lib/promscrape: add -promscrape.minResponseSizeForStreamParse command-line option for automatic switching to stream parsing mode when scraping targets with big responses

This should reduce memory usage when vmagent scrapes targets with non-uniform response sizes.
This is common case in Kubernetes monitoring.

README.md

@@ -1540,6 +1540,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
     	The maximum number of CPU cores to use for big merges. Default value is used if set to 0
   -csvTrimTimestamp duration
     	Trim timestamps when importing csv data to this duration. Minimum practical duration is 1ms. Higher duration (i.e. 1s) may be used for reducing disk space usage for timestamp data (default 1ms)
+  -datadog.maxInsertRequestSize size
+    	The maximum size in bytes of a single DataDog POST request to /api/v1/series
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 67108864)
   -dedup.minScrapeInterval duration
     	Leave only the first sample in every time series per each discrete interval equal to -dedup.minScrapeInterval > 0. See https://docs.victoriametrics.com/#deduplication for details
   -deleteAuthKey string
@@ -1705,6 +1708,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
   -promscrape.maxScrapeSize size
     	The maximum size of scrape response in bytes to process from Prometheus targets. Bigger responses are rejected
     	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 16777216)
+  -promscrape.minResponseSizeForStreamParse size
+    	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
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 1000000)
   -promscrape.noStaleMarkers
     	Whether to disable sending Prometheus stale markers for metrics when scrape target disappears. This option may reduce memory usage if stale markers aren't needed for your setup. See also https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode
   -promscrape.openstackSDCheckInterval duration
@@ -1718,7 +1724,7 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
   -promscrape.suppressScrapeErrors
     	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
   -relabelConfig string
-    	Optional path to a file with relabeling rules, which are applied to all the ingested metrics. See https://docs.victoriametrics.com/#relabeling for details
+    	Optional path to a file with relabeling rules, which are applied to all the ingested metrics. See https://docs.victoriametrics.com/#relabeling for details. The config is reloaded on SIGHUP signal
   -relabelDebug
     	Whether to log metrics before and after relabeling with -relabelConfig. If the -relabelDebug is enabled, then the metrics aren't sent to storage. This is useful for debugging the relabeling configs
   -retentionPeriod value
@@ -1795,6 +1801,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
     	The maximum number of unique series can be added to the storage during the last 24 hours. Excess series are logged and dropped. This can be useful for limiting series churn rate. See also -storage.maxHourlySeries
   -storage.maxHourlySeries int
     	The maximum number of unique series can be added to the storage during the last hour. Excess series are logged and dropped. This can be useful for limiting series cardinality. See also -storage.maxDailySeries
+  -storage.minFreeDiskSpaceBytes size
+    	The minimum free disk space at -storageDataPath after which the storage stops accepting new data
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 10000000)
   -storageDataPath string
     	Path to storage data (default "victoria-metrics-data")
   -tls

app/vmagent/README.md

@@ -307,7 +307,9 @@ Prometheus staleness markers aren't sent to `-remoteWrite.url` in [stream parsin
 
 ## Stream parsing mode
 
-By default `vmagent` reads the full response from scrape target into memory, then parses it, applies [relabeling](#relabeling) and then pushes the resulting metrics to the configured `-remoteWrite.url`. This mode works good for the majority of cases when the scrape target exposes small number of metrics (e.g. less than 10 thousand). But this mode may take big amounts of memory when the scrape target exposes big number of metrics. In this case it is recommended enabling stream parsing mode. When this mode is enabled, then `vmagent` reads response from scrape target in chunks, then immediately processes every chunk and pushes the processed metrics to remote storage. This allows saving memory when scraping targets that expose millions of metrics. Stream parsing mode may be enabled in the following places:
+By default `vmagent` reads the full response body from scrape target into memory, then parses it, applies [relabeling](#relabeling) and then pushes the resulting metrics to the configured `-remoteWrite.url`. This mode works good for the majority of cases when the scrape target exposes small number of metrics (e.g. less than 10 thousand). But this mode may take big amounts of memory when the scrape target exposes big number of metrics. In this case it is recommended enabling stream parsing mode. When this mode is enabled, then `vmagent` reads response from scrape target in chunks, then immediately processes every chunk and pushes the processed metrics to remote storage. This allows saving memory when scraping targets that expose millions of metrics.
+
+Stream parsing mode is automatically enabled for scrape targets returning response bodies with sizes bigger than the `-promscrape.minResponseSizeForStreamParse` command-line flag value. Additionally, the stream parsing mode can be explicitly enabled in the following places:
 
 - Via `-promscrape.streamParse` command-line flag. In this case all the scrape targets defined in the file pointed by `-promscrape.config` are scraped in stream parsing mode.
 - Via `stream_parse: true` option at `scrape_configs` section. In this case all the scrape targets defined in this section are scraped in stream parsing mode.
@@ -706,6 +708,9 @@ See the docs at https://docs.victoriametrics.com/vmagent.html .
 
   -csvTrimTimestamp duration
     	Trim timestamps when importing csv data to this duration. Minimum practical duration is 1ms. Higher duration (i.e. 1s) may be used for reducing disk space usage for timestamp data (default 1ms)
+  -datadog.maxInsertRequestSize size
+    	The maximum size in bytes of a single DataDog POST request to /api/v1/series
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 67108864)
   -dryRun
     	Whether to check only config files without running vmagent. The following files are checked: -promscrape.config, -remoteWrite.relabelConfig, -remoteWrite.urlRelabelConfig . Unknown config entries are allowed in -promscrape.config by default. This can be changed with -promscrape.config.strictParse
   -enableTCP6
@@ -853,6 +858,9 @@ See the docs at https://docs.victoriametrics.com/vmagent.html .
   -promscrape.maxScrapeSize size
     	The maximum size of scrape response in bytes to process from Prometheus targets. Bigger responses are rejected
     	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 16777216)
+  -promscrape.minResponseSizeForStreamParse size
+    	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
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 1000000)
   -promscrape.noStaleMarkers
     	Whether to disable sending Prometheus stale markers for metrics when scrape target disappears. This option may reduce memory usage if stale markers aren't needed for your setup. See also https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode
   -promscrape.openstackSDCheckInterval duration

docs/CHANGELOG.md

@@ -9,6 +9,7 @@ sort: 15
 * FEATURE: vmagent: expose `-promscrape.config` contents at `/config` page as Prometheus does. See [this feature request](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1695).
 * FEATURE: vmagent: add `show original labels` button per each scrape target displayed at `http://vmagent;8429/targets` page. This should improve debuggability for service discovery and relabeling issues similar to [this one](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1664). See [this feature request](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1698).
 * FEATURE: vmagent: shard targets among cluster nodes after the relabeling is applied. This should guarantee that targets with the same set of labels go to the same `vmagent` node in the cluster. See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1687).
+* FEATURE: vmagent: add `-promscrape.minResponseSizeForStreamParse` command-line flag, which can be used for instructing `vmagent` to automatically switch to [stream parsing mode](https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode) for scrape targets with big responses. This should reduce memory usage when `vmagent` scrapes targets with non-uniform response sizes (this is the case in Kubernetes monitoring).
 * FEATURE: vmagent: return error if `sample_limit` or `series_limit` options are set when [stream parsing mode](https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode) is enabled, since these limits cannot be applied in stream parsing mode.
 * FEATURE: add trigonometric functions, which are going to be added in [Prometheus 2.31](https://github.com/prometheus/prometheus/pull/9239): [acosh](https://docs.victoriametrics.com/MetricsQL.html#acosh), [asinh](https://docs.victoriametrics.com/MetricsQL.html#asinh), [atan](https://docs.victoriametrics.com/MetricsQL.html#atan), [atanh](https://docs.victoriametrics.com/MetricsQL.html#atanh), [cosh](https://docs.victoriametrics.com/MetricsQL.html#cosh), [deg](https://docs.victoriametrics.com/MetricsQL.html#deg), [rad](https://docs.victoriametrics.com/MetricsQL.html#rad), [sinh](https://docs.victoriametrics.com/MetricsQL.html#sinh), [tan](https://docs.victoriametrics.com/MetricsQL.html#tan), [tanh](https://docs.victoriametrics.com/MetricsQL.html#tanh). Also add `atan2` binary operator. See [this pull request](https://github.com/prometheus/prometheus/pull/9248).
 * FEATURE: consistently return the same set of time series from [limitk](https://docs.victoriametrics.com/MetricsQL.html#limitk) function. This improves the usability of periodically refreshed graphs.

docs/README.md

@@ -1540,6 +1540,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
     	The maximum number of CPU cores to use for big merges. Default value is used if set to 0
   -csvTrimTimestamp duration
     	Trim timestamps when importing csv data to this duration. Minimum practical duration is 1ms. Higher duration (i.e. 1s) may be used for reducing disk space usage for timestamp data (default 1ms)
+  -datadog.maxInsertRequestSize size
+    	The maximum size in bytes of a single DataDog POST request to /api/v1/series
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 67108864)
   -dedup.minScrapeInterval duration
     	Leave only the first sample in every time series per each discrete interval equal to -dedup.minScrapeInterval > 0. See https://docs.victoriametrics.com/#deduplication for details
   -deleteAuthKey string
@@ -1705,6 +1708,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
   -promscrape.maxScrapeSize size
     	The maximum size of scrape response in bytes to process from Prometheus targets. Bigger responses are rejected
     	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 16777216)
+  -promscrape.minResponseSizeForStreamParse size
+    	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
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 1000000)
   -promscrape.noStaleMarkers
     	Whether to disable sending Prometheus stale markers for metrics when scrape target disappears. This option may reduce memory usage if stale markers aren't needed for your setup. See also https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode
   -promscrape.openstackSDCheckInterval duration
@@ -1718,7 +1724,7 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
   -promscrape.suppressScrapeErrors
     	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
   -relabelConfig string
-    	Optional path to a file with relabeling rules, which are applied to all the ingested metrics. See https://docs.victoriametrics.com/#relabeling for details
+    	Optional path to a file with relabeling rules, which are applied to all the ingested metrics. See https://docs.victoriametrics.com/#relabeling for details. The config is reloaded on SIGHUP signal
   -relabelDebug
     	Whether to log metrics before and after relabeling with -relabelConfig. If the -relabelDebug is enabled, then the metrics aren't sent to storage. This is useful for debugging the relabeling configs
   -retentionPeriod value
@@ -1795,6 +1801,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
     	The maximum number of unique series can be added to the storage during the last 24 hours. Excess series are logged and dropped. This can be useful for limiting series churn rate. See also -storage.maxHourlySeries
   -storage.maxHourlySeries int
     	The maximum number of unique series can be added to the storage during the last hour. Excess series are logged and dropped. This can be useful for limiting series cardinality. See also -storage.maxDailySeries
+  -storage.minFreeDiskSpaceBytes size
+    	The minimum free disk space at -storageDataPath after which the storage stops accepting new data
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 10000000)
   -storageDataPath string
     	Path to storage data (default "victoria-metrics-data")
   -tls

docs/Single-server-VictoriaMetrics.md

@@ -1544,6 +1544,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
     	The maximum number of CPU cores to use for big merges. Default value is used if set to 0
   -csvTrimTimestamp duration
     	Trim timestamps when importing csv data to this duration. Minimum practical duration is 1ms. Higher duration (i.e. 1s) may be used for reducing disk space usage for timestamp data (default 1ms)
+  -datadog.maxInsertRequestSize size
+    	The maximum size in bytes of a single DataDog POST request to /api/v1/series
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 67108864)
   -dedup.minScrapeInterval duration
     	Leave only the first sample in every time series per each discrete interval equal to -dedup.minScrapeInterval > 0. See https://docs.victoriametrics.com/#deduplication for details
   -deleteAuthKey string
@@ -1709,6 +1712,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
   -promscrape.maxScrapeSize size
     	The maximum size of scrape response in bytes to process from Prometheus targets. Bigger responses are rejected
     	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 16777216)
+  -promscrape.minResponseSizeForStreamParse size
+    	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
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 1000000)
   -promscrape.noStaleMarkers
     	Whether to disable sending Prometheus stale markers for metrics when scrape target disappears. This option may reduce memory usage if stale markers aren't needed for your setup. See also https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode
   -promscrape.openstackSDCheckInterval duration
@@ -1722,7 +1728,7 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
   -promscrape.suppressScrapeErrors
     	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
   -relabelConfig string
-    	Optional path to a file with relabeling rules, which are applied to all the ingested metrics. See https://docs.victoriametrics.com/#relabeling for details
+    	Optional path to a file with relabeling rules, which are applied to all the ingested metrics. See https://docs.victoriametrics.com/#relabeling for details. The config is reloaded on SIGHUP signal
   -relabelDebug
     	Whether to log metrics before and after relabeling with -relabelConfig. If the -relabelDebug is enabled, then the metrics aren't sent to storage. This is useful for debugging the relabeling configs
   -retentionPeriod value
@@ -1799,6 +1805,9 @@ Pass `-help` to VictoriaMetrics in order to see the list of supported command-li
     	The maximum number of unique series can be added to the storage during the last 24 hours. Excess series are logged and dropped. This can be useful for limiting series churn rate. See also -storage.maxHourlySeries
   -storage.maxHourlySeries int
     	The maximum number of unique series can be added to the storage during the last hour. Excess series are logged and dropped. This can be useful for limiting series cardinality. See also -storage.maxDailySeries
+  -storage.minFreeDiskSpaceBytes size
+    	The minimum free disk space at -storageDataPath after which the storage stops accepting new data
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 10000000)
   -storageDataPath string
     	Path to storage data (default "victoria-metrics-data")
   -tls

docs/vmagent.md

@@ -311,7 +311,9 @@ Prometheus staleness markers aren't sent to `-remoteWrite.url` in [stream parsin
 
 ## Stream parsing mode
 
-By default `vmagent` reads the full response from scrape target into memory, then parses it, applies [relabeling](#relabeling) and then pushes the resulting metrics to the configured `-remoteWrite.url`. This mode works good for the majority of cases when the scrape target exposes small number of metrics (e.g. less than 10 thousand). But this mode may take big amounts of memory when the scrape target exposes big number of metrics. In this case it is recommended enabling stream parsing mode. When this mode is enabled, then `vmagent` reads response from scrape target in chunks, then immediately processes every chunk and pushes the processed metrics to remote storage. This allows saving memory when scraping targets that expose millions of metrics. Stream parsing mode may be enabled in the following places:
+By default `vmagent` reads the full response body from scrape target into memory, then parses it, applies [relabeling](#relabeling) and then pushes the resulting metrics to the configured `-remoteWrite.url`. This mode works good for the majority of cases when the scrape target exposes small number of metrics (e.g. less than 10 thousand). But this mode may take big amounts of memory when the scrape target exposes big number of metrics. In this case it is recommended enabling stream parsing mode. When this mode is enabled, then `vmagent` reads response from scrape target in chunks, then immediately processes every chunk and pushes the processed metrics to remote storage. This allows saving memory when scraping targets that expose millions of metrics.
+
+Stream parsing mode is automatically enabled for scrape targets returning response bodies with sizes bigger than the `-promscrape.minResponseSizeForStreamParse` command-line flag value. Additionally, the stream parsing mode can be explicitly enabled in the following places:
 
 - Via `-promscrape.streamParse` command-line flag. In this case all the scrape targets defined in the file pointed by `-promscrape.config` are scraped in stream parsing mode.
 - Via `stream_parse: true` option at `scrape_configs` section. In this case all the scrape targets defined in this section are scraped in stream parsing mode.
@@ -710,6 +712,9 @@ See the docs at https://docs.victoriametrics.com/vmagent.html .
 
   -csvTrimTimestamp duration
     	Trim timestamps when importing csv data to this duration. Minimum practical duration is 1ms. Higher duration (i.e. 1s) may be used for reducing disk space usage for timestamp data (default 1ms)
+  -datadog.maxInsertRequestSize size
+    	The maximum size in bytes of a single DataDog POST request to /api/v1/series
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 67108864)
   -dryRun
     	Whether to check only config files without running vmagent. The following files are checked: -promscrape.config, -remoteWrite.relabelConfig, -remoteWrite.urlRelabelConfig . Unknown config entries are allowed in -promscrape.config by default. This can be changed with -promscrape.config.strictParse
   -enableTCP6
@@ -857,6 +862,9 @@ See the docs at https://docs.victoriametrics.com/vmagent.html .
   -promscrape.maxScrapeSize size
     	The maximum size of scrape response in bytes to process from Prometheus targets. Bigger responses are rejected
     	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 16777216)
+  -promscrape.minResponseSizeForStreamParse size
+    	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
+    	Supports the following optional suffixes for size values: KB, MB, GB, KiB, MiB, GiB (default 1000000)
   -promscrape.noStaleMarkers
     	Whether to disable sending Prometheus stale markers for metrics when scrape target disappears. This option may reduce memory usage if stale markers aren't needed for your setup. See also https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode
   -promscrape.openstackSDCheckInterval duration

lib/promscrape/scrapework.go

@@ -13,6 +13,7 @@ import (
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bloomfilter"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/flagutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/leveledbytebufferpool"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/promauth"
@@ -30,6 +31,7 @@ var (
 		"The last error for each target is always available at '/targets' page even if scrape errors logging is suppressed")
 	noStaleMarkers                = flag.Bool("promscrape.noStaleMarkers", false, "Whether to disable sending Prometheus stale markers for metrics when scrape target disappears. This option may reduce memory usage if stale markers aren't needed for your setup. See also https://docs.victoriametrics.com/vmagent.html#stream-parsing-mode")
 	seriesLimitPerTarget          = flag.Int("promscrape.seriesLimitPerTarget", 0, "Optional limit on the number of unique time series a single scrape target can expose. See https://docs.victoriametrics.com/vmagent.html#cardinality-limiter for more info")
+	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.
@@ -115,6 +117,11 @@ type ScrapeWork struct {
 	jobNameOriginal string
 }
 
+func (sw *ScrapeWork) canUseStreamParse() bool {
+	// Stream parsing mode cannot be used if SampleLimit or SeriesLimit is set.
+	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.
@@ -291,9 +298,9 @@ var (
 )
 
 func (sw *scrapeWork) scrapeInternal(scrapeTimestamp, realTimestamp int64) error {
-	if *streamParse || sw.Config.StreamParse {
+	if (sw.Config.canUseStreamParse() && sw.prevBodyLen >= minResponseSizeForStreamParse.N) || *streamParse || sw.Config.StreamParse {
 		// Read data from scrape targets in streaming manner.
-		// This case is optimized for targets exposing millions and more of metrics per target.
+		// This case is optimized for targets exposing more than ten thousand of metrics per target.
 		return sw.scrapeStream(scrapeTimestamp, realTimestamp)
 	}
 
@@ -355,15 +362,23 @@ func (sw *scrapeWork) scrapeInternal(scrapeTimestamp, realTimestamp int64) error
 	sw.addAutoTimeseries(wc, "scrape_timeout_seconds", sw.Config.ScrapeTimeout.Seconds(), scrapeTimestamp)
 	sw.pushData(&wc.writeRequest)
 	sw.prevLabelsLen = len(wc.labels)
-	wc.reset()
-	writeRequestCtxPool.Put(wc)
-	// body must be released only after wc is released, since wc refers to body.
 	sw.prevBodyLen = len(body.B)
+	wc.reset()
+	if len(body.B) < minResponseSizeForStreamParse.N {
+		// 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 {
 		sw.sendStaleSeries(bodyString, scrapeTimestamp, false)
 	}
+	if len(body.B) < minResponseSizeForStreamParse.N {
+		// Save body to sw.lastScrape and return it to the pool only if its size is smaller than -promscrape.minResponseSizeForStreamParse
+		// This should reduce memory usage when scraping targets which return big responses.
 		sw.lastScrape = append(sw.lastScrape[:0], bodyString...)
 		leveledbytebufferpool.Put(body)
+	}
 	tsmGlobal.Update(sw.Config, sw.ScrapeGroup, up == 1, realTimestamp, int64(duration*1000), samplesScraped, err)
 	return err
 }
@@ -432,6 +447,7 @@ func (sw *scrapeWork) scrapeStream(scrapeTimestamp, realTimestamp int64) error {
 	sw.addAutoTimeseries(wc, "scrape_timeout_seconds", sw.Config.ScrapeTimeout.Seconds(), scrapeTimestamp)
 	sw.pushData(&wc.writeRequest)
 	sw.prevLabelsLen = len(wc.labels)
+	sw.prevBodyLen = int(responseSize)
 	wc.reset()
 	writeRequestCtxPool.Put(wc)
 	tsmGlobal.Update(sw.Config, sw.ScrapeGroup, up == 1, realTimestamp, int64(duration*1000), samplesScraped, err)