How to Generate Parquet File Using Pure Java (Including Date & Decimal Types) And Upload to S3 [Windows] (No HDFS)

Question

I recently had a requirement where I needed to generate Parquet files that could be read by Apache Spark using only Java (Using no additional software installations such as: Apache Drill, Hive, Spark, etc.). The files needed to be saved to S3 so I will be sharing details on how to do both.

There were no simple to follow guides on how to do this. I'm also not a Java programmer so the concepts of using Maven, Hadoop, etc. were all foreign to me. So it took me nearly two weeks to get this working. I'd like to share my personal guide below on how I achieved this

Solution

Disclaimer: The code samples below in no way represent best practices and are only presented as a rough how-to.

Dependencies:

• parquet-avro (1.9.0) : https://mvnrepository.com/artifact/org.apache.parquet/parquet-avro/1.9.0 (We use 1.9.0 because this version uses Avro 1.8+ which supports Decimals and Dates)
• hadoop-aws (2.8.2) [If you don't plan on writing to S3 you won't need this but you will need to add several other dependencies that normally get added thanks to this. I will not cover that scenario. So even if you're going to generate Parquet files only on your local disk, you can still add this to your project as a dependency]: https://mvnrepository.com/artifact/org.apache.hadoop/hadoop-aws/2.8.2 (We use this because it was the latest version at the time)
• Hadoop 2.8.1: https://github.com/steveloughran/winutils/tree/master/hadoop-2.8.1 (We use 2.8.X because it needs to match the hadoop libraries used in the parquet-avro and hadoop-aws dependencies)

I'll be using NetBeans as my IDE.

Some info regarding parquet in Java (For noobs such as me):

• In order to serialize your data into parquet, you must choose one of the popular Java data serialization frameworks: Avro, Protocol Buffers or Thrift (I'll be using Avro (1.8.0), as can be seen from our parquet-avro dependency)
• You will need to use an IDE that supports Maven. This is because the dependencies above have a lot of dependencies of their own. Maven will automatically download those for you (like NuGet for VisualStudio)

Pre-requisite:

You must have hadoop on the windows machine that will be running the Java code. The good news is you don't need to install the entire hadoop software, rather you need only two files:

• hadoop.dll
• winutils.exe

These can be downloaded here. You will need version 2.8.1 for this example (due to parquet-avro 1.9.0).

1. Copy these files to C:\hadoop-2.8.1\bin on the target machine.

2. Add a new System Variable (not user variable) called: HADOOP_HOME with the value C:\hadoop-2.8.1

   

3. Modify the System Path variable (not user variable) and add the following to the end: %HADOOP_HOME%\bin

4. Restart the machine for changes to take affect.

If this config was not done properly you will get the following error at run-time: java.lang.UnsatisfiedLinkError: org.apache.hadoop.io.nativeio.NativeIO$Windows.access0(Ljava/lang/String;I)Z

Getting Started with Coding:

• First create a new empty Maven Project and add parquet-avro 1.9.0 and hadoop-aws 2.8.2 as dependencies:
• Create your main class where you can write some code

• First thing is you need to generate a Schema. Now as far as I can tell there is no way you can generate a schema programmatically at run-time. the Schema.Parser class' parse() method only takes a file or a string literal as a parameter and doesn't let you modify the schema once it is created. To circumvent this I am generating my Schema JSON at run time and parsing that. Below is an example Schema:

  String schema = "{\"namespace\": \"org.myorganization.mynamespace\"," //Not used in Parquet, can put anything
      + "\"type\": \"record\"," //Must be set as record
      + "\"name\": \"myrecordname\"," //Not used in Parquet, can put anything
      + "\"fields\": ["
      + " {\"name\": \"myInteger\", \"type\": \"int\"}," //Required field
      + " {\"name\": \"myString\",  \"type\": [\"string\", \"null\"]},"
      + " {\"name\": \"myDecimal\", \"type\": [{\"type\": \"fixed\", \"size\":16, \"logicalType\": \"decimal\", \"name\": \"mydecimaltype1\", \"precision\": 32, \"scale\": 4}, \"null\"]},"
      + " {\"name\": \"myDate\", \"type\": [{\"type\": \"int\", \"logicalType\" : \"date\"}, \"null\"]}"
      + " ]}";
  Parser parser = new Schema.Parser().setValidate(true);
  Schema avroSchema = parser.parse(schema);
  

  Details on Avro schema can be found here: https://avro.apache.org/docs/1.8.0/spec.html

• Next we can start generating records (Avro primitive types are simple):

  GenericData.Record record = new GenericData.Record(avroSchema);
  record.put("myInteger", 1);
  record.put("myString", "string value 1");
  

  • In order to generate a decimal logical type a fixed or bytes primitive type must be used as the actual data type for storage. The current Parquet format only supports Fixed length byte arrays (aka: fixed_len_byte_array). So we have to use fixed in our case as well (as can be seen in the schema). In Java we must use BigDecimal in order to truly handle decimals. And I've identified that a Decimal(32,4) will not take more than 16 bytes no matter the value. So we will use a standard byte array size of 16 in our serialization below (and in the schema above):

  BigDecimal myDecimalValue = new BigDecimal("99.9999");
  
  //First we need to make sure the BigDecimal matches our schema scale:
  myDecimalValue = myDecimalValue.setScale(4, RoundingMode.HALF_UP);
  
  //Next we get the decimal value as one BigInteger (like there was no decimal point)
  BigInteger myUnscaledDecimalValue = myDecimalValue.unscaledValue();
  
  //Finally we serialize the integer
  byte[] decimalBytes = myUnscaledDecimalValue.toByteArray();
  
  //We need to create an Avro 'Fixed' type and pass the decimal schema once more here:
  GenericData.Fixed fixed = new GenericData.Fixed(new Schema.Parser().parse("{\"type\": \"fixed\", \"size\":16, \"precision\": 32, \"scale\": 4, \"name\":\"mydecimaltype1\"}"));
  
  byte[] myDecimalBuffer = new byte[16];
  if (myDecimalBuffer.length >= decimalBytes.length) {            
      //Because we set our fixed byte array size as 16 bytes, we need to
      //pad-left our original value's bytes with zeros
      int myDecimalBufferIndex = myDecimalBuffer.length - 1;
      for(int i = decimalBytes.length - 1; i >= 0; i--){
          myDecimalBuffer[myDecimalBufferIndex] = decimalBytes[i];
          myDecimalBufferIndex--;
      }
      //Save result
      fixed.bytes(myDecimalBuffer);
  } else {
      throw new IllegalArgumentException(String.format("Decimal size: %d was greater than the allowed max: %d", decimalBytes.length, myDecimalBuffer.length));
  }
  
  //We can finally write our decimal to our record
  record.put("myDecimal", fixed);
  

• For Date values, Avro specifies that we need to save the number of days since EPOCH as an integer. (If you need the time component as well, such as an actual DateTime type, you need to use the Timestamp Avro type, which I will not cover). The easiest way I found to get the number of days since epoch is using the joda-time library. If you added the hadoop-aws dependency to your project you should already have this library. If not you will need to add it yourself:

  //Get epoch value
  MutableDateTime epoch = new MutableDateTime(0l, DateTimeZone.UTC);
  
  DateTime currentDate = new DateTime(); //Can take Java Date in constructor
  Days days = Days.daysBetween(epoch, currentDate);
  
  //We can write number of days since epoch into the record
  record.put("myDate", days.getDays());
  

• We finally can start writing our parquet file as such

  try {
     Configuration conf = new Configuration();
     conf.set("fs.s3a.access.key", "ACCESSKEY");
     conf.set("fs.s3a.secret.key", "SECRETKEY");
     //Below are some other helpful settings
     //conf.set("fs.s3a.endpoint", "s3.amazonaws.com");
     //conf.set("fs.s3a.aws.credentials.provider", "org.apache.hadoop.fs.s3a.SimpleAWSCredentialsProvider");
     //conf.set("fs.hdfs.impl", org.apache.hadoop.hdfs.DistributedFileSystem.class.getName()); // Not needed unless you reference the hadoop-hdfs library.
     //conf.set("fs.file.impl", org.apache.hadoop.fs.LocalFileSystem.class.getName()); // Uncomment if you get "No FileSystem for scheme: file" errors
  
     Path path = new Path("s3a://your-bucket-name/examplefolder/data.parquet");
  
     //Use path below to save to local file system instead
     //Path path = new Path("data.parquet");
  
     try (ParquetWriter writer = AvroParquetWriter.builder(path)
             .withSchema(avroSchema)
             .withCompressionCodec(CompressionCodecName.GZIP)
             .withConf(conf)
             .withPageSize(4 * 1024 * 1024) //For compression
             .withRowGroupSize(16 * 1024 * 1024) //For write buffering (Page size)
             .build()) {
         //We only have one record to write in our example
         writer.write(record);
     }
  } catch (Exception ex) { ex.printStackTrace(System.out); }

• Here is the data loaded into Apache Spark (2.2.0):

And for your convenience, the entire source code:

package com.mycompany.stackoverflow;

import java.math.BigDecimal;
import java.math.BigInteger;
import java.math.RoundingMode;
import org.apache.avro.Schema;
import org.apache.avro.generic.GenericData;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.Path;
import org.apache.parquet.avro.AvroParquetWriter;
import org.apache.parquet.hadoop.ParquetWriter;
import org.apache.parquet.hadoop.metadata.CompressionCodecName;
import org.joda.time.DateTime;
import org.joda.time.DateTimeZone;
import org.joda.time.Days;
import org.joda.time.MutableDateTime;

public class Main {
    public static void main(String[] args) {
        System.out.println("Start");

        String schema = "{\"namespace\": \"org.myorganization.mynamespace\"," //Not used in Parquet, can put anything
                + "\"type\": \"record\"," //Must be set as record
                + "\"name\": \"myrecordname\"," //Not used in Parquet, can put anything
                + "\"fields\": ["
                + " {\"name\": \"myInteger\", \"type\": \"int\"}," //Required field
                + " {\"name\": \"myString\",  \"type\": [\"string\", \"null\"]},"
                + " {\"name\": \"myDecimal\", \"type\": [{\"type\": \"fixed\", \"size\":16, \"logicalType\": \"decimal\", \"name\": \"mydecimaltype1\", \"precision\": 32, \"scale\": 4}, \"null\"]},"
                + " {\"name\": \"myDate\", \"type\": [{\"type\": \"int\", \"logicalType\" : \"date\"}, \"null\"]}"
                + " ]}";

        Schema.Parser parser = new Schema.Parser().setValidate(true);
        Schema avroSchema = parser.parse(schema);

        GenericData.Record record = new GenericData.Record(avroSchema);
        record.put("myInteger", 1);
        record.put("myString", "string value 1");

        BigDecimal myDecimalValue = new BigDecimal("99.9999");

        //First we need to make sure the huge decimal matches our schema scale:
        myDecimalValue = myDecimalValue.setScale(4, RoundingMode.HALF_UP);

        //Next we get the decimal value as one BigInteger (like there was no decimal point)
        BigInteger myUnscaledDecimalValue = myDecimalValue.unscaledValue();

        //Finally we serialize the integer
        byte[] decimalBytes = myUnscaledDecimalValue.toByteArray();

        //We need to create an Avro 'Fixed' type and pass the decimal schema once more here:
        GenericData.Fixed fixed = new GenericData.Fixed(new Schema.Parser().parse("{\"type\": \"fixed\", \"size\":16, \"precision\": 32, \"scale\": 4, \"name\":\"mydecimaltype1\"}"));

        byte[] myDecimalBuffer = new byte[16];
        if (myDecimalBuffer.length >= decimalBytes.length) {            
            //Because we set our fixed byte array size as 16 bytes, we need to
            //pad-left our original value's bytes with zeros
            int myDecimalBufferIndex = myDecimalBuffer.length - 1;
            for(int i = decimalBytes.length - 1; i >= 0; i--){
                myDecimalBuffer[myDecimalBufferIndex] = decimalBytes[i];
                myDecimalBufferIndex--;
            }

            //Save result
            fixed.bytes(myDecimalBuffer);
        } else {
            throw new IllegalArgumentException(String.format("Decimal size: %d was greater than the allowed max: %d", decimalBytes.length, myDecimalBuffer.length));
        }

        //We can finally write our decimal to our record
        record.put("myDecimal", fixed);

        //Get epoch value
        MutableDateTime epoch = new MutableDateTime(0l, DateTimeZone.UTC);

        DateTime currentDate = new DateTime(); //Can take Java Date in constructor
        Days days = Days.daysBetween(epoch, currentDate);

        //We can write number of days since epoch into the record
        record.put("myDate", days.getDays());

        try {
           Configuration conf = new Configuration();
           conf.set("fs.s3a.access.key", "ACCESSKEY");
           conf.set("fs.s3a.secret.key", "SECRETKEY");
           //Below are some other helpful settings
           //conf.set("fs.s3a.endpoint", "s3.amazonaws.com");
           //conf.set("fs.s3a.aws.credentials.provider", "org.apache.hadoop.fs.s3a.SimpleAWSCredentialsProvider");
           //conf.set("fs.hdfs.impl", org.apache.hadoop.hdfs.DistributedFileSystem.class.getName()); // Not needed unless you reference the hadoop-hdfs library.
           //conf.set("fs.file.impl", org.apache.hadoop.fs.LocalFileSystem.class.getName()); // Uncomment if you get "No FileSystem for scheme: file" errors.

           Path path = new Path("s3a://your-bucket-name/examplefolder/data.parquet");

           //Use path below to save to local file system instead
           //Path path = new Path("data.parquet");

           try (ParquetWriter<GenericData.Record> writer = AvroParquetWriter.<GenericData.Record>builder(path)
                   .withSchema(avroSchema)
                   .withCompressionCodec(CompressionCodecName.GZIP)
                   .withConf(conf)
                   .withPageSize(4 * 1024 * 1024) //For compression
                   .withRowGroupSize(16 * 1024 * 1024) //For write buffering (Page size)
                   .build()) {

               //We only have one record to write in our example
               writer.write(record);
           }
        } catch (Exception ex) { 
            ex.printStackTrace(System.out);
        }
    }
}