diff --git a/scrape/scrape_test.go b/scrape/scrape_test.go
index 47829bf20e..61eace1bab 100644
--- a/scrape/scrape_test.go
+++ b/scrape/scrape_test.go
@@ -5568,3 +5568,219 @@ func TestScrapeAppendWithParseError(t *testing.T) {
 	}
 	requireEqual(t, want, capp.resultFloats, "Appended samples not as expected:\n%s", capp)
 }
+
+// This test covers a case where there's a target with sample_limit set and the some of exporter samples
+// changes between scrapes.
+func TestScrapeLoopAppendSampleLimitWithDisappearingSeries(t *testing.T) {
+	const sampleLimit = 4
+	resApp := &collectResultAppender{}
+	sl := newBasicScrapeLoop(t, context.Background(), nil, func(_ context.Context) storage.Appender {
+		return resApp
+	}, 0)
+	sl.sampleLimit = sampleLimit
+
+	now := time.Now()
+	slApp := sl.appender(context.Background())
+	samplesScraped, samplesAfterRelabel, createdSeries, err := sl.append(
+		slApp,
+		// Start with 3 samples, all accepted.
+		[]byte("metric_a 1\nmetric_b 1\nmetric_c 1\n"),
+		"text/plain",
+		now,
+	)
+	require.NoError(t, err)
+	require.NoError(t, slApp.Commit())
+	require.Equal(t, 3, samplesScraped)      // All on scrape.
+	require.Equal(t, 3, samplesAfterRelabel) // This is series after relabeling.
+	require.Equal(t, 3, createdSeries)       // Newly added to TSDB.
+	want := []floatSample{
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_a"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_b"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_c"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+	}
+	requireEqual(t, want, resApp.resultFloats, "Appended samples not as expected:\n%s", slApp)
+
+	now = now.Add(time.Minute)
+	slApp = sl.appender(context.Background())
+	samplesScraped, samplesAfterRelabel, createdSeries, err = sl.append(
+		slApp,
+		// Start exporting 3 more samples, so we're over the limit now.
+		[]byte("metric_a 1\nmetric_b 1\nmetric_c 1\nmetric_d 1\nmetric_e 1\nmetric_f 1\n"),
+		"text/plain",
+		now,
+	)
+	require.ErrorIs(t, err, errSampleLimit)
+	require.NoError(t, slApp.Rollback())
+	require.Equal(t, 6, samplesScraped)
+	require.Equal(t, 6, samplesAfterRelabel)
+	require.Equal(t, 1, createdSeries) // We've added one series before hitting the limit.
+	requireEqual(t, want, resApp.resultFloats, "Appended samples not as expected:\n%s", slApp)
+	sl.cache.iterDone(false)
+
+	now = now.Add(time.Minute)
+	slApp = sl.appender(context.Background())
+	samplesScraped, samplesAfterRelabel, createdSeries, err = sl.append(
+		slApp,
+		// Remove all samples except first 2.
+		[]byte("metric_a 1\nmetric_b 1\n"),
+		"text/plain",
+		now,
+	)
+	require.NoError(t, err)
+	require.NoError(t, slApp.Commit())
+	require.Equal(t, 2, samplesScraped)
+	require.Equal(t, 2, samplesAfterRelabel)
+	require.Equal(t, 0, createdSeries)
+	// This is where important things happen. We should now see:
+	// - Appends for samples from metric_a & metric_b.
+	// - Append with stale markers for metric_c - this series was added during first scrape but disappeared during last scrape.
+	// - Append with stale marker for metric_d - this series was added during second scrape before we hit the sample_limit.
+	// We should NOT see:
+	// - Appends with stale markers for metric_e & metric_f - both over the limit during second scrape and so they never made it into TSDB.
+	want = append(want, []floatSample{
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_a"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_b"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_c"),
+			t:      timestamp.FromTime(now),
+			f:      math.Float64frombits(value.StaleNaN),
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_d"),
+			t:      timestamp.FromTime(now),
+			f:      math.Float64frombits(value.StaleNaN),
+		},
+	}...)
+	requireEqual(t, want, resApp.resultFloats, "Appended samples not as expected:\n%s", slApp)
+}
+
+// This test covers a case where there's a target with sample_limit set and each scrape sees a completely
+// different set of samples.
+func TestScrapeLoopAppendSampleLimitReplaceAllSamples(t *testing.T) {
+	const sampleLimit = 4
+	resApp := &collectResultAppender{}
+	sl := newBasicScrapeLoop(t, context.Background(), nil, func(_ context.Context) storage.Appender {
+		return resApp
+	}, 0)
+	sl.sampleLimit = sampleLimit
+
+	now := time.Now()
+	slApp := sl.appender(context.Background())
+	samplesScraped, samplesAfterRelabel, createdSeries, err := sl.append(
+		slApp,
+		// Start with 4 samples, all accepted.
+		[]byte("metric_a 1\nmetric_b 1\nmetric_c 1\nmetric_d 1\n"),
+		"text/plain",
+		now,
+	)
+	require.NoError(t, err)
+	require.NoError(t, slApp.Commit())
+	require.Equal(t, 4, samplesScraped)      // All on scrape.
+	require.Equal(t, 4, samplesAfterRelabel) // This is series after relabeling.
+	require.Equal(t, 4, createdSeries)       // Newly added to TSDB.
+	want := []floatSample{
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_a"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_b"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_c"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_d"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+	}
+	requireEqual(t, want, resApp.resultFloats, "Appended samples not as expected:\n%s", slApp)
+
+	now = now.Add(time.Minute)
+	slApp = sl.appender(context.Background())
+	samplesScraped, samplesAfterRelabel, createdSeries, err = sl.append(
+		slApp,
+		// Replace all samples with new time series.
+		[]byte("metric_e 1\nmetric_f 1\nmetric_g 1\nmetric_h 1\n"),
+		"text/plain",
+		now,
+	)
+	require.NoError(t, err)
+	require.NoError(t, slApp.Commit())
+	require.Equal(t, 4, samplesScraped)
+	require.Equal(t, 4, samplesAfterRelabel)
+	require.Equal(t, 4, createdSeries)
+	// We replaced all samples from first scrape with new set of samples.
+	// We expect to see:
+	// - 4 appends for new samples.
+	// - 4 appends with staleness markers for old samples.
+	want = append(want, []floatSample{
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_e"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_f"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_g"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_h"),
+			t:      timestamp.FromTime(now),
+			f:      1,
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_a"),
+			t:      timestamp.FromTime(now),
+			f:      math.Float64frombits(value.StaleNaN),
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_b"),
+			t:      timestamp.FromTime(now),
+			f:      math.Float64frombits(value.StaleNaN),
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_c"),
+			t:      timestamp.FromTime(now),
+			f:      math.Float64frombits(value.StaleNaN),
+		},
+		{
+			metric: labels.FromStrings(model.MetricNameLabel, "metric_d"),
+			t:      timestamp.FromTime(now),
+			f:      math.Float64frombits(value.StaleNaN),
+		},
+	}...)
+	requireEqual(t, want, resApp.resultFloats, "Appended samples not as expected:\n%s", slApp)
+}