scrapework.go

package promscrape

import (
	"bytes"
	"flag"
	"fmt"
	"io"
	"math"
	"math/bits"
	"strings"
	"sync"
	"time"

	"github.com/VictoriaMetrics/VictoriaMetrics/lib/auth"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bloomfilter"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/cgroup"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/flagutil"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/leveledbytebufferpool"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promauth"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promrelabel"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promutils"
	parser "github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/prometheus"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/protoparser/prometheus/stream"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/proxy"
	"github.com/VictoriaMetrics/VictoriaMetrics/lib/timerpool"
	"github.com/VictoriaMetrics/metrics"
	"github.com/cespare/xxhash/v2"
)

var (
	suppressScrapeErrors = flag.Bool("promscrape.suppressScrapeErrors", false, "Whether to suppress scrape errors logging. "+
		"The last error for each target is always available at '/targets' page even if scrape errors logging is suppressed. "+
		"See also -promscrape.suppressScrapeErrorsDelay")
	suppressScrapeErrorsDelay = flag.Duration("promscrape.suppressScrapeErrorsDelay", 0, "The delay for suppressing repeated scrape errors logging per each scrape targets. "+
		"This may be used for reducing the number of log lines related to scrape errors. See also -promscrape.suppressScrapeErrors")
	minResponseSizeForStreamParse = flagutil.NewBytes("promscrape.minResponseSizeForStreamParse", 1e6, "The minimum target response size for automatic switching to stream parsing mode, which can reduce memory usage. See https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode")
)

// ScrapeWork represents a unit of work for scraping Prometheus metrics.
//
// It must be immutable during its lifetime, since it is read from concurrently running goroutines.
type ScrapeWork struct {
	// Full URL (including query args) for the scrape.
	ScrapeURL string

	// Interval for scraping the ScrapeURL.
	ScrapeInterval time.Duration

	// Timeout for scraping the ScrapeURL.
	ScrapeTimeout time.Duration

	// How to deal with conflicting labels.
	// See https://prometheus.io/docs/prometheus/latest/configuration/configuration/#scrape_config
	HonorLabels bool

	// How to deal with scraped timestamps.
	// See https://prometheus.io/docs/prometheus/latest/configuration/configuration/#scrape_config
	HonorTimestamps bool

	// Whether to deny redirects during requests to scrape config.
	DenyRedirects bool

	// Do not support enable_http2 option because of the following reasons:
	//
	// - http2 is used very rarely comparing to http for Prometheus metrics exposition and service discovery
	// - http2 is much harder to debug than http
	// - http2 has very bad security record because of its complexity - see https://portswigger.net/research/http2
	//
	// VictoriaMetrics components are compiled with nethttpomithttp2 tag because of these issues.
	//
	// EnableHTTP2 bool

	// OriginalLabels contains original labels before relabeling.
	//
	// These labels are needed for relabeling troubleshooting at /targets page.
	//
	// OriginalLabels are sorted by name.
	OriginalLabels *promutils.Labels

	// Labels to add to the scraped metrics.
	//
	// The list contains at least the following labels according to https://www.robustperception.io/life-of-a-label/
	//
	//     * job
	//     * instance
	//     * user-defined labels set via `relabel_configs` section in `scrape_config`
	//
	// See also https://prometheus.io/docs/concepts/jobs_instances/
	//
	// Labels are sorted by name.
	Labels *promutils.Labels

	// ExternalLabels contains labels from global->external_labels section of -promscrape.config
	//
	// These labels are added to scraped metrics after the relabeling.
	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3137
	//
	// ExternalLabels are sorted by name.
	ExternalLabels *promutils.Labels

	// ProxyURL HTTP proxy url
	ProxyURL *proxy.URL

	// Auth config for ProxyUR:
	ProxyAuthConfig *promauth.Config

	// Auth config
	AuthConfig *promauth.Config

	// Optional `relabel_configs`.
	RelabelConfigs *promrelabel.ParsedConfigs

	// Optional `metric_relabel_configs`.
	MetricRelabelConfigs *promrelabel.ParsedConfigs

	// The maximum number of metrics to scrape after relabeling.
	SampleLimit int

	// Whether to disable response compression when querying ScrapeURL.
	DisableCompression bool

	// Whether to disable HTTP keep-alive when querying ScrapeURL.
	DisableKeepAlive bool

	// Whether to parse target responses in a streaming manner.
	StreamParse bool

	// The interval for aligning the first scrape.
	ScrapeAlignInterval time.Duration

	// The offset for the first scrape.
	ScrapeOffset time.Duration

	// Optional limit on the number of unique series the scrape target can expose.
	SeriesLimit int

	// Whether to process stale markers for the given target.
	// See https://docs.victoriametrics.com/vmagent.html#prometheus-staleness-markers
	NoStaleMarkers bool

	// The Tenant Info
	AuthToken *auth.Token

	// The original 'job_name'
	jobNameOriginal string
}

func (sw *ScrapeWork) canSwitchToStreamParseMode() bool {
	// Deny switching to stream parse mode if `sample_limit` or `series_limit` options are set,
	// since these limits cannot be applied in stream parsing mode.
	return sw.SampleLimit <= 0 && sw.SeriesLimit <= 0
}

// key returns unique identifier for the given sw.
//
// It can be used for comparing for equality for two ScrapeWork objects.
func (sw *ScrapeWork) key() string {
	// Do not take into account OriginalLabels, since they can be changed with relabeling.
	// Do not take into account RelabelConfigs, since it is already applied to Labels.
	// Take into account JobNameOriginal in order to capture the case when the original job_name is changed via relabeling.
	key := fmt.Sprintf("JobNameOriginal=%s, ScrapeURL=%s, ScrapeInterval=%s, ScrapeTimeout=%s, HonorLabels=%v, HonorTimestamps=%v, DenyRedirects=%v, Labels=%s, "+
		"ExternalLabels=%s, "+
		"ProxyURL=%s, ProxyAuthConfig=%s, AuthConfig=%s, MetricRelabelConfigs=%q, "+
		"SampleLimit=%d, DisableCompression=%v, DisableKeepAlive=%v, StreamParse=%v, "+
		"ScrapeAlignInterval=%s, ScrapeOffset=%s, SeriesLimit=%d, NoStaleMarkers=%v",
		sw.jobNameOriginal, sw.ScrapeURL, sw.ScrapeInterval, sw.ScrapeTimeout, sw.HonorLabels, sw.HonorTimestamps, sw.DenyRedirects, sw.Labels.String(),
		sw.ExternalLabels.String(),
		sw.ProxyURL.String(), sw.ProxyAuthConfig.String(), sw.AuthConfig.String(), sw.MetricRelabelConfigs.String(),
		sw.SampleLimit, sw.DisableCompression, sw.DisableKeepAlive, sw.StreamParse,
		sw.ScrapeAlignInterval, sw.ScrapeOffset, sw.SeriesLimit, sw.NoStaleMarkers)
	return key
}

// Job returns job for the ScrapeWork
func (sw *ScrapeWork) Job() string {
	return sw.Labels.Get("job")
}

type scrapeWork struct {
	// Config for the scrape.
	Config *ScrapeWork

	// ReadData is called for reading the data.
	ReadData func(dst []byte) ([]byte, error)

	// GetStreamReader is called if Config.StreamParse is set.
	GetStreamReader func() (*streamReader, error)

	// PushData is called for pushing collected data.
	PushData func(at *auth.Token, wr *prompbmarshal.WriteRequest)

	// ScrapeGroup is name of ScrapeGroup that
	// scrapeWork belongs to
	ScrapeGroup string

	tmpRow parser.Row

	// This flag is set to true if series_limit is exceeded.
	seriesLimitExceeded bool

	// labelsHashBuf is used for calculating the hash on series labels
	labelsHashBuf []byte

	// Optional limiter on the number of unique series per scrape target.
	seriesLimiter *bloomfilter.Limiter

	// prevBodyLen contains the previous response body length for the given scrape work.
	// It is used as a hint in order to reduce memory usage for body buffers.
	prevBodyLen int

	// prevLabelsLen contains the number labels scraped during the previous scrape.
	// It is used as a hint in order to reduce memory usage when parsing scrape responses.
	prevLabelsLen int

	// lastScrape holds the last response from scrape target.
	// It is used for staleness tracking and for populating scrape_series_added metric.
	// The lastScrape isn't populated if -promscrape.noStaleMarkers is set. This reduces memory usage.
	lastScrape []byte

	// lastScrapeCompressed is used for storing the compressed lastScrape between scrapes
	// in stream parsing mode in order to reduce memory usage when the lastScrape size
	// equals to or exceeds -promscrape.minResponseSizeForStreamParse
	lastScrapeCompressed []byte

	// nextErrorLogTime is the timestamp in millisecond when the next scrape error should be logged.
	nextErrorLogTime int64

	// failureRequestsCount is the number of suppressed scrape errors during the last suppressScrapeErrorsDelay
	failureRequestsCount int

	// successRequestsCount is the number of success requests during the last suppressScrapeErrorsDelay
	successRequestsCount int
}

func (sw *scrapeWork) loadLastScrape() string {
	if len(sw.lastScrapeCompressed) > 0 {
		b, err := encoding.DecompressZSTD(sw.lastScrape[:0], sw.lastScrapeCompressed)
		if err != nil {
			logger.Panicf("BUG: cannot unpack compressed previous response: %s", err)
		}
		sw.lastScrape = b
	}
	return bytesutil.ToUnsafeString(sw.lastScrape)
}

func (sw *scrapeWork) storeLastScrape(lastScrape []byte) {
	mustCompress := minResponseSizeForStreamParse.N > 0 && len(lastScrape) >= minResponseSizeForStreamParse.IntN()
	if mustCompress {
		sw.lastScrapeCompressed = encoding.CompressZSTDLevel(sw.lastScrapeCompressed[:0], lastScrape, 1)
		sw.lastScrape = nil
	} else {
		sw.lastScrape = append(sw.lastScrape[:0], lastScrape...)
		sw.lastScrapeCompressed = nil
	}
}

func (sw *scrapeWork) finalizeLastScrape() {
	if len(sw.lastScrapeCompressed) > 0 {
		// The compressed lastScrape is available in sw.lastScrapeCompressed.
		// Release the memory occupied by sw.lastScrape, so it won't be occupied between scrapes.
		sw.lastScrape = nil
	}
	if len(sw.lastScrape) > 0 {
		// Release the memory occupied by sw.lastScrapeCompressed, so it won't be occupied between scrapes.
		sw.lastScrapeCompressed = nil
	}
}

func (sw *scrapeWork) run(stopCh <-chan struct{}, globalStopCh <-chan struct{}) {
	var randSleep uint64
	scrapeInterval := sw.Config.ScrapeInterval
	scrapeAlignInterval := sw.Config.ScrapeAlignInterval
	scrapeOffset := sw.Config.ScrapeOffset
	if scrapeOffset > 0 {
		scrapeAlignInterval = scrapeInterval
	}
	if scrapeAlignInterval <= 0 {
		// Calculate start time for the first scrape from ScrapeURL and labels.
		// This should spread load when scraping many targets with different
		// scrape urls and labels.
		// This also makes consistent scrape times across restarts
		// for a target with the same ScrapeURL and labels.
		//
		// Include clusterName to the key in order to guarantee that the same
		// scrape target is scraped at different offsets per each cluster.
		// This guarantees that the deduplication consistently leaves samples received from the same vmagent.
		// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2679
		//
		// Include clusterMemberID to the key in order to guarantee that each member in vmagent cluster
		// scrapes replicated targets at different time offsets. This guarantees that the deduplication consistently leaves samples
		// received from the same vmagent replica.
		// See https://docs.victoriametrics.com/vmagent.html#scraping-big-number-of-targets
		key := fmt.Sprintf("clusterName=%s, clusterMemberID=%d, ScrapeURL=%s, Labels=%s", *clusterName, clusterMemberID, sw.Config.ScrapeURL, sw.Config.Labels.String())
		h := xxhash.Sum64(bytesutil.ToUnsafeBytes(key))
		randSleep = uint64(float64(scrapeInterval) * (float64(h) / (1 << 64)))
		sleepOffset := uint64(time.Now().UnixNano()) % uint64(scrapeInterval)
		if randSleep < sleepOffset {
			randSleep += uint64(scrapeInterval)
		}
		randSleep -= sleepOffset
	} else {
		d := uint64(scrapeAlignInterval)
		randSleep = d - uint64(time.Now().UnixNano())%d
		if scrapeOffset > 0 {
			randSleep += uint64(scrapeOffset)
		}
		randSleep %= uint64(scrapeInterval)
	}
	timer := timerpool.Get(time.Duration(randSleep))
	var timestamp int64
	var ticker *time.Ticker
	select {
	case <-stopCh:
		timerpool.Put(timer)
		return
	case <-timer.C:
		timerpool.Put(timer)
		ticker = time.NewTicker(scrapeInterval)
		timestamp = time.Now().UnixNano() / 1e6
		sw.scrapeAndLogError(timestamp, timestamp)
	}
	defer ticker.Stop()
	for {
		timestamp += scrapeInterval.Milliseconds()
		select {
		case <-stopCh:
			t := time.Now().UnixNano() / 1e6
			lastScrape := sw.loadLastScrape()
			select {
			case <-globalStopCh:
				// Do not send staleness markers on graceful shutdown as Prometheus does.
				// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2013#issuecomment-1006994079
			default:
				// Send staleness markers to all the metrics scraped last time from the target
				// when the given target disappears as Prometheus does.
				// Use the current real timestamp for staleness markers, so queries
				// stop returning data just after the time the target disappears.
				sw.sendStaleSeries(lastScrape, "", t, true)
			}
			if sw.seriesLimiter != nil {
				sw.seriesLimiter.MustStop()
				sw.seriesLimiter = nil
			}
			return
		case tt := <-ticker.C:
			t := tt.UnixNano() / 1e6
			if d := math.Abs(float64(t - timestamp)); d > 0 && d/float64(scrapeInterval.Milliseconds()) > 0.1 {
				// Too big jitter. Adjust timestamp
				timestamp = t
			}
			sw.scrapeAndLogError(timestamp, t)
		}
	}
}

func (sw *scrapeWork) logError(s string) {
	if !*suppressScrapeErrors {
		logger.ErrorfSkipframes(1, "error when scraping %q from job %q with labels %s: %s; "+
			"scrape errors can be disabled by -promscrape.suppressScrapeErrors command-line flag",
			sw.Config.ScrapeURL, sw.Config.Job(), sw.Config.Labels.String(), s)
	}
}

func (sw *scrapeWork) scrapeAndLogError(scrapeTimestamp, realTimestamp int64) {
	err := sw.scrapeInternal(scrapeTimestamp, realTimestamp)
	if *suppressScrapeErrors {
		return
	}
	if err == nil {
		sw.successRequestsCount++
		return
	}
	sw.failureRequestsCount++
	if sw.nextErrorLogTime == 0 {
		sw.nextErrorLogTime = realTimestamp + suppressScrapeErrorsDelay.Milliseconds()
	}
	if realTimestamp < sw.nextErrorLogTime {
		return
	}
	totalRequests := sw.failureRequestsCount + sw.successRequestsCount
	logger.Warnf("cannot scrape target %q (%s) %d out of %d times during -promscrape.suppressScrapeErrorsDelay=%s; the last error: %s",
		sw.Config.ScrapeURL, sw.Config.Labels.String(), sw.failureRequestsCount, totalRequests, *suppressScrapeErrorsDelay, err)
	sw.nextErrorLogTime = realTimestamp + suppressScrapeErrorsDelay.Milliseconds()
	sw.failureRequestsCount = 0
	sw.successRequestsCount = 0
}

var (
	scrapeDuration              = metrics.NewHistogram("vm_promscrape_scrape_duration_seconds")
	scrapeResponseSize          = metrics.NewHistogram("vm_promscrape_scrape_response_size_bytes")
	scrapedSamples              = metrics.NewHistogram("vm_promscrape_scraped_samples")
	scrapesSkippedBySampleLimit = metrics.NewCounter("vm_promscrape_scrapes_skipped_by_sample_limit_total")
	scrapesFailed               = metrics.NewCounter("vm_promscrape_scrapes_failed_total")
	pushDataDuration            = metrics.NewHistogram("vm_promscrape_push_data_duration_seconds")
)

func (sw *scrapeWork) mustSwitchToStreamParseMode(responseSize int) bool {
	if minResponseSizeForStreamParse.N <= 0 {
		return false
	}
	return sw.Config.canSwitchToStreamParseMode() && responseSize >= minResponseSizeForStreamParse.IntN()
}

// getTargetResponse() fetches response from sw target in the same way as when scraping the target.
func (sw *scrapeWork) getTargetResponse() ([]byte, error) {
	// use stream reader when stream mode enabled
	if *streamParse || sw.Config.StreamParse || sw.mustSwitchToStreamParseMode(sw.prevBodyLen) {
		// Read the response in stream mode.
		sr, err := sw.GetStreamReader()
		if err != nil {
			return nil, err
		}
		data, err := io.ReadAll(sr)
		sr.MustClose()
		return data, err
	}
	// Read the response in usual mode.
	return sw.ReadData(nil)
}

func (sw *scrapeWork) scrapeInternal(scrapeTimestamp, realTimestamp int64) error {
	if *streamParse || sw.Config.StreamParse || sw.mustSwitchToStreamParseMode(sw.prevBodyLen) {
		// Read data from scrape targets in streaming manner.
		// This case is optimized for targets exposing more than ten thousand of metrics per target.
		return sw.scrapeStream(scrapeTimestamp, realTimestamp)
	}

	// Common case: read all the data from scrape target to memory (body) and then process it.
	// This case should work more optimally than stream parse code for common case when scrape target exposes
	// up to a few thousand metrics.
	body := leveledbytebufferpool.Get(sw.prevBodyLen)
	var err error
	body.B, err = sw.ReadData(body.B[:0])
	releaseBody, err := sw.processScrapedData(scrapeTimestamp, realTimestamp, body, err)
	if releaseBody {
		leveledbytebufferpool.Put(body)
	}
	return err
}

var processScrapedDataConcurrencyLimitCh = make(chan struct{}, cgroup.AvailableCPUs())

func (sw *scrapeWork) processScrapedData(scrapeTimestamp, realTimestamp int64, body *bytesutil.ByteBuffer, err error) (bool, error) {
	// This function is CPU-bound, while it may allocate big amounts of memory.
	// That's why it is a good idea to limit the number of concurrent calls to this function
	// in order to limit memory usage under high load without sacrificing the performance.
	processScrapedDataConcurrencyLimitCh <- struct{}{}
	defer func() {
		<-processScrapedDataConcurrencyLimitCh
	}()
	endTimestamp := time.Now().UnixNano() / 1e6
	duration := float64(endTimestamp-realTimestamp) / 1e3
	scrapeDuration.Update(duration)
	scrapeResponseSize.Update(float64(len(body.B)))
	up := 1
	wc := writeRequestCtxPool.Get(sw.prevLabelsLen)
	lastScrape := sw.loadLastScrape()
	bodyString := bytesutil.ToUnsafeString(body.B)
	areIdenticalSeries := sw.areIdenticalSeries(lastScrape, bodyString)
	if err != nil {
		up = 0
		scrapesFailed.Inc()
	} else {
		wc.rows.UnmarshalWithErrLogger(bodyString, sw.logError)
	}
	srcRows := wc.rows.Rows
	samplesScraped := len(srcRows)
	scrapedSamples.Update(float64(samplesScraped))
	for i := range srcRows {
		sw.addRowToTimeseries(wc, &srcRows[i], scrapeTimestamp, true)
	}
	samplesPostRelabeling := len(wc.writeRequest.Timeseries)
	if sw.Config.SampleLimit > 0 && samplesPostRelabeling > sw.Config.SampleLimit {
		wc.resetNoRows()
		up = 0
		scrapesSkippedBySampleLimit.Inc()
		err = fmt.Errorf("the response from %q exceeds sample_limit=%d; "+
			"either reduce the sample count for the target or increase sample_limit", sw.Config.ScrapeURL, sw.Config.SampleLimit)
	}
	if up == 0 {
		bodyString = ""
	}
	seriesAdded := 0
	if !areIdenticalSeries {
		// The returned value for seriesAdded may be bigger than the real number of added series
		// if some series were removed during relabeling.
		// This is a trade-off between performance and accuracy.
		seriesAdded = sw.getSeriesAdded(lastScrape, bodyString)
	}
	samplesDropped := 0
	if sw.seriesLimitExceeded || !areIdenticalSeries {
		samplesDropped = sw.applySeriesLimit(wc)
	}
	am := &autoMetrics{
		up:                        up,
		scrapeDurationSeconds:     duration,
		samplesScraped:            samplesScraped,
		samplesPostRelabeling:     samplesPostRelabeling,
		seriesAdded:               seriesAdded,
		seriesLimitSamplesDropped: samplesDropped,
	}
	sw.addAutoMetrics(am, wc, scrapeTimestamp)
	sw.pushData(sw.Config.AuthToken, &wc.writeRequest)
	sw.prevLabelsLen = len(wc.labels)
	sw.prevBodyLen = len(bodyString)
	wc.reset()
	mustSwitchToStreamParse := sw.mustSwitchToStreamParseMode(len(bodyString))
	if !mustSwitchToStreamParse {
		// Return wc to the pool if the parsed response size was smaller than -promscrape.minResponseSizeForStreamParse
		// This should reduce memory usage when scraping targets with big responses.
		writeRequestCtxPool.Put(wc)
	}
	// body must be released only after wc is released, since wc refers to body.
	if !areIdenticalSeries {
		// Send stale markers for disappeared metrics with the real scrape timestamp
		// in order to guarantee that query doesn't return data after this time for the disappeared metrics.
		sw.sendStaleSeries(lastScrape, bodyString, realTimestamp, false)
		sw.storeLastScrape(body.B)
	}
	sw.finalizeLastScrape()
	tsmGlobal.Update(sw, up == 1, realTimestamp, int64(duration*1000), samplesScraped, err)
	return !mustSwitchToStreamParse, err
}

func (sw *scrapeWork) pushData(at *auth.Token, wr *prompbmarshal.WriteRequest) {
	startTime := time.Now()
	sw.PushData(at, wr)
	pushDataDuration.UpdateDuration(startTime)
}

type streamBodyReader struct {
	body       []byte
	bodyLen    int
	readOffset int
}

func (sbr *streamBodyReader) Init(sr *streamReader) error {
	sbr.body = nil
	sbr.bodyLen = 0
	sbr.readOffset = 0
	// Read the whole response body in memory before parsing it in stream mode.
	// This minimizes the time needed for reading response body from scrape target.
	startTime := fasttime.UnixTimestamp()
	body, err := io.ReadAll(sr)
	if err != nil {
		d := fasttime.UnixTimestamp() - startTime
		return fmt.Errorf("cannot read stream body in %d seconds: %w", d, err)
	}
	sbr.body = body
	sbr.bodyLen = len(body)
	return nil
}

func (sbr *streamBodyReader) Read(b []byte) (int, error) {
	if sbr.readOffset >= len(sbr.body) {
		return 0, io.EOF
	}
	n := copy(b, sbr.body[sbr.readOffset:])
	sbr.readOffset += n
	return n, nil
}

func (sw *scrapeWork) scrapeStream(scrapeTimestamp, realTimestamp int64) error {
	samplesScraped := 0
	samplesPostRelabeling := 0
	wc := writeRequestCtxPool.Get(sw.prevLabelsLen)
	// Do not pool sbr and do not pre-allocate sbr.body in order to reduce memory usage when scraping big responses.
	var sbr streamBodyReader

	lastScrape := sw.loadLastScrape()
	bodyString := ""
	areIdenticalSeries := true
	samplesDropped := 0
	sr, err := sw.GetStreamReader()
	if err != nil {
		err = fmt.Errorf("cannot read data: %w", err)
	} else {
		var mu sync.Mutex
		err = sbr.Init(sr)
		if err == nil {
			bodyString = bytesutil.ToUnsafeString(sbr.body)
			areIdenticalSeries = sw.areIdenticalSeries(lastScrape, bodyString)
			err = stream.Parse(&sbr, scrapeTimestamp, false, false, func(rows []parser.Row) error {
				mu.Lock()
				defer mu.Unlock()
				samplesScraped += len(rows)
				for i := range rows {
					sw.addRowToTimeseries(wc, &rows[i], scrapeTimestamp, true)
				}
				samplesPostRelabeling += len(wc.writeRequest.Timeseries)
				if sw.Config.SampleLimit > 0 && samplesPostRelabeling > sw.Config.SampleLimit {
					wc.resetNoRows()
					scrapesSkippedBySampleLimit.Inc()
					return fmt.Errorf("the response from %q exceeds sample_limit=%d; "+
						"either reduce the sample count for the target or increase sample_limit", sw.Config.ScrapeURL, sw.Config.SampleLimit)
				}
				if sw.seriesLimitExceeded || !areIdenticalSeries {
					samplesDropped += sw.applySeriesLimit(wc)
				}
				// Push the collected rows to sw before returning from the callback, since they cannot be held
				// after returning from the callback - this will result in data race.
				// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/825#issuecomment-723198247
				sw.pushData(sw.Config.AuthToken, &wc.writeRequest)
				wc.resetNoRows()
				return nil
			}, sw.logError)
		}
		sr.MustClose()
	}

	scrapedSamples.Update(float64(samplesScraped))
	endTimestamp := time.Now().UnixNano() / 1e6
	duration := float64(endTimestamp-realTimestamp) / 1e3
	scrapeDuration.Update(duration)
	scrapeResponseSize.Update(float64(sbr.bodyLen))
	up := 1
	if err != nil {
		// Mark the scrape as failed even if it already read and pushed some samples
		// to remote storage. This makes the logic compatible with Prometheus.
		up = 0
		scrapesFailed.Inc()
	}
	seriesAdded := 0
	if !areIdenticalSeries {
		// The returned value for seriesAdded may be bigger than the real number of added series
		// if some series were removed during relabeling.
		// This is a trade-off between performance and accuracy.
		seriesAdded = sw.getSeriesAdded(lastScrape, bodyString)
	}
	am := &autoMetrics{
		up:                        up,
		scrapeDurationSeconds:     duration,
		samplesScraped:            samplesScraped,
		samplesPostRelabeling:     samplesPostRelabeling,
		seriesAdded:               seriesAdded,
		seriesLimitSamplesDropped: samplesDropped,
	}
	sw.addAutoMetrics(am, wc, scrapeTimestamp)
	sw.pushData(sw.Config.AuthToken, &wc.writeRequest)
	sw.prevLabelsLen = len(wc.labels)
	sw.prevBodyLen = sbr.bodyLen
	wc.reset()
	writeRequestCtxPool.Put(wc)
	if !areIdenticalSeries {
		// Send stale markers for disappeared metrics with the real scrape timestamp
		// in order to guarantee that query doesn't return data after this time for the disappeared metrics.
		sw.sendStaleSeries(lastScrape, bodyString, realTimestamp, false)
		sw.storeLastScrape(sbr.body)
	}
	sw.finalizeLastScrape()
	tsmGlobal.Update(sw, up == 1, realTimestamp, int64(duration*1000), samplesScraped, err)
	// Do not track active series in streaming mode, since this may need too big amounts of memory
	// when the target exports too big number of metrics.
	return err
}

func (sw *scrapeWork) areIdenticalSeries(prevData, currData string) bool {
	if sw.Config.NoStaleMarkers && sw.Config.SeriesLimit <= 0 {
		// Do not spend CPU time on tracking the changes in series if stale markers are disabled.
		// The check for series_limit is needed for https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3660
		return true
	}
	return parser.AreIdenticalSeriesFast(prevData, currData)
}

// leveledWriteRequestCtxPool allows reducing memory usage when writeRequesCtx
// structs contain mixed number of labels.
//
// Its logic has been copied from leveledbytebufferpool.
type leveledWriteRequestCtxPool struct {
	pools [13]sync.Pool
}

func (lwp *leveledWriteRequestCtxPool) Get(labelsCapacity int) *writeRequestCtx {
	id, capacityNeeded := lwp.getPoolIDAndCapacity(labelsCapacity)
	for i := 0; i < 2; i++ {
		if id < 0 || id >= len(lwp.pools) {
			break
		}
		if v := lwp.pools[id].Get(); v != nil {
			return v.(*writeRequestCtx)
		}
		id++
	}
	return &writeRequestCtx{
		labels: make([]prompbmarshal.Label, 0, capacityNeeded),
	}
}

func (lwp *leveledWriteRequestCtxPool) Put(wc *writeRequestCtx) {
	capacity := cap(wc.labels)
	id, poolCapacity := lwp.getPoolIDAndCapacity(capacity)
	if capacity <= poolCapacity {
		wc.reset()
		lwp.pools[id].Put(wc)
	}
}

func (lwp *leveledWriteRequestCtxPool) getPoolIDAndCapacity(size int) (int, int) {
	size--
	if size < 0 {
		size = 0
	}
	size >>= 3
	id := bits.Len(uint(size))
	if id >= len(lwp.pools) {
		id = len(lwp.pools) - 1
	}
	return id, (1 << (id + 3))
}

type writeRequestCtx struct {
	rows         parser.Rows
	writeRequest prompbmarshal.WriteRequest
	labels       []prompbmarshal.Label
	samples      []prompbmarshal.Sample
}

func (wc *writeRequestCtx) reset() {
	wc.rows.Reset()
	wc.resetNoRows()
}

func (wc *writeRequestCtx) resetNoRows() {
	prompbmarshal.ResetWriteRequest(&wc.writeRequest)

	labels := wc.labels
	for i := range labels {
		label := &labels[i]
		label.Name = ""
		label.Value = ""
	}
	wc.labels = wc.labels[:0]

	wc.samples = wc.samples[:0]
}

var writeRequestCtxPool leveledWriteRequestCtxPool

func (sw *scrapeWork) getSeriesAdded(lastScrape, currScrape string) int {
	if currScrape == "" {
		return 0
	}
	bodyString := parser.GetRowsDiff(currScrape, lastScrape)
	return strings.Count(bodyString, "\n")
}

func (sw *scrapeWork) applySeriesLimit(wc *writeRequestCtx) int {
	if sw.Config.SeriesLimit <= 0 {
		return 0
	}
	if sw.seriesLimiter == nil {
		sw.seriesLimiter = bloomfilter.NewLimiter(sw.Config.SeriesLimit, 24*time.Hour)
	}
	sl := sw.seriesLimiter
	dstSeries := wc.writeRequest.Timeseries[:0]
	samplesDropped := 0
	for _, ts := range wc.writeRequest.Timeseries {
		h := sw.getLabelsHash(ts.Labels)
		if !sl.Add(h) {
			samplesDropped++
			continue
		}
		dstSeries = append(dstSeries, ts)
	}
	prompbmarshal.ResetTimeSeries(wc.writeRequest.Timeseries[len(dstSeries):])
	wc.writeRequest.Timeseries = dstSeries
	if samplesDropped > 0 && !sw.seriesLimitExceeded {
		sw.seriesLimitExceeded = true
	}
	return samplesDropped
}

var sendStaleSeriesConcurrencyLimitCh = make(chan struct{}, cgroup.AvailableCPUs())

func (sw *scrapeWork) sendStaleSeries(lastScrape, currScrape string, timestamp int64, addAutoSeries bool) {
	// This function is CPU-bound, while it may allocate big amounts of memory.
	// That's why it is a good idea to limit the number of concurrent calls to this function
	// in order to limit memory usage under high load without sacrificing the performance.
	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3668
	sendStaleSeriesConcurrencyLimitCh <- struct{}{}
	defer func() {
		<-sendStaleSeriesConcurrencyLimitCh
	}()
	if sw.Config.NoStaleMarkers {
		return
	}
	bodyString := lastScrape
	if currScrape != "" {
		bodyString = parser.GetRowsDiff(lastScrape, currScrape)
	}
	wc := writeRequestCtxPool.Get(sw.prevLabelsLen)
	defer func() {
		wc.reset()
		writeRequestCtxPool.Put(wc)
	}()
	if bodyString != "" {
		// Send stale markers in streaming mode in order to reduce memory usage
		// when stale markers for targets exposing big number of metrics must be generated.
		// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3668
		// and https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3675
		var mu sync.Mutex
		br := bytes.NewBufferString(bodyString)
		err := stream.Parse(br, timestamp, false, false, func(rows []parser.Row) error {
			mu.Lock()
			defer mu.Unlock()
			for i := range rows {
				sw.addRowToTimeseries(wc, &rows[i], timestamp, true)
			}
			// Apply series limit to stale markers in order to prevent sending stale markers for newly created series.
			// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3660
			if sw.seriesLimitExceeded {
				sw.applySeriesLimit(wc)
			}
			// Push the collected rows to sw before returning from the callback, since they cannot be held
			// after returning from the callback - this will result in data race.
			// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/825#issuecomment-723198247
			setStaleMarkersForRows(wc.writeRequest.Timeseries)
			sw.pushData(sw.Config.AuthToken, &wc.writeRequest)
			wc.resetNoRows()
			return nil
		}, sw.logError)
		if err != nil {
			sw.logError(fmt.Errorf("cannot send stale markers: %w", err).Error())
		}
	}
	if addAutoSeries {
		am := &autoMetrics{}
		sw.addAutoMetrics(am, wc, timestamp)
	}
	setStaleMarkersForRows(wc.writeRequest.Timeseries)
	sw.pushData(sw.Config.AuthToken, &wc.writeRequest)
}

func setStaleMarkersForRows(series []prompbmarshal.TimeSeries) {
	for _, tss := range series {
		samples := tss.Samples
		for i := range samples {
			samples[i].Value = decimal.StaleNaN
		}
		staleSamplesCreated.Add(len(samples))
	}
}

var staleSamplesCreated = metrics.NewCounter(`vm_promscrape_stale_samples_created_total`)

func (sw *scrapeWork) getLabelsHash(labels []prompbmarshal.Label) uint64 {
	// It is OK if there will be hash collisions for distinct sets of labels,
	// since the accuracy for `scrape_series_added` metric may be lower than 100%.
	b := sw.labelsHashBuf[:0]
	for _, label := range labels {
		b = append(b, label.Name...)
		b = append(b, label.Value...)
	}
	sw.labelsHashBuf = b
	return xxhash.Sum64(b)
}

type autoMetrics struct {
	up                        int
	scrapeDurationSeconds     float64
	samplesScraped            int
	samplesPostRelabeling     int
	seriesAdded               int
	seriesLimitSamplesDropped int
}

func isAutoMetric(s string) bool {
	switch s {
	case "up", "scrape_duration_seconds", "scrape_samples_scraped",
		"scrape_samples_post_metric_relabeling", "scrape_series_added",
		"scrape_timeout_seconds", "scrape_samples_limit",
		"scrape_series_limit_samples_dropped", "scrape_series_limit",
		"scrape_series_current":
		return true
	}
	return false
}

func (sw *scrapeWork) addAutoMetrics(am *autoMetrics, wc *writeRequestCtx, timestamp int64) {
	sw.addAutoTimeseries(wc, "up", float64(am.up), timestamp)
	sw.addAutoTimeseries(wc, "scrape_duration_seconds", am.scrapeDurationSeconds, timestamp)
	sw.addAutoTimeseries(wc, "scrape_samples_scraped", float64(am.samplesScraped), timestamp)
	sw.addAutoTimeseries(wc, "scrape_samples_post_metric_relabeling", float64(am.samplesPostRelabeling), timestamp)
	sw.addAutoTimeseries(wc, "scrape_series_added", float64(am.seriesAdded), timestamp)
	sw.addAutoTimeseries(wc, "scrape_timeout_seconds", sw.Config.ScrapeTimeout.Seconds(), timestamp)
	if sampleLimit := sw.Config.SampleLimit; sampleLimit > 0 {
		// Expose scrape_samples_limit metric if sample_limt config is set for the target.
		// See https://github.com/VictoriaMetrics/operator/issues/497
		sw.addAutoTimeseries(wc, "scrape_samples_limit", float64(sampleLimit), timestamp)
	}
	if sl := sw.seriesLimiter; sl != nil {
		sw.addAutoTimeseries(wc, "scrape_series_limit_samples_dropped", float64(am.seriesLimitSamplesDropped), timestamp)
		sw.addAutoTimeseries(wc, "scrape_series_limit", float64(sl.MaxItems()), timestamp)
		sw.addAutoTimeseries(wc, "scrape_series_current", float64(sl.CurrentItems()), timestamp)
	}
}

// addAutoTimeseries adds automatically generated time series with the given name, value and timestamp.
//
// See https://prometheus.io/docs/concepts/jobs_instances/#automatically-generated-labels-and-time-series
func (sw *scrapeWork) addAutoTimeseries(wc *writeRequestCtx, name string, value float64, timestamp int64) {
	sw.tmpRow.Metric = name
	sw.tmpRow.Tags = nil
	sw.tmpRow.Value = value
	sw.tmpRow.Timestamp = timestamp
	sw.addRowToTimeseries(wc, &sw.tmpRow, timestamp, false)
}

func (sw *scrapeWork) addRowToTimeseries(wc *writeRequestCtx, r *parser.Row, timestamp int64, needRelabel bool) {
	metric := r.Metric

	// Add `exported_` prefix to metrics, which clash with the automatically generated
	// metric names only if the following conditions are met:
	//
	// - The `honor_labels` option isn't set to true in the scrape_config.
	//   If `honor_labels: true`, then the scraped metric name must remain unchanged
	//   because the user explicitly asked about it in the config.
	// - The metric has no labels (tags). If it has labels, then the metric value
	//   will be written into a separate time series comparing to automatically generated time series.
	//
	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3557
	// and https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3406
	if needRelabel && !sw.Config.HonorLabels && len(r.Tags) == 0 && isAutoMetric(metric) {
		bb := bbPool.Get()
		bb.B = append(bb.B, "exported_"...)
		bb.B = append(bb.B, metric...)
		metric = bytesutil.InternBytes(bb.B)
		bbPool.Put(bb)
	}
	labelsLen := len(wc.labels)
	targetLabels := sw.Config.Labels.GetLabels()
	wc.labels = appendLabels(wc.labels, metric, r.Tags, targetLabels, sw.Config.HonorLabels)
	if needRelabel {
		wc.labels = sw.Config.MetricRelabelConfigs.Apply(wc.labels, labelsLen)
	}
	wc.labels = promrelabel.FinalizeLabels(wc.labels[:labelsLen], wc.labels[labelsLen:])
	if len(wc.labels) == labelsLen {
		// Skip row without labels.
		return
	}
	// Add labels from `global->external_labels` section after the relabeling like Prometheus does.
	// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3137
	externalLabels := sw.Config.ExternalLabels.GetLabels()
	wc.labels = appendExtraLabels(wc.labels, externalLabels, labelsLen, sw.Config.HonorLabels)
	sampleTimestamp := r.Timestamp
	if !sw.Config.HonorTimestamps || sampleTimestamp == 0 {
		sampleTimestamp = timestamp
	}
	wc.samples = append(wc.samples, prompbmarshal.Sample{
		Value:     r.Value,
		Timestamp: sampleTimestamp,
	})
	wr := &wc.writeRequest
	wr.Timeseries = append(wr.Timeseries, prompbmarshal.TimeSeries{
		Labels:  wc.labels[labelsLen:],
		Samples: wc.samples[len(wc.samples)-1:],
	})
}

var bbPool bytesutil.ByteBufferPool

func appendLabels(dst []prompbmarshal.Label, metric string, src []parser.Tag, extraLabels []prompbmarshal.Label, honorLabels bool) []prompbmarshal.Label {
	dstLen := len(dst)
	dst = append(dst, prompbmarshal.Label{
		Name:  "__name__",
		Value: metric,
	})
	for i := range src {
		tag := &src[i]
		dst = append(dst, prompbmarshal.Label{
			Name:  tag.Key,
			Value: tag.Value,
		})
	}
	return appendExtraLabels(dst, extraLabels, dstLen, honorLabels)
}

func appendExtraLabels(dst, extraLabels []prompbmarshal.Label, offset int, honorLabels bool) []prompbmarshal.Label {
	// Add extraLabels to labels.
	// Handle duplicates in the same way as Prometheus does.
	if len(dst) == offset {
		// Fast path - add extraLabels to dst without the need to de-duplicate.
		dst = append(dst, extraLabels...)
		return dst
	}
	offsetEnd := len(dst)
	for _, label := range extraLabels {
		labels := dst[offset:offsetEnd]
		prevLabel := promrelabel.GetLabelByName(labels, label.Name)
		if prevLabel == nil {
			// Fast path - the label doesn't exist in labels, so just add it to dst.
			dst = append(dst, label)
			continue
		}
		if honorLabels {
			// Skip the extra label with the same name.
			continue
		}
		// Rename the prevLabel to "exported_" + label.Name
		// See https://prometheus.io/docs/prometheus/latest/configuration/configuration/#scrape_config
		exportedName := "exported_" + label.Name
		exportedLabel := promrelabel.GetLabelByName(labels, exportedName)
		if exportedLabel != nil {
			// The label with the name exported_<label.Name> already exists.
			// Add yet another 'exported_' prefix to it.
			exportedLabel.Name = "exported_" + exportedName
		}
		prevLabel.Name = exportedName
		dst = append(dst, label)
	}
	return dst
}