fix: nats timing flakes (#26944)

coderd/x/nats/metrics_test.go

@@ -58,8 +58,13 @@ func TestPubsub_Metrics(t *testing.T) {
 			testutil.PromCounterHasValue(t, metrics, gatherCount, "coder_nats_pubsub_messages_total", "normal") &&
 			testutil.PromCounterHasValue(t, metrics, float64(len(data))+latencyBytes, "coder_nats_pubsub_received_bytes_total") &&
 			testutil.PromCounterHasValue(t, metrics, float64(len(data))+latencyBytes, "coder_nats_pubsub_published_bytes_total") &&
-			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in > 0 }, "coder_nats_pubsub_send_latency_seconds") &&
+			// The latency gauges can be exactly 0 on Windows, whose monotonic
-			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in > 0 }, "coder_nats_pubsub_receive_latency_seconds") &&
+			// clock advances in coarse (up to 15.6ms) ticks; a sub-tick publish
+			// measures a zero duration. An absent gauge reads as -1, so >= 0
+			// still proves the gauge was gathered, and the measures_total/errs
+			// counters below prove measurement ran and succeeded.
+			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in >= 0 }, "coder_nats_pubsub_send_latency_seconds") &&
+			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in >= 0 }, "coder_nats_pubsub_receive_latency_seconds") &&
 			testutil.PromCounterHasValue(t, metrics, gatherCount, "coder_nats_pubsub_latency_measures_total") &&
 			!testutil.PromCounterGathered(t, metrics, "coder_nats_pubsub_latency_measure_errs_total")
 	}, testutil.WaitShort, testutil.IntervalFast)
@@ -95,8 +100,10 @@ func TestPubsub_Metrics(t *testing.T) {
 			testutil.PromCounterHasValue(t, metrics, 1, "coder_nats_pubsub_messages_total", "colossal") &&
 			testutil.PromCounterHasValue(t, metrics, float64(colossalSize+len(data))+latencyBytes, "coder_nats_pubsub_received_bytes_total") &&
 			testutil.PromCounterHasValue(t, metrics, float64(colossalSize+len(data))+latencyBytes, "coder_nats_pubsub_published_bytes_total") &&
-			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in > 0 }, "coder_nats_pubsub_send_latency_seconds") &&
+			// Zero latency is accepted for coarse clocks; see the comment on
-			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in > 0 }, "coder_nats_pubsub_receive_latency_seconds") &&
+			// the first Eventually above.
+			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in >= 0 }, "coder_nats_pubsub_send_latency_seconds") &&
+			testutil.PromGaugeAssertion(t, metrics, func(in float64) bool { return in >= 0 }, "coder_nats_pubsub_receive_latency_seconds") &&
 			testutil.PromCounterHasValue(t, metrics, gatherCount, "coder_nats_pubsub_latency_measures_total") &&
 			!testutil.PromCounterGathered(t, metrics, "coder_nats_pubsub_latency_measure_errs_total")
 	}, testutil.WaitShort, testutil.IntervalFast)

coderd/x/nats/natsbench/natsbench_internal_test.go

@@ -2,7 +2,6 @@ package main
 
 import (
 	"testing"
-	"time"
 
 	"github.com/stretchr/testify/require"
 
@@ -66,9 +65,22 @@ func TestRunSingleNode(t *testing.T) {
 	require.EqualValues(t, pl.totalExpected, res.Delivered)
 	require.Greater(t, res.Delivered, res.Published, "fan-out must exceed publishes")
 	require.Zero(t, res.Drops)
-	require.Greater(t, res.PubsPerSec, 0.0)
+	// The exact count assertions above are the authoritative correctness
-	require.Greater(t, res.DeliveriesPerSec, 0.0)
+	// checks; a run that published or delivered nothing fails there. The
-	require.Greater(t, res.DeliverDuration, time.Duration(0))
+	// rates below are derived (count / duration), and Windows' monotonic
+	// clock advances in coarse (up to 15.6ms) ticks, so a sub-tick phase
+	// measures a 0 duration and must yield a 0 rate rather than a
+	// fabricated one. On fine-grained clocks the strict branch applies.
+	if res.PublishDuration > 0 {
+		require.Greater(t, res.PubsPerSec, 0.0)
+	} else {
+		require.Zero(t, res.PubsPerSec)
+	}
+	if res.DeliverDuration > 0 {
+		require.Greater(t, res.DeliveriesPerSec, 0.0)
+	} else {
+		require.Zero(t, res.DeliveriesPerSec)
+	}
 	require.GreaterOrEqual(t, res.DeliverDuration, res.PublishDuration)
 }
 
@@ -101,6 +113,16 @@ func TestRunCluster(t *testing.T) {
 	require.EqualValues(t, pl.totalExpected, res.Expected)
 	require.EqualValues(t, pl.totalExpected, res.Delivered)
 	require.Zero(t, res.Drops)
-	require.Greater(t, res.PubsPerSec, 0.0)
+	// See TestRunSingleNode: counts above are the correctness checks;
-	require.Greater(t, res.DeliveriesPerSec, 0.0)
+	// coarse clocks can quantize a phase duration, and thus its rate, to 0.
+	if res.PublishDuration > 0 {
+		require.Greater(t, res.PubsPerSec, 0.0)
+	} else {
+		require.Zero(t, res.PubsPerSec)
+	}
+	if res.DeliverDuration > 0 {
+		require.Greater(t, res.DeliveriesPerSec, 0.0)
+	} else {
+		require.Zero(t, res.DeliveriesPerSec)
+	}
 }