parquet-go

parquet-go

parquet-go is a pure-go implementation of reading and writing the parquet format file.

• Support Read/Write Nested/Flat Parquet File
• Simple to use
• High performance

Install

Add the parquet-go library to your $GOPATH/src and install dependencies:

go get github.com/xitongsys/parquet-go

Examples

The example/ directory contains several examples.

The local_flat.go example creates some data and writes it out to the example/output/flat.parquet file.

cd $GOPATH/src/github.com/xitongsys/parquet-go/example
go run local_flat.go

The local_flat.go code shows how it's easy to output structs from Go programs to Parquet files.

Type

There are two types in Parquet: Primitive Type and Logical Type. Logical types are stored as primitive types.

Primitive Type

Logical Type

Tips

• Parquet-go supports type alias such type MyString string. But the base type must follow the table instructions.

• Some type convert functions: converter.go

Encoding

PLAIN:

All types

PLAIN_DICTIONARY/RLE_DICTIONARY:

All types

DELTA_BINARY_PACKED:

INT32, INT64, INT_8, INT_16, INT_32, INT_64, UINT_8, UINT_16, UINT_32, UINT_64, TIME_MILLIS, TIME_MICROS, TIMESTAMP_MILLIS, TIMESTAMP_MICROS

DELTA_BYTE_ARRAY:

BYTE_ARRAY, UTF8

DELTA_LENGTH_BYTE_ARRAY:

BYTE_ARRAY, UTF8

Tips

• Some platforms don't support all kinds of encodings. If you are not sure, just use PLAIN and PLAIN_DICTIONARY.
• If the fields have many different values, please don't use PLAIN_DICTIONARY encoding. Because it will record all the different values in a map which will use a lot of memory. Actually it use a 32-bit integer to store the index. It can not used if your unique values number is larger than 32-bit.
• Large array values may be duplicated as min and max values in page stats, significantly increasing file size. If stats are not useful for such a field, they can be omitted from written files by adding omitstats=true to a field tag.

Repetition Type

There are three repetition types in Parquet: REQUIRED, OPTIONAL, REPEATED.

Tips

• The difference between a List and a REPEATED variable is the 'repetitiontype' in tags. Although both of them are stored as slice in go, they are different in parquet. You can find the detail of List in parquet at here. I suggest just use a List.
• For LIST and MAP, some existed parquet files use some nonstandard formats(see here). For standard format, parquet-go will convert them to go slice and go map. For nonstandard formats, parquet-go will convert them to corresponding structs.

Example of Type and Encoding

…

Compression Type

ParquetFile

Read/Write a parquet file need a ParquetFile interface implemented

type ParquetFile interface {
	io.Seeker
	io.Reader
	io.Writer
	io.Closer
	Open(name string) (ParquetFile, error)
	Create(name string) (ParquetFile, error)
}

Using this interface, parquet-go can read/write parquet file on different platforms. All the file sources are at parquet-go-source. Now it supports(local/hdfs/s3/gcs/memory).

Writer

Four Writers are supported: ParquetWriter, JSONWriter, CSVWriter, ArrowWriter.

• ParquetWriter is used to write predefined Golang structs. Example of ParquetWriter

• JSONWriter is used to write JSON strings Example of JSONWriter

• CSVWriter is used to write data format similar with CSV(not nested) Example of CSVWriter

• ArrowWriter is used to write parquet files using Arrow Schemas Example of ArrowWriter

Reader

Two Readers are supported: ParquetReader, ColumnReader

• ParquetReader is used to read predefined Golang structs Example of ParquetReader

• ColumnReader is used to read raw column data. The read function return 3 slices([value], [RepetitionLevel], [DefinitionLevel]) of the records. Example of ColumnReader

Tips

• If the parquet file is very big (even the size of parquet file is small, the uncompressed size may be very large), please don't read all rows at one time, which may induce the OOM. You can read a small portion of the data at a time like a stream-oriented file.

• RowGroupSize and PageSize may influence the final parquet file size. You can find the details from here. You can reset them in ParquetWriter

	pw.RowGroupSize = 128 * 1024 * 1024 // default 128M
	pw.PageSize = 8 * 1024 // default 8K

Schema

There are four methods to define the schema: go struct tags, Json, CSV, Arrow metadata. Only items in schema will be written and others will be ignored.

Tag

type Student struct {
	Name    string  `parquet:"name=name, type=BYTE_ARRAY, convertedtype=UTF8, encoding=PLAIN_DICTIONARY"`
	Age     int32   `parquet:"name=age, type=INT32, encoding=PLAIN"`
	Id      int64   `parquet:"name=id, type=INT64"`
	Weight  float32 `parquet:"name=weight, type=FLOAT"`
	Sex     bool    `parquet:"name=sex, type=BOOLEAN"`
	Day     int32   `parquet:"name=day, type=INT32, convertedtype=DATE"`
	Ignored int32   //without parquet tag and won't write
}

Example of tags

JSON

JSON schema can be used to define some complicated schema, which can't be defined by tag.

…

Example of JSON schema

CSV metadata

	md := []string{
		"name=Name, type=BYTE_ARRAY, convertedtype=UTF8, encoding=PLAIN_DICTIONARY",
		"name=Age, type=INT32",
		"name=Id, type=INT64",
		"name=Weight, type=FLOAT",
		"name=Sex, type=BOOLEAN",
	}

Example of CSV metadata

Arrow metadata

	schema := arrow.NewSchema(
		[]arrow.Field{
			{Name: "int64", Type: arrow.PrimitiveTypes.Int64},
			{Name: "float64", Type: arrow.PrimitiveTypes.Float64},
			{Name: "str", Type: arrow.BinaryTypes.String},
		},
		nil,
	)

Example of Arrow metadata

Tips

• Parquet-go reads data as an object in Golang and every field must be a public field, which start with an upper letter. This field name we call it InName. Field name in parquet file we call it ExName. Function common.HeadToUpper converts ExName to InName. There are some restriction:

1. It's not allowed if two field names are only different at their first letter case. Such as name and Name.
2. PARGO_PREFIX_ is a reserved string, which you'd better not use it as a name prefix. (#294)
3. Use \x01 as the delimiter of fields to support . in some field name.(dot_in_name.go, #349)

Concurrency

Marshal/Unmarshal is the most time consuming process in writing/reading. To improve the performance, parquet-go can use multiple goroutines to marshal/unmarshal the objects. You can set the concurrent number parameter np in the Read/Write initial functions.

func NewParquetReader(pFile ParquetFile.ParquetFile, obj interface{}, np int64) (*ParquetReader, error)
func NewParquetWriter(pFile ParquetFile.ParquetFile, obj interface{}, np int64) (*ParquetWriter, error)
func NewJSONWriter(jsonSchema string, pfile ParquetFile.ParquetFile, np int64) (*JSONWriter, error)
func NewCSVWriter(md []string, pfile ParquetFile.ParquetFile, np int64) (*CSVWriter, error)
func NewArrowWriter(arrowSchema *arrow.Schema, pfile source.ParquetFile, np int64) (*ArrowWriter error)

Examples