VictoriaMetrics/lib/protoparser/native/streamparser.go
Aliaksandr Valialkin c63755c316
lib/writeconcurrencylimiter: improve the logic behind -maxConcurrentInserts limit
Previously the -maxConcurrentInserts was limiting the number of established client connections,
which write data to VictoriaMetrics. Some of these connections could be idle.
Such connections do not consume big amounts of CPU and RAM, so there is a little sense in limiting
the number of such connections. So now the -maxConcurrentInserts command-line option
limits the number of concurrently executed insert requests, not including idle connections.

It is recommended removing -maxConcurrentInserts command-line option, since the default value
for this option should work good for most cases.
2023-01-06 22:20:19 -08:00

235 lines
6.2 KiB
Go

package native
import (
"bufio"
"fmt"
"io"
"sync"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/common"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/writeconcurrencylimiter"
"github.com/VictoriaMetrics/metrics"
)
// ParseStream parses /api/v1/import/native lines from req and calls callback for parsed blocks.
//
// The callback can be called concurrently multiple times for streamed data from r.
//
// callback shouldn't hold block after returning.
func ParseStream(r io.Reader, isGzip bool, callback func(block *Block) error) error {
wcr := writeconcurrencylimiter.GetReader(r)
defer writeconcurrencylimiter.PutReader(wcr)
r = wcr
if isGzip {
zr, err := common.GetGzipReader(r)
if err != nil {
return fmt.Errorf("cannot read gzipped vmimport data: %w", err)
}
defer common.PutGzipReader(zr)
r = zr
}
br := getBufferedReader(r)
defer putBufferedReader(br)
// Read time range (tr)
trBuf := make([]byte, 16)
var tr storage.TimeRange
if _, err := io.ReadFull(br, trBuf); err != nil {
readErrors.Inc()
return fmt.Errorf("cannot read time range: %w", err)
}
tr.MinTimestamp = encoding.UnmarshalInt64(trBuf)
tr.MaxTimestamp = encoding.UnmarshalInt64(trBuf[8:])
// Read native blocks and feed workers with work.
sizeBuf := make([]byte, 4)
ctx := &streamContext{}
for {
uw := getUnmarshalWork()
uw.tr = tr
uw.ctx = ctx
uw.callback = callback
// Read uw.metricNameBuf
if _, err := io.ReadFull(br, sizeBuf); err != nil {
if err == io.EOF {
// End of stream
putUnmarshalWork(uw)
ctx.wg.Wait()
return ctx.err
}
readErrors.Inc()
ctx.wg.Wait()
return fmt.Errorf("cannot read metricName size: %w", err)
}
readCalls.Inc()
bufSize := encoding.UnmarshalUint32(sizeBuf)
if bufSize > 1024*1024 {
parseErrors.Inc()
ctx.wg.Wait()
return fmt.Errorf("too big metricName size; got %d; shouldn't exceed %d", bufSize, 1024*1024)
}
uw.metricNameBuf = bytesutil.ResizeNoCopyMayOverallocate(uw.metricNameBuf, int(bufSize))
if _, err := io.ReadFull(br, uw.metricNameBuf); err != nil {
readErrors.Inc()
ctx.wg.Wait()
return fmt.Errorf("cannot read metricName with size %d bytes: %w", bufSize, err)
}
readCalls.Inc()
// Read uw.blockBuf
if _, err := io.ReadFull(br, sizeBuf); err != nil {
readErrors.Inc()
ctx.wg.Wait()
return fmt.Errorf("cannot read native block size: %w", err)
}
readCalls.Inc()
bufSize = encoding.UnmarshalUint32(sizeBuf)
if bufSize > 1024*1024 {
parseErrors.Inc()
ctx.wg.Wait()
return fmt.Errorf("too big native block size; got %d; shouldn't exceed %d", bufSize, 1024*1024)
}
uw.blockBuf = bytesutil.ResizeNoCopyMayOverallocate(uw.blockBuf, int(bufSize))
if _, err := io.ReadFull(br, uw.blockBuf); err != nil {
readErrors.Inc()
ctx.wg.Wait()
return fmt.Errorf("cannot read native block with size %d bytes: %w", bufSize, err)
}
readCalls.Inc()
blocksRead.Inc()
ctx.wg.Add(1)
common.ScheduleUnmarshalWork(uw)
wcr.DecConcurrency()
}
}
type streamContext struct {
wg sync.WaitGroup
errLock sync.Mutex
err error
}
// Block is a single block from `/api/v1/import/native` request.
type Block struct {
MetricName storage.MetricName
Values []float64
Timestamps []int64
}
func (b *Block) reset() {
b.MetricName.Reset()
b.Values = b.Values[:0]
b.Timestamps = b.Timestamps[:0]
}
var (
readCalls = metrics.NewCounter(`vm_protoparser_read_calls_total{type="native"}`)
readErrors = metrics.NewCounter(`vm_protoparser_read_errors_total{type="native"}`)
rowsRead = metrics.NewCounter(`vm_protoparser_rows_read_total{type="native"}`)
blocksRead = metrics.NewCounter(`vm_protoparser_blocks_read_total{type="native"}`)
parseErrors = metrics.NewCounter(`vm_protoparser_parse_errors_total{type="native"}`)
processErrors = metrics.NewCounter(`vm_protoparser_process_errors_total{type="native"}`)
)
type unmarshalWork struct {
tr storage.TimeRange
ctx *streamContext
callback func(block *Block) error
metricNameBuf []byte
blockBuf []byte
block Block
}
func (uw *unmarshalWork) reset() {
uw.ctx = nil
uw.callback = nil
uw.metricNameBuf = uw.metricNameBuf[:0]
uw.blockBuf = uw.blockBuf[:0]
uw.block.reset()
}
// Unmarshal implements common.UnmarshalWork
func (uw *unmarshalWork) Unmarshal() {
err := uw.unmarshal()
if err != nil {
parseErrors.Inc()
} else {
err = uw.callback(&uw.block)
}
ctx := uw.ctx
if err != nil {
processErrors.Inc()
ctx.errLock.Lock()
if ctx.err == nil {
ctx.err = fmt.Errorf("error when processing native block: %w", err)
}
ctx.errLock.Unlock()
}
ctx.wg.Done()
putUnmarshalWork(uw)
}
func (uw *unmarshalWork) unmarshal() error {
block := &uw.block
if err := block.MetricName.Unmarshal(uw.metricNameBuf); err != nil {
return fmt.Errorf("cannot unmarshal metricName from %d bytes: %w", len(uw.metricNameBuf), err)
}
tmpBlock := blockPool.Get().(*storage.Block)
defer blockPool.Put(tmpBlock)
tail, err := tmpBlock.UnmarshalPortable(uw.blockBuf)
if err != nil {
return fmt.Errorf("cannot unmarshal native block from %d bytes: %w", len(uw.blockBuf), err)
}
if len(tail) > 0 {
return fmt.Errorf("unexpected non-empty tail left after unmarshaling native block from %d bytes; len(tail)=%d bytes", len(uw.blockBuf), len(tail))
}
block.Timestamps, block.Values = tmpBlock.AppendRowsWithTimeRangeFilter(block.Timestamps[:0], block.Values[:0], uw.tr)
rowsRead.Add(len(block.Timestamps))
return nil
}
var blockPool = &sync.Pool{
New: func() interface{} {
return &storage.Block{}
},
}
func getUnmarshalWork() *unmarshalWork {
v := unmarshalWorkPool.Get()
if v == nil {
return &unmarshalWork{}
}
return v.(*unmarshalWork)
}
func putUnmarshalWork(uw *unmarshalWork) {
uw.reset()
unmarshalWorkPool.Put(uw)
}
var unmarshalWorkPool sync.Pool
func getBufferedReader(r io.Reader) *bufio.Reader {
v := bufferedReaderPool.Get()
if v == nil {
return bufio.NewReaderSize(r, 64*1024)
}
br := v.(*bufio.Reader)
br.Reset(r)
return br
}
func putBufferedReader(br *bufio.Reader) {
br.Reset(nil)
bufferedReaderPool.Put(br)
}
var bufferedReaderPool sync.Pool