feat(system): process heartbeat + /system/health, engine runs business fields (W9-I)

- process_heartbeat / process_heartbeat_sample (0034): api / worker / worker:agent / all upsert every 15s (event loop lag last/max, concurrency, running runs, SSE connections, memory) - GET /system/health (admin): processes with 45s stale flag, SSE, DB latency/size/connections, disks (DATA_DIR / library roots / download dirs, cache thresholds), mount sentinel checks with IO timeout, engine pending/oldest/24h failure rate, schedules - GET /system/health/history?hours=: bucketed lag / running / SSE series - /engine/runs and /engine/runs/{id}: business {kind, refId, title, triggeredBy} joined from run input / task / agent_run / tasks_run_meta / outbox; businessKind / businessRefId filters - docs/runbook/ops.md: field meanings and alert thresholds Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-09-28 15:23:37 +08:00
1 parent 46d37d87fe
commit e8745115f4
@@ -0,0 +1,437 @@
+import { readdir, stat } from "node:fs/promises";
+import { checkRoot, DEFAULT_ROOT_SENTINEL, type DiskUsage, diskUsage } from "@umt/adapters";
+import type {
+  HealthLevel,
+  SystemDisk,
+  SystemHealth,
+  SystemHealthHistory,
+  SystemMount,
+  SystemProcess,
+} from "@umt/contracts";
+import { type DB, type Kysely, sql } from "@umt/db";
+import { HEARTBEAT_STALE_MS } from "./heartbeat.ts";
+
+// GET /system/health 各段的取数(W9-I)。每段独立 try/catch:一段失败只把该段标 unknown,不让整页 500。
+// 慢 IO(statfs / 哨兵检查)都带超时:挂载卡死时只标 timeout,不拖住请求。
+
+export interface HealthDeps {
+  db: Kysely<DB>;
+  engineSchema: string;
+  dataDir: string;
+  /** 本进程(处理请求的 api)信息 */
+  self: { version: string; startedAt: Date; sseConnections: number; instanceId?: string };
+  /** 磁盘阈值(百分比):warn / critical */
+  thresholds: { warnPercent: number; criticalPercent: number };
+  /** 测试注入 */
+  now?: () => Date;
+  diskUsage?: (path: string) => Promise<DiskUsage>;
+  checkRoot?: typeof checkRoot;
+  /** 单个路径 IO 检查超时(ms),缺省 3000 */
+  ioTimeoutMs?: number;
+}
+
+const LEVEL_ORDER: HealthLevel[] = ["ok", "warn", "critical"];
+export function worst(levels: HealthLevel[]): HealthLevel {
+  let w: HealthLevel = "ok";
+  for (const l of levels) if (l !== "unknown" && LEVEL_ORDER.indexOf(l) > LEVEL_ORDER.indexOf(w)) w = l;
+  return w;
+}
+
+class Timeout extends Error {}
+function withTimeout<T>(p: Promise<T>, ms: number): Promise<T> {
+  let t: ReturnType<typeof setTimeout> | undefined;
+  return Promise.race([
+    p,
+    new Promise<T>((_, reject) => {
+      t = setTimeout(() => reject(new Timeout("timeout")), ms);
+    }),
+  ]).finally(() => clearTimeout(t));
+}
+
+const toDate = (v: Date | string) => (v instanceof Date ? v : new Date(v));
+const num = (v: unknown) => (v == null ? null : Number(v));
+
+// ── 进程 ──
+
+export async function loadProcesses(db: Kysely<DB>, now: Date): Promise<SystemProcess[]> {
+  const rows = await db
+    .selectFrom("process_heartbeat")
+    .selectAll()
+    .where("last_beat_at", ">", new Date(now.getTime() - 24 * 3600_000))
+    .orderBy("process_kind")
+    .orderBy("instance_id")
+    .execute();
+  return rows.map((r) => {
+    const beat = toDate(r.last_beat_at);
+    const age = Math.max(0, now.getTime() - beat.getTime());
+    const extra = (r.extra ?? {}) as Record<string, unknown>;
+    const conc = (r.concurrency ?? {}) as { io?: number; agent?: number };
+    const sse = extra.sseConnections;
+    return {
+      instanceId: r.instance_id,
+      processKind: r.process_kind,
+      pid: r.pid,
+      host: r.host,
+      version: r.version,
+      startedAt: toDate(r.started_at).toISOString(),
+      lastBeatAt: beat.toISOString(),
+      beatAgeSec: Math.round(age / 100) / 10,
+      stale: age > HEARTBEAT_STALE_MS,
+      eventLoopLagMs: r.event_loop_lag_ms,
+      eventLoopLagMaxMs: r.event_loop_lag_max_ms,
+      concurrency: {
+        ...(typeof conc.io === "number" ? { io: conc.io } : {}),
+        ...(typeof conc.agent === "number" ? { agent: conc.agent } : {}),
+      },
+      runningRuns: r.running_runs,
+      sseConnections: typeof sse === "number" ? sse : null,
+      extra,
+    };
+  });
+}
+
+/** 进程段的档位:有 stale 或延迟 ≥ 5s 为 critical;延迟 ≥ 1s 为 warn;没有任何 worker 心跳为 warn */
+export function processLevel(ps: SystemProcess[]): HealthLevel {
+  const live = ps.filter((p) => !p.stale);
+  const levels: HealthLevel[] = [];
+  if (ps.some((p) => p.stale)) levels.push("warn");
+  if (!live.some((p) => p.processKind !== "api")) levels.push("warn");
+  for (const p of live) {
+    if (p.eventLoopLagMaxMs >= 5000) levels.push("critical");
+    else if (p.eventLoopLagMaxMs >= 1000) levels.push("warn");
+  }
+  return worst(levels);
+}
+
+// ── DB ──
+
+export async function loadDb(db: Kysely<DB>): Promise<SystemHealth["db"]> {
+  try {
+    const t0 = performance.now();
+    await sql`select 1`.execute(db);
+    const latencyMs = Math.round((performance.now() - t0) * 10) / 10;
+    const size = await sql<{ size: string }>`select pg_database_size(current_database())::text as size`.execute(db);
+    const conns = await sql<{ total: number; active: number; idle: number; iit: number; max: number }>`
+      select count(*)::int as total,
+             count(*) filter (where state = 'active')::int as active,
+             count(*) filter (where state = 'idle')::int as idle,
+             count(*) filter (where state like 'idle in transaction%')::int as iit,
+             current_setting('max_connections')::int as max
+        from pg_stat_activity where datname = current_database()`.execute(db);
+    const c = conns.rows[0]!;
+    const usage = c.max > 0 ? c.total / c.max : 0;
+    const level: HealthLevel =
+      latencyMs >= 500 || usage >= 0.9 ? "critical" : latencyMs >= 100 || usage >= 0.75 || c.iit > 5 ? "warn" : "ok";
+    return {
+      level,
+      latencyMs,
+      sizeBytes: Number(size.rows[0]?.size ?? 0),
+      connections: { total: c.total, active: c.active, idle: c.idle, idleInTransaction: c.iit, max: c.max },
+      error: null,
+    };
+  } catch (err) {
+    return { level: "critical", latencyMs: null, sizeBytes: null, connections: null, error: (err as Error).message };
+  }
+}
+
+// ── 磁盘与挂载根 ──
+
+interface PathTarget {
+  scope: SystemDisk["scope"];
+  id: string | null;
+  name: string;
+  path: string;
+}
+
+export async function loadPathTargets(db: Kysely<DB>) {
+  const roots = await db
+    .selectFrom("library_root")
+    .select(["id", "name", "local_path", "sentinel_file"])
+    .where("enabled", "=", true)
+    .orderBy("name")
+    .execute();
+  const maps = await db
+    .selectFrom("path_mapping")
+    .leftJoin("downloader_instance", "downloader_instance.id", "path_mapping.downloader_instance_id")
+    .select([
+      "path_mapping.id as id",
+      "path_mapping.remote_prefix as remote",
+      "path_mapping.local_prefix as local",
+      "downloader_instance.name as downloader",
+    ])
+    .where("path_mapping.enabled", "=", true)
+    .where("path_mapping.view_kind", "=", "downloader")
+    .orderBy("path_mapping.local_prefix")
+    .execute();
+  return { roots, maps };
+}
+
+export function diskLevel(ratio: number, t: HealthDeps["thresholds"]): HealthLevel {
+  const pct = ratio * 100;
+  return pct >= t.criticalPercent ? "critical" : pct >= t.warnPercent ? "warn" : "ok";
+}
+
+export async function loadDisksAndMounts(d: HealthDeps): Promise<{ disks: SystemDisk[]; mounts: SystemMount[] }> {
+  const usage = d.diskUsage ?? diskUsage;
+  const check = d.checkRoot ?? checkRoot;
+  const ms = d.ioTimeoutMs ?? 3000;
+  const { roots, maps } = await loadPathTargets(d.db);
+
+  const targets: PathTarget[] = [
+    { scope: "data_dir", id: null, name: "DATA_DIR", path: d.dataDir },
+    ...roots.map((r) => ({ scope: "library_root" as const, id: r.id, name: r.name, path: r.local_path })),
+    ...maps.map((m) => ({
+      scope: "download_dir" as const,
+      id: m.id,
+      name: m.downloader ? `${m.downloader} ${m.remote}` : m.remote,
+      path: m.local,
+    })),
+  ];
+  const disks = await Promise.all(
+    targets.map(async (t): Promise<SystemDisk> => {
+      const base = { scope: t.scope, id: t.id, name: t.name, path: t.path };
+      try {
+        const u = await withTimeout(usage(t.path), ms);
+        if (u.status !== "ok")
+          return {
+            ...base,
+            level: "unknown",
+            totalBytes: null,
+            usedBytes: null,
+            availableBytes: null,
+            usedRatio: null,
+            error: u.code,
+          };
+        return {
+          ...base,
+          level: diskLevel(u.usedRatio, d.thresholds),
+          totalBytes: u.totalBytes,
+          usedBytes: u.usedBytes,
+          availableBytes: u.availableBytes,
+          usedRatio: Math.min(1, Math.max(0, u.usedRatio)),
+          error: null,
+        };
+      } catch (err) {
+        return {
+          ...base,
+          level: err instanceof Timeout ? "critical" : "unknown",
+          totalBytes: null,
+          usedBytes: null,
+          availableBytes: null,
+          usedRatio: null,
+          error: err instanceof Timeout ? "timeout" : (err as Error).message,
+        };
+      }
+    }),
+  );
+
+  const mounts = await Promise.all([
+    ...roots.map(async (r): Promise<SystemMount> => {
+      const base = { kind: "library_root" as const, id: r.id, name: r.name, path: r.local_path, sentinelChecked: true };
+      try {
+        const h = await withTimeout(check(r.local_path, r.sentinel_file ?? DEFAULT_ROOT_SENTINEL), ms);
+        return h.kind === "online"
+          ? { ...base, status: "online", reason: null }
+          : { ...base, status: "offline", reason: h.reason };
+      } catch (err) {
+        return err instanceof Timeout
+          ? { ...base, status: "timeout", reason: "timeout" }
+          : { ...base, status: "offline", reason: "error" };
+      }
+    }),
+    ...maps.map(async (m): Promise<SystemMount> => {
+      const base = {
+        kind: "path_mapping" as const,
+        id: m.id,
+        name: m.downloader ? `${m.downloader} ${m.remote}` : m.remote,
+        path: m.local,
+        sentinelChecked: false,
+      };
+      try {
+        await withTimeout(
+          (async () => {
+            const st = await stat(m.local);
+            if (!st.isDirectory()) throw Object.assign(new Error("不是目录"), { code: "ENOTDIR" });
+            await readdir(m.local);
+          })(),
+          ms,
+        );
+        return { ...base, status: "online", reason: null };
+      } catch (err) {
+        if (err instanceof Timeout) return { ...base, status: "timeout", reason: "timeout" };
+        const code = (err as { code?: string }).code;
+        const reason =
+          code === "ENOENT" ? "root_missing" : code === "EACCES" || code === "EPERM" ? "permission_denied" : "error";
+        return { ...base, status: "offline", reason };
+      }
+    }),
+  ]);
+  return { disks, mounts };
+}
+
+// ── 引擎 ──
+
+const schemaTable = (schema: string, table: string) => {
+  if (!/^[a-z_][a-z0-9_]*$/.test(schema)) throw new Error(`非法 schema 名 ${schema}`);
+  return sql.raw(`"${schema}".${table}`);
+};
+
+export async function loadEngine(db: Kysely<DB>, schema: string, now: Date): Promise<SystemHealth["engine"]> {
+  const t = schemaTable(schema, "workflow_runs");
+  const since = new Date(now.getTime() - 24 * 3600_000);
+  const r = await sql<{
+    pending: number;
+    running: number;
+    oldest: Date | null;
+    finished: number;
+    failed: number;
+  }>`select
+      count(*) filter (where status = 'pending')::int as pending,
+      count(*) filter (where status = 'running' and worker_id is not null)::int as running,
+      min(coalesce(available_at, created_at)) filter (where status = 'pending' and coalesce(available_at, created_at) <= ${now}) as oldest,
+      count(*) filter (where status in ('completed', 'succeeded', 'failed') and finished_at >= ${since})::int as finished,
+      count(*) filter (where status = 'failed' and finished_at >= ${since})::int as failed
+    from ${t} where namespace_id = 'default'`.execute(db);
+  const row = r.rows[0]!;
+  const oldestAge = row.oldest ? Math.max(0, (now.getTime() - toDate(row.oldest).getTime()) / 1000) : null;
+  const rate = row.finished > 0 ? row.failed / row.finished : null;
+  const level: HealthLevel = worst([
+    oldestAge != null && oldestAge >= 1800 ? "critical" : oldestAge != null && oldestAge >= 300 ? "warn" : "ok",
+    rate != null && row.finished >= 10 && rate >= 0.3 ? "critical" : rate != null && rate >= 0.1 ? "warn" : "ok",
+  ]);
+  return {
+    level,
+    pending: row.pending,
+    running: row.running,
+    oldestPendingAgeSec: oldestAge == null ? null : Math.round(oldestAge),
+    finished24h: row.finished,
+    failed24h: row.failed,
+    failureRate24h: rate,
+  };
+}
+
+// ── 定时任务 ──
+
+export async function loadSchedules(db: Kysely<DB>, now: Date): Promise<SystemHealth["schedules"]> {
+  const jobs = await db
+    .selectFrom("scheduled_job")
+    .select(["id", "name", "enabled", "available", "last_status", "last_run_at", "last_error", "running_since"])
+    .orderBy("id")
+    .execute();
+  const failedRuns = await sql<{ n: number }>`
+    select count(*)::int as n from scheduled_job_run
+     where status = 'failed' and started_at >= ${new Date(now.getTime() - 24 * 3600_000)}`.execute(db);
+  const enabled = jobs.filter((j) => j.enabled && j.available);
+  const failing = enabled
+    .filter((j) => j.last_status === "failed")
+    .map((j) => ({
+      id: j.id,
+      name: j.name,
+      lastRunAt: j.last_run_at ? toDate(j.last_run_at).toISOString() : null,
+      lastError: j.last_error,
+    }));
+  return {
+    level: failing.length ? "warn" : "ok",
+    total: jobs.length,
+    enabled: enabled.length,
+    running: jobs.filter((j) => j.running_since).length,
+    failing,
+    failedRuns24h: failedRuns.rows[0]?.n ?? 0,
+  };
+}
+
+// ── 汇总 ──
+
+export async function buildHealth(d: HealthDeps): Promise<SystemHealth> {
+  const now = d.now?.() ?? new Date();
+  const [processes, db, paths, engine, schedules] = await Promise.all([
+    loadProcesses(d.db, now).catch(() => [] as SystemProcess[]),
+    loadDb(d.db),
+    loadDisksAndMounts(d).catch(() => ({ disks: [] as SystemDisk[], mounts: [] as SystemMount[] })),
+    loadEngine(d.db, d.engineSchema, now).catch((): SystemHealth["engine"] => ({
+      level: "unknown",
+      pending: 0,
+      running: 0,
+      oldestPendingAgeSec: null,
+      finished24h: 0,
+      failed24h: 0,
+      failureRate24h: null,
+    })),
+    loadSchedules(d.db, now).catch((): SystemHealth["schedules"] => ({
+      level: "unknown",
+      total: 0,
+      enabled: 0,
+      running: 0,
+      failing: [],
+      failedRuns24h: 0,
+    })),
+  ]);
+  // SSE:存活的 api/all 进程上报的连接数合计;本进程若还没写过心跳,用实时值
+  const selfCounted = processes.some((p) => !p.stale && p.instanceId === d.self.instanceId && p.sseConnections != null);
+  const sse =
+    processes.filter((p) => !p.stale).reduce((s, p) => s + (p.sseConnections ?? 0), 0) +
+    (selfCounted ? 0 : d.self.sseConnections);
+  const mountLevel: HealthLevel = paths.mounts.some((m) => m.status !== "online") ? "critical" : "ok";
+  return {
+    level: worst([
+      processLevel(processes),
+      db.level,
+      ...paths.disks.map((x) => x.level),
+      mountLevel,
+      engine.level,
+      schedules.level,
+    ]),
+    checkedAt: now.toISOString(),
+    api: {
+      version: d.self.version,
+      startedAt: d.self.startedAt.toISOString(),
+      uptimeSec: Math.max(0, Math.round((now.getTime() - d.self.startedAt.getTime()) / 1000)),
+    },
+    processes,
+    sse: { connections: sse },
+    db,
+    disks: paths.disks,
+    diskThresholds: d.thresholds,
+    mounts: paths.mounts,
+    engine,
+    schedules,
+  };
+}
+
+// ── 历史 ──
+
+export async function loadHistory(db: Kysely<DB>, hours: number, now = new Date()): Promise<SystemHealthHistory> {
+  const from = new Date(now.getTime() - hours * 3600_000);
+  // 约 360 个点;最细 15s(一个心跳周期)
+  const bucketSec = Math.max(15, Math.ceil((hours * 3600) / 360 / 15) * 15);
+  const r = await sql<{
+    instance_id: string;
+    process_kind: string;
+    bucket: Date;
+    lag: number;
+    lag_max: number;
+    running: number | null;
+    sse: number | null;
+  }>`select instance_id, max(process_kind) as process_kind,
+        date_bin(make_interval(secs => ${bucketSec}), at, timestamptz 'epoch') as bucket,
+        avg(event_loop_lag_ms)::float8 as lag, max(event_loop_lag_max_ms)::float8 as lag_max,
+        max(running_runs)::float8 as running, max(sse_connections)::float8 as sse
+      from process_heartbeat_sample where at >= ${from} and at <= ${now}
+      group by instance_id, bucket order by instance_id, bucket`.execute(db);
+  const series = new Map<string, SystemHealthHistory["series"][number]>();
+  for (const row of r.rows) {
+    let s = series.get(row.instance_id);
+    if (!s) {
+      s = { instanceId: row.instance_id, processKind: row.process_kind, points: [] };
+      series.set(row.instance_id, s);
+    }
+    s.points.push({
+      at: toDate(row.bucket).toISOString(),
+      lagMs: Math.round(Number(row.lag) * 10) / 10,
+      lagMaxMs: Number(row.lag_max),
+      runningRuns: num(row.running),
+      sseConnections: num(row.sse),
+    });
+  }
+  return { from: from.toISOString(), to: now.toISOString(), bucketSec, series: [...series.values()] };
+}