parquet.js

fully asynchronous, pure node.js implementation of the Parquet file format

This package contains a fully asynchronous, pure JavaScript implementation of the Parquet file format. The implementation conforms with the Parquet specification and is tested for compatibility with Apache's Java reference implementation.

As a fork of parquetjs-lite, this project aims to conform more with microservices (eg. AWS Lambda, AWS Athena, etc.) than it aims to be a robust implementation. For example, Brotli (available in parquetjs-lite) has been stripped from this fork, as while Brotli's compression rates are high, its speed (both in terms of cold-start loading times and compression/decompression times) are slow. Its resource usage (memory/cpu) is also high relative to other solutions which breaks the goal of using it within the context of "snappy" serverless microservices (if you will).

What is Parquet?: Parquet is a column-oriented file format; it allows you to write a large amount of structured data to a file, compress it and then read parts of it back out efficiently. The Parquet format is based on Google's Dremel paper.

Installation

To use parquet.js with node.js, install it using npm:

  $ npm install @entitycs/parquetjs-micro

parquet.js requires node.js >= 7.6.0

Usage: Writing files

Once you have installed the parquet.js library, you can import it as a single module:

var parquet = require('@entitycs/parquetjs-micro');

Parquet files have a strict schema, similar to tables in a SQL database. So, in order to produce a Parquet file we first need to declare a new schema. Here is a simple example that shows how to instantiate a ParquetSchema object:

// declare a schema for the `fruits` table
var schema = new parquet.ParquetSchema({
  name: { type: 'UTF8' },
  quantity: { type: 'INT64' },
  price: { type: 'DOUBLE' },
  date: { type: 'TIMESTAMP_MILLIS' },
  in_stock: { type: 'BOOLEAN' }
});

Note that the Parquet schema supports nesting, so you can store complex, arbitrarily nested records into a single row (more on that later) while still maintaining good compression.

Once we have a schema, we can create a ParquetWriter object. The writer will take input rows as JSON objects, convert them to the Parquet format and store them on disk.

// create new ParquetWriter that writes to 'fruits.parquet`
var writer = await parquet.ParquetWriter.openFile(schema, 'fruits.parquet');

// append a few rows to the file
await writer.appendRow({name: 'apples', quantity: 10, price: 2.5, date: new Date(), in_stock: true});
await writer.appendRow({name: 'oranges', quantity: 10, price: 2.5, date: new Date(), in_stock: true});

Once we are finished adding rows to the file, we have to tell the writer object to flush the metadata to disk and close the file by calling the close() method:

Usage: Reading files

A parquet reader allows retrieving the rows from a parquet file in order. The basic usage is to create a reader and then retrieve a cursor/iterator which allows you to consume row after row until all rows have been read.

You may open more than one cursor and use them concurrently. All cursors become invalid once close() is called on the reader object.

// create new ParquetReader that reads from 'fruits.parquet`
let reader = await parquet.ParquetReader.openFile('fruits.parquet');

// create a new cursor
let cursor = reader.getCursor();

// read all records from the file and print them
let record = null;
while (record = await cursor.next()) {
  console.log(record);
}

When creating a cursor, you can optionally request that only a subset of the columns should be read from disk. For example:

// create a new cursor that will only return the `name` and `price` columns
let cursor = reader.getCursor(['name', 'price']);

It is important that you call close() after you are finished reading the file to avoid leaking file descriptors.

await reader.close();

Reading data from a url

Parquet files can be read from a url without having to download the whole file. You will have to supply the request library as a first argument and the request parameters as a second argument to the function parquetReader.openUrl.

const request = require('request');
let reader = await parquet.ParquetReader.openUrl(request,'https://domain/fruits.parquet');

Reading data from S3

Parquet files can be read from an S3 object without having to download the whole file. You will have to supply the aws-sdk client as first argument and the bucket/key information as second argument to the function parquetReader.openS3.

If using version 3 of the aws-sdk for your S3 client, supply as the client argument an object containing an already constructed S3Client along with the plain HeadObjectCommand and GetObjectCommand modules.

const params = {
  Bucket: 'xxxxxxxxxxx',
  Key: 'xxxxxxxxxxx'
};
// v2 example
const AWS = require('aws-sdk');
const client = new AWS.S3({
  accessKeyId: 'xxxxxxxxxxx',
  secretAccessKey: 'xxxxxxxxxxx'
});
let reader = await parquet.ParquetReader.openS3(client,params);

//v3 example
const {S3Client, HeadObjectCommand, GetObjectCommand} = require('@aws-sdk/client-s3');
const client = new S3Client({region:"us-east-1"});
let reader = await parquet.ParquetReader.openS3(
  {S3Client:client, HeadObjectCommand, GetObjectCommand},
  params
);

Reading data from a buffer

If the complete parquet file is in buffer it can be read directly from memory without incurring any additional I/O.

const file = fs.readFileSync('fruits.parquet');
let reader = await parquet.ParquetReader.openBuffer(file);

Encodings

Internally, the Parquet format will store values from each field as consecutive arrays which can be compressed/encoded using a number of schemes.

Plain Encoding (PLAIN)

The most simple encoding scheme is the PLAIN encoding. It simply stores the values as they are without any compression. The PLAIN encoding is currently the default for all types except BOOLEAN:

var schema = new parquet.ParquetSchema({
  name: { type: 'UTF8', encoding: 'PLAIN' },
});

Run Length Encoding (RLE)

The Parquet hybrid run length and bitpacking encoding allows to compress runs of numbers very efficiently. Note that the RLE encoding can only be used in combination with the BOOLEAN, INT32 and INT64 types. The RLE encoding requires an additional bitWidth parameter that contains the maximum number of bits required to store the largest value of the field.

var schema = new parquet.ParquetSchema({
  age: { type: 'UINT_32', encoding: 'RLE', bitWidth: 7 },
});

Optional Fields

By default, all fields are required to be present in each row. You can also mark a field as 'optional' which will let you store rows with that field missing:

var schema = new parquet.ParquetSchema({
  name: { type: 'UTF8' },
  quantity: { type: 'INT64', optional: true },
});

var writer = await parquet.ParquetWriter.openFile(schema, 'fruits.parquet');
await writer.appendRow({name: 'apples', quantity: 10 });
await writer.appendRow({name: 'banana' }); // not in stock

Nested Rows & Arrays

Parquet supports nested schemas that allow you to store rows that have a more complex structure than a simple tuple of scalar values. To declare a schema with a nested field, omit the type in the column definition and add a fields list instead:

Consider this example, which allows us to store a more advanced "fruits" table where each row contains a name, a list of colours and a list of "stock" objects.

// advanced fruits table
var schema = new parquet.ParquetSchema({
  name: { type: 'UTF8' },
  colours: { type: 'UTF8', repeated: true },
  stock: {
    repeated: true,
    fields: {
      price: { type: 'DOUBLE' },
      quantity: { type: 'INT64' },
    }
  }
});

// the above schema allows us to store the following rows:
var writer = await parquet.ParquetWriter.openFile(schema, 'fruits.parquet');

await writer.appendRow({
  name: 'banana',
  colours: ['yellow'],
  stock: [
    { price: 2.45, quantity: 16 },
    { price: 2.60, quantity: 420 }
  ]
});

await writer.appendRow({
  name: 'apple',
  colours: ['red', 'green'],
  stock: [
    { price: 1.20, quantity: 42 },
    { price: 1.30, quantity: 230 }
  ]
});

await writer.close();

// reading nested rows with a list of explicit columns
let reader = await parquet.ParquetReader.openFile('fruits.parquet');

let cursor = reader.getCursor([['name'], ['stock', 'price']]);
let record = null;
while (record = await cursor.next()) {
  console.log(record);
}

await reader.close();

It might not be obvious why one would want to implement or use such a feature when the same can - in principle - be achieved by serializing the record using JSON (or a similar scheme) and then storing it into a UTF8 field:

Putting aside the philosophical discussion on the merits of strict typing, knowing about the structure and subtypes of all records (globally) means we do not have to duplicate this metadata (i.e. the field names) for every record. On top of that, knowing about the type of a field allows us to compress the remaining data more efficiently.

Nested Lists for Hive / Athena

Lists have to be annotated to be queriable with AWS Athena. See parquet-format for more detail and a full working example with comments in the test directory (test/list.js)

List of Supported Types & Encodings

We aim to be feature-complete and add new features as they are added to the Parquet specification; this is the list of currently implemented data types and encodings:

Buffering & Row Group Size

When writing a Parquet file, the ParquetWriter will buffer rows in memory until a row group is complete (or close() is called) and then write out the row group to disk.

The size of a row group is configurable by the user and controls the maximum number of rows that are buffered in memory at any given time as well as the number of rows that are co-located on disk:

var writer = await parquet.ParquetWriter.openFile(schema, 'fruits.parquet');
writer.setRowGroupSize(8192);

Dependencies

Parquet uses thrift to encode the schema and other metadata, but the actual data does not use thrift.

Notes

Currently parquet-cpp doesn't fully support DATA_PAGE_V2. You can work around this by setting the useDataPageV2 option to false.

Contributions

Please make sure you sign the contributor license agreement in order for us to be able to accept your contribution. We thank you very much!

License

Copyright (c) 2017 ironSource Ltd.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.