diff --git a/scrape/scrape.go b/scrape/scrape.go
index 2039a7b82..979fa4cf6 100644
--- a/scrape/scrape.go
+++ b/scrape/scrape.go
@@ -125,6 +125,12 @@ var (
 			Help: "Total number of samples rejected due to timestamp falling outside of the time bounds",
 		},
 	)
+	targetScrapeCacheFlushForced = prometheus.NewCounter(
+		prometheus.CounterOpts{
+			Name: "prometheus_target_scrapes_cache_flush_forced_total",
+			Help: "How many times a scrape cache was flushed due to getting big while scrapes are failing.",
+		},
+	)
 )
 
 func init() {
@@ -140,6 +146,7 @@ func init() {
 	prometheus.MustRegister(targetScrapeSampleDuplicate)
 	prometheus.MustRegister(targetScrapeSampleOutOfOrder)
 	prometheus.MustRegister(targetScrapeSampleOutOfBounds)
+	prometheus.MustRegister(targetScrapeCacheFlushForced)
 }
 
 // scrapePool manages scrapes for sets of targets.
@@ -606,6 +613,9 @@ type scrapeLoop struct {
 type scrapeCache struct {
 	iter uint64 // Current scrape iteration.
 
+	// How many series and metadata entries there were at the last success.
+	successfulCount int
+
 	// Parsed string to an entry with information about the actual label set
 	// and its storage reference.
 	series map[string]*cacheEntry
@@ -643,8 +653,24 @@ func newScrapeCache() *scrapeCache {
 	}
 }
 
-func (c *scrapeCache) iterDone(cleanCache bool) {
-	if cleanCache {
+func (c *scrapeCache) iterDone(flushCache bool) {
+	c.metaMtx.Lock()
+	count := len(c.series) + len(c.droppedSeries) + len(c.metadata)
+	c.metaMtx.Unlock()
+
+	if flushCache {
+		c.successfulCount = count
+	} else if count > c.successfulCount*2+1000 {
+		// If a target had varying labels in scrapes that ultimately failed,
+		// the caches would grow indefinitely. Force a flush when this happens.
+		// We use the heuristic that this is a doubling of the cache size
+		// since the last scrape, and allow an additional 1000 in case
+		// initial scrapes all fail.
+		flushCache = true
+		targetScrapeCacheFlushForced.Inc()
+	}
+
+	if flushCache {
 		// All caches may grow over time through series churn
 		// or multiple string representations of the same metric. Clean up entries
 		// that haven't appeared in the last scrape.
@@ -1185,6 +1211,8 @@ loop:
 		return total, added, err
 	}
 
+	// Only perform cache cleaning if the scrape was not empty.
+	// An empty scrape (usually) is used to indicate a failed scrape.
 	sl.cache.iterDone(len(b) > 0)
 
 	return total, added, nil
diff --git a/scrape/scrape_test.go b/scrape/scrape_test.go
index 64a211996..bb5f51c4b 100644
--- a/scrape/scrape_test.go
+++ b/scrape/scrape_test.go
@@ -859,6 +859,66 @@ func TestScrapeLoopCache(t *testing.T) {
 	}
 }
 
+func TestScrapeLoopCacheMemoryExhaustionProtection(t *testing.T) {
+	s := testutil.NewStorage(t)
+	defer s.Close()
+
+	sapp, err := s.Appender()
+	if err != nil {
+		t.Error(err)
+	}
+	appender := &collectResultAppender{next: sapp}
+	var (
+		signal  = make(chan struct{})
+		scraper = &testScraper{}
+		app     = func() storage.Appender { return appender }
+	)
+	defer close(signal)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	sl := newScrapeLoop(ctx,
+		scraper,
+		nil, nil,
+		nopMutator,
+		nopMutator,
+		app,
+		nil,
+		0,
+		true,
+	)
+
+	numScrapes := 0
+
+	scraper.scrapeFunc = func(ctx context.Context, w io.Writer) error {
+		numScrapes++
+		if numScrapes < 5 {
+			s := ""
+			for i := 0; i < 500; i++ {
+				s = fmt.Sprintf("%smetric_%d_%d 42\n", s, i, numScrapes)
+			}
+			w.Write([]byte(fmt.Sprintf(s + "&")))
+		} else {
+			cancel()
+		}
+		return nil
+	}
+
+	go func() {
+		sl.run(10*time.Millisecond, time.Hour, nil)
+		signal <- struct{}{}
+	}()
+
+	select {
+	case <-signal:
+	case <-time.After(5 * time.Second):
+		t.Fatalf("Scrape wasn't stopped.")
+	}
+
+	if len(sl.cache.series) > 2000 {
+		t.Fatalf("More than 2000 series cached. Got: %d", len(sl.cache.series))
+	}
+}
+
 func TestScrapeLoopAppend(t *testing.T) {
 
 	tests := []struct {