fix(usage): calculate overview TPS from valid post-TTFT generation time

2026-08-29 00:04:27 +08:00
1 parent 001a1a5149
commit 2e0b2aef89
@@ -258,6 +258,7 @@ pub(crate) struct UsageOverview {
    rpm: f64,
    tpm: f64,
    tps: f64,
+    tps_sample_count: u64,
    average_latency_ms: f64,
    cache_hit_rate: f64,
    estimated_cost: f64,
@@ -2129,10 +2130,42 @@ fn load_usage_overview(
            COALESCE(SUM(cache_creation_tokens), 0),
            COALESCE(SUM(total_tokens), 0),
            COALESCE(SUM(latency_ms), 0),
-            COALESCE(AVG(CASE
-                WHEN output_tokens > 0 AND latency_ms > 0
-                THEN CAST(output_tokens AS REAL) * 1000.0 / latency_ms
-            END), 0.0),
+            COALESCE(
+                SUM(CASE
+                    WHEN generate != 0
+                     AND failed = 0
+                     AND canceled = 0
+                     AND output_tokens > 0
+                     AND ttft_ms IS NOT NULL
+                     AND ttft_ms > 0
+                     AND latency_ms > ttft_ms
+                    THEN output_tokens
+                    ELSE 0
+                END) * 1000.0
+                / NULLIF(SUM(CASE
+                    WHEN generate != 0
+                     AND failed = 0
+                     AND canceled = 0
+                     AND output_tokens > 0
+                     AND ttft_ms IS NOT NULL
+                     AND ttft_ms > 0
+                     AND latency_ms > ttft_ms
+                    THEN latency_ms - ttft_ms
+                    ELSE 0
+                END), 0),
+                0.0
+            ),
+            COALESCE(SUM(CASE
+                WHEN generate != 0
+                 AND failed = 0
+                 AND canceled = 0
+                 AND output_tokens > 0
+                 AND ttft_ms IS NOT NULL
+                 AND ttft_ms > 0
+                 AND latency_ms > ttft_ms
+                THEN 1
+                ELSE 0
+            END), 0),
            MIN(timestamp_ms),
            MAX(timestamp_ms)
        FROM usage_events{}
@@ -2157,8 +2190,9 @@ fn load_usage_overview(
                    row.get::<_, i64>(9)?,
                    row.get::<_, i64>(10)?,
                    row.get::<_, f64>(11)?,
-                    row.get::<_, Option<i64>>(12)?,
+                    row.get::<_, i64>(12)?,
                    row.get::<_, Option<i64>>(13)?,
+                    row.get::<_, Option<i64>>(14)?,
                ))
            },
        )
@@ -2226,11 +2260,12 @@ fn load_usage_overview(
        overview.average_latency_ms =
            from_sql_i64(summary.10) as f64 / overview.total_requests as f64;
        overview.tps = summary.11;
+        overview.tps_sample_count = from_sql_i64(summary.12);
        if overview.input_tokens > 0 {
            overview.cache_hit_rate =
                (overview.cache_read_tokens as f64 / overview.input_tokens as f64).min(1.0);
        }
-        let minutes = query_window_minutes(query, summary.12, summary.13);
+        let minutes = query_window_minutes(query, summary.13, summary.14);
        overview.rpm = overview.total_requests as f64 / minutes;
        overview.tpm = overview.total_tokens as f64 / minutes;
    }
@@ -4174,7 +4209,8 @@ mod tests {

        assert_eq!(overview.total_requests, 1);
        assert_eq!(overview.failure_count, 1);
-        assert_eq!(overview.tps, 200.0);
+        assert_eq!(overview.tps, 0.0);
+        assert_eq!(overview.tps_sample_count, 0);
        assert_eq!(overview.cache_hit_rate, 0.2);
        assert!((overview.estimated_cost - 0.0003205).abs() < 0.0000001);
        assert_eq!(overview.priced_requests, 1);
@@ -4185,6 +4221,65 @@ mod tests {
        fs::remove_dir_all(root).unwrap();
    }

+    #[test]
+    fn overview_tps_uses_weighted_generation_time_and_ignores_invalid_records() {
+        let root = test_root("tps-overview");
+        let mut connection = open_test_database(&root);
+
+        let mut first = sample_record("tps-1", "2026-07-17T20:30:00+08:00", "gpt-a");
+        first.latency_ms = 1_000;
+        first.ttft_ms = Some(200);
+        first.tokens.output_tokens = 80;
+
+        let mut second = sample_record("tps-2", "2026-07-17T20:31:00+08:00", "gpt-a");
+        second.latency_ms = 2_000;
+        second.ttft_ms = Some(1_000);
+        second.tokens.output_tokens = 20;
+
+        let mut missing_ttft = sample_record("tps-3", "2026-07-17T20:32:00+08:00", "gpt-a");
+        missing_ttft.latency_ms = 500;
+        missing_ttft.ttft_ms = None;
+        missing_ttft.tokens.output_tokens = 100;
+
+        let mut equal_latency = sample_record("tps-4", "2026-07-17T20:33:00+08:00", "gpt-a");
+        equal_latency.latency_ms = 500;
+        equal_latency.ttft_ms = Some(500);
+        equal_latency.tokens.output_tokens = 100;
+
+        let mut failed = sample_record("tps-5", "2026-07-17T20:34:00+08:00", "gpt-a");
+        failed.failed = true;
+        failed.tokens.output_tokens = 100;
+
+        let mut canceled = sample_record("tps-6", "2026-07-17T20:35:00+08:00", "gpt-a");
+        canceled.canceled = true;
+        canceled.tokens.output_tokens = 100;
+
+        let mut non_generation = sample_record("tps-7", "2026-07-17T20:36:00+08:00", "gpt-a");
+        non_generation.generate = false;
+        non_generation.tokens.output_tokens = 100;
+
+        insert_usage_records(
+            &mut connection,
+            &[
+                first,
+                second,
+                missing_ttft,
+                equal_latency,
+                failed,
+                canceled,
+                non_generation,
+            ],
+        )
+        .unwrap();
+
+        let overview = load_usage_overview(&connection, &UsageQuery::default()).unwrap();
+
+        assert!((overview.tps - (100.0 * 1_000.0 / 1_800.0)).abs() < f64::EPSILON);
+        assert_eq!(overview.tps_sample_count, 2);
+        drop(connection);
+        fs::remove_dir_all(root).unwrap();
+    }
+
    #[test]
    fn estimates_gpt56_cost_with_cache_and_service_tier_rules() {
        let terra = official_model_price("openai/gpt-5.6-terra").unwrap();
@@ -150,8 +150,8 @@ export const jaOverrides = {
  'usage.stat.successMetaTitle': '成功 {success} 件、失敗 {failed} 件、キャンセル {canceled} 件',
  'usage.stat.averageLatency': '平均レイテンシ',
  'usage.stat.tps': 'TPS',
-  'usage.stat.performanceMeta': 'RPM {rpm} · 平均レイテンシ {latency} ms',
-  'usage.stat.performanceMetaTitle': 'TPS {tps}、RPM {rpm}、平均レイテンシ {latency} ms',
+  'usage.stat.performanceMeta': '有効サンプル {samples} 件 · RPM {rpm}',
+  'usage.stat.performanceMetaTitle': 'TPS {tps}、有効サンプル {samples} 件、RPM {rpm}、平均レイテンシ {latency} ms',
  'usage.stat.cacheHitRate': 'キャッシュヒット率',
  'usage.stat.cacheHitMeta': 'キャッシュ読取 {hit} · 入力 {input}',
  'usage.stat.cacheHitMetaTitle': 'キャッシュヒット率 {rate}%、読取 {hit} Token、入力 {input} Token',
@@ -150,8 +150,8 @@ export const en: Record<MessageKey, string> = {
  'usage.stat.successMetaTitle': '{success} succeeded, {failed} failed, {canceled} canceled',
  'usage.stat.averageLatency': 'Average Latency',
  'usage.stat.tps': 'TPS',
-  'usage.stat.performanceMeta': 'RPM {rpm} · Avg. latency {latency} ms',
-  'usage.stat.performanceMetaTitle': 'TPS {tps}, RPM {rpm}, average latency {latency} ms',
+  'usage.stat.performanceMeta': '{samples} samples · RPM {rpm}',
+  'usage.stat.performanceMetaTitle': 'TPS {tps}, {samples} valid samples, RPM {rpm}, average latency {latency} ms',
  'usage.stat.cacheHitRate': 'Cache Hit Rate',
  'usage.stat.cacheHitMeta': 'Cache read {hit} · Input {input}',
  'usage.stat.cacheHitMetaTitle': 'Cache hit rate {rate}%, cache read {hit} tokens, input {input} tokens',
@@ -148,8 +148,8 @@ export const zhCN = {
  'usage.stat.successMetaTitle': '成功 {success} 次，失败 {failed} 次，取消 {canceled} 次',
  'usage.stat.averageLatency': '平均延迟',
  'usage.stat.tps': 'TPS',
-  'usage.stat.performanceMeta': 'RPM {rpm} · 平均延迟 {latency} ms',
-  'usage.stat.performanceMetaTitle': 'TPS {tps}，RPM {rpm}，平均延迟 {latency} ms',
+  'usage.stat.performanceMeta': '{samples} 条有效样本 · RPM {rpm}',
+  'usage.stat.performanceMetaTitle': 'TPS {tps}，{samples} 条有效样本，RPM {rpm}，平均延迟 {latency} ms',
  'usage.stat.cacheHitRate': '缓存命中率',
  'usage.stat.cacheHitMeta': '缓存读取 {hit} · 输入 {input}',
  'usage.stat.cacheHitMetaTitle': '缓存命中率 {rate}%，缓存读取 {hit} Token，输入 {input} Token',
@@ -64,6 +64,7 @@ type UsageOverview = {
  rpm: number;
  tpm: number;
  tps: number;
+  tpsSampleCount: number;
  averageLatencyMs: number;
  cacheHitRate: number;
  estimatedCost: number;
@@ -818,13 +819,15 @@ function OverviewView({ overview }: { overview: UsageOverview }) {
    },
    {
      label: t('usage.stat.tps'),
-      value: `${overview.tps.toFixed(1)} TPS`,
+      value: overview.tpsSampleCount > 0 ? `${overview.tps.toFixed(1)} TPS` : '—',
      meta: t('usage.stat.performanceMeta', {
+        samples: compactNumber(overview.tpsSampleCount),
        rpm: overview.rpm.toFixed(2),
        latency: Math.round(overview.averageLatencyMs),
      }),
      metaTitle: t('usage.stat.performanceMetaTitle', {
-        tps: overview.tps.toFixed(1),
+        tps: overview.tpsSampleCount > 0 ? overview.tps.toFixed(1) : '—',
+        samples: compactNumber(overview.tpsSampleCount),
        rpm: overview.rpm.toFixed(2),
        latency: Math.round(overview.averageLatencyMs),
      }),