fix: align model card summary with status bar hours (#7847)

pkg/perf_metrics/metrics.go

@@ -199,11 +199,7 @@ func QuerySummaryAll(hours int, groups []string, recentModels ...string) (Summar
 	startTs, endTs := queryWindow(now, hours)
 	recentStartTs, _ := queryWindow(now, recentSuccessHistoryHours)
 	allowedGroups := allowedGroupSet(groups)
-
-	rows, err := model.GetPerfMetricsSummaryBucketsAll(startTs, endTs, groups)
-	if err != nil {
-		return SummaryAllResult{}, err
-	}
+	queryStartTs := startTs
 
 	recentBuckets := make(map[string]map[int64]counters, len(recentModels))
 	for _, name := range recentModels {
@@ -219,6 +215,7 @@ func QuerySummaryAll(hours int, groups []string, recentModels ...string) (Summar
 			return SummaryAllResult{}, err
 		}
 		for _, bucket := range buckets {
+			queryStartTs = min(queryStartTs, bucket.BucketTs)
 			mergeModelBucket(recentBuckets, name, bucket.BucketTs, counters{
 				requestCount: bucket.RequestCount,
 				successCount: bucket.SuccessCount,
@@ -226,7 +223,11 @@ func QuerySummaryAll(hours int, groups []string, recentModels ...string) (Summar
 		}
 	}
 
-	totals := map[string]counters{}
+	rows, err := model.GetPerfMetricsSummaryBucketsAll(queryStartTs, endTs, groups)
+	if err != nil {
+		return SummaryAllResult{}, err
+	}
+
 	modelBuckets := map[string]map[int64]counters{}
 	for _, row := range rows {
 		value := counters{
@@ -236,7 +237,6 @@ func QuerySummaryAll(hours int, groups []string, recentModels ...string) (Summar
 			outputTokens:   row.OutputTokens,
 			generationMs:   row.GenerationMs,
 		}
-		mergeModelTotals(totals, row.ModelName, value)
 		mergeModelBucket(modelBuckets, row.ModelName, row.BucketTs, value)
 	}
 
@@ -257,14 +257,38 @@ func QuerySummaryAll(hours int, groups []string, recentModels ...string) (Summar
 		if _, requested := recentBuckets[k.model]; requested && k.bucketTs >= recentStartTs {
 			mergeModelBucket(recentBuckets, k.model, k.bucketTs, snap)
 		}
-		if k.bucketTs < startTs {
+		if k.bucketTs < queryStartTs {
 			return true
 		}
-		mergeModelTotals(totals, k.model, snap)
 		mergeModelBucket(modelBuckets, k.model, k.bucketTs, snap)
 		return true
 	})
 
+	recentSeries := make(map[string][]SuccessRatePoint, len(recentBuckets))
+	for name, buckets := range recentBuckets {
+		points := recentSuccessSeries(buckets)
+		if len(points) == 0 {
+			continue
+		}
+		recentSeries[name] = points[max(0, len(points)-recentSuccessSampleLimit):]
+	}
+
+	// Requested models are summarized over the same hours as their status samples.
+	totals := map[string]counters{}
+	for name, buckets := range modelBuckets {
+		fromTs := startTs
+		if points, ok := recentSeries[name]; ok {
+			fromTs = points[0].Ts
+		}
+		for bucketTs, value := range buckets {
+			if bucketTs < fromTs {
+				delete(buckets, bucketTs)
+				continue
+			}
+			mergeModelTotals(totals, name, value)
+		}
+	}
+
 	all := counters{}
 	models := make([]ModelSummary, 0, len(totals))
 	for name, total := range totals {
@@ -295,14 +319,6 @@ func QuerySummaryAll(hours int, groups []string, recentModels ...string) (Summar
 		return models[i].RequestCount > models[j].RequestCount
 	})
 
-	recentSeries := make(map[string][]SuccessRatePoint, len(recentBuckets))
-	for name, buckets := range recentBuckets {
-		points := recentSuccessSeries(buckets)
-		if len(points) == 0 {
-			continue
-		}
-		recentSeries[name] = points[max(0, len(points)-recentSuccessSampleLimit):]
-	}
 	return SummaryAllResult{
 		Summary: summarize(all), WindowStart: startTs, WindowEnd: endTs,
 		Models: models, RecentSuccessSeries: recentSeries,

pkg/perf_metrics/metrics_test.go

@@ -12,6 +12,7 @@ import (
 	"github.com/QuantumNous/new-api/model"
 	relaycommon "github.com/QuantumNous/new-api/relay/common"
 	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/samber/lo"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -262,7 +263,7 @@ func TestPerformanceAggregationAndFlush(t *testing.T) {
 			assert.Equal(t, 98.04, combined.Summary.SuccessRate)
 			assert.Equal(t, 99.01, combined.Models[0].SuccessRate)
 
-			t.Run("recent history does not change window summaries", func(t *testing.T) {
+			t.Run("recent models are summarized over their status sample hours", func(t *testing.T) {
 				for i := range 26 {
 					require.NoError(t, model.UpsertPerfMetric(&model.PerfMetric{
 						ModelName: "history-model", Group: "a", BucketTs: hour - int64(i+1)*6*3600,
@@ -293,9 +294,13 @@ func TestPerformanceAggregationAndFlush(t *testing.T) {
 				require.NoError(t, err)
 				history, err := QuerySummaryAll(24, groups, "history-model", "history-idle", "history-hidden", "history-hot", "missing")
 				require.NoError(t, err)
-				assert.Equal(t, window.Summary, history.Summary)
-				assert.ElementsMatch(t, window.Models, history.Models)
 				assert.Equal(t, window.WindowStart, history.WindowStart)
+				summaries := lo.KeyBy(history.Models, func(m ModelSummary) string { return m.ModelName })
+				assert.Equal(t, int64(124), summaries["history-model"].RequestCount)
+				assert.Equal(t, 99.19, summaries["history-model"].SuccessRate)
+				assert.Equal(t, int64(1), summaries["history-idle"].RequestCount)
+				assert.Equal(t, 0.0, summaries["history-idle"].SuccessRate)
+				assert.NotContains(t, summaries, "history-hidden")
 				points := history.RecentSuccessSeries["history-model"]
 				require.Len(t, points, 24)
 				expected := make([]SuccessRatePoint, 0, 24)