mirror of
https://github.com/minio/minio.git
synced 2025-11-20 18:06:10 -05:00
Concurrent CSV parsing and reduce S3 select allocations (#8200)
``` CSV parsing, BEFORE: BenchmarkReaderBasic-12 2842 407533 ns/op 397860 B/op 957 allocs/op BenchmarkReaderReplace-12 2718 429914 ns/op 397844 B/op 957 allocs/op BenchmarkReaderReplaceTwo-12 2718 435556 ns/op 397855 B/op 957 allocs/op BenchmarkAggregateCount_100K-12 171 6798974 ns/op 16667102 B/op 308077 allocs/op BenchmarkAggregateCount_1M-12 19 65657411 ns/op 168057743 B/op 3146610 allocs/op BenchmarkSelectAll_10M-12 1 20882119900 ns/op 2758799896 B/op 41978762 allocs/op CSV parsing, AFTER: BenchmarkReaderBasic-12 3721 312549 ns/op 101920 B/op 338 allocs/op BenchmarkReaderReplace-12 3776 318810 ns/op 101993 B/op 340 allocs/op BenchmarkReaderReplaceTwo-12 3610 330967 ns/op 102012 B/op 341 allocs/op BenchmarkAggregateCount_100K-12 295 4149588 ns/op 3553623 B/op 103261 allocs/op BenchmarkAggregateCount_1M-12 30 37746503 ns/op 33827931 B/op 1049435 allocs/op BenchmarkSelectAll_10M-12 1 17608495800 ns/op 1416504040 B/op 21007082 allocs/op ~ benchcmp old.txt new.txt benchmark old ns/op new ns/op delta BenchmarkReaderBasic-12 407533 312549 -23.31% BenchmarkReaderReplace-12 429914 318810 -25.84% BenchmarkReaderReplaceTwo-12 435556 330967 -24.01% BenchmarkAggregateCount_100K-12 6798974 4149588 -38.97% BenchmarkAggregateCount_1M-12 65657411 37746503 -42.51% BenchmarkSelectAll_10M-12 20882119900 17608495800 -15.68% benchmark old allocs new allocs delta BenchmarkReaderBasic-12 957 338 -64.68% BenchmarkReaderReplace-12 957 340 -64.47% BenchmarkReaderReplaceTwo-12 957 341 -64.37% BenchmarkAggregateCount_100K-12 308077 103261 -66.48% BenchmarkAggregateCount_1M-12 3146610 1049435 -66.65% BenchmarkSelectAll_10M-12 41978762 21007082 -49.96% benchmark old bytes new bytes delta BenchmarkReaderBasic-12 397860 101920 -74.38% BenchmarkReaderReplace-12 397844 101993 -74.36% BenchmarkReaderReplaceTwo-12 397855 102012 -74.36% BenchmarkAggregateCount_100K-12 16667102 3553623 -78.68% BenchmarkAggregateCount_1M-12 168057743 33827931 -79.87% BenchmarkSelectAll_10M-12 2758799896 1416504040 -48.66% ``` ``` BenchmarkReaderHuge/97K-12 2200 540840 ns/op 184.32 MB/s 1604450 B/op 687 allocs/op BenchmarkReaderHuge/194K-12 1522 752257 ns/op 265.04 MB/s 2143135 B/op 1335 allocs/op BenchmarkReaderHuge/389K-12 1190 947858 ns/op 420.69 MB/s 3221831 B/op 2630 allocs/op BenchmarkReaderHuge/778K-12 806 1472486 ns/op 541.61 MB/s 5201856 B/op 5187 allocs/op BenchmarkReaderHuge/1557K-12 426 2575269 ns/op 619.36 MB/s 9101330 B/op 10233 allocs/op BenchmarkReaderHuge/3115K-12 286 4034656 ns/op 790.66 MB/s 12397968 B/op 16099 allocs/op BenchmarkReaderHuge/6230K-12 172 6830563 ns/op 934.05 MB/s 16008416 B/op 26844 allocs/op BenchmarkReaderHuge/12461K-12 100 11409467 ns/op 1118.39 MB/s 22655163 B/op 48107 allocs/op BenchmarkReaderHuge/24922K-12 66 19780395 ns/op 1290.19 MB/s 35158559 B/op 90216 allocs/op BenchmarkReaderHuge/49844K-12 34 37282559 ns/op 1369.03 MB/s 60528624 B/op 174497 allocs/op ```
This commit is contained in:
Klaus Post
committed by
Harshavardhana
Harshavardhana
parent
e7f491a14b
commit
ddea0bdf11
@@ -17,89 +17,66 @@
|
||||
package csv
|
||||
|
||||
import (
|
||||
"bufio"
|
||||
"bytes"
|
||||
"encoding/csv"
|
||||
"fmt"
|
||||
"io"
|
||||
"runtime"
|
||||
"sync"
|
||||
|
||||
"github.com/minio/minio/pkg/s3select/sql"
|
||||
)
|
||||
|
||||
type recordReader struct {
|
||||
reader io.Reader
|
||||
recordDelimiter []byte
|
||||
oneByte []byte
|
||||
useOneByte bool
|
||||
}
|
||||
|
||||
func (rr *recordReader) Read(p []byte) (n int, err error) {
|
||||
if rr.useOneByte {
|
||||
p[0] = rr.oneByte[0]
|
||||
rr.useOneByte = false
|
||||
n, err = rr.reader.Read(p[1:])
|
||||
n++
|
||||
} else {
|
||||
n, err = rr.reader.Read(p)
|
||||
}
|
||||
|
||||
if err != nil {
|
||||
return 0, err
|
||||
}
|
||||
|
||||
if string(rr.recordDelimiter) == "\n" {
|
||||
return n, nil
|
||||
}
|
||||
|
||||
for {
|
||||
i := bytes.Index(p, rr.recordDelimiter)
|
||||
if i < 0 {
|
||||
break
|
||||
}
|
||||
|
||||
p[i] = '\n'
|
||||
if len(rr.recordDelimiter) > 1 {
|
||||
p = append(p[:i+1], p[i+len(rr.recordDelimiter):]...)
|
||||
n--
|
||||
}
|
||||
}
|
||||
|
||||
if len(rr.recordDelimiter) == 1 || p[n-1] != rr.recordDelimiter[0] {
|
||||
return n, nil
|
||||
}
|
||||
|
||||
if _, err = rr.reader.Read(rr.oneByte); err != nil {
|
||||
return 0, err
|
||||
}
|
||||
|
||||
if rr.oneByte[0] == rr.recordDelimiter[1] {
|
||||
p[n-1] = '\n'
|
||||
return n, nil
|
||||
}
|
||||
|
||||
rr.useOneByte = true
|
||||
return n, nil
|
||||
}
|
||||
|
||||
// Reader - CSV record reader for S3Select.
|
||||
type Reader struct {
|
||||
args *ReaderArgs
|
||||
readCloser io.ReadCloser
|
||||
csvReader *csv.Reader
|
||||
columnNames []string
|
||||
nameIndexMap map[string]int64
|
||||
readCloser io.ReadCloser // raw input
|
||||
buf *bufio.Reader // input to the splitter
|
||||
columnNames []string // names of columns
|
||||
nameIndexMap map[string]int64 // name to column index
|
||||
current [][]string // current block of results to be returned
|
||||
recordsRead int // number of records read in current slice
|
||||
input chan *queueItem // input for workers
|
||||
queue chan *queueItem // output from workers in order
|
||||
err error // global error state, only touched by Reader.Read
|
||||
bufferPool sync.Pool // pool of []byte objects for input
|
||||
csvDstPool sync.Pool // pool of [][]string used for output
|
||||
close chan struct{} // used for shutting down the splitter before end of stream
|
||||
readerWg sync.WaitGroup // used to keep track of async reader.
|
||||
}
|
||||
|
||||
// queueItem is an item in the queue.
|
||||
type queueItem struct {
|
||||
input []byte // raw input sent to the worker
|
||||
dst chan [][]string // result of block decode
|
||||
err error // any error encountered will be set here
|
||||
}
|
||||
|
||||
// Read - reads single record.
|
||||
func (r *Reader) Read() (sql.Record, error) {
|
||||
csvRecord, err := r.csvReader.Read()
|
||||
if err != nil {
|
||||
if err != io.EOF {
|
||||
return nil, errCSVParsingError(err)
|
||||
// Once Read is called the previous record should no longer be referenced.
|
||||
func (r *Reader) Read(dst sql.Record) (sql.Record, error) {
|
||||
// If we have have any records left, return these before any error.
|
||||
for len(r.current) <= r.recordsRead {
|
||||
if r.err != nil {
|
||||
return nil, r.err
|
||||
}
|
||||
|
||||
return nil, err
|
||||
// Move to next block
|
||||
item, ok := <-r.queue
|
||||
if !ok {
|
||||
r.err = io.EOF
|
||||
return nil, r.err
|
||||
}
|
||||
//lint:ignore SA6002 Using pointer would allocate more since we would have to copy slice header before taking a pointer.
|
||||
r.csvDstPool.Put(r.current)
|
||||
r.current = <-item.dst
|
||||
r.err = item.err
|
||||
r.recordsRead = 0
|
||||
}
|
||||
csvRecord := r.current[r.recordsRead]
|
||||
r.recordsRead++
|
||||
|
||||
// If no column names are set, use _(index)
|
||||
if r.columnNames == nil {
|
||||
r.columnNames = make([]string, len(csvRecord))
|
||||
for i := range csvRecord {
|
||||
@@ -107,67 +84,225 @@ func (r *Reader) Read() (sql.Record, error) {
|
||||
}
|
||||
}
|
||||
|
||||
// If no index max, add that.
|
||||
if r.nameIndexMap == nil {
|
||||
r.nameIndexMap = make(map[string]int64)
|
||||
for i := range r.columnNames {
|
||||
r.nameIndexMap[r.columnNames[i]] = int64(i)
|
||||
}
|
||||
}
|
||||
dstRec, ok := dst.(*Record)
|
||||
if !ok {
|
||||
dstRec = &Record{}
|
||||
}
|
||||
dstRec.columnNames = r.columnNames
|
||||
dstRec.csvRecord = csvRecord
|
||||
dstRec.nameIndexMap = r.nameIndexMap
|
||||
|
||||
return &Record{
|
||||
columnNames: r.columnNames,
|
||||
csvRecord: csvRecord,
|
||||
nameIndexMap: r.nameIndexMap,
|
||||
}, nil
|
||||
return dstRec, nil
|
||||
}
|
||||
|
||||
// Close - closes underlaying reader.
|
||||
// Close - closes underlying reader.
|
||||
func (r *Reader) Close() error {
|
||||
if r.close != nil {
|
||||
close(r.close)
|
||||
r.close = nil
|
||||
r.readerWg.Wait()
|
||||
}
|
||||
r.recordsRead = len(r.current)
|
||||
if r.err == nil {
|
||||
r.err = io.EOF
|
||||
}
|
||||
return r.readCloser.Close()
|
||||
}
|
||||
|
||||
// nextSplit will attempt to skip a number of bytes and
|
||||
// return the buffer until the next newline occurs.
|
||||
// The last block will be sent along with an io.EOF.
|
||||
func (r *Reader) nextSplit(skip int, dst []byte) ([]byte, error) {
|
||||
if cap(dst) < skip {
|
||||
dst = make([]byte, 0, skip+1024)
|
||||
}
|
||||
dst = dst[:skip]
|
||||
if skip > 0 {
|
||||
n, err := io.ReadFull(r.buf, dst)
|
||||
if err != nil && err != io.ErrUnexpectedEOF {
|
||||
// If an EOF happens after reading some but not all the bytes,
|
||||
// ReadFull returns ErrUnexpectedEOF.
|
||||
return dst[:n], err
|
||||
}
|
||||
dst = dst[:n]
|
||||
if err == io.ErrUnexpectedEOF {
|
||||
return dst, io.EOF
|
||||
}
|
||||
}
|
||||
// Read until next line.
|
||||
in, err := r.buf.ReadBytes('\n')
|
||||
dst = append(dst, in...)
|
||||
return dst, err
|
||||
}
|
||||
|
||||
// csvSplitSize is the size of each block.
|
||||
// Blocks will read this much and find the first following newline.
|
||||
// 128KB appears to be a very reasonable default.
|
||||
const csvSplitSize = 128 << 10
|
||||
|
||||
// startReaders will read the header if needed and spin up a parser
|
||||
// and a number of workers based on GOMAXPROCS.
|
||||
// If an error is returned no goroutines have been started and r.err will have been set.
|
||||
func (r *Reader) startReaders(in io.Reader, newReader func(io.Reader) *csv.Reader) error {
|
||||
if r.args.FileHeaderInfo != none {
|
||||
// Read column names
|
||||
// Get one line.
|
||||
b, err := r.nextSplit(0, nil)
|
||||
if err != nil {
|
||||
r.err = err
|
||||
return err
|
||||
}
|
||||
reader := newReader(bytes.NewReader(b))
|
||||
record, err := reader.Read()
|
||||
if err != nil {
|
||||
r.err = err
|
||||
if err != io.EOF {
|
||||
r.err = errCSVParsingError(err)
|
||||
return errCSVParsingError(err)
|
||||
}
|
||||
return err
|
||||
}
|
||||
|
||||
if r.args.FileHeaderInfo == use {
|
||||
// Copy column names since records will be reused.
|
||||
columns := append(make([]string, 0, len(record)), record...)
|
||||
r.columnNames = columns
|
||||
}
|
||||
}
|
||||
|
||||
r.bufferPool.New = func() interface{} {
|
||||
return make([]byte, csvSplitSize+1024)
|
||||
}
|
||||
|
||||
// Create queue
|
||||
r.queue = make(chan *queueItem, runtime.GOMAXPROCS(0))
|
||||
r.input = make(chan *queueItem, runtime.GOMAXPROCS(0))
|
||||
r.readerWg.Add(1)
|
||||
|
||||
// Start splitter
|
||||
go func() {
|
||||
defer close(r.input)
|
||||
defer close(r.queue)
|
||||
defer r.readerWg.Done()
|
||||
for {
|
||||
next, err := r.nextSplit(csvSplitSize, r.bufferPool.Get().([]byte))
|
||||
q := queueItem{
|
||||
input: next,
|
||||
dst: make(chan [][]string, 1),
|
||||
err: err,
|
||||
}
|
||||
select {
|
||||
case <-r.close:
|
||||
return
|
||||
case r.queue <- &q:
|
||||
}
|
||||
|
||||
select {
|
||||
case <-r.close:
|
||||
return
|
||||
case r.input <- &q:
|
||||
}
|
||||
if err != nil {
|
||||
// Exit on any error.
|
||||
return
|
||||
}
|
||||
}
|
||||
}()
|
||||
|
||||
// Start parsers
|
||||
for i := 0; i < runtime.GOMAXPROCS(0); i++ {
|
||||
go func() {
|
||||
for in := range r.input {
|
||||
if len(in.input) == 0 {
|
||||
in.dst <- nil
|
||||
continue
|
||||
}
|
||||
dst, ok := r.csvDstPool.Get().([][]string)
|
||||
if !ok {
|
||||
dst = make([][]string, 0, 1000)
|
||||
}
|
||||
|
||||
cr := newReader(bytes.NewBuffer(in.input))
|
||||
all := dst[:0]
|
||||
err := func() error {
|
||||
// Read all records until EOF or another error.
|
||||
for {
|
||||
record, err := cr.Read()
|
||||
if err == io.EOF {
|
||||
return nil
|
||||
}
|
||||
if err != nil {
|
||||
return errCSVParsingError(err)
|
||||
}
|
||||
var recDst []string
|
||||
if len(dst) > len(all) {
|
||||
recDst = dst[len(all)]
|
||||
}
|
||||
if cap(recDst) < len(record) {
|
||||
recDst = make([]string, len(record))
|
||||
}
|
||||
recDst = recDst[:len(record)]
|
||||
copy(recDst, record)
|
||||
all = append(all, recDst)
|
||||
}
|
||||
}()
|
||||
if err != nil {
|
||||
in.err = err
|
||||
}
|
||||
// We don't need the input any more.
|
||||
//lint:ignore SA6002 Using pointer would allocate more since we would have to copy slice header before taking a pointer.
|
||||
r.bufferPool.Put(in.input)
|
||||
in.input = nil
|
||||
in.dst <- all
|
||||
}
|
||||
}()
|
||||
}
|
||||
return nil
|
||||
|
||||
}
|
||||
|
||||
// NewReader - creates new CSV reader using readCloser.
|
||||
func NewReader(readCloser io.ReadCloser, args *ReaderArgs) (*Reader, error) {
|
||||
if args == nil || args.IsEmpty() {
|
||||
panic(fmt.Errorf("empty args passed %v", args))
|
||||
}
|
||||
|
||||
csvReader := csv.NewReader(&recordReader{
|
||||
reader: readCloser,
|
||||
recordDelimiter: []byte(args.RecordDelimiter),
|
||||
oneByte: []byte{0},
|
||||
})
|
||||
csvReader.Comma = []rune(args.FieldDelimiter)[0]
|
||||
csvReader.Comment = []rune(args.CommentCharacter)[0]
|
||||
csvReader.FieldsPerRecord = -1
|
||||
// If LazyQuotes is true, a quote may appear in an unquoted field and a
|
||||
// non-doubled quote may appear in a quoted field.
|
||||
csvReader.LazyQuotes = true
|
||||
// We do not trim leading space to keep consistent with s3.
|
||||
csvReader.TrimLeadingSpace = false
|
||||
csvIn := io.Reader(readCloser)
|
||||
if args.RecordDelimiter != "\n" {
|
||||
csvIn = &recordTransform{
|
||||
reader: readCloser,
|
||||
recordDelimiter: []byte(args.RecordDelimiter),
|
||||
oneByte: make([]byte, len(args.RecordDelimiter)-1),
|
||||
}
|
||||
}
|
||||
|
||||
r := &Reader{
|
||||
args: args,
|
||||
buf: bufio.NewReaderSize(csvIn, csvSplitSize*2),
|
||||
readCloser: readCloser,
|
||||
csvReader: csvReader,
|
||||
close: make(chan struct{}),
|
||||
}
|
||||
|
||||
if args.FileHeaderInfo == none {
|
||||
return r, nil
|
||||
// Assume args are validated by ReaderArgs.UnmarshalXML()
|
||||
newCsvReader := func(r io.Reader) *csv.Reader {
|
||||
ret := csv.NewReader(r)
|
||||
ret.Comma = []rune(args.FieldDelimiter)[0]
|
||||
ret.Comment = []rune(args.CommentCharacter)[0]
|
||||
ret.FieldsPerRecord = -1
|
||||
// If LazyQuotes is true, a quote may appear in an unquoted field and a
|
||||
// non-doubled quote may appear in a quoted field.
|
||||
ret.LazyQuotes = true
|
||||
// We do not trim leading space to keep consistent with s3.
|
||||
ret.TrimLeadingSpace = false
|
||||
ret.ReuseRecord = true
|
||||
return ret
|
||||
}
|
||||
|
||||
record, err := csvReader.Read()
|
||||
if err != nil {
|
||||
if err != io.EOF {
|
||||
return nil, errCSVParsingError(err)
|
||||
}
|
||||
|
||||
return nil, err
|
||||
}
|
||||
|
||||
if args.FileHeaderInfo == use {
|
||||
r.columnNames = record
|
||||
}
|
||||
|
||||
return r, nil
|
||||
return r, r.startReaders(csvIn, newCsvReader)
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user