fix(dashboard): 修复仪表盘与明细统计数值不一致并重建 cost_savings 聚合

- 统一 dashboard 聚合与 raw 查询的 token 计算，拆分今日节省和周期节省 - cache savings 改用 input price 估算未命中成本，修复历史 cost_savings 偏高 - raw 查询 total_tokens 改用 effective_input + output + cache_creation + cache_read 公式 - 新增独立 backfill 重建历史 cost_savings 聚合表，保留已发布 backfill 不变
2026-09-02 01:10:23 +08:00 · 2026-05-05 17:33:16 +08:00
parent 2f915b33c7
commit 53fa33b0c5
10 changed files with 487 additions and 113 deletions
--- a/crates/aether-data/src/backfill.rs
+++ b/crates/aether-data/src/backfill.rs
@@ -549,6 +549,7 @@ mod tests {
                cache_read_cost_usd,
                total_cost_usd,
                actual_total_cost_usd,
+                input_price_per_1m,
                output_price_per_1m,
                status,
                billing_status,
@@ -576,6 +577,7 @@ mod tests {
                1.25,
                1.10,
                50.0,
+                50.0,
                'completed',
                'settled',
                TIMESTAMPTZ '2024-05-06 07:18:09+00',
@@ -595,9 +597,10 @@ mod tests {
        let pending_before = pending_backfills(&pool)
            .await
            .expect("pending backfills should load");
-        assert_eq!(pending_before.len(), 2);
+        assert_eq!(pending_before.len(), 3);
        assert_eq!(pending_before[0].version, 20260422110000);
        assert_eq!(pending_before[1].version, 20260422120000);
+        assert_eq!(pending_before[2].version, 20260504120000);

        run_backfills(&pool)
            .await
@@ -613,7 +616,10 @@ mod tests {
                .fetch_all(&pool)
                .await
                .expect("applied backfill versions should load");
-        assert_eq!(applied_versions, vec![20260422110000, 20260422120000]);
+        assert_eq!(
+            applied_versions,
+            vec![20260422110000, 20260422120000, 20260504120000]
+        );

        let api_key_total_requests: i64 = query_scalar(
            "SELECT COALESCE(total_requests, 0)::BIGINT FROM public.api_keys WHERE id = 'api-key-backfill-1'",
@@ -1265,6 +1271,6 @@ mod tests {
                .fetch_one(&pool)
                .await
                .expect("backfill count should load");
-        assert_eq!(applied_count, 1);
+        assert_eq!(applied_count, 3);
    }
 }
--- a/crates/aether-data/src/repository/usage/memory.rs
+++ b/crates/aether-data/src/repository/usage/memory.rs
@@ -1915,7 +1915,7 @@ impl UsageReadRepository for InMemoryUsageReadRepository {
                .saturating_add(item.cache_read_input_tokens);
            summary.cache_read_cost_usd += item.cache_read_cost_usd;
            summary.cache_creation_cost_usd += item.cache_creation_cost_usd;
-            summary.estimated_full_cost_usd += item.settlement_output_price_per_1m().unwrap_or(0.0)
+            summary.estimated_full_cost_usd += item.settlement_input_price_per_1m().unwrap_or(0.0)
                * item.cache_read_input_tokens as f64
                / 1_000_000.0;
        }
--- a/crates/aether-data/src/repository/usage/sql/mod.rs
+++ b/crates/aether-data/src/repository/usage/sql/mod.rs
@@ -1394,7 +1394,7 @@ SELECT
  COALESCE(SUM(input_tokens), 0)::BIGINT AS input_tokens,
  COALESCE(SUM(effective_input_tokens), 0)::BIGINT AS effective_input_tokens,
  COALESCE(SUM(output_tokens), 0)::BIGINT AS output_tokens,
-  COALESCE(SUM(input_tokens + output_tokens), 0)::BIGINT AS total_tokens,
+  COALESCE(SUM(effective_input_tokens + output_tokens + cache_creation_tokens + cache_read_tokens), 0)::BIGINT AS total_tokens,
  COALESCE(SUM(cache_creation_tokens), 0)::BIGINT AS cache_creation_tokens,
  COALESCE(SUM(cache_read_tokens), 0)::BIGINT AS cache_read_tokens,
  COALESCE(SUM(total_input_context), 0)::BIGINT AS total_input_context,
@@ -1425,7 +1425,7 @@ SELECT
  COALESCE(SUM(input_tokens), 0)::BIGINT AS input_tokens,
  COALESCE(SUM(effective_input_tokens), 0)::BIGINT AS effective_input_tokens,
  COALESCE(SUM(output_tokens), 0)::BIGINT AS output_tokens,
-  COALESCE(SUM(input_tokens + output_tokens), 0)::BIGINT AS total_tokens,
+  COALESCE(SUM(effective_input_tokens + output_tokens + cache_creation_tokens + cache_read_tokens), 0)::BIGINT AS total_tokens,
  COALESCE(SUM(cache_creation_tokens), 0)::BIGINT AS cache_creation_tokens,
  COALESCE(SUM(cache_read_tokens), 0)::BIGINT AS cache_read_tokens,
  COALESCE(SUM(total_input_context), 0)::BIGINT AS total_input_context,
@@ -1482,7 +1482,33 @@ SELECT
    END
  ), 0)::BIGINT AS effective_input_tokens,
  COALESCE(SUM(GREATEST(COALESCE("usage".output_tokens, 0), 0)), 0)::BIGINT AS output_tokens,
-  COALESCE(SUM(GREATEST(COALESCE("usage".total_tokens, 0), 0)), 0)::BIGINT AS total_tokens,
+  COALESCE(SUM(
+    CASE
+      WHEN GREATEST(COALESCE("usage".input_tokens, 0), 0) <= 0 THEN 0
+      WHEN GREATEST(COALESCE("usage".cache_read_input_tokens, 0), 0) <= 0
+      THEN GREATEST(COALESCE("usage".input_tokens, 0), 0)
+      WHEN split_part(lower(COALESCE(COALESCE("usage".endpoint_api_format, "usage".api_format), '')), ':', 1)
+           IN ('openai', 'gemini', 'google')
+      THEN GREATEST(
+        GREATEST(COALESCE("usage".input_tokens, 0), 0)
+          - GREATEST(COALESCE("usage".cache_read_input_tokens, 0), 0),
+        0
+      )
+      ELSE GREATEST(COALESCE("usage".input_tokens, 0), 0)
+    END
+    + GREATEST(COALESCE("usage".output_tokens, 0), 0)
+    + CASE
+        WHEN COALESCE("usage".cache_creation_input_tokens, 0) = 0
+             AND (
+               COALESCE("usage".cache_creation_input_tokens_5m, 0)
+               + COALESCE("usage".cache_creation_input_tokens_1h, 0)
+             ) > 0
+        THEN COALESCE("usage".cache_creation_input_tokens_5m, 0)
+           + COALESCE("usage".cache_creation_input_tokens_1h, 0)
+        ELSE COALESCE("usage".cache_creation_input_tokens, 0)
+      END
+    + GREATEST(COALESCE("usage".cache_read_input_tokens, 0), 0)
+  ), 0)::BIGINT AS total_tokens,
  COALESCE(SUM(
    CASE
      WHEN COALESCE("usage".cache_creation_input_tokens, 0) = 0
@@ -3740,7 +3766,33 @@ SELECT
  "usage".model AS model,
  "usage".provider_name AS provider,
  COUNT(*)::BIGINT AS requests,
-  COALESCE(SUM(GREATEST(COALESCE("usage".total_tokens, 0), 0)), 0)::BIGINT AS total_tokens,
+  COALESCE(SUM(
+    CASE
+      WHEN GREATEST(COALESCE("usage".input_tokens, 0), 0) <= 0 THEN 0
+      WHEN GREATEST(COALESCE("usage".cache_read_input_tokens, 0), 0) <= 0
+      THEN GREATEST(COALESCE("usage".input_tokens, 0), 0)
+      WHEN split_part(lower(COALESCE(COALESCE("usage".endpoint_api_format, "usage".api_format), '')), ':', 1)
+           IN ('openai', 'gemini', 'google')
+      THEN GREATEST(
+        GREATEST(COALESCE("usage".input_tokens, 0), 0)
+          - GREATEST(COALESCE("usage".cache_read_input_tokens, 0), 0),
+        0
+      )
+      ELSE GREATEST(COALESCE("usage".input_tokens, 0), 0)
+    END
+    + GREATEST(COALESCE("usage".output_tokens, 0), 0)
+    + CASE
+        WHEN COALESCE("usage".cache_creation_input_tokens, 0) = 0
+             AND (
+               COALESCE("usage".cache_creation_input_tokens_5m, 0)
+               + COALESCE("usage".cache_creation_input_tokens_1h, 0)
+             ) > 0
+        THEN COALESCE("usage".cache_creation_input_tokens_5m, 0)
+           + COALESCE("usage".cache_creation_input_tokens_1h, 0)
+        ELSE COALESCE("usage".cache_creation_input_tokens, 0)
+      END
+    + GREATEST(COALESCE("usage".cache_read_input_tokens, 0), 0)
+  ), 0)::BIGINT AS total_tokens,
  COALESCE(SUM(COALESCE(CAST("usage".total_cost_usd AS DOUBLE PRECISION), 0)), 0)
    AS total_cost_usd,
  COALESCE(SUM(
@@ -3944,7 +3996,7 @@ SELECT
  {group_column} AS group_key,
  COALESCE(SUM(total_requests), 0)::BIGINT AS request_count,
  COALESCE(SUM(input_tokens), 0)::BIGINT AS input_tokens,
-  COALESCE(SUM(total_tokens), 0)::BIGINT AS total_tokens,
+  COALESCE(SUM(effective_input_tokens + output_tokens + cache_creation_tokens + cache_read_tokens), 0)::BIGINT AS total_tokens,
  COALESCE(SUM(output_tokens), 0)::BIGINT AS output_tokens,
  COALESCE(SUM(effective_input_tokens), 0)::BIGINT AS effective_input_tokens,
  COALESCE(SUM(total_input_context), 0)::BIGINT AS total_input_context,
@@ -4951,7 +5003,7 @@ SELECT
    AS cache_creation_cost_usd,
  COALESCE(SUM(
    COALESCE(
-      CAST("usage".output_price_per_1m AS DOUBLE PRECISION),
+      CAST("usage".input_price_per_1m AS DOUBLE PRECISION),
      0
    ) * GREATEST(COALESCE("usage".cache_read_input_tokens, 0), 0)::DOUBLE PRECISION / 1000000.0
  ), 0) AS estimated_full_cost_usd