parquet-dotnet
parquet-dotnet copied to clipboard
Fully managed Apache Parquet implementation
Apache Parquet for .NET
![Nuget](https://img.shields.io/nuget/dt/Parquet.Net)
Fully portable, managed .NET library to read and write Apache Parquet files. Supports .NET 6.0
, .NET 5.0
, .NET Core 3.1 (LTS)
, .NET Core 2.1 (LTS)
, .NET Standard 1.4
and up.
Runs everywhere .NET runs Linux, MacOS, Windows, iOS, Android, Tizen, Xbox, PS4, Raspberry Pi, Samsung TVs and much more.
Support Web Assembly is coming (email me if you are interested in details).
Why
Parquet is a de facto physical storage format in big data applications, including Apache Spark, as well as newly emerging Delta Lake and lakehouse architectures. It's really easy to read and write data if you are using one of those platforms, however in standalone mode it's almost impossible or involves using heavy engines. Parquet.Net is very small, fast, pure managed implementation that is crazy fast.
Index
- Getting Started
- Reading Data
- Writing Data
- Declaring Schema
- Complex Types
- Row-Based API
- Fast Automatic Serialisation
You can track the amount of features we have implemented so far.
Getting started
Parquet.Net is redistributed as a NuGet package. All the code is managed and doesn't have any native dependencies, therefore you are ready to go after referencing the package. This also means the library works on Windows, Linux and MacOS X (including M1).
General
This intro is covering only basic use cases. Parquet format is more complicated when it comes to complex types like structures, lists, maps and arrays, therefore you should read this page if you are planning to use them.
Reading files
In order to read a parquet file you need to open a stream first. Due to the fact that Parquet utilises file seeking extensively, the input stream must be readable and seekable. You cannot stream parquet data! This somewhat limits the amount of streaming you can do, for instance you can't read a parquet file from a network stream as we need to jump around it, therefore you have to download it locally to disk and then open.
For instance, to read a file c:\test.parquet
you would normally write the following code:
using System.Collections.Generic;
using System.IO;
using System.Linq;
using Parquet.Data;
// open file stream
using (Stream fileStream = System.IO.File.OpenRead("c:\\test.parquet"))
{
// open parquet file reader
using (var parquetReader = new ParquetReader(fileStream))
{
// get file schema (available straight after opening parquet reader)
// however, get only data fields as only they contain data values
DataField[] dataFields = parquetReader.Schema.GetDataFields();
// enumerate through row groups in this file
for(int i = 0; i < parquetReader.RowGroupCount; i++)
{
// create row group reader
using (ParquetRowGroupReader groupReader = parquetReader.OpenRowGroupReader(i))
{
// read all columns inside each row group (you have an option to read only
// required columns if you need to.
DataColumn[] columns = dataFields.Select(groupReader.ReadColumn).ToArray();
// get first column, for instance
DataColumn firstColumn = columns[0];
// .Data member contains a typed array of column data you can cast to the type of the column
Array data = firstColumn.Data;
int[] ids = (int[])data;
}
}
}
}
Writing files
Writing operates on streams, therefore you need to create it first. The following example shows how to create a file on disk with two columns - id
and city
.
//create data columns with schema metadata and the data you need
var idColumn = new DataColumn(
new DataField<int>("id"),
new int[] { 1, 2 });
var cityColumn = new DataColumn(
new DataField<string>("city"),
new string[] { "London", "Derby" });
// create file schema
var schema = new Schema(idColumn.Field, cityColumn.Field);
using (Stream fileStream = System.IO.File.Create("c:\\test.parquet"))
{
using (var parquetWriter = new ParquetWriter(schema, fileStream))
{
// create a new row group in the file
using (ParquetRowGroupWriter groupWriter = parquetWriter.CreateRowGroup())
{
groupWriter.WriteColumn(idColumn);
groupWriter.WriteColumn(cityColumn);
}
}
}
Row-Based Access
There are API for row-based access that simplify parquet programming at the expense of memory, speed and flexibility. We recommend using column based approach when you can (examples above) however if not possible use these API as we constantly optimise for speed and use them internally ourselves in certain situations.
Who?
- ML.NET.
- Some parts of the popular RavenDB NoSQL database engine.
- Native Windows Parquet Viewer app.
- Recfluence YouTube analytics.
and many more. Want to be listed here? Just raise a PR.
Contributions
Are welcome in any form - documentation, code, reviews, donations.