erigon-pulse/common/etl/collector.go
Alex Sharov 0e253e7336
lmdb transactions of unlimited size (#918)
* add logging to loader

* use pure tx in etl loading, logs in mutation commit

* clean

* bletter logging and more cleanup

* bletter logging and more cleanup

* increase batch size to 500M

* better batch commit logging

* async fsync

* sync fsync

* sync fsync

* unify logging

* fix corner-case when etl can use empty bucket name

* fix tests

* better logging

* better logging

* rebase master

* remove lmdb.NoMetaSync flag for now

* consistent walk and multi-walk

* clean

* sub tx

* add consistent multi-put

* implement dupsort support in one new cursor method

* clear
2020-08-17 07:45:52 +01:00

213 lines
5.6 KiB
Go

package etl
import (
"bytes"
"container/heap"
"fmt"
"io"
"runtime"
"time"
"github.com/ledgerwatch/turbo-geth/common"
"github.com/ledgerwatch/turbo-geth/ethdb"
"github.com/ledgerwatch/turbo-geth/log"
"github.com/ugorji/go/codec"
)
type LoadNextFunc func(originalK, k, v []byte) error
type LoadFunc func(k []byte, value []byte, state State, next LoadNextFunc) error
// Collector performs the job of ETL Transform, but can also be used without "E" (Extract) part
// as a Collect Transform Load
type Collector struct {
extractNextFunc ExtractNextFunc
flushBuffer func([]byte, bool) error
dataProviders []dataProvider
allFlushed bool
}
func NewCollector(datadir string, sortableBuffer Buffer) *Collector {
c := &Collector{}
encoder := codec.NewEncoder(nil, &cbor)
c.flushBuffer = func(currentKey []byte, canStoreInRam bool) error {
if sortableBuffer.Len() == 0 {
return nil
}
var provider dataProvider
var err error
sortableBuffer.Sort()
if canStoreInRam && len(c.dataProviders) == 0 {
provider = KeepInRAM(sortableBuffer)
c.allFlushed = true
} else {
provider, err = FlushToDisk(encoder, currentKey, sortableBuffer, datadir)
}
if err != nil {
return err
}
if provider != nil {
c.dataProviders = append(c.dataProviders, provider)
}
return nil
}
c.extractNextFunc = func(originalK, k []byte, v []byte) error {
sortableBuffer.Put(common.CopyBytes(k), common.CopyBytes(v))
if sortableBuffer.CheckFlushSize() {
if err := c.flushBuffer(originalK, false); err != nil {
return err
}
}
return nil
}
return c
}
func (c *Collector) Collect(k, v []byte) error {
return c.extractNextFunc(k, k, v)
}
func (c *Collector) Load(db ethdb.Database, toBucket string, loadFunc LoadFunc, args TransformArgs) error {
defer func() {
disposeProviders(c.dataProviders)
}()
if !c.allFlushed {
if err := c.flushBuffer(nil, true); err != nil {
return err
}
}
return loadFilesIntoBucket(db, toBucket, c.dataProviders, loadFunc, args)
}
func loadFilesIntoBucket(db ethdb.Database, bucket string, providers []dataProvider, loadFunc LoadFunc, args TransformArgs) error {
decoder := codec.NewDecoder(nil, &cbor)
var m runtime.MemStats
h := &Heap{}
heap.Init(h)
for i, provider := range providers {
if key, value, err := provider.Next(decoder); err == nil {
he := HeapElem{key, i, value}
heap.Push(h, he)
} else /* we must have at least one entry per file */ {
eee := fmt.Errorf("error reading first readers: n=%d current=%d provider=%s err=%v",
len(providers), i, provider, err)
panic(eee)
}
}
batch, err := db.Begin()
if err != nil {
return err
}
defer batch.Rollback()
state := &bucketState{batch, bucket, args.Quit}
haveSortingGuaranties := isIdentityLoadFunc(loadFunc) // user-defined loadFunc may change ordering
var lastKey []byte
if bucket != "" { // passing empty bucket name is valid case for etl when DB modification is not expected
var errLast error
lastKey, _, errLast = batch.Last(bucket)
if errLast != nil {
return errLast
}
}
var canUseAppend bool
putTimer := time.Now()
i := 0
loadNextFunc := func(originalK, k, v []byte) error {
if i == 0 {
isEndOfBucket := lastKey == nil || bytes.Compare(lastKey, k) == -1
canUseAppend = haveSortingGuaranties && isEndOfBucket
}
i++
if i%1_000_000 == 0 && time.Since(putTimer) > 30*time.Second {
putTimer = time.Now()
runtime.ReadMemStats(&m)
log.Info(
"Loading into bucket",
"bucket", bucket,
"size", common.StorageSize(batch.BatchSize()),
"keys", fmt.Sprintf("%.1fM", float64(i)/1_000_000),
"use append", canUseAppend,
"current key", makeCurrentKeyStr(originalK),
"alloc", common.StorageSize(m.Alloc), "sys", common.StorageSize(m.Sys), "numGC", int(m.NumGC))
}
if canUseAppend && len(v) == 0 {
return nil // nothing to delete after end of bucket
}
if len(v) == 0 {
if err := batch.Delete(bucket, k); err != nil {
return err
}
return nil
}
if canUseAppend {
if err := batch.(*ethdb.TxDb).Append(bucket, k, v); err != nil {
return err
}
return nil
}
if err := batch.Put(bucket, k, v); err != nil {
return err
}
return nil
}
// Main loading loop
for h.Len() > 0 {
if err := common.Stopped(args.Quit); err != nil {
return err
}
element := (heap.Pop(h)).(HeapElem)
provider := providers[element.TimeIdx]
err := loadFunc(element.Key, element.Value, state, loadNextFunc)
if err != nil {
return err
}
if element.Key, element.Value, err = provider.Next(decoder); err == nil {
heap.Push(h, element)
} else if err != io.EOF {
return fmt.Errorf("error while reading next element from disk: %v", err)
}
}
// Final commit
if args.OnLoadCommit != nil {
if err := args.OnLoadCommit(batch, []byte{}, true); err != nil {
return err
}
}
commitTimer := time.Now()
if _, err := batch.Commit(); err != nil {
return err
}
commitTook := time.Since(commitTimer)
runtime.ReadMemStats(&m)
log.Debug(
"Committed batch",
"bucket", bucket,
"commit", commitTook,
"size", common.StorageSize(batch.BatchSize()),
"current key", makeCurrentKeyStr(nil),
"alloc", common.StorageSize(m.Alloc), "sys", common.StorageSize(m.Sys), "numGC", int(m.NumGC))
return nil
}
func makeCurrentKeyStr(k []byte) string {
var currentKeyStr string
if k == nil {
currentKeyStr = "final"
} else if len(k) < 4 {
currentKeyStr = fmt.Sprintf("%x", k)
} else if k[0] == 0 && k[1] == 0 && k[2] == 0 && k[3] == 0 && len(k) >= 8 { // if key has leading zeroes, show a bit more info
currentKeyStr = fmt.Sprintf("%x...", k[:8])
} else {
currentKeyStr = fmt.Sprintf("%x...", k[:4])
}
return currentKeyStr
}