featurebase/ctl/presort.go
Matthew Jaffee 8f1f3c6d06
Fix pre sort cmd (#2222)
* make ndjson pre_sort parallel, fix several bugs, test

* remove json tags (not needed), rename pre_sort -> presort
2023-01-25 11:01:58 -06:00

296 lines
7 KiB
Go

package ctl
import (
"bufio"
"context"
"encoding/csv"
"encoding/json"
"fmt"
"hash/fnv"
"io"
"io/fs"
"os"
"path/filepath"
"sync"
"github.com/featurebasedb/featurebase/v3/logger"
"github.com/pkg/errors"
"golang.org/x/sync/errgroup"
)
type PreSortCommand struct {
// Optional Index filter
File string
Type string
Table string
PrimaryKeyFields []string
PrimaryKeySeparator string
PartitionN int
JobSize int
NumWorkers int
// Path to write sorted files to
OutputDir string
// Standard input/output
logDest logger.Logger
outputLocks []sync.Mutex
outputFiles map[int]*os.File
outputEncoders map[int]*json.Encoder
outputWriters map[int]*csv.Writer
}
// NewPreSortCommand returns a new instance of PreSortCommand.
func NewPreSortCommand(logdest logger.Logger) *PreSortCommand {
return &PreSortCommand{
logDest: logdest,
Type: "ndjson",
OutputDir: "presorted_files",
PartitionN: 256,
JobSize: 1000,
NumWorkers: 4,
PrimaryKeySeparator: "|",
outputFiles: make(map[int]*os.File),
}
}
// Run executes the main program execution.
func (cmd *PreSortCommand) Run(ctx context.Context) error {
if cmd.File == "" {
return errors.New("must set a file or directory to read from")
}
if len(cmd.PrimaryKeyFields) == 0 {
return errors.New("must define primary-key-fields")
}
if cmd.Table == "" {
return errors.New("must set a table name")
}
cmd.outputLocks = make([]sync.Mutex, cmd.PartitionN)
filelist := []string{}
walkfunc := func(path string, d fs.DirEntry, err error) error {
if !d.IsDir() {
filelist = append(filelist, path)
}
return nil
}
err := filepath.WalkDir(cmd.File, walkfunc)
if err != nil {
return errors.Wrapf(err, "walking %s", cmd.File)
}
if len(filelist) > 3 {
cmd.logDest.Printf("%d files: %v...", len(filelist), filelist[:3])
} else {
cmd.logDest.Printf("Files: %v", filelist)
}
if err := os.MkdirAll(cmd.OutputDir, 0755); err != nil {
return errors.Wrapf(err, "creating output directory: '%s'", cmd.OutputDir)
}
for i := 0; i < cmd.PartitionN; i++ {
name := filepath.Join(cmd.OutputDir, fmt.Sprintf("output_%d", i))
f, err := os.Create(name)
if err != nil {
return errors.Wrapf(err, "couldn't open output file %s", name)
}
cmd.outputFiles[i] = f
}
defer func() {
for _, f := range cmd.outputFiles {
_ = f.Sync()
_ = f.Close()
}
}()
linech := make(chan [][]byte, 16)
var eg *errgroup.Group
switch cmd.Type {
case "ndjson":
eg = cmd.setupNDJSON(linech)
case "csv":
// CSV ends here
return errors.Wrap(cmd.RunCSV(filelist), "running CSV command")
default:
return errors.Errorf("unsupported type %s, try ndjson or csv", cmd.Type)
}
// The rest of this is just for NDJSON... CSV doesn't currently
// support multiple workers. It needs to do the file reading
// differently in order to handle potential newlines within
// fields.
for _, fname := range filelist {
f, err := os.Open(fname)
if err != nil {
return errors.Wrapf(err, "opening %s", fname)
}
sc := bufio.NewScanner(f)
batch:
for {
lines := make([][]byte, 0, cmd.JobSize)
for i := 0; i < cmd.JobSize; i++ {
if !sc.Scan() {
linech <- lines
break batch
}
line := sc.Bytes()
// must copy line as scanner will reuse the buffer it's given us
lc := make([]byte, len(line))
copy(lc, line)
lines = append(lines, lc)
}
linech <- lines
}
if err := sc.Err(); err != nil {
return errors.Wrap(err, "scanning file")
}
}
close(linech)
if err := eg.Wait(); err != nil {
return err
}
return nil
}
func (cmd *PreSortCommand) setupNDJSON(linech chan [][]byte) *errgroup.Group {
cmd.outputEncoders = make(map[int]*json.Encoder)
for i, f := range cmd.outputFiles {
cmd.outputEncoders[i] = json.NewEncoder(f)
cmd.outputEncoders[i].SetIndent("", "")
cmd.outputEncoders[i].SetEscapeHTML(false)
}
// start workers
eg := &errgroup.Group{}
for i := 0; i < cmd.NumWorkers; i++ {
eg.Go(func() error {
return cmd.ndjsonWorker(linech)
})
}
return eg
}
func (cmd *PreSortCommand) ndjsonWorker(linech chan [][]byte) error {
for lines := range linech {
rec := map[string]interface{}{}
for _, line := range lines {
err := json.Unmarshal(line, &rec)
if err != nil {
return errors.Wrapf(err, "unmarshaling '%s'", line)
}
partition, err := cmd.ndjsonPartition(rec)
if err != nil {
return err
}
if err := cmd.outputNDJSON(partition, rec); err != nil {
return errors.Wrap(err, "outputting ndjson")
}
for k := range rec {
delete(rec, k)
}
}
}
return nil
}
func (cmd *PreSortCommand) outputNDJSON(partition int, record map[string]interface{}) error {
cmd.outputLocks[partition].Lock()
defer cmd.outputLocks[partition].Unlock()
return cmd.outputEncoders[partition].Encode(record)
}
func (cmd *PreSortCommand) ndjsonPartition(rec map[string]interface{}) (int, error) {
h := fnv.New64a()
_, _ = h.Write([]byte(cmd.Table))
for i, name := range cmd.PrimaryKeyFields {
val, ok := rec[name]
if !ok {
return 0, errors.Errorf("couldn't find primary key part '%s' in record: %+v", name, rec)
}
h.Write([]byte(toString(val)))
if i < len(cmd.PrimaryKeyFields)-1 {
h.Write([]byte(cmd.PrimaryKeySeparator))
}
}
return int(h.Sum64() % uint64(cmd.PartitionN)), nil
}
func toString(val interface{}) string {
switch valt := val.(type) {
case string:
return valt
case float64, int, int64, uint64, float32, uint:
return fmt.Sprintf("%d", valt)
default:
return fmt.Sprintf("%v", valt)
}
}
func (cmd *PreSortCommand) RunCSV(filelist []string) error {
outputWriters := make(map[int]*csv.Writer)
for i, f := range cmd.outputFiles {
outputWriters[i] = csv.NewWriter(f)
}
for _, fname := range filelist {
f, err := os.Open(fname)
if err != nil {
cmd.logDest.Warnf("Could not open %s, skipping", fname)
}
reader := csv.NewReader(f)
reader.ReuseRecord = true
rec, err := reader.Read()
if err != nil {
return errors.Wrap(err, "")
}
header := make(map[string]int)
for i, key := range rec {
header[key] = i
}
for rec, err = reader.Read(); err == nil; rec, err = reader.Read() {
partition, err := cmd.csvPartition(header, rec)
if err != nil {
return err
}
err = outputWriters[partition].Write(rec)
if err != nil {
return errors.Wrapf(err, "writing record to partition %d", partition)
}
}
if err != io.EOF && err != nil {
return errors.Wrapf(err, "error attempting to decode json from %s", fname)
}
}
for _, w := range outputWriters {
w.Flush()
}
return nil
}
func (cmd *PreSortCommand) csvPartition(header map[string]int, rec []string) (int, error) {
h := fnv.New64a()
_, _ = h.Write([]byte(cmd.Table))
for i, name := range cmd.PrimaryKeyFields {
pos, ok := header[name]
if !ok {
return 0, errors.Errorf("couldn't find primary key part '%s' in header: %+v", name, header)
}
h.Write([]byte(rec[pos]))
if i < len(cmd.PrimaryKeyFields)-1 {
h.Write([]byte(cmd.PrimaryKeySeparator))
}
}
return int(h.Sum64() % uint64(cmd.PartitionN)), nil
}