parquet-go is a pure-go implementation of reading and writing the parquet format file.
- Support Read/Write Nested/Flat Parquet File
- Support all Types in Parquet
- Very simple to use (like json marshal/unmarshal)
- High performance
- git.apache.org/thrift.git/lib/go/thrift
- github.com/golang/snappy
There are two Types in Parquet: Base Type and Logical Type They are defined in ParquetType.go as following:
//base type
type BOOLEAN bool
type INT32 int32
type INT64 int64
type INT96 string // length=96
type FLOAT float32
type DOUBLE float64
type BYTE_ARRAY string
type FIXED_LEN_BYTE_ARRAY string
//logical type
type UTF8 string
type INT_8 int32
type INT_16 int32
type INT_32 int32
type INT_64 int64
type UINT_8 uint32
type UINT_16 uint32
type UINT_32 uint32
type UINT_64 uint64
type DATE int32
type TIME_MILLIS int32
type TIME_MICROS int64
type TIMESTAMP_MILLIS int64
type TIMESTAMP_MICROS int64
type INTERVAL string // length=12
type DECIMAL string
The variables which will read/write from/to a parquet file must be declared as these types. OPTIONAL variables are declared as pointers.
The core data structure named "Table":
type Table struct {
Repetition_Type parquet.FieldRepetitionType
Type parquet.Type
Path []string
MaxDefinitionLevel int32
MaxRepetitionLevel int32
Values []interface{}
DefinitionLevels []int32
RepetitionLevels []int32
}
Values is the column data; RepetitionLevels is the repetition levels of the values; DefinitionLevels is the definition levels of the values. The architecture of the data struct is following:
Table -> Page
Pages -> Chunk
Chunks -> RowGroup
RowGroups -> ParquetFile
Marshal/Unmarshal functions are used to encode/decode the parquet file.
Marshl convert a struct slice to a *map[string]*Table
Unmarshal convert a *map[string]*Table to a struct slice
stus := make([]Student, 0)
stus = append(stus, stu01, stu02)
src := Marshal(stus, 0, len(stus), schemaHandler)
dst := make([]Student, 0)
Unmarshal(src, &dst, schemaHandler)
read/write a parquet file need a ParquetFile interface implemented
type ParquetFile interface {
Seek(offset int, pos int) (int64, error)
Read(b []byte) (n int, err error)
Write(b []byte) (n int, err error)
Close()
Open(name string) (ParquetFile, error)
Create(name string) (ParquetFile, error)
}
Using this interface, parquet-go can read/write parquet file on any plantform(local/hdfs/s3...)
- Open(name string) (ParquetFile, error) is used for read parquet. If name is "", it should return a new file handler of the same file.
The read and unmarshal processes can be separated and an example is shown in example/benchmark/ReadParquet.go
In reading process, Unmarshal is a very time-consuming function. If this process is not needed, you can just get the table map and values by yourself.
The following is a simple example of read/write parquet file on local disk. It can be found in example directory:
package main
import (
. "github.com/xitongsys/parquet-go/ParquetHandler"
. "github.com/xitongsys/parquet-go/ParquetType"
"log"
"os"
)
type Student struct {
Name UTF8
Age INT32
Id INT64
Weight FLOAT
Sex BOOLEAN
}
type MyFile struct {
FilePath string
File *os.File
}
func (self *MyFile) Create(name string) (ParquetFile, error) {
file, err := os.Create(name)
myFile := new(MyFile)
myFile.File = file
return myFile, err
}
func (self *MyFile) Open(name string) (ParquetFile, error) {
var (
err error
)
if name == "" {
name = self.FilePath
}
myFile := new(MyFile)
myFile.FilePath = name
myFile.File, err = os.Open(name)
return myFile, err
}
func (self *MyFile) Seek(offset int, pos int) (int64, error) {
return self.File.Seek(int64(offset), pos)
}
func (self *MyFile) Read(b []byte) (n int, err error) {
return self.File.Read(b)
}
func (self *MyFile) Write(b []byte) (n int, err error) {
return self.File.Write(b)
}
func (self *MyFile) Close() {
self.File.Close()
}
func main() {
var f ParquetFile
f = &MyFile{}
//write flat
f, _ = f.Create("flat.parquet")
ph := NewParquetHandler()
ph.WriteInit(f, new(Student), 4, 30)
num := 10
for i := 0; i < num; i++ {
stu := Student{
Name: UTF8("StudentName"),
Age: INT32(20 + i%5),
Id: INT64(i),
Weight: FLOAT(50.0 + float32(i)*0.1),
Sex: BOOLEAN(i%2 == 0),
}
ph.Write(stu)
}
ph.WriteStop()
log.Println("Write Finished")
f.Close()
///read flat
f, _ = f.Open("flat.parquet")
ph = NewParquetHandler()
rowGroupNum := ph.ReadInit(f, 10)
for i := 0; i < rowGroupNum; i++ {
stus := make([]Student, 0)
ph.ReadOneRowGroupAndUnmarshal(&stus)
log.Println(stus)
}
f.Close()
}
Read/Write initial functions have a parallel parameters np which is the number of goroutines in reading/writing.
func (self *ParquetHandler) ReadInit(pfile ParquetFile, np int64)
func (self *ParquetHandler) WriteInit(pfile ParquetFile, obj interface{}, np int64)
Plugin is used for some special purpose and will be added gradually.
This plugin is used for data format similar with CSV(not nested).
func main() {
md := []MetadataType{
{Type: "UTF8", Name: "Name"},
{Type: "INT32", Name: "Age"},
{Type: "INT64", Name: "Id"},
{Type: "FLOAT", Name: "Weight"},
{Type: "BOOLEAN", Name: "Sex"},
}
var f ParquetFile
f = &MyFile{}
//write flat
f, _ = f.Create("csv.parquet")
ph := NewCSVWriterHandler()
ph.WriteInit(md, f, 10, 30)
num := 10
for i := 0; i < num; i++ {
data := []string{
"StudentName",
fmt.Sprintf("%d", 20+i%5),
fmt.Sprintf("%d", i),
fmt.Sprintf("%f", 50.0+float32(i)*0.1),
fmt.Sprintf("%t", i%2 == 0),
}
rec := make([]*string, len(data))
for j := 0; j < len(data); j++ {
rec[j] = &data[j]
}
ph.Write(rec)
}
ph.WriteStop()
log.Println("Write Finished")
f.Close()
}
A very simple performance test of writing/reading parquet did on Linux host (JRE 1.8.0, Golang 1.7.5, 23GB, 24 Cores). It is faster than java :)
Write Test Results
Read Test Results
