fix: update incorrect metrics in cache

Signed-off-by: CYJiang <[email protected]>

Author: googs1025

pkg/cache/cache_metrics.go

@@ -353,17 +353,21 @@ func (c *Store) fetchMetrics(pod *Pod, allMetrics map[string]*dto.MetricFamily,
 		klog.V(4).Infof(err.Error())
 		engineType = defaultEngineLabelValue
 	}
-	rawMetricName, ok := metric.EngineMetricsNameMapping[engineType]
-	if !ok {
-		klog.V(4).Infof("Cannot find engine type %v mapping for metrics %v", engineType, labelMetricName)
+	candidates, ok := metric.EngineMetricsNameMapping[engineType]
+	if !ok || len(candidates) == 0 {
+		klog.V(4).Infof("No metric name candidates found for engine type %v and metric %v", engineType, labelMetricName)
 		return nil, false
 	}
-	metricFamily, exists := allMetrics[rawMetricName]
-	if !exists {
-		klog.V(4).Infof("Cannot find raw metrics %v, engine type %v", rawMetricName, engineType)
-		return nil, false
+
+	for _, rawMetricName := range candidates {
+		if metricFamily, exists := allMetrics[rawMetricName]; exists {
+			return metricFamily, true
+		}
 	}
-	return metricFamily, true
+
+	klog.V(4).Infof("None of the candidate raw metrics %v found for engine %v and metric %v", candidates, engineType, labelMetricName)
+	return nil, false
+
 }
 
 // Update `PodMetrics` and `PodModelMetrics` according to the metric scope

pkg/metrics/engine_fetcher.go

@@ -100,14 +100,14 @@ func (ef *EngineMetricsFetcher) FetchTypedMetric(ctx context.Context, endpoint,
 		return nil, fmt.Errorf("metric %s is not a raw pod metric, use FetchAllTypedMetrics for complex queries", metricName)
 	}
 
-	// Get raw metric name for this engine
-	rawMetricName, exists := metricDef.EngineMetricsNameMapping[engineType]
-	if !exists {
+	// Get raw metric name candidates for this engine
+	candidates, exists := metricDef.EngineMetricsNameMapping[engineType]
+	if !exists || len(candidates) == 0 {
 		return nil, fmt.Errorf("metric %s not supported for engine type %s", metricName, engineType)
 	}
 
 	url := fmt.Sprintf("http://%s/metrics", endpoint)
-
+	var lastErr error
 	// Fetch with retry logic
 	for attempt := 0; attempt <= ef.config.MaxRetries; attempt++ {
 		if attempt > 0 {
@@ -130,21 +130,37 @@ func (ef *EngineMetricsFetcher) FetchTypedMetric(ctx context.Context, endpoint,
 			continue
 		}
 
-		// Parse the specific metric we need
-		metricValue, err := ef.parseMetricFromFamily(allMetrics, rawMetricName, metricDef)
-		if err != nil {
-			klog.V(4).InfoS("Failed to parse metric from engine endpoint",
-				"attempt", attempt+1, "identifier", identifier, "metric", metricName, "error", err)
-			continue
+		// Try each candidate until one exists and can be parsed
+		for _, rawMetricName := range candidates {
+			if _, ok := allMetrics[rawMetricName]; !ok {
+				continue // skip if not present
+			}
+
+			metricValue, err := ef.parseMetricFromFamily(allMetrics, rawMetricName, metricDef)
+			if err != nil {
+				lastErr = err
+				klog.V(5).InfoS("Failed to parse candidate metric", "candidate", rawMetricName, "error", err)
+				continue
+			}
+
+			klog.V(4).InfoS("Successfully fetched typed metric from engine endpoint",
+				"identifier", identifier, "metric", metricName, "rawMetric", rawMetricName, "value", metricValue, "attempt", attempt+1)
+			return metricValue, nil
 		}
 
-		klog.V(4).InfoS("Successfully fetched typed metric from engine endpoint",
-			"identifier", identifier, "metric", metricName, "value", metricValue, "attempt", attempt+1)
-		return metricValue, nil
+		klog.V(4).InfoS("Failed to find valid metric among candidates",
+			"candidates", candidates, "identifier", identifier, "metric", metricName)
+		// Continue to next retry if any
 	}
 
-	return nil, fmt.Errorf("failed to fetch typed metric %s from engine endpoint %s after %d attempts",
-		metricName, identifier, ef.config.MaxRetries+1)
+	// If we get here, none of the candidates worked
+	errMsg := "none of the candidate metric names found or parsed successfully"
+	if lastErr != nil {
+		errMsg += ": " + lastErr.Error()
+	}
+
+	return nil, fmt.Errorf("failed to fetch typed metric %s from engine endpoint %s after %d attempts: %w",
+		metricName, identifier, ef.config.MaxRetries+1, lastErr)
 }
 
 // FetchAllTypedMetrics fetches all available typed metrics from an engine endpoint
@@ -215,10 +231,26 @@ func (ef *EngineMetricsFetcher) FetchAllTypedMetrics(ctx context.Context, endpoi
 			continue
 		}
 
-		// Get raw metric name for this engine
-		rawMetricName, exists := metricDef.EngineMetricsNameMapping[result.EngineType]
-		if !exists {
-			klog.V(5).InfoS("Metric not supported for engine type", "metric", metricName, "engine", result.EngineType)
+		// Get raw metric name candidates for this engine
+		candidates, exists := metricDef.EngineMetricsNameMapping[result.EngineType]
+		if !exists || len(candidates) == 0 {
+			klog.V(5).InfoS("No raw metric names defined for metric and engine type",
+				"metric", metricName, "engine", result.EngineType)
+			continue
+		}
+
+		// Find the first candidate that exists in allMetrics
+		var rawMetricName string
+		for _, name := range candidates {
+			if _, ok := allMetrics[name]; ok {
+				rawMetricName = name
+				break
+			}
+		}
+
+		if rawMetricName == "" {
+			klog.V(5).InfoS("None of the candidate raw metrics found in endpoint response",
+				"metric", metricName, "engine", result.EngineType, "candidates", candidates)
 			continue
 		}
 

pkg/metrics/engine_fetcher_test.go

@@ -77,9 +77,9 @@ func setupMockMetrics() {
 	Metrics["running_requests"] = Metric{
 		MetricSource: PodRawMetrics,
 		MetricType:   MetricType{Raw: Gauge},
-		EngineMetricsNameMapping: map[string]string{
-			"vllm":   "vllm_num_requests_running",
-			"sglang": "sglang_running_requests",
+		EngineMetricsNameMapping: map[string][]string{
+			"vllm":   {"vllm_num_requests_running"},
+			"sglang": {"sglang_running_requests"},
 		},
 		Description: "Number of running requests",
 		MetricScope: PodModelMetricScope,
@@ -88,9 +88,9 @@ func setupMockMetrics() {
 	Metrics["waiting_requests"] = Metric{
 		MetricSource: PodRawMetrics,
 		MetricType:   MetricType{Raw: Gauge},
-		EngineMetricsNameMapping: map[string]string{
-			"vllm":   "vllm_num_requests_waiting",
-			"sglang": "sglang_waiting_requests",
+		EngineMetricsNameMapping: map[string][]string{
+			"vllm":   {"vllm_num_requests_waiting"},
+			"sglang": {"sglang_waiting_requests"},
 		},
 		Description: "Number of waiting requests",
 		MetricScope: PodModelMetricScope,
@@ -99,9 +99,9 @@ func setupMockMetrics() {
 	Metrics["cache_usage"] = Metric{
 		MetricSource: PodRawMetrics,
 		MetricType:   MetricType{Raw: Gauge},
-		EngineMetricsNameMapping: map[string]string{
-			"vllm":   "vllm_gpu_cache_usage_perc",
-			"sglang": "sglang_cache_usage",
+		EngineMetricsNameMapping: map[string][]string{
+			"vllm":   {"vllm_gpu_cache_usage_perc"},
+			"sglang": {"sglang_cache_usage"},
 		},
 		Description: "Cache usage percentage",
 		MetricScope: PodMetricScope,
@@ -110,8 +110,8 @@ func setupMockMetrics() {
 	Metrics["time_to_first_token"] = Metric{
 		MetricSource: PodRawMetrics,
 		MetricType:   MetricType{Raw: Histogram},
-		EngineMetricsNameMapping: map[string]string{
-			"vllm": "vllm_time_to_first_token_seconds",
+		EngineMetricsNameMapping: map[string][]string{
+			"vllm": {"vllm_time_to_first_token_seconds"},
 		},
 		Description: "Time to first token histogram",
 		MetricScope: PodModelMetricScope,