fix: 治理 Prometheus 指标基数爆炸和内存缓存无界增长

- 移除 token/latency Prometheus 指标的 model 标签，避免 provider x model 笛卡尔积
- HealthMonitor 滑动窗口从 DB JSON 迁移至进程内存，减少写放大
- ModelCostService 三层缓存增加 500 条上限，超限时清空
- StickyPriority 粘性缓存和健康状态字典增加容量淘汰
- AffinityManager 请求锁字典增加 500 条上限，淘汰空闲锁
- 配额刷新/探测查询使用 defer/load_only 避免加载大 JSON 列
- Alembic 迁移清理 DB 中遗留的 request_results_window 数据
- 同步更新测试适配 batch_get_cooldowns 返回值和批量删除异步化

src/plugins/load_balancer/sticky_priority.py

@@ -78,10 +78,12 @@ class StickyPriorityStrategy(LoadBalancerStrategy):
                 "total_failures": 0,
             }
         )
+        self._max_provider_health_entries: int = 200  # 上限，防止无界增长
 
         # 当前粘性提供商缓存 {cache_key: provider_id}
         # cache_key 可以是 api_key_id 或者其他标识
         self._sticky_providers: dict[str, str] = {}
+        self._max_sticky_entries: int = 2000  # 上限，防止无界增长
 
         # 统计信息
         self._stats = {
@@ -260,6 +262,12 @@ class StickyPriorityStrategy(LoadBalancerStrategy):
 
         # 没有缓存或缓存失效，选择权重最大的
         sticky_candidate = max(candidates, key=lambda c: c.weight)
+        # 超出上限时清理不在当前候选列表中的旧条目
+        if len(self._sticky_providers) >= self._max_sticky_entries:
+            active_ids = {str(c.provider.id) for c in candidates}
+            stale = [k for k, v in self._sticky_providers.items() if v not in active_ids]
+            for k in stale[: len(stale) // 2 or len(stale)]:
+                del self._sticky_providers[k]
         self._sticky_providers[cache_key] = str(sticky_candidate.provider.id)
 
         logger.info(f"Set new sticky provider {sticky_candidate.provider.name}")
@@ -358,6 +366,17 @@ class StickyPriorityStrategy(LoadBalancerStrategy):
             error: 错误信息（如果失败）
         """
         provider_id = str(provider.id)
+        # 超出上限时淘汰健康且请求量最少的条目
+        if (
+            provider_id not in self._provider_health
+            and len(self._provider_health) >= self._max_provider_health_entries
+        ):
+            healthy_keys = [
+                k for k, v in self._provider_health.items() if v["is_healthy"] and k != provider_id
+            ]
+            if healthy_keys:
+                victim = min(healthy_keys, key=lambda k: self._provider_health[k]["total_requests"])
+                del self._provider_health[victim]
         health_info = self._provider_health[provider_id]
 
         health_info["total_requests"] += 1

src/plugins/monitor/base.py

@@ -209,7 +209,7 @@ class MonitorPlugin(BasePlugin):
     def record_token_usage(
         self,
         provider: str,
-        model: str,
+        model: str,  # noqa: ARG002 - 保留签名兼容性，不再用于 Prometheus 标签
         input_tokens: int,
         output_tokens: int,
         cost: float | None = None,
@@ -219,12 +219,12 @@ class MonitorPlugin(BasePlugin):
 
         Args:
             provider: 提供商名称
-            model: 模型名称
+            model: 模型名称（保留签名兼容性，不再作为 Prometheus 标签）
             input_tokens: 输入token数
             output_tokens: 输出token数
             cost: 费用
         """
-        labels = {"provider": provider, "model": model}
+        labels = {"provider": provider}
 
         import asyncio
 

src/plugins/monitor/prometheus.py

@@ -74,21 +74,19 @@ class PrometheusPlugin(MonitorPlugin):
             http_label_names,
         )
 
-        # Token使用指标
+        # Token使用指标（仅按 provider 维度聚合，避免 provider x model 笛卡尔积导致基数爆炸）
         self._metrics["tokens_input_total"] = Counter(
-            "tokens_input_total", "Total input tokens", ["provider", "model"]
+            "tokens_input_total", "Total input tokens", ["provider"]
         )
 
         self._metrics["tokens_output_total"] = Counter(
-            "tokens_output_total", "Total output tokens", ["provider", "model"]
+            "tokens_output_total", "Total output tokens", ["provider"]
         )
 
-        self._metrics["tokens_total"] = Counter(
-            "tokens_total", "Total tokens", ["provider", "model"]
-        )
+        self._metrics["tokens_total"] = Counter("tokens_total", "Total tokens", ["provider"])
 
         self._metrics["usage_cost_total"] = Counter(
-            "usage_cost_total", "Total usage cost in USD", ["provider", "model"]
+            "usage_cost_total", "Total usage cost in USD", ["provider"]
         )
 
         # 系统指标
@@ -112,7 +110,7 @@ class PrometheusPlugin(MonitorPlugin):
         self._metrics["provider_latency_seconds"] = Histogram(
             "provider_latency_seconds",
             "Provider response latency in seconds",
-            ["provider", "model"],
+            ["provider"],
             buckets=(0.1, 0.25, 0.5, 1, 2.5, 5, 10, 30),
         )