")


You can compare the original and converted CSV files.

You can provide parquet files to your Hadoop based applications rather than providing plain CSV files.

For more information about parquet schema and performance comparisons with other storage formats you can refer to Parquet official docs.


Related Articles


NHibernate, Linq and Lazy Collections

For the past few months we have been busy building a services framework for one of our clients so that they can build all of their enterprise web services without ever having to worry about the cross cutting concerns and... Read more

Page Redirects using Spring.NET

Who is responsible for page redirects in ASPNET MVP – The View or the Presenter None of the above, it is you :) On a serious note, it is the View as one shouldn’t pollute the Presenter with the page navigation... Read more
  • Thomas Spicer

    Thank you for great tips! Hard to find info on the topic. Parquet is also useful with serverless query technologies like Amazon Athena and Spectrum. Using a columnar format like Parquet matters for cost and performance benefits. I highlighted some resons here:

  • Shamdooniz

    Thanks for the post. The only hiccup I had while using it was the mixed data types. In my case I had a combination of IntegerType and StringType so I had to modify the map to sth like: map(lambda line: [int(i) if i.isdigit() else i for i in line.split(“,”)])
    in order to avoid TypeError exception.