@@ -450,7 +461,12 @@ async function loadWorkerStatus() {
async function loadPerformanceHistory() {
try {
const res = await apiGet('/admin/api/performance-history', { params: { minutes: 30 } });
- historyData.value = res?.data?.snapshots || [];
+ const snapshots = res?.data?.snapshots || [];
+ // Pre-process: convert worker_rss_bytes to worker_rss_mb for trend chart
+ historyData.value = snapshots.map((s) => ({
+ ...s,
+ worker_rss_mb: s.worker_rss_bytes ? Math.round(s.worker_rss_bytes / 1048576 * 10) / 10 : 0,
+ }));
} catch (e) {
console.error('Failed to load performance history:', e);
}
@@ -460,6 +476,7 @@ async function loadPerformanceHistory() {
const poolTrendSeries = [
{ name: '飽和度', key: 'pool_saturation', color: '#6366f1' },
{ name: '使用中', key: 'pool_checked_out', color: '#f59e0b' },
+ { name: '慢查詢執行中', key: 'slow_query_active', color: '#ef4444' },
];
const latencyTrendSeries = [
@@ -478,6 +495,10 @@ const hitRateTrendSeries = [
{ name: 'L2 命中率', key: 'rc_l2_hit_rate', color: '#f59e0b' },
];
+const memoryTrendSeries = [
+ { name: 'RSS (MB)', key: 'worker_rss_mb', color: '#8b5cf6' },
+];
+
async function refreshAll() {
loading.value = true;
try {
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/.openspec.yaml b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/.openspec.yaml
new file mode 100644
index 0000000..85ae75c
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/.openspec.yaml
@@ -0,0 +1,2 @@
+schema: spec-driven
+created: 2026-02-26
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/design.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/design.md
new file mode 100644
index 0000000..c2f3680
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/design.md
@@ -0,0 +1,81 @@
+## Context
+
+2026-02-25 的 server crash 暴露出 pool 隔離架構變更後的監控盲區。event_fetcher 和 lineage_engine 已遷移到 `read_sql_df_slow`（獨立連線 + semaphore），但 metrics_history 快照只記錄 pool 相關指標，slow query 並行數、排隊數、Worker RSS 完全無歷史紀錄。
+
+同時 `/admin/performance` 仍保留 1249 行 Jinja template 作為 Vue SPA fallback，但 SPA 已是唯一使用的版本（build artifact 存在於 `static/dist/admin-performance.html`），兩套 UI 增加維護成本且 Jinja 版功能遠不及 SPA。
+
+`logs/archive/` 目錄累積 rotated log 檔案無自動清理，是唯一會無限增長的儲存。
+
+## Goals / Non-Goals
+
+**Goals:**
+- 移除 Jinja fallback，統一為 Vue SPA 單一架構
+- 讓 slow query 並行數、排隊數、Worker RSS 成為可觀測的歷史趨勢指標
+- 讓 P50/P95/P99 反映所有查詢路徑（pool + slow path）
+- 解決 archive log 無限增長問題
+
+**Non-Goals:**
+- 不修改 `/admin/pages`（仍為 Jinja template）
+- 不新增 async job queue 面板（P1，後續 change 處理）
+- 不新增 event cache hit/miss 計數器（P2）
+- 不增加即時告警或 webhook 通知機制
+
+## Decisions
+
+### D1：SQLite schema migration 策略
+
+**選擇**：啟動時執行 `ALTER TABLE ADD COLUMN IF NOT EXISTS`（容錯 "duplicate column" error）
+
+**替代方案**：version table + migration script → 過度工程，SQLite 只有 3 天保留，加欄是向後相容的
+
+**理由**：新欄位 nullable，舊 row 自動為 NULL，不影響既有查詢。MetricsHistoryStore.initialize() 已在啟動時執行 CREATE TABLE IF NOT EXISTS，加入 ALTER TABLE 語句自然整合。
+
+### D2：RSS 記憶體取得方式
+
+**選擇**：`resource.getrusage(resource.RUSAGE_SELF).ru_maxrss * 1024`（Python stdlib，Linux 上單位為 KB）
+
+**替代方案 A**：讀取 `/proc/self/status` VmRSS → 平台相依，解析 overhead
+**替代方案 B**：`psutil.Process().memory_info().rss` → 需新增外部依賴
+
+**理由**：`resource` 模組為 Python 標準庫，無需額外依賴。`ru_maxrss` 在 Linux 上返回 KB，乘以 1024 轉為 bytes。
+
+### D3：Semaphore 排隊計數器實作
+
+**選擇**：在 `read_sql_df_slow()` 的 semaphore.acquire() 前後遞增/遞減 `_SLOW_QUERY_WAITING` atomic counter
+
+**流程**：
+```
+_SLOW_QUERY_WAITING += 1
+acquired = semaphore.acquire(timeout=60)
+_SLOW_QUERY_WAITING -= 1
+if not acquired: raise RuntimeError
+_SLOW_QUERY_ACTIVE += 1
+... execute query ...
+_SLOW_QUERY_ACTIVE -= 1
+```
+
+**理由**：與既有 `_SLOW_QUERY_ACTIVE` 模式一致，使用 threading.Lock 保護。
+
+### D4：Archive log cleanup 整合位置
+
+**選擇**：整合到 `MetricsHistoryCollector._run()` 的 cleanup cycle（每 ~100 intervals ≈ 50 分鐘）
+
+**替代方案**：獨立 cron job → 需額外 crontab 配置，不自包含
+
+**理由**：已有 daemon thread 定期 cleanup SQLite，加入 archive cleanup 邏輯一致且自包含。
+
+### D5：移除 Jinja fallback 的安全性
+
+**選擇**：直接移除 fallback，admin_routes.py 改為只 `send_from_directory(dist_dir, "admin-performance.html")`
+
+**理由**：
+- Vue SPA build artifact 已存在（`static/dist/admin-performance.html`，2026-02-26 更新）
+- `frontend/package.json` build script 已包含 admin-performance entry
+- CI/deploy 流程必包含 `npx vite build`
+- 若 build 失敗，`/health/frontend-shell` 已有 asset readiness 檢查可偵測
+
+## Risks / Trade-offs
+
+- **[Risk] Build 失敗時 /admin/performance 返回 404** → 既有 `/health/frontend-shell` 檢查 + deploy script 驗證。移除 fallback 反而讓問題更早暴露。
+- **[Risk] ALTER TABLE 在 SQLite 大表上可能慢** → metrics_history 最多 50K rows，ALTER TABLE 即時完成。
+- **[Trade-off] `ru_maxrss` 是 peak RSS，非 current RSS** → 在 Linux 上 `ru_maxrss` 是 process lifetime 的 max RSS。改用 `/proc/self/status` 的 VmRSS 可取得 current，但需 file I/O。鑑於每 30 秒收集一次且 max RSS 更能反映記憶體壓力，接受此 trade-off。若日後需要 current RSS，可改讀 `/proc/self/status`。
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/proposal.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/proposal.md
new file mode 100644
index 0000000..d58c08f
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/proposal.md
@@ -0,0 +1,34 @@
+## Why
+
+2026-02-25 server crash 暴露出管理員效能監控頁面在 pool 隔離架構變更後的關鍵盲區：slow query 並行數、slow-path 延遲、Worker 記憶體等核心指標既未收集也未顯示，導致 crash 前完全無法觀測系統真實負載。同時，`/admin/performance` 仍保留 1249 行的 Jinja template 作為 fallback，與已完成的 Vue SPA 遷移架構不一致，增加維護成本。
+
+## What Changes
+
+- **移除** Jinja template `templates/admin/performance.html`，`/admin/performance` 路由直接服務 Vue SPA（`static/dist/admin-performance.html`），不再有 fallback 邏輯
+- **新增** `slow_query_active`、`slow_query_waiting`、`worker_rss_bytes` 三個欄位到 `metrics_history.sqlite` 快照，含 SQLite schema migration
+- **新增** semaphore 排隊計數器（`_SLOW_QUERY_WAITING`），追蹤等待 slow query semaphore 的 thread 數量
+- **修正** `read_sql_df_slow()` 和 `read_sql_df_slow_iter()` 將查詢延遲記錄到 `QueryMetrics`，使 P50/P95/P99 反映所有查詢路徑
+- **新增** Vue SPA 連線池區塊顯示「慢查詢執行中」「慢查詢排隊中」指標 + 連線池趨勢圖加入 slow_query_active 線 + Worker 記憶體趨勢圖
+- **新增** archive log 自動清理機制，整合到既有 `MetricsHistoryCollector` 的 cleanup cycle
+
+## Capabilities
+
+### New Capabilities
+
+- `slow-query-observability`: 追蹤 slow query 並行數、排隊數、延遲，寫入 metrics history 並在前端顯示趨勢
+- `worker-memory-tracking`: 追蹤 Worker RSS 記憶體，寫入 metrics history 並在前端顯示趨勢
+- `archive-log-rotation`: logs/archive/ 目錄的自動清理機制，防止檔案無限增長
+
+### Modified Capabilities
+
+- `admin-performance-spa`: 移除 Jinja template fallback，完全遷移至 Vue SPA，新增 slow query 與記憶體監控面板
+- `metrics-history-trending`: 擴充 snapshot schema 加入 slow_query_active、slow_query_waiting、worker_rss_bytes
+- `connection-pool-monitoring`: 新增 semaphore 排隊計數器，slow-path 延遲納入 QueryMetrics
+
+## Impact
+
+- **後端**：`core/database.py`（排隊計數器 + latency 記錄）、`core/metrics_history.py`（schema 擴充 + archive cleanup）、`routes/admin_routes.py`（移除 fallback）
+- **前端**：`frontend/src/admin-performance/App.vue`（新面板 + 趨勢圖）→ 需 rebuild
+- **刪除**：`templates/admin/performance.html`（1249 行）
+- **資料**：既有 `metrics_history.sqlite` 需 ALTER TABLE 加欄（向後相容，新欄位 nullable）
+- **測試**：既有 `test_performance_integration.py` 已測試 SPA 路徑，無需修改；需新增 schema migration 測試
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/admin-performance-spa/spec.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/admin-performance-spa/spec.md
new file mode 100644
index 0000000..843dfc2
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/admin-performance-spa/spec.md
@@ -0,0 +1,37 @@
+## MODIFIED Requirements
+
+### Requirement: Vue 3 SPA page replaces Jinja2 template
+The `/admin/performance` route SHALL serve the Vite-built `admin-performance.html` static file directly. The Jinja2 template fallback SHALL be removed. If the SPA build artifact does not exist, the server SHALL return a standard HTTP error (no fallback rendering).
+
+#### Scenario: Page loads as Vue SPA
+- **WHEN** user navigates to `/admin/performance`
+- **THEN** the server SHALL return the Vite-built `admin-performance.html` static file via `send_from_directory`
+
+#### Scenario: Portal-shell integration
+- **WHEN** the portal-shell renders `/admin/performance`
+- **THEN** it SHALL load the page as a native Vue SPA (not an external iframe)
+
+#### Scenario: Build artifact missing
+- **WHEN** the SPA build artifact `admin-performance.html` does not exist in `static/dist/`
+- **THEN** the server SHALL return an HTTP error (no Jinja2 fallback)
+
+### Requirement: Connection pool panel
+The dashboard SHALL display connection pool saturation as a GaugeBar and stat cards showing checked_out, checked_in, overflow, max_capacity, pool_size, pool_recycle, pool_timeout, direct connection count, slow_query_active, and slow_query_waiting.
+
+#### Scenario: Pool under normal load
+- **WHEN** pool saturation is below 80%
+- **THEN** the GaugeBar SHALL display in a normal color (green/blue)
+
+#### Scenario: Pool near saturation
+- **WHEN** pool saturation exceeds 80%
+- **THEN** the GaugeBar SHALL display in a warning color (yellow/orange/red)
+
+#### Scenario: Slow query metrics displayed
+- **WHEN** `db_pool.status` includes `slow_query_active` and `slow_query_waiting`
+- **THEN** the panel SHALL display StatCards for both values
+
+## REMOVED Requirements
+
+### Requirement: Jinja2 template fallback for performance page
+**Reason**: The Vue SPA is the sole UI. Maintaining a 1249-line Jinja template as fallback adds maintenance burden and feature divergence.
+**Migration**: Delete `templates/admin/performance.html`. The route handler serves the SPA directly.
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/archive-log-rotation/spec.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/archive-log-rotation/spec.md
new file mode 100644
index 0000000..a8bbbeb
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/archive-log-rotation/spec.md
@@ -0,0 +1,30 @@
+## ADDED Requirements
+
+### Requirement: Automatic archive log cleanup
+The system SHALL provide a `cleanup_archive_logs()` function in `core/metrics_history.py` that deletes old rotated log files from `logs/archive/`, keeping the most recent N files per log type (access, error, watchdog, rq_worker, startup).
+
+#### Scenario: Cleanup keeps recent files
+- **WHEN** `cleanup_archive_logs()` is called with `keep_per_type=20` and there are 30 access_*.log files
+- **THEN** 10 oldest access_*.log files SHALL be deleted, keeping the 20 most recent by modification time
+
+#### Scenario: No excess files
+- **WHEN** `cleanup_archive_logs()` is called and each type has fewer than `keep_per_type` files
+- **THEN** no files SHALL be deleted
+
+#### Scenario: Archive directory missing
+- **WHEN** `cleanup_archive_logs()` is called and the archive directory does not exist
+- **THEN** the function SHALL return 0 without error
+
+### Requirement: Archive cleanup integrated into collector cycle
+The `MetricsHistoryCollector` SHALL call `cleanup_archive_logs()` alongside the existing SQLite cleanup, running approximately every 50 minutes (every 100 collection intervals).
+
+#### Scenario: Periodic cleanup executes
+- **WHEN** the cleanup counter reaches 100 intervals
+- **THEN** both SQLite metrics cleanup and archive log cleanup SHALL execute
+
+### Requirement: Archive cleanup configuration
+The archive log cleanup SHALL be configurable via environment variables: `ARCHIVE_LOG_DIR` (default: `logs/archive`) and `ARCHIVE_LOG_KEEP_COUNT` (default: 20).
+
+#### Scenario: Custom keep count
+- **WHEN** `ARCHIVE_LOG_KEEP_COUNT=10` is set
+- **THEN** cleanup SHALL keep only the 10 most recent files per type
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/connection-pool-monitoring/spec.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/connection-pool-monitoring/spec.md
new file mode 100644
index 0000000..4902f47
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/connection-pool-monitoring/spec.md
@@ -0,0 +1,29 @@
+## MODIFIED Requirements
+
+### Requirement: Connection pool status in performance detail
+The performance-detail API SHALL include `db_pool` section with `status` (checked_out, checked_in, overflow, max_capacity, saturation, slow_query_active, slow_query_waiting) from `get_pool_status()` and `config` (pool_size, max_overflow, pool_timeout, pool_recycle) from `get_pool_runtime_config()`.
+
+#### Scenario: Pool status retrieved
+- **WHEN** the API is called
+- **THEN** `db_pool.status` SHALL contain current pool utilization metrics including `slow_query_active` and `slow_query_waiting`, and `db_pool.config` SHALL contain the pool configuration values
+
+#### Scenario: Saturation calculation
+- **WHEN** the pool has 8 checked_out connections and max_capacity is 30
+- **THEN** saturation SHALL be reported as approximately 26.7%
+
+#### Scenario: Slow query waiting included
+- **WHEN** 2 threads are waiting for the slow query semaphore
+- **THEN** `db_pool.status.slow_query_waiting` SHALL be 2
+
+## ADDED Requirements
+
+### Requirement: Slow-path query latency included in QueryMetrics
+The `read_sql_df_slow()` and `read_sql_df_slow_iter()` functions SHALL call `record_query_latency()` with the total elapsed time upon completion, ensuring P50/P95/P99 percentiles reflect queries from all paths (pooled and slow/direct).
+
+#### Scenario: Slow query latency recorded
+- **WHEN** `read_sql_df_slow()` completes a query in 8.5 seconds
+- **THEN** `record_query_latency(8.5)` SHALL be called and the value SHALL appear in subsequent `get_percentiles()` results
+
+#### Scenario: Slow iter latency recorded
+- **WHEN** `read_sql_df_slow_iter()` completes streaming in 45 seconds
+- **THEN** `record_query_latency(45.0)` SHALL be called in the finally block
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/metrics-history-trending/spec.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/metrics-history-trending/spec.md
new file mode 100644
index 0000000..5e1c4b6
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/metrics-history-trending/spec.md
@@ -0,0 +1,54 @@
+## MODIFIED Requirements
+
+### Requirement: SQLite metrics history store
+The system SHALL provide a `MetricsHistoryStore` class in `core/metrics_history.py` that persists metrics snapshots to a SQLite database (`logs/metrics_history.sqlite` by default). The store SHALL use thread-local connections and a write lock, following the `LogStore` pattern in `core/log_store.py`. The schema SHALL include columns for `slow_query_active` (INTEGER), `slow_query_waiting` (INTEGER), and `worker_rss_bytes` (INTEGER) in addition to the existing pool, Redis, route cache, and latency columns.
+
+#### Scenario: Write and query snapshots
+- **WHEN** `write_snapshot(data)` is called with pool/redis/route_cache/latency/slow_query/memory metrics
+- **THEN** a row SHALL be inserted into `metrics_snapshots` with the current ISO 8601 timestamp, worker PID, and all metric columns
+
+#### Scenario: Query by time range
+- **WHEN** `query_snapshots(minutes=30)` is called
+- **THEN** it SHALL return all rows from the last 30 minutes, ordered by timestamp ascending, including the new columns
+
+#### Scenario: Retention cleanup
+- **WHEN** `cleanup()` is called
+- **THEN** rows older than `METRICS_HISTORY_RETENTION_DAYS` (default 3) SHALL be deleted, and total rows SHALL be capped at `METRICS_HISTORY_MAX_ROWS` (default 50000)
+
+#### Scenario: Thread safety
+- **WHEN** multiple threads write snapshots concurrently
+- **THEN** the write lock SHALL serialize writes and prevent database corruption
+
+#### Scenario: Schema migration for existing databases
+- **WHEN** the store initializes on an existing database without the new columns
+- **THEN** it SHALL execute ALTER TABLE ADD COLUMN for each missing column, tolerating "duplicate column" errors
+
+### Requirement: Background metrics collector
+The system SHALL provide a `MetricsHistoryCollector` class that runs a daemon thread collecting metrics snapshots at a configurable interval (default 30 seconds, via `METRICS_HISTORY_INTERVAL` env var). The collector SHALL include `slow_query_active`, `slow_query_waiting`, and `worker_rss_bytes` in each snapshot.
+
+#### Scenario: Automatic collection
+- **WHEN** the collector is started via `start_metrics_history(app)`
+- **THEN** it SHALL collect pool status (including slow_query_active and slow_query_waiting), Redis info, route cache status, query latency metrics, and worker RSS memory every interval and write them to the store
+
+#### Scenario: Graceful shutdown
+- **WHEN** `stop_metrics_history()` is called
+- **THEN** the collector thread SHALL stop within one interval period
+
+#### Scenario: Subsystem unavailability
+- **WHEN** a subsystem (e.g., Redis) is unavailable during collection
+- **THEN** the collector SHALL write null/0 for those fields and continue collecting other metrics
+
+### Requirement: Frontend trend charts
+The system SHALL display 5 trend chart panels in the admin performance dashboard using vue-echarts VChart line/area charts: connection pool saturation, query latency (P50/P95/P99), Redis memory, cache hit rates, and worker memory.
+
+#### Scenario: Trend charts with data
+- **WHEN** historical snapshots contain more than 1 data point
+- **THEN** the dashboard SHALL display trend charts for: connection pool saturation (including slow_query_active), query latency (P50/P95/P99), Redis memory, cache hit rates, and worker memory (RSS in MB)
+
+#### Scenario: Trend charts without data
+- **WHEN** historical snapshots are empty or contain only 1 data point
+- **THEN** the trend charts SHALL NOT be displayed (hidden via `v-if`)
+
+#### Scenario: Auto-refresh
+- **WHEN** the dashboard auto-refreshes
+- **THEN** historical data SHALL also be refreshed alongside real-time metrics
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/slow-query-observability/spec.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/slow-query-observability/spec.md
new file mode 100644
index 0000000..fb3e130
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/slow-query-observability/spec.md
@@ -0,0 +1,49 @@
+## ADDED Requirements
+
+### Requirement: Slow query active count in metrics history snapshots
+The `MetricsHistoryCollector` SHALL include `slow_query_active` in each 30-second snapshot, recording the number of slow queries currently executing via dedicated connections.
+
+#### Scenario: Snapshot includes slow_query_active
+- **WHEN** the collector writes a snapshot while 3 slow queries are executing
+- **THEN** the `slow_query_active` column SHALL contain the value 3
+
+#### Scenario: No slow queries active
+- **WHEN** the collector writes a snapshot while no slow queries are executing
+- **THEN** the `slow_query_active` column SHALL contain the value 0
+
+### Requirement: Slow query waiting count tracked and persisted
+The system SHALL maintain a thread-safe counter `_SLOW_QUERY_WAITING` in `database.py` that tracks the number of threads currently waiting to acquire the slow query semaphore. This counter SHALL be included in `get_pool_status()` and persisted to metrics history snapshots.
+
+#### Scenario: Counter increments on semaphore wait
+- **WHEN** a thread enters `read_sql_df_slow()` and the semaphore is full
+- **THEN** `_SLOW_QUERY_WAITING` SHALL be incremented before `semaphore.acquire()` and decremented after acquire completes (success or timeout)
+
+#### Scenario: Counter in pool status API
+- **WHEN** `get_pool_status()` is called
+- **THEN** the returned dict SHALL include `slow_query_waiting` with the current waiting thread count
+
+#### Scenario: Counter persisted to metrics history
+- **WHEN** the collector writes a snapshot
+- **THEN** the `slow_query_waiting` column SHALL reflect the count at snapshot time
+
+### Requirement: Slow-path query latency recorded in QueryMetrics
+The `read_sql_df_slow()` and `read_sql_df_slow_iter()` functions SHALL call `record_query_latency()` with the elapsed query time, so that P50/P95/P99 metrics reflect all query paths (pool + slow).
+
+#### Scenario: Slow query latency appears in percentiles
+- **WHEN** a `read_sql_df_slow()` call completes in 5.2 seconds
+- **THEN** `record_query_latency(5.2)` SHALL be called and the latency SHALL appear in subsequent `get_percentiles()` results
+
+#### Scenario: Slow iter latency recorded on completion
+- **WHEN** a `read_sql_df_slow_iter()` generator completes after yielding all batches in 120 seconds total
+- **THEN** `record_query_latency(120.0)` SHALL be called in the finally block
+
+### Requirement: Slow query metrics displayed in Vue SPA
+The admin performance Vue SPA SHALL display `slow_query_active` and `slow_query_waiting` as StatCards in the connection pool panel, and include `slow_query_active` as a trend line in the connection pool trend chart.
+
+#### Scenario: StatCards display current values
+- **WHEN** the performance-detail API returns `db_pool.status.slow_query_active = 4` and `db_pool.status.slow_query_waiting = 2`
+- **THEN** the connection pool panel SHALL display StatCards showing "慢查詢執行中: 4" and "慢查詢排隊中: 2"
+
+#### Scenario: Trend chart includes slow_query_active
+- **WHEN** historical snapshots contain `slow_query_active` data points
+- **THEN** the connection pool trend chart SHALL include a "慢查詢執行中" line series
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/worker-memory-tracking/spec.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/worker-memory-tracking/spec.md
new file mode 100644
index 0000000..73c3716
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/specs/worker-memory-tracking/spec.md
@@ -0,0 +1,23 @@
+## ADDED Requirements
+
+### Requirement: Worker RSS memory in metrics history snapshots
+The `MetricsHistoryCollector` SHALL include `worker_rss_bytes` in each 30-second snapshot, recording the current worker process peak RSS memory using Python's `resource.getrusage()`.
+
+#### Scenario: RSS recorded in snapshot
+- **WHEN** the collector writes a snapshot and the worker process has 256 MB peak RSS
+- **THEN** the `worker_rss_bytes` column SHALL contain approximately 268435456
+
+#### Scenario: RSS collection failure
+- **WHEN** `resource.getrusage()` raises an exception
+- **THEN** the collector SHALL write NULL for `worker_rss_bytes` and continue collecting other metrics
+
+### Requirement: Worker memory trend chart in Vue SPA
+The admin performance Vue SPA SHALL display a "Worker 記憶體趨勢" TrendChart showing RSS memory over time in megabytes.
+
+#### Scenario: Memory trend displayed
+- **WHEN** historical snapshots contain `worker_rss_bytes` data with more than 1 data point
+- **THEN** the dashboard SHALL display a TrendChart with RSS values converted to MB
+
+#### Scenario: No memory data
+- **WHEN** historical snapshots do not contain `worker_rss_bytes` data (all NULL)
+- **THEN** the trend chart SHALL show "趨勢資料不足" message
diff --git a/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/tasks.md b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/tasks.md
new file mode 100644
index 0000000..e35cfd7
--- /dev/null
+++ b/openspec/changes/archive/2026-02-26-admin-perf-vue-migration-monitoring-gaps/tasks.md
@@ -0,0 +1,37 @@
+## 1. 後端：Semaphore 排隊計數器 + Slow-path latency
+
+- [x] 1.1 在 `src/mes_dashboard/core/database.py` 新增 `_SLOW_QUERY_WAITING` counter 和 `get_slow_query_waiting_count()` 函數
+- [x] 1.2 修改 `read_sql_df_slow()` 在 semaphore.acquire() 前後遞增/遞減 `_SLOW_QUERY_WAITING`
+- [x] 1.3 修改 `read_sql_df_slow_iter()` 同上加入 waiting counter 邏輯
+- [x] 1.4 修改 `get_pool_status()` 回傳中加入 `slow_query_waiting` 欄位
+- [x] 1.5 在 `read_sql_df_slow()` finally block 呼叫 `record_query_latency(elapsed)`
+- [x] 1.6 在 `read_sql_df_slow_iter()` finally block 呼叫 `record_query_latency(elapsed)`
+
+## 2. 後端：metrics_history schema 擴充 + archive cleanup
+
+- [x] 2.1 在 `src/mes_dashboard/core/metrics_history.py` 的 schema 新增 `slow_query_active INTEGER`, `slow_query_waiting INTEGER`, `worker_rss_bytes INTEGER` 欄位
+- [x] 2.2 在 `MetricsHistoryStore.initialize()` 加入 ALTER TABLE ADD COLUMN migration（容錯 duplicate column）
+- [x] 2.3 更新 `COLUMNS` list 加入新欄位
+- [x] 2.4 更新 `write_snapshot()` 加入新欄位的讀取和 INSERT
+- [x] 2.5 更新 `_collect_snapshot()` 收集 `slow_query_active`、`slow_query_waiting`（從 `get_pool_status()`）和 `worker_rss_bytes`（從 `resource.getrusage()`）
+- [x] 2.6 新增 `cleanup_archive_logs(archive_dir, keep_per_type)` 函數，含 `ARCHIVE_LOG_DIR` 和 `ARCHIVE_LOG_KEEP_COUNT` env var 配置
+- [x] 2.7 在 `MetricsHistoryCollector._run()` 的 cleanup cycle 呼叫 `cleanup_archive_logs()`
+
+## 3. 後端：移除 Jinja fallback
+
+- [x] 3.1 修改 `src/mes_dashboard/routes/admin_routes.py` 的 `performance()` 路由，移除 Jinja fallback 邏輯（改為直接 `send_from_directory`）
+- [x] 3.2 刪除 `src/mes_dashboard/templates/admin/performance.html`
+- [x] 3.3 更新 `scripts/check_full_modernization_gates.py` 將 `/admin/performance` 的 gate check 從 template 路徑改為 `frontend/src/admin-performance/style.css`
+
+## 4. 前端：Vue SPA 新增監控面板
+
+- [x] 4.1 在 `frontend/src/admin-performance/App.vue` 連線池 section 新增 `slow_query_active` 和 `slow_query_waiting` StatCards
+- [x] 4.2 在 `poolTrendSeries` 加入 `slow_query_active` 趨勢線
+- [x] 4.3 新增 `memoryTrendSeries` 定義和 Worker 記憶體 TrendChart 組件
+- [x] 4.4 新增 `historyData` 預處理邏輯：將 `worker_rss_bytes` 轉為 `worker_rss_mb`
+
+## 5. Build + 測試驗證
+
+- [x] 5.1 執行 `cd frontend && npx vite build` 確認 build 成功
+- [x] 5.2 執行 `python -m pytest tests/ -v --tb=short` 確認既有測試通過
+- [x] 5.3 確認 `test_performance_page_loads` 測試通過（SPA 路徑驗證）
diff --git a/openspec/specs/admin-performance-spa/spec.md b/openspec/specs/admin-performance-spa/spec.md
index 431b7a0..843dfc2 100644
--- a/openspec/specs/admin-performance-spa/spec.md
+++ b/openspec/specs/admin-performance-spa/spec.md
@@ -1,100 +1,37 @@
-## ADDED Requirements
-
-### Requirement: Vue 3 SPA page replaces Jinja2 template
-The `/admin/performance` route SHALL serve a Vue 3 SPA page built by Vite, replacing the existing Jinja2 server-rendered template. The SPA SHALL be registered as a Vite entry point and integrated into the portal-shell navigation as a `renderMode: 'native'` route.
-
-#### Scenario: Page loads as Vue SPA
-- **WHEN** user navigates to `/admin/performance`
-- **THEN** the server SHALL return the Vite-built `admin-performance.html` static file (not a Jinja2 rendered template)
-
-#### Scenario: Portal-shell integration
-- **WHEN** the portal-shell renders `/admin/performance`
-- **THEN** it SHALL load the page as a native Vue SPA (not an external iframe)
-
-### Requirement: Status cards display system health
-The dashboard SHALL display 4 status cards in a horizontal grid: Database, Redis, Circuit Breaker, and Worker PID. Each card SHALL show a StatusDot indicator (healthy/degraded/error/disabled) with the current status value.
-
-#### Scenario: All systems healthy
-- **WHEN** all backend systems report healthy status via `/admin/api/system-status`
-- **THEN** all 4 status cards SHALL display green StatusDot indicators with their respective values
-
-#### Scenario: Redis disabled
-- **WHEN** Redis is disabled (`REDIS_ENABLED=false`)
-- **THEN** the Redis status card SHALL display a disabled StatusDot indicator and the Redis cache panel SHALL show a graceful degradation message
-
-### Requirement: Query performance panel with ECharts
-The dashboard SHALL display query performance metrics (P50, P95, P99 latencies, total queries, slow queries) and an ECharts latency distribution chart, replacing the existing Chart.js implementation.
-
-#### Scenario: Metrics loaded successfully
-- **WHEN** `/admin/api/metrics` returns valid performance data
-- **THEN** the panel SHALL display P50/P95/P99 latency values and render an ECharts bar chart showing latency distribution
-
-#### Scenario: No metrics data
-- **WHEN** `/admin/api/metrics` returns empty or null metrics
-- **THEN** the panel SHALL display placeholder text indicating no data available
-
-### Requirement: Redis cache detail panel
-The dashboard SHALL display a Redis cache detail panel showing memory usage (as a GaugeBar), connected clients, hit rate percentage, peak memory, and a namespace key distribution table.
-
-#### Scenario: Redis active with data
-- **WHEN** `/admin/api/performance-detail` returns Redis data with namespace key counts
-- **THEN** the panel SHALL display a memory GaugeBar, hit rate, client count, and a table listing each namespace with its key count
-
-#### Scenario: Redis disabled
-- **WHEN** Redis is disabled
-- **THEN** the Redis detail panel SHALL display a disabled state message without errors
-
-### Requirement: Memory cache panel
-The dashboard SHALL display ProcessLevelCache statistics as grid cards (showing entries/max_size as a mini gauge and TTL) plus Route Cache telemetry (L1 hit rate, L2 hit rate, miss rate, total reads).
-
-#### Scenario: Multiple caches registered
-- **WHEN** `/admin/api/performance-detail` returns process_caches with multiple entries
-- **THEN** the panel SHALL render one card per cache instance showing entries, max_size, TTL, and description
-
-#### Scenario: Route cache telemetry
-- **WHEN** `/admin/api/performance-detail` returns route_cache data
-- **THEN** the panel SHALL display L1 hit rate, L2 hit rate, miss rate, and total reads
-
-### Requirement: Connection pool panel
-The dashboard SHALL display connection pool saturation as a GaugeBar and stat cards showing checked_out, checked_in, overflow, max_capacity, pool_size, pool_recycle, pool_timeout, and direct connection count.
-
-#### Scenario: Pool under normal load
-- **WHEN** pool saturation is below 80%
-- **THEN** the GaugeBar SHALL display in a normal color (green/blue)
-
-#### Scenario: Pool near saturation
-- **WHEN** pool saturation exceeds 80%
-- **THEN** the GaugeBar SHALL display in a warning color (yellow/orange/red)
-
-### Requirement: Worker control panel
-The dashboard SHALL display worker PID, uptime, cooldown status, and provide a restart button with a confirmation modal.
-
-#### Scenario: Restart worker
-- **WHEN** user clicks the restart button and confirms in the modal
-- **THEN** the system SHALL POST to `/admin/api/worker/restart` and display the result
-
-#### Scenario: Restart during cooldown
-- **WHEN** worker is in cooldown period
-- **THEN** the restart button SHALL be disabled with a cooldown indicator
-
-### Requirement: System logs panel with filtering and pagination
-The dashboard SHALL display system logs with level filtering, text search, and pagination controls.
-
-#### Scenario: Filter by log level
-- **WHEN** user selects a specific log level filter
-- **THEN** only logs matching that level SHALL be displayed
-
-#### Scenario: Paginate logs
-- **WHEN** logs exceed the page size
-- **THEN** pagination controls SHALL allow navigating between pages
-
-### Requirement: Auto-refresh with toggle
-The dashboard SHALL auto-refresh all panels every 30 seconds using `useAutoRefresh`. The user SHALL be able to toggle auto-refresh on/off and manually trigger a refresh.
-
-#### Scenario: Auto-refresh enabled
-- **WHEN** auto-refresh is enabled (default)
-- **THEN** all panels SHALL refresh their data every 30 seconds via `Promise.all` parallel fetch
-
-#### Scenario: Manual refresh
-- **WHEN** user clicks the manual refresh button
-- **THEN** all panels SHALL immediately refresh their data
+## MODIFIED Requirements
+
+### Requirement: Vue 3 SPA page replaces Jinja2 template
+The `/admin/performance` route SHALL serve the Vite-built `admin-performance.html` static file directly. The Jinja2 template fallback SHALL be removed. If the SPA build artifact does not exist, the server SHALL return a standard HTTP error (no fallback rendering).
+
+#### Scenario: Page loads as Vue SPA
+- **WHEN** user navigates to `/admin/performance`
+- **THEN** the server SHALL return the Vite-built `admin-performance.html` static file via `send_from_directory`
+
+#### Scenario: Portal-shell integration
+- **WHEN** the portal-shell renders `/admin/performance`
+- **THEN** it SHALL load the page as a native Vue SPA (not an external iframe)
+
+#### Scenario: Build artifact missing
+- **WHEN** the SPA build artifact `admin-performance.html` does not exist in `static/dist/`
+- **THEN** the server SHALL return an HTTP error (no Jinja2 fallback)
+
+### Requirement: Connection pool panel
+The dashboard SHALL display connection pool saturation as a GaugeBar and stat cards showing checked_out, checked_in, overflow, max_capacity, pool_size, pool_recycle, pool_timeout, direct connection count, slow_query_active, and slow_query_waiting.
+
+#### Scenario: Pool under normal load
+- **WHEN** pool saturation is below 80%
+- **THEN** the GaugeBar SHALL display in a normal color (green/blue)
+
+#### Scenario: Pool near saturation
+- **WHEN** pool saturation exceeds 80%
+- **THEN** the GaugeBar SHALL display in a warning color (yellow/orange/red)
+
+#### Scenario: Slow query metrics displayed
+- **WHEN** `db_pool.status` includes `slow_query_active` and `slow_query_waiting`
+- **THEN** the panel SHALL display StatCards for both values
+
+## REMOVED Requirements
+
+### Requirement: Jinja2 template fallback for performance page
+**Reason**: The Vue SPA is the sole UI. Maintaining a 1249-line Jinja template as fallback adds maintenance burden and feature divergence.
+**Migration**: Delete `templates/admin/performance.html`. The route handler serves the SPA directly.
diff --git a/openspec/specs/archive-log-rotation/spec.md b/openspec/specs/archive-log-rotation/spec.md
new file mode 100644
index 0000000..a8bbbeb
--- /dev/null
+++ b/openspec/specs/archive-log-rotation/spec.md
@@ -0,0 +1,30 @@
+## ADDED Requirements
+
+### Requirement: Automatic archive log cleanup
+The system SHALL provide a `cleanup_archive_logs()` function in `core/metrics_history.py` that deletes old rotated log files from `logs/archive/`, keeping the most recent N files per log type (access, error, watchdog, rq_worker, startup).
+
+#### Scenario: Cleanup keeps recent files
+- **WHEN** `cleanup_archive_logs()` is called with `keep_per_type=20` and there are 30 access_*.log files
+- **THEN** 10 oldest access_*.log files SHALL be deleted, keeping the 20 most recent by modification time
+
+#### Scenario: No excess files
+- **WHEN** `cleanup_archive_logs()` is called and each type has fewer than `keep_per_type` files
+- **THEN** no files SHALL be deleted
+
+#### Scenario: Archive directory missing
+- **WHEN** `cleanup_archive_logs()` is called and the archive directory does not exist
+- **THEN** the function SHALL return 0 without error
+
+### Requirement: Archive cleanup integrated into collector cycle
+The `MetricsHistoryCollector` SHALL call `cleanup_archive_logs()` alongside the existing SQLite cleanup, running approximately every 50 minutes (every 100 collection intervals).
+
+#### Scenario: Periodic cleanup executes
+- **WHEN** the cleanup counter reaches 100 intervals
+- **THEN** both SQLite metrics cleanup and archive log cleanup SHALL execute
+
+### Requirement: Archive cleanup configuration
+The archive log cleanup SHALL be configurable via environment variables: `ARCHIVE_LOG_DIR` (default: `logs/archive`) and `ARCHIVE_LOG_KEEP_COUNT` (default: 20).
+
+#### Scenario: Custom keep count
+- **WHEN** `ARCHIVE_LOG_KEEP_COUNT=10` is set
+- **THEN** cleanup SHALL keep only the 10 most recent files per type
diff --git a/openspec/specs/connection-pool-monitoring/spec.md b/openspec/specs/connection-pool-monitoring/spec.md
index ae71a86..4902f47 100644
--- a/openspec/specs/connection-pool-monitoring/spec.md
+++ b/openspec/specs/connection-pool-monitoring/spec.md
@@ -1,27 +1,29 @@
-## ADDED Requirements
-
-### Requirement: Connection pool status in performance detail
-The performance-detail API SHALL include `db_pool` section with `status` (checked_out, checked_in, overflow, max_capacity, saturation) from `get_pool_status()` and `config` (pool_size, max_overflow, pool_timeout, pool_recycle) from `get_pool_runtime_config()`.
-
-#### Scenario: Pool status retrieved
-- **WHEN** the API is called
-- **THEN** `db_pool.status` SHALL contain current pool utilization metrics and `db_pool.config` SHALL contain the pool configuration values
-
-#### Scenario: Saturation calculation
-- **WHEN** the pool has 8 checked_out connections and max_capacity is 30
-- **THEN** saturation SHALL be reported as approximately 26.7%
-
-### Requirement: Direct Oracle connection counter
-The system SHALL maintain a thread-safe monotonic counter in `database.py` that increments each time `get_db_connection()` or `read_sql_df_slow()` successfully creates a direct (non-pooled) Oracle connection.
-
-#### Scenario: Counter increments on direct connection
-- **WHEN** `get_db_connection()` successfully creates a connection
-- **THEN** the direct connection counter SHALL increment by 1
-
-#### Scenario: Counter in performance detail
-- **WHEN** the performance-detail API is called
-- **THEN** `direct_connections` SHALL contain `total_since_start` (counter value) and `worker_pid` (current process PID)
-
-#### Scenario: Counter is per-worker
-- **WHEN** multiple gunicorn workers are running
-- **THEN** each worker SHALL maintain its own independent counter, and the API SHALL return the counter for the responding worker
+## MODIFIED Requirements
+
+### Requirement: Connection pool status in performance detail
+The performance-detail API SHALL include `db_pool` section with `status` (checked_out, checked_in, overflow, max_capacity, saturation, slow_query_active, slow_query_waiting) from `get_pool_status()` and `config` (pool_size, max_overflow, pool_timeout, pool_recycle) from `get_pool_runtime_config()`.
+
+#### Scenario: Pool status retrieved
+- **WHEN** the API is called
+- **THEN** `db_pool.status` SHALL contain current pool utilization metrics including `slow_query_active` and `slow_query_waiting`, and `db_pool.config` SHALL contain the pool configuration values
+
+#### Scenario: Saturation calculation
+- **WHEN** the pool has 8 checked_out connections and max_capacity is 30
+- **THEN** saturation SHALL be reported as approximately 26.7%
+
+#### Scenario: Slow query waiting included
+- **WHEN** 2 threads are waiting for the slow query semaphore
+- **THEN** `db_pool.status.slow_query_waiting` SHALL be 2
+
+## ADDED Requirements
+
+### Requirement: Slow-path query latency included in QueryMetrics
+The `read_sql_df_slow()` and `read_sql_df_slow_iter()` functions SHALL call `record_query_latency()` with the total elapsed time upon completion, ensuring P50/P95/P99 percentiles reflect queries from all paths (pooled and slow/direct).
+
+#### Scenario: Slow query latency recorded
+- **WHEN** `read_sql_df_slow()` completes a query in 8.5 seconds
+- **THEN** `record_query_latency(8.5)` SHALL be called and the value SHALL appear in subsequent `get_percentiles()` results
+
+#### Scenario: Slow iter latency recorded
+- **WHEN** `read_sql_df_slow_iter()` completes streaming in 45 seconds
+- **THEN** `record_query_latency(45.0)` SHALL be called in the finally block
diff --git a/openspec/specs/metrics-history-trending/spec.md b/openspec/specs/metrics-history-trending/spec.md
index c13633b..5e1c4b6 100644
--- a/openspec/specs/metrics-history-trending/spec.md
+++ b/openspec/specs/metrics-history-trending/spec.md
@@ -1,65 +1,54 @@
-## ADDED Requirements
-
-### Requirement: SQLite metrics history store
-The system SHALL provide a `MetricsHistoryStore` class in `core/metrics_history.py` that persists metrics snapshots to a SQLite database (`logs/metrics_history.sqlite` by default). The store SHALL use thread-local connections and a write lock, following the `LogStore` pattern in `core/log_store.py`.
-
-#### Scenario: Write and query snapshots
-- **WHEN** `write_snapshot(data)` is called with pool/redis/route_cache/latency metrics
-- **THEN** a row SHALL be inserted into `metrics_snapshots` with the current ISO 8601 timestamp and worker PID
-
-#### Scenario: Query by time range
-- **WHEN** `query_snapshots(minutes=30)` is called
-- **THEN** it SHALL return all rows from the last 30 minutes, ordered by timestamp ascending
-
-#### Scenario: Retention cleanup
-- **WHEN** `cleanup()` is called
-- **THEN** rows older than `METRICS_HISTORY_RETENTION_DAYS` (default 3) SHALL be deleted, and total rows SHALL be capped at `METRICS_HISTORY_MAX_ROWS` (default 50000)
-
-#### Scenario: Thread safety
-- **WHEN** multiple threads write snapshots concurrently
-- **THEN** the write lock SHALL serialize writes and prevent database corruption
-
-### Requirement: Background metrics collector
-The system SHALL provide a `MetricsHistoryCollector` class that runs a daemon thread collecting metrics snapshots at a configurable interval (default 30 seconds, via `METRICS_HISTORY_INTERVAL` env var).
-
-#### Scenario: Automatic collection
-- **WHEN** the collector is started via `start_metrics_history(app)`
-- **THEN** it SHALL collect pool status, Redis info, route cache status, and query latency metrics every interval and write them to the store
-
-#### Scenario: Graceful shutdown
-- **WHEN** `stop_metrics_history()` is called
-- **THEN** the collector thread SHALL stop within one interval period
-
-#### Scenario: Subsystem unavailability
-- **WHEN** a subsystem (e.g., Redis) is unavailable during collection
-- **THEN** the collector SHALL write null/0 for those fields and continue collecting other metrics
-
-### Requirement: Performance history API endpoint
-The system SHALL expose `GET /admin/api/performance-history` that returns historical metrics snapshots.
-
-#### Scenario: Query with time range
-- **WHEN** the API is called with `?minutes=30`
-- **THEN** it SHALL return `{"success": true, "data": {"snapshots": [...], "count": N}}`
-
-#### Scenario: Time range bounds
-- **WHEN** `minutes` is less than 1 or greater than 180
-- **THEN** it SHALL be clamped to the range [1, 180]
-
-#### Scenario: Admin authentication
-- **WHEN** the API is called without admin authentication
-- **THEN** it SHALL be rejected by the `@admin_required` decorator
-
-### Requirement: Frontend trend charts
-The system SHALL display 4 trend chart panels in the admin performance dashboard using vue-echarts VChart line/area charts.
-
-#### Scenario: Trend charts with data
-- **WHEN** historical snapshots contain more than 1 data point
-- **THEN** the dashboard SHALL display trend charts for: connection pool saturation, query latency (P50/P95/P99), Redis memory, and cache hit rates
-
-#### Scenario: Trend charts without data
-- **WHEN** historical snapshots are empty or contain only 1 data point
-- **THEN** the trend charts SHALL NOT be displayed (hidden via `v-if`)
-
-#### Scenario: Auto-refresh
-- **WHEN** the dashboard auto-refreshes
-- **THEN** historical data SHALL also be refreshed alongside real-time metrics
+## MODIFIED Requirements
+
+### Requirement: SQLite metrics history store
+The system SHALL provide a `MetricsHistoryStore` class in `core/metrics_history.py` that persists metrics snapshots to a SQLite database (`logs/metrics_history.sqlite` by default). The store SHALL use thread-local connections and a write lock, following the `LogStore` pattern in `core/log_store.py`. The schema SHALL include columns for `slow_query_active` (INTEGER), `slow_query_waiting` (INTEGER), and `worker_rss_bytes` (INTEGER) in addition to the existing pool, Redis, route cache, and latency columns.
+
+#### Scenario: Write and query snapshots
+- **WHEN** `write_snapshot(data)` is called with pool/redis/route_cache/latency/slow_query/memory metrics
+- **THEN** a row SHALL be inserted into `metrics_snapshots` with the current ISO 8601 timestamp, worker PID, and all metric columns
+
+#### Scenario: Query by time range
+- **WHEN** `query_snapshots(minutes=30)` is called
+- **THEN** it SHALL return all rows from the last 30 minutes, ordered by timestamp ascending, including the new columns
+
+#### Scenario: Retention cleanup
+- **WHEN** `cleanup()` is called
+- **THEN** rows older than `METRICS_HISTORY_RETENTION_DAYS` (default 3) SHALL be deleted, and total rows SHALL be capped at `METRICS_HISTORY_MAX_ROWS` (default 50000)
+
+#### Scenario: Thread safety
+- **WHEN** multiple threads write snapshots concurrently
+- **THEN** the write lock SHALL serialize writes and prevent database corruption
+
+#### Scenario: Schema migration for existing databases
+- **WHEN** the store initializes on an existing database without the new columns
+- **THEN** it SHALL execute ALTER TABLE ADD COLUMN for each missing column, tolerating "duplicate column" errors
+
+### Requirement: Background metrics collector
+The system SHALL provide a `MetricsHistoryCollector` class that runs a daemon thread collecting metrics snapshots at a configurable interval (default 30 seconds, via `METRICS_HISTORY_INTERVAL` env var). The collector SHALL include `slow_query_active`, `slow_query_waiting`, and `worker_rss_bytes` in each snapshot.
+
+#### Scenario: Automatic collection
+- **WHEN** the collector is started via `start_metrics_history(app)`
+- **THEN** it SHALL collect pool status (including slow_query_active and slow_query_waiting), Redis info, route cache status, query latency metrics, and worker RSS memory every interval and write them to the store
+
+#### Scenario: Graceful shutdown
+- **WHEN** `stop_metrics_history()` is called
+- **THEN** the collector thread SHALL stop within one interval period
+
+#### Scenario: Subsystem unavailability
+- **WHEN** a subsystem (e.g., Redis) is unavailable during collection
+- **THEN** the collector SHALL write null/0 for those fields and continue collecting other metrics
+
+### Requirement: Frontend trend charts
+The system SHALL display 5 trend chart panels in the admin performance dashboard using vue-echarts VChart line/area charts: connection pool saturation, query latency (P50/P95/P99), Redis memory, cache hit rates, and worker memory.
+
+#### Scenario: Trend charts with data
+- **WHEN** historical snapshots contain more than 1 data point
+- **THEN** the dashboard SHALL display trend charts for: connection pool saturation (including slow_query_active), query latency (P50/P95/P99), Redis memory, cache hit rates, and worker memory (RSS in MB)
+
+#### Scenario: Trend charts without data
+- **WHEN** historical snapshots are empty or contain only 1 data point
+- **THEN** the trend charts SHALL NOT be displayed (hidden via `v-if`)
+
+#### Scenario: Auto-refresh
+- **WHEN** the dashboard auto-refreshes
+- **THEN** historical data SHALL also be refreshed alongside real-time metrics
diff --git a/openspec/specs/slow-query-observability/spec.md b/openspec/specs/slow-query-observability/spec.md
new file mode 100644
index 0000000..fb3e130
--- /dev/null
+++ b/openspec/specs/slow-query-observability/spec.md
@@ -0,0 +1,49 @@
+## ADDED Requirements
+
+### Requirement: Slow query active count in metrics history snapshots
+The `MetricsHistoryCollector` SHALL include `slow_query_active` in each 30-second snapshot, recording the number of slow queries currently executing via dedicated connections.
+
+#### Scenario: Snapshot includes slow_query_active
+- **WHEN** the collector writes a snapshot while 3 slow queries are executing
+- **THEN** the `slow_query_active` column SHALL contain the value 3
+
+#### Scenario: No slow queries active
+- **WHEN** the collector writes a snapshot while no slow queries are executing
+- **THEN** the `slow_query_active` column SHALL contain the value 0
+
+### Requirement: Slow query waiting count tracked and persisted
+The system SHALL maintain a thread-safe counter `_SLOW_QUERY_WAITING` in `database.py` that tracks the number of threads currently waiting to acquire the slow query semaphore. This counter SHALL be included in `get_pool_status()` and persisted to metrics history snapshots.
+
+#### Scenario: Counter increments on semaphore wait
+- **WHEN** a thread enters `read_sql_df_slow()` and the semaphore is full
+- **THEN** `_SLOW_QUERY_WAITING` SHALL be incremented before `semaphore.acquire()` and decremented after acquire completes (success or timeout)
+
+#### Scenario: Counter in pool status API
+- **WHEN** `get_pool_status()` is called
+- **THEN** the returned dict SHALL include `slow_query_waiting` with the current waiting thread count
+
+#### Scenario: Counter persisted to metrics history
+- **WHEN** the collector writes a snapshot
+- **THEN** the `slow_query_waiting` column SHALL reflect the count at snapshot time
+
+### Requirement: Slow-path query latency recorded in QueryMetrics
+The `read_sql_df_slow()` and `read_sql_df_slow_iter()` functions SHALL call `record_query_latency()` with the elapsed query time, so that P50/P95/P99 metrics reflect all query paths (pool + slow).
+
+#### Scenario: Slow query latency appears in percentiles
+- **WHEN** a `read_sql_df_slow()` call completes in 5.2 seconds
+- **THEN** `record_query_latency(5.2)` SHALL be called and the latency SHALL appear in subsequent `get_percentiles()` results
+
+#### Scenario: Slow iter latency recorded on completion
+- **WHEN** a `read_sql_df_slow_iter()` generator completes after yielding all batches in 120 seconds total
+- **THEN** `record_query_latency(120.0)` SHALL be called in the finally block
+
+### Requirement: Slow query metrics displayed in Vue SPA
+The admin performance Vue SPA SHALL display `slow_query_active` and `slow_query_waiting` as StatCards in the connection pool panel, and include `slow_query_active` as a trend line in the connection pool trend chart.
+
+#### Scenario: StatCards display current values
+- **WHEN** the performance-detail API returns `db_pool.status.slow_query_active = 4` and `db_pool.status.slow_query_waiting = 2`
+- **THEN** the connection pool panel SHALL display StatCards showing "慢查詢執行中: 4" and "慢查詢排隊中: 2"
+
+#### Scenario: Trend chart includes slow_query_active
+- **WHEN** historical snapshots contain `slow_query_active` data points
+- **THEN** the connection pool trend chart SHALL include a "慢查詢執行中" line series
diff --git a/openspec/specs/worker-memory-tracking/spec.md b/openspec/specs/worker-memory-tracking/spec.md
new file mode 100644
index 0000000..73c3716
--- /dev/null
+++ b/openspec/specs/worker-memory-tracking/spec.md
@@ -0,0 +1,23 @@
+## ADDED Requirements
+
+### Requirement: Worker RSS memory in metrics history snapshots
+The `MetricsHistoryCollector` SHALL include `worker_rss_bytes` in each 30-second snapshot, recording the current worker process peak RSS memory using Python's `resource.getrusage()`.
+
+#### Scenario: RSS recorded in snapshot
+- **WHEN** the collector writes a snapshot and the worker process has 256 MB peak RSS
+- **THEN** the `worker_rss_bytes` column SHALL contain approximately 268435456
+
+#### Scenario: RSS collection failure
+- **WHEN** `resource.getrusage()` raises an exception
+- **THEN** the collector SHALL write NULL for `worker_rss_bytes` and continue collecting other metrics
+
+### Requirement: Worker memory trend chart in Vue SPA
+The admin performance Vue SPA SHALL display a "Worker 記憶體趨勢" TrendChart showing RSS memory over time in megabytes.
+
+#### Scenario: Memory trend displayed
+- **WHEN** historical snapshots contain `worker_rss_bytes` data with more than 1 data point
+- **THEN** the dashboard SHALL display a TrendChart with RSS values converted to MB
+
+#### Scenario: No memory data
+- **WHEN** historical snapshots do not contain `worker_rss_bytes` data (all NULL)
+- **THEN** the trend chart SHALL show "趨勢資料不足" message
diff --git a/scripts/check_full_modernization_gates.py b/scripts/check_full_modernization_gates.py
index 887ad1e..b4d6514 100755
--- a/scripts/check_full_modernization_gates.py
+++ b/scripts/check_full_modernization_gates.py
@@ -86,7 +86,7 @@ def _route_css_targets() -> dict[str, list[Path]]:
"/qc-gate": [ROOT / "frontend/src/qc-gate/style.css"],
"/job-query": [ROOT / "frontend/src/job-query/style.css"],
"/admin/pages": [ROOT / "src/mes_dashboard/templates/admin/pages.html"],
- "/admin/performance": [ROOT / "src/mes_dashboard/templates/admin/performance.html"],
+ "/admin/performance": [ROOT / "frontend/src/admin-performance/style.css"],
"/tables": [ROOT / "frontend/src/tables/style.css"],
"/excel-query": [ROOT / "frontend/src/excel-query/style.css"],
"/query-tool": [ROOT / "frontend/src/query-tool/style.css"],
diff --git a/src/mes_dashboard/core/database.py b/src/mes_dashboard/core/database.py
index cbe8dc0..5d3dca1 100644
--- a/src/mes_dashboard/core/database.py
+++ b/src/mes_dashboard/core/database.py
@@ -233,6 +233,7 @@ def get_pool_status() -> Dict[str, Any]:
"max_capacity": max_capacity,
"saturation": saturation,
"slow_query_active": get_slow_query_active_count(),
+ "slow_query_waiting": get_slow_query_waiting_count(),
}
@@ -436,6 +437,7 @@ _DIRECT_CONN_LOCK = threading.Lock()
# Slow-query concurrency control
_SLOW_QUERY_SEMAPHORE: Optional[threading.Semaphore] = None
_SLOW_QUERY_ACTIVE = 0
+_SLOW_QUERY_WAITING = 0
_SLOW_QUERY_LOCK = threading.Lock()
@@ -454,6 +456,11 @@ def get_slow_query_active_count() -> int:
return _SLOW_QUERY_ACTIVE
+def get_slow_query_waiting_count() -> int:
+ """Return the number of threads waiting for the slow-query semaphore."""
+ return _SLOW_QUERY_WAITING
+
+
def get_direct_connection_count() -> int:
"""Return total direct (non-pooled) connections since worker start."""
return _DIRECT_CONN_COUNTER
@@ -627,7 +634,8 @@ def read_sql_df_slow(
Returns:
DataFrame with query results.
"""
- global _SLOW_QUERY_ACTIVE
+ global _SLOW_QUERY_ACTIVE, _SLOW_QUERY_WAITING
+ from mes_dashboard.core.metrics import record_query_latency
runtime = get_db_runtime_config()
if timeout_seconds is None:
@@ -636,7 +644,13 @@ def read_sql_df_slow(
timeout_ms = timeout_seconds * 1000
sem = _get_slow_query_semaphore()
- acquired = sem.acquire(timeout=60)
+ with _SLOW_QUERY_LOCK:
+ _SLOW_QUERY_WAITING += 1
+ try:
+ acquired = sem.acquire(timeout=60)
+ finally:
+ with _SLOW_QUERY_LOCK:
+ _SLOW_QUERY_WAITING -= 1
if not acquired:
raise RuntimeError(
"Slow-query concurrency limit reached; try again later"
@@ -690,6 +704,8 @@ def read_sql_df_slow(
)
raise
finally:
+ elapsed = time.time() - start_time
+ record_query_latency(elapsed)
if conn:
try:
conn.close()
@@ -718,7 +734,8 @@ def read_sql_df_slow_iter(
batch_size: Number of rows per fetchmany call. None = use
DB_SLOW_FETCHMANY_SIZE from config (default 5000).
"""
- global _SLOW_QUERY_ACTIVE
+ global _SLOW_QUERY_ACTIVE, _SLOW_QUERY_WAITING
+ from mes_dashboard.core.metrics import record_query_latency
runtime = get_db_runtime_config()
if timeout_seconds is None:
@@ -730,7 +747,13 @@ def read_sql_df_slow_iter(
batch_size = runtime["slow_fetchmany_size"]
sem = _get_slow_query_semaphore()
- acquired = sem.acquire(timeout=60)
+ with _SLOW_QUERY_LOCK:
+ _SLOW_QUERY_WAITING += 1
+ try:
+ acquired = sem.acquire(timeout=60)
+ finally:
+ with _SLOW_QUERY_LOCK:
+ _SLOW_QUERY_WAITING -= 1
if not acquired:
raise RuntimeError(
"Slow-query concurrency limit reached; try again later"
@@ -788,6 +811,8 @@ def read_sql_df_slow_iter(
)
raise
finally:
+ elapsed = time.time() - start_time
+ record_query_latency(elapsed)
if conn:
try:
conn.close()
diff --git a/src/mes_dashboard/core/metrics_history.py b/src/mes_dashboard/core/metrics_history.py
index bdb08e2..99cd964 100644
--- a/src/mes_dashboard/core/metrics_history.py
+++ b/src/mes_dashboard/core/metrics_history.py
@@ -32,6 +32,9 @@ METRICS_HISTORY_INTERVAL = int(os.getenv('METRICS_HISTORY_INTERVAL', '30'))
METRICS_HISTORY_RETENTION_DAYS = int(os.getenv('METRICS_HISTORY_RETENTION_DAYS', '3'))
METRICS_HISTORY_MAX_ROWS = int(os.getenv('METRICS_HISTORY_MAX_ROWS', '50000'))
+ARCHIVE_LOG_DIR = os.getenv('ARCHIVE_LOG_DIR', 'logs/archive')
+ARCHIVE_LOG_KEEP_COUNT = int(os.getenv('ARCHIVE_LOG_KEEP_COUNT', '20'))
+
# ============================================================
# Database Schema
# ============================================================
@@ -54,10 +57,20 @@ CREATE TABLE IF NOT EXISTS metrics_snapshots (
latency_p50_ms REAL,
latency_p95_ms REAL,
latency_p99_ms REAL,
- latency_count INTEGER
+ latency_count INTEGER,
+ slow_query_active INTEGER,
+ slow_query_waiting INTEGER,
+ worker_rss_bytes INTEGER
);
"""
+# New columns added after initial schema — used for ALTER TABLE migration.
+_MIGRATION_COLUMNS = [
+ ("slow_query_active", "INTEGER"),
+ ("slow_query_waiting", "INTEGER"),
+ ("worker_rss_bytes", "INTEGER"),
+]
+
CREATE_INDEX_SQL = (
"CREATE INDEX IF NOT EXISTS idx_metrics_ts ON metrics_snapshots(ts);"
)
@@ -69,9 +82,49 @@ COLUMNS = [
"redis_used_memory", "redis_hit_rate",
"rc_l1_hit_rate", "rc_l2_hit_rate", "rc_miss_rate",
"latency_p50_ms", "latency_p95_ms", "latency_p99_ms", "latency_count",
+ "slow_query_active", "slow_query_waiting", "worker_rss_bytes",
]
+# ============================================================
+# Archive Log Cleanup
+# ============================================================
+
+_ARCHIVE_LOG_PREFIXES = ("access_", "error_", "watchdog_", "rq_worker_", "startup_")
+
+
+def cleanup_archive_logs(
+ archive_dir: str = ARCHIVE_LOG_DIR,
+ keep_per_type: int = ARCHIVE_LOG_KEEP_COUNT,
+) -> int:
+ """Delete old rotated log files from the archive directory.
+
+ Keeps the most recent *keep_per_type* files per log type (by mtime).
+ Returns the total number of files deleted.
+ """
+ archive_path = Path(archive_dir)
+ if not archive_path.is_dir():
+ return 0
+
+ deleted = 0
+ for prefix in _ARCHIVE_LOG_PREFIXES:
+ files = sorted(
+ (f for f in archive_path.iterdir() if f.name.startswith(prefix) and f.is_file()),
+ key=lambda f: f.stat().st_mtime,
+ reverse=True,
+ )
+ for old_file in files[keep_per_type:]:
+ try:
+ old_file.unlink()
+ deleted += 1
+ except OSError as exc:
+ logger.warning("Failed to delete archive log %s: %s", old_file, exc)
+
+ if deleted > 0:
+ logger.info("Cleaned up %d archive log files from %s", deleted, archive_dir)
+ return deleted
+
+
# ============================================================
# Metrics History Store
# ============================================================
@@ -94,6 +147,14 @@ class MetricsHistoryStore:
cursor = conn.cursor()
cursor.execute(CREATE_TABLE_SQL)
cursor.execute(CREATE_INDEX_SQL)
+ # Migrate existing databases: add new columns if missing.
+ for col_name, col_type in _MIGRATION_COLUMNS:
+ try:
+ cursor.execute(
+ f"ALTER TABLE metrics_snapshots ADD COLUMN {col_name} {col_type}"
+ )
+ except sqlite3.OperationalError:
+ pass # Column already exists — tolerate duplicate column error.
conn.commit()
self._initialized = True
logger.info("Metrics history store initialized at %s", self.db_path)
@@ -136,8 +197,9 @@ class MetricsHistoryStore:
pool_overflow, pool_max_capacity,
redis_used_memory, redis_hit_rate,
rc_l1_hit_rate, rc_l2_hit_rate, rc_miss_rate,
- latency_p50_ms, latency_p95_ms, latency_p99_ms, latency_count)
- VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)
+ latency_p50_ms, latency_p95_ms, latency_p99_ms, latency_count,
+ slow_query_active, slow_query_waiting, worker_rss_bytes)
+ VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)
""",
(
ts, pid,
@@ -155,6 +217,9 @@ class MetricsHistoryStore:
lat.get("p95_ms"),
lat.get("p99_ms"),
lat.get("count"),
+ data.get("slow_query_active"),
+ data.get("slow_query_waiting"),
+ data.get("worker_rss_bytes"),
),
)
conn.commit()
@@ -263,17 +328,36 @@ class MetricsHistoryCollector:
if self._cleanup_counter >= 100:
self._cleanup_counter = 0
self._store.cleanup()
+ try:
+ cleanup_archive_logs()
+ except Exception as exc:
+ logger.debug("Archive log cleanup failed: %s", exc)
def _collect_snapshot(self) -> None:
try:
data: Dict[str, Any] = {}
- # Pool status
+ # Pool status (includes slow_query_active and slow_query_waiting)
try:
from mes_dashboard.core.database import get_pool_status
- data["pool"] = get_pool_status()
+ pool_status = get_pool_status()
+ data["pool"] = pool_status
+ data["slow_query_active"] = pool_status.get("slow_query_active", 0)
+ data["slow_query_waiting"] = pool_status.get("slow_query_waiting", 0)
except Exception:
data["pool"] = {}
+ data["slow_query_active"] = 0
+ data["slow_query_waiting"] = 0
+
+ # Worker RSS memory
+ try:
+ import resource
+ # ru_maxrss is in KB on Linux
+ data["worker_rss_bytes"] = resource.getrusage(
+ resource.RUSAGE_SELF
+ ).ru_maxrss * 1024
+ except Exception:
+ data["worker_rss_bytes"] = 0
# Redis
try:
diff --git a/src/mes_dashboard/routes/admin_routes.py b/src/mes_dashboard/routes/admin_routes.py
index 9dfc1b5..ef7f2fe 100644
--- a/src/mes_dashboard/routes/admin_routes.py
+++ b/src/mes_dashboard/routes/admin_routes.py
@@ -71,10 +71,7 @@ _last_restart_request: float = 0.0
def performance():
"""Performance monitoring dashboard (Vue SPA)."""
dist_dir = os.path.join(current_app.static_folder or "", "dist")
- dist_html = os.path.join(dist_dir, "admin-performance.html")
- if os.path.exists(dist_html):
- return send_from_directory(dist_dir, "admin-performance.html")
- return render_template("admin/performance.html")
+ return send_from_directory(dist_dir, "admin-performance.html")
@admin_bp.route("/api/system-status", methods=["GET"])
diff --git a/src/mes_dashboard/templates/admin/performance.html b/src/mes_dashboard/templates/admin/performance.html
deleted file mode 100644
index 54dfd0c..0000000
--- a/src/mes_dashboard/templates/admin/performance.html
+++ /dev/null
@@ -1,1249 +0,0 @@
-{% extends "_base.html" %}
-
-{% block title %}效能監控 - MES Dashboard{% endblock %}
-
-{% block head_extra %}
-
-
-{% endblock %}
-
-{% block content %}
-