hyparquet

hyparquet

Dependency free since 2023!

What is hyparquet?

Hyparquet is a lightweight, pure JavaScript library for parsing Apache Parquet files. Apache Parquet is a popular columnar storage format that is widely used in data engineering, data science, and machine learning applications for efficiently storing and processing large datasets.

Hyparquet allows you to read and extract data from Parquet files directly in JavaScript environments, both in Node.js and in the browser. It is designed to be fast, memory-efficient, and easy to use.

Why hyparquet?

1. Performant: Designed to efficiently process large datasets by only loading the required data, making it suitable for big data and machine learning applications.
2. Browser-native: Built to work seamlessly in the browser, opening up new possibilities for web-based data applications and visualizations.
3. Dependency-free: Hyparquet has zero dependencies, making it lightweight and easy to install and use in any JavaScript project.
4. TypeScript support: The library is written in jsdoc-typed JavaScript and provides TypeScript definitions out of the box.
5. Flexible data access: Hyparquet allows you to read specific subsets of data by specifying row and column ranges, giving fine-grained control over what data is fetched and loaded.

Features

• Designed to work with huge ML datasets (like starcoder)
• Can load metadata separately from data
• Data can be filtered by row and column ranges
• Only fetches the data needed
• Written in JavaScript, checked with TypeScript
• Fast data loading for large scale ML applications
• Bring data visualization closer to the user, in the browser

Why make a new parquet parser? First, existing libraries like parquetjs are officially "inactive". Importantly, they do not support the kind of stream processing needed to make a really performant parser in the browser. And finally, no dependencies means that hyparquet is lean, and easy to package and deploy.

Demo

Online parquet file reader demo available at:

https://hyparam.github.io/hyparquet/

Demo source: index.html

Usage

Install the hyparquet package from npm:

npm install hyparquet

If you're in a node.js environment, you can load a parquet file with the following example:

const { parquetMetadata } = await import('hyparquet')
const fs = await import('fs')

const buffer = fs.readFileSync('example.parquet')
const arrayBuffer = new Uint8Array(buffer).buffer
const metadata = parquetMetadata(arrayBuffer)

If you're in a browser environment, you'll probably get parquet file data from either a drag-and-dropped file from the user, or downloaded from the web.

To load parquet data in the browser from a remote server using fetch:

import { parquetMetadata } from 'hyparquet'

const res = await fetch(url)
const arrayBuffer = await res.arrayBuffer()
const metadata = parquetMetadata(arrayBuffer)

To parse parquet files from a user drag-and-drop action, see example in index.html.

Reading Data

To read the entire contents of a parquet file in a browser environment:

const { parquetRead } = await import("https://cdn.jsdelivr.net/npm/hyparquet/src/hyparquet.min.js")
const res = await fetch(url)
const arrayBuffer = await res.arrayBuffer()
await parquetRead({
  file: arrayBuffer,
  onComplete: data => console.log(data)
})

Filtering

To read large parquet files, it is recommended that you filter by row and column. Hyparquet is designed to load only the minimal amount of data needed to fulfill a query. You can filter rows by number, or columns by name:

import { parquetRead } from 'hyparquet'

await parquetRead({
  file,
  columns: ['colA', 'colB'], // include columns colA and colB
  rowStart: 100,
  rowEnd: 200,
  onComplete: data => console.log(data),
})

Async

Hyparquet supports asynchronous fetching of parquet files over a network. You can provide an AsyncBuffer which is like a js ArrayBuffer but the slice method returns Promise<ArrayBuffer>.

interface AsyncBuffer {
  byteLength: number
  slice(start: number, end?: number): Promise<ArrayBuffer>
}

You can read parquet files asynchronously using HTTP Range requests so that only the necessary byte ranges from a url will be fetched:

import { parquetRead } from 'hyparquet'

const url = 'https://...'
await parquetRead({
  file: { // AsyncBuffer
    byteLength,
    async slice(start, end) {
      const headers = new Headers()
      headers.set('Range', `bytes=${start}-${end - 1}`)
      const res = await fetch(url, { headers })
      if (!res.ok || !res.body) throw new Error('fetch failed')
      return readableStreamToArrayBuffer(res.body)
    },
  }
  onComplete: data => console.log(data),
})

Supported Parquet Files

The parquet format is known to be a sprawling format which includes options for a wide array of compression schemes, encoding types, and data structures.

Hyparquet does not support 100% of all parquet files. Supporting every possible compression codec available in parquet would blow up the size of the hyparquet library. In practice, most parquet files use snappy compression.

Parquet compression types supported by default:

• Uncompressed
• Snappy
• GZip
• LZO
• Brotli
• LZ4
• ZSTD
• LZ4_RAW

You can extend support for other compression codecs using the compressors option.

import { parquetRead } from 'hyparquet'
import { gunzipSync } from 'zlib'

parquetRead({ file, compressors: {
  GZIP: (input, output) => output.set(gunzipSync(input)), // add gzip support
}})

Parquet encodings:

• PLAIN
• PLAIN_DICTIONARY
• RLE_DICTIONARY
• RLE
• BIT_PACKED
• DELTA_BINARY_PACKED
• DELTA_BYTE_ARRAY
• DELTA_LENGTH_BYTE_ARRAY
• BYTE_STREAM_SPLIT

Hysnappy

The most common compression codec used in parquet is snappy compression. Hyparquet includes a built-in snappy decompressor written in javascript.

We developed hysnappy to make parquet parsing even faster. Hysnappy is a snappy decompression codec written in C, compiled to WASM.

To use hysnappy for faster parsing of large parquet files, override the SNAPPY compressor for hyparquet:

import { parquetRead } from 'hyparquet'
import { snappyUncompressor } from 'hysnappy'

parquetRead({ file, compressors: {
  SNAPPY: snappyUncompressor(),
}})

Parsing a 420mb wikipedia parquet file using hysnappy reduces parsing time by 40% (4.1s to 2.3s).

References

• https://github.com/apache/parquet-format
• https://github.com/apache/parquet-testing
• https://github.com/apache/thrift
• https://github.com/apache/arrow
• https://github.com/dask/fastparquet
• https://github.com/duckdb/duckdb
• https://github.com/google/snappy
• https://github.com/ironSource/parquetjs
• https://github.com/zhipeng-jia/snappyjs