From 1ca15e90255d9dba73e14759f637574a62f43ebc Mon Sep 17 00:00:00 2001 From: yml2213 Date: Thu, 1 Oct 2026 12:55:09 +0800 Subject: [PATCH] =?UTF-8?q?perf(db):=20=E7=BB=9F=E8=AE=A1=E9=87=91?= =?UTF-8?q?=E9=A2=9D=E8=90=BD=E5=88=B0=E5=88=97=E4=B8=8A=E3=80=81=E4=BC=9A?= =?UTF-8?q?=E8=AF=9D=E5=88=B7=E6=96=B0=E8=8A=82=E6=B5=81=E3=80=81=E8=A1=A5?= =?UTF-8?q?=E9=BD=90=E4=BF=9D=E7=95=99=E4=B8=8E=E5=86=85=E5=AD=98=E6=B2=BB?= =?UTF-8?q?=E7=90=86?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 背景：生产 32 天内 >500ms 慢查询从 2.4k/天涨到 1.3 万/天（09-26 尖峰 14.3 万）， pg_stat_statements 显示后台工单列表族占 57%、打手端 visible_orders 占 27%； 同时 backend 容器十天从 194MB 涨到 754MB(768MB 限额的 98%)，贴顶后整体变慢。 1. 新增迁移 083：work_orders 增加派生列 source_order_amount_fen，并从 material_json.source.orderAmountFen 回填（只认 15 位以内整数，与 JS/SQL 口径一致）。 统计查询原来每行要解析 1.2KB JSON（全表约 132MB）且同一表达式求值两次， 生产实测 3.5~4.4 秒；改走派生列后 COALESCE 短路，未命中列的行才回退 JSON/orders.total_amount， 口径与改造前完全一致。写入点：createWorkOrder / updateWorkOrder 各一处，并加静态守卫。 2. worker_sessions.last_seen_at 从"每个请求都写"改为"5 分钟才写一次"： 32 天被写了 3300 万次（占 DB 总时间 2.8%，并堆出 14% 死元组）。仍用一条语句返回鉴权有效性 （EXISTS target AS valid），不能因为不写就让上层 401。 3. 保留策略补上 work_order_events（71.7 万行 / 269MB，此前完全没纳入清理，只增不减）， 新增配置 retention.workOrderEventDays（默认 180 天）与 WORK_ORDER_EVENT_RETENTION_DAYS 出口。 4. 内存治理：compose 后端限额 768m→1g 并显式设置 V8 老生代上限（--max-old-space-size=768）， 避免贴 cgroup 上限后 GC 抖动；新增每 10 分钟的进程内存快照日志（rss/heapUsed/heapLimit）， 用于区分"正常 GC 曲线"与"真泄漏"。 5. 限流桶加 5 万条硬上限：key 里含请求体字段（collect 的 orderNo）时可被构造成高基数， 只按过期清理挡不住内存增长。 验证：后端 530 项单测全绿（新增 11 项，含派生列/会话/保留的静态守卫与解析函数单测）。 --- .env.server.example | 1 + apps/backend/src/config/defaults.ts | 1 + .../src/config/env-documentation.test.ts | 2 + apps/backend/src/config/env-overrides.ts | 1 + .../083_work_order_source_amount_column.sql | 27 ++++++++ .../backend/src/middleware/rate-limit.test.ts | 8 +++ apps/backend/src/middleware/rate-limit.ts | 15 ++++ .../worker-platform/shared.test.ts | 69 +++++++++++++++++++ .../repositories/worker-platform/shared.ts | 31 +++++++++ .../work-order-management-repo.ts | 14 ++-- .../worker-platform/work-order-query-repo.ts | 32 +++++++-- .../worker-platform/worker-session-repo.ts | 37 ++++++++-- .../scheduler/data-retention-job.test.ts | 48 +++++++++++++ .../services/scheduler/data-retention-job.ts | 7 +- .../worker-platform-service.test.ts | 56 +++++++++++++++ apps/backend/src/startup/bootstrap.ts | 2 + apps/backend/src/startup/memory-reporter.ts | 40 +++++++++++ apps/backend/src/types/runtime-config.ts | 2 + docker-compose.yml | 8 ++- 19 files changed, 382 insertions(+), 19 deletions(-) create mode 100644 apps/backend/src/db/migrations/083_work_order_source_amount_column.sql create mode 100644 apps/backend/src/repositories/worker-platform/shared.test.ts create mode 100644 apps/backend/src/services/scheduler/data-retention-job.test.ts create mode 100644 apps/backend/src/startup/memory-reporter.ts diff --git a/.env.server.example b/.env.server.example index 01573881..116012d9 100644 --- a/.env.server.example +++ b/.env.server.example @@ -62,6 +62,7 @@ RAW_PAYLOAD_RETENTION_DAYS=180 TASK_EVENT_RETENTION_DAYS=180 WEBHOOK_EVENT_RETENTION_DAYS=180 AUDIT_LOG_RETENTION_DAYS=365 +WORK_ORDER_EVENT_RETENTION_DAYS=180 # 文件存储：生产使用阿里云 OSS（生产编排已移除 MinIO，仅本地开发使用）。 # STORAGE_MODE 取值 minio | dual | oss；dual 仅在 MinIO→OSS 迁移/回滚期间使用。 diff --git a/apps/backend/src/config/defaults.ts b/apps/backend/src/config/defaults.ts index e73e187c..761c2382 100644 --- a/apps/backend/src/config/defaults.ts +++ b/apps/backend/src/config/defaults.ts @@ -29,6 +29,7 @@ export function createDefaultRuntimeConfig(projectRoot: string): RuntimeConfig { taskEventDays: 180, webhookEventDays: 180, auditLogDays: 365, + workOrderEventDays: 180, }, database: { diff --git a/apps/backend/src/config/env-documentation.test.ts b/apps/backend/src/config/env-documentation.test.ts index 92b75868..a8e180f4 100644 --- a/apps/backend/src/config/env-documentation.test.ts +++ b/apps/backend/src/config/env-documentation.test.ts @@ -35,6 +35,8 @@ const DEPLOYMENT_ONLY_ENV_NAMES = [ 'POSTGRES_MAINTENANCE_WORK_MEM', 'BACKEND_MEM_LIMIT', 'BACKEND_CPUS', + 'BACKEND_NODE_OPTIONS', + 'NODE_OPTIONS', 'WEB_MEM_LIMIT', 'WEB_CPUS', 'POSTGRES_PASSWORD', diff --git a/apps/backend/src/config/env-overrides.ts b/apps/backend/src/config/env-overrides.ts index 400b3395..a46557fa 100644 --- a/apps/backend/src/config/env-overrides.ts +++ b/apps/backend/src/config/env-overrides.ts @@ -34,6 +34,7 @@ export const ENV_OVERRIDES: readonly EnvOverride[] = [ integerEnv('TASK_EVENT_RETENTION_DAYS', ['retention', 'taskEventDays']), integerEnv('WEBHOOK_EVENT_RETENTION_DAYS', ['retention', 'webhookEventDays']), integerEnv('AUDIT_LOG_RETENTION_DAYS', ['retention', 'auditLogDays']), + integerEnv('WORK_ORDER_EVENT_RETENTION_DAYS', ['retention', 'workOrderEventDays']), stringEnv('DATABASE_URL', ['database', 'url']), booleanEnv('DATABASE_SSL', ['database', 'ssl']), integerEnv('DATABASE_MAX_CONNECTIONS', ['database', 'maxConnections']), diff --git a/apps/backend/src/db/migrations/083_work_order_source_amount_column.sql b/apps/backend/src/db/migrations/083_work_order_source_amount_column.sql new file mode 100644 index 00000000..0d27956d --- /dev/null +++ b/apps/backend/src/db/migrations/083_work_order_source_amount_column.sql @@ -0,0 +1,27 @@ +-- 083_work_order_source_amount_column.sql —— 把「用户付款金额」从 material_json 落到列上。 +-- +-- 背景：后台接单工单列表的「本次筛选统计」每次刷新都要算 +-- SUM(material_json->'source'->>'orderAmountFen' ...) +-- 该表达式带正则校验、每行要被求值两次（总付款 + 退款），而 material_json 平均 1.2KB， +-- work_orders 9.6 万行时单次查询要解析约 132MB JSON，实测平均 2.8 秒（pg_stat_statements）。 +-- 这条查询随工单量线性变慢，是"系统越跑越慢"的主因之一。 +-- +-- 处理：新增 source_order_amount_fen 派生列。 +-- - 有合法值 → 写入该列，统计时直接读列（不再碰 JSON）； +-- - 没有合法值 → 留 NULL，统计逻辑回退 orders.total_amount（与改造前语义完全一致）。 +-- NULL 表示"来源没给"，不是"金额为 0"，因此不能用 0 当缺省值。 + +ALTER TABLE work_orders + ADD COLUMN IF NOT EXISTS source_order_amount_fen BIGINT; + +COMMENT ON COLUMN work_orders.source_order_amount_fen IS + '来源订单用户付款金额（分）：取自 material_json.source.orderAmountFen；NULL 表示来源未提供合法值，统计回退 orders.total_amount。由 createWorkOrder / updateWorkOrder 同步维护。'; + +-- 回填历史数据。只认 15 位以内的整数，三处口径必须一致（本迁移、统计 SQL、 +-- repositories/worker-platform/shared.ts 的 resolveSourceOrderAmountFen）： +-- 正则放行超长数字时 ::bigint 会直接报 22003，让整条迁移失败回滚； +-- 15 位以内也保证 JS 侧 Number 能精确表示。 +UPDATE work_orders +SET source_order_amount_fen = (material_json->'source'->>'orderAmountFen')::bigint +WHERE source_order_amount_fen IS NULL + AND COALESCE(material_json->'source'->>'orderAmountFen', '') ~ '^-?[0-9]{1,15}$'; diff --git a/apps/backend/src/middleware/rate-limit.test.ts b/apps/backend/src/middleware/rate-limit.test.ts index 154fd086..14722cd9 100644 --- a/apps/backend/src/middleware/rate-limit.test.ts +++ b/apps/backend/src/middleware/rate-limit.test.ts @@ -136,3 +136,11 @@ function createMockResponse(): Response & { body: any } } + +test('限流桶有数量硬上限：key 含用户可控字段时不能无限膨胀', async () => { + const { readFile } = await import('node:fs/promises') + const src = await readFile(new URL('./rate-limit.ts', import.meta.url), 'utf8') + // collect 路由的限流 key 里带请求体 orderNo（可被构造），只按过期清理挡不住高基数灌键 + assert.ok(src.includes('MAX_RATE_LIMIT_BUCKETS'), '缺少限流桶数量上限') + assert.ok(src.includes('buckets.size > MAX_RATE_LIMIT_BUCKETS'), '必须在清理时按上限淘汰最老的桶') +}) diff --git a/apps/backend/src/middleware/rate-limit.ts b/apps/backend/src/middleware/rate-limit.ts index ef6218c0..1355208c 100644 --- a/apps/backend/src/middleware/rate-limit.ts +++ b/apps/backend/src/middleware/rate-limit.ts @@ -25,6 +25,8 @@ type RateLimitBucket = { const buckets = new Map() let lastCleanupAt = 0 +/** 限流桶数量硬上限：key 含请求体字段时可被构造成高基数，必须兜底防止内存无限增长。 */ +const MAX_RATE_LIMIT_BUCKETS = 50_000 export function createRateLimitMiddleware({ scope, @@ -120,4 +122,17 @@ function cleanupExpiredBuckets(now: number): void { buckets.delete(key) } } + + // 兜底：限流 key 里含请求体字段（如 collect 路由的 orderNo），是可被构造的高基数键， + // 只按"过期"清理挡不住短时间内灌进来的海量新键。这里再加一道硬上限， + // 超限时按插入顺序淘汰最老的桶（Map 保持插入序），保证内存不随扫描流量无限涨。 + if (buckets.size > MAX_RATE_LIMIT_BUCKETS) { + const overflow = buckets.size - MAX_RATE_LIMIT_BUCKETS + let removed = 0 + for (const key of buckets.keys()) { + buckets.delete(key) + removed += 1 + if (removed >= overflow) break + } + } } diff --git a/apps/backend/src/repositories/worker-platform/shared.test.ts b/apps/backend/src/repositories/worker-platform/shared.test.ts new file mode 100644 index 00000000..3e5eb355 --- /dev/null +++ b/apps/backend/src/repositories/worker-platform/shared.test.ts @@ -0,0 +1,69 @@ +import assert from 'node:assert/strict' +import test from 'node:test' + +import { resolveSourceOrderAmountFen } from './shared.js' + +test('resolveSourceOrderAmountFen：正常取值（字符串/数字都认）', () => { + assert.equal(resolveSourceOrderAmountFen({ source: { orderAmountFen: 12345 } }), 12345) + assert.equal(resolveSourceOrderAmountFen({ source: { orderAmountFen: '12345' } }), 12345) + // JSON 字符串形态（createWorkOrder 收到的就是字符串） + assert.equal( + resolveSourceOrderAmountFen(JSON.stringify({ source: { orderAmountFen: '8800' } })), + 8800, + ) +}) + +test('resolveSourceOrderAmountFen：0 是合法金额，必须与"没有值"区分开', () => { + // 统计口径是 COALESCE(列, ...)：0 会直接命中列，NULL 才会回退 orders.total_amount， + // 所以这里必须返回 0 而不是 null + assert.equal(resolveSourceOrderAmountFen({ source: { orderAmountFen: 0 } }), 0) + assert.equal(resolveSourceOrderAmountFen({ source: { orderAmountFen: '0' } }), 0) +}) + +test('resolveSourceOrderAmountFen：没有合法整数时返回 null（交给 SQL 回退）', () => { + const invalidCases: unknown[] = [ + undefined, + null, + {}, + { source: {} }, + { source: { orderAmountFen: null } }, + { source: { orderAmountFen: '' } }, + { source: { orderAmountFen: ' ' } }, + { source: { orderAmountFen: '12.5' } }, + { source: { orderAmountFen: '12a' } }, + { source: { orderAmountFen: '￥120' } }, + { source: { orderAmountFen: {} } }, + { source: { orderAmountFen: [] } }, + 'not-json', + ] + for (const value of invalidCases) { + assert.equal( + resolveSourceOrderAmountFen(value), + null, + `${JSON.stringify(value)} 应判为"来源没给合法值"`, + ) + } +}) + +test('resolveSourceOrderAmountFen：位数上限与迁移 083 / 统计 SQL 保持一致（1~15 位）', () => { + // 15 位是安全上限：Number 能精确表示，::bigint 也不会溢出 + assert.equal( + resolveSourceOrderAmountFen({ source: { orderAmountFen: '999999999999999' } }), + 999999999999999, + ) + // 16 位起一律判为非法（不写列 → SQL 回退 orders.total_amount），避免与 SQL 口径分叉 + assert.equal( + resolveSourceOrderAmountFen({ source: { orderAmountFen: '1000000000000000' } }), + null, + ) + assert.equal( + resolveSourceOrderAmountFen({ source: { orderAmountFen: '12345678901234567890' } }), + null, + ) +}) + +test('resolveSourceOrderAmountFen：material 结构异常不抛错', () => { + for (const weird of [0, false, '', '[]', [], '{}']) { + assert.doesNotThrow(() => resolveSourceOrderAmountFen(weird)) + } +}) diff --git a/apps/backend/src/repositories/worker-platform/shared.ts b/apps/backend/src/repositories/worker-platform/shared.ts index 0bf114d6..45fd3280 100644 --- a/apps/backend/src/repositories/worker-platform/shared.ts +++ b/apps/backend/src/repositories/worker-platform/shared.ts @@ -1,7 +1,38 @@ import type { PoolClient } from 'pg' +import { parseJsonObject } from '../../utils/json.js' +import type { JsonObject } from '../../types/json.js' import type { WorkerWalletRow } from './types.js' +/** + * 解析工单 `material.source.orderAmountFen`（来源订单的用户付款金额，分），用于派生列 + * `work_orders.source_order_amount_fen`。 + * + * 统计口径要求：来源给了合法整数就用它，否则回退 `orders.total_amount`。 + * 所以**没有合法值时必须返回 null 而不是 0**——0 是"确实付了 0 元"，NULL 才是"来源没给"。 + * + * 位数上限与迁移 083 的回填、统计 SQL 里的正则保持一致（1~15 位）： + * 15 位以内一定能被 Number 精确表示且不会撑爆 bigint。 + */ +export function resolveSourceOrderAmountFen(materialJson: unknown): number | null { + const material = + typeof materialJson === 'string' + ? parseJsonObject(materialJson) + : (materialJson as JsonObject | null | undefined) + if (!material || typeof material !== 'object') return null + + const source = (material as JsonObject).source + if (!source || typeof source !== 'object' || Array.isArray(source)) return null + + const raw = (source as JsonObject).orderAmountFen + if (raw === null || raw === undefined) return null + + const text = String(raw).trim() + if (!/^-?[0-9]{1,15}$/.test(text)) return null + + return Number(text) +} + export async function ensureWorkerWalletWithClient( client: PoolClient, workerId: number, diff --git a/apps/backend/src/repositories/worker-platform/work-order-management-repo.ts b/apps/backend/src/repositories/worker-platform/work-order-management-repo.ts index 88652ebb..dd33e00d 100644 --- a/apps/backend/src/repositories/worker-platform/work-order-management-repo.ts +++ b/apps/backend/src/repositories/worker-platform/work-order-management-repo.ts @@ -1,5 +1,5 @@ import { query, withTransaction } from '../../db/client.js' -import { toJsonString, toPositiveInteger } from './shared.js' +import { resolveSourceOrderAmountFen, toJsonString, toPositiveInteger } from './shared.js' import { getWorkOrderById, getWorkOrderByIdWithClient } from './work-order-query-repo.js' import { createWorkOrderEventWithClient } from './work-order-event-repo.js' import { @@ -22,12 +22,12 @@ export async function createWorkOrder(input: CreateWorkOrderInput): Promise'source'->>'orderAmountFen', '') ~ '^-?[0-9]{1,15}$' + THEN (wo.material_json->'source'->>'orderAmountFen')::bigint + END, + COALESCE(o.total_amount, 0) + ) +` + export async function getWorkOrderStatistics(input: ListInput = {}): Promise { const { joins, whereClause, params } = buildWorkOrderWhere(input) const vipEvidencePendingWhere = ` @@ -346,13 +370,7 @@ export async function getWorkOrderStatistics(input: ListInput = {}): Promise'source'->>'orderAmountFen', '') ~ '^-?[0-9]+$' - THEN (wo.material_json->'source'->>'orderAmountFen')::bigint - ELSE COALESCE(o.total_amount, 0) - END - ` + const paymentExpression = WORK_ORDER_PAYMENT_AMOUNT_SQL const result = await query( ` SELECT diff --git a/apps/backend/src/repositories/worker-platform/worker-session-repo.ts b/apps/backend/src/repositories/worker-platform/worker-session-repo.ts index fa1eaf1b..2177392a 100644 --- a/apps/backend/src/repositories/worker-platform/worker-session-repo.ts +++ b/apps/backend/src/repositories/worker-platform/worker-session-repo.ts @@ -83,14 +83,41 @@ export async function getWorkerSessionBySessionId( return result.rows[0] || null } +/** + * 校验会话仍然有效，并顺带刷新 last_seen_at。 + * + * 这里必须同时满足两个约束，所以不能用"直接跳过"的写法： + * 1. 返回值是**鉴权有效性**——上层 `!(await touchWorkerSession(...))` 为真就直接 401 + * 「登录设备已被删除」，跳过查询会把所有正常请求打成未登录； + * 2. 也不能像以前那样每个请求都真写一次：32 天里 last_seen_at 被写了 3300 万次， + * 占数据库总执行时间的 2.8%，还在 worker_sessions 上堆出 14% 的死元组。 + * + * 因此改成「一条语句拿到有效性 + 只在超过 5 分钟没刷新时才写」。 + * 副作用只有 last_seen_at 精度变粗（用于设备列表排序与展示）， + * 不影响会话有效期：expires_at 在建会话时确定，不做滑动续期。 + */ +const WORKER_SESSION_TOUCH_MIN_INTERVAL = '5 minutes' + export async function touchWorkerSession(sessionId: string, now: string): Promise { - const result = await query( - `UPDATE worker_sessions - SET last_seen_at = $2, updated_at = $2 - WHERE session_id = $1 AND status = 'active' AND expires_at > $2`, + const result = await query<{ valid: boolean }>( + ` + WITH target AS ( + SELECT session_id + FROM worker_sessions + WHERE session_id = $1 AND status = 'active' AND expires_at > $2 + ), touched AS ( + UPDATE worker_sessions ws + SET last_seen_at = $2, updated_at = $2 + FROM target + WHERE ws.session_id = target.session_id + AND ws.last_seen_at < ($2::timestamptz - INTERVAL '${WORKER_SESSION_TOUCH_MIN_INTERVAL}') + RETURNING ws.session_id + ) + SELECT EXISTS (SELECT 1 FROM target) AS valid + `, [String(sessionId || '').trim(), now], ) - return (result.rowCount || 0) > 0 + return Boolean(result.rows[0]?.valid) } export async function revokeWorkerSession( diff --git a/apps/backend/src/services/scheduler/data-retention-job.test.ts b/apps/backend/src/services/scheduler/data-retention-job.test.ts new file mode 100644 index 00000000..7e92e1a0 --- /dev/null +++ b/apps/backend/src/services/scheduler/data-retention-job.test.ts @@ -0,0 +1,48 @@ +import assert from 'node:assert/strict' +import { readFile } from 'node:fs/promises' +import test from 'node:test' + +async function readJobSource() { + return readFile(new URL('./data-retention-job.ts', import.meta.url), 'utf8') +} + +test('保留清理必须覆盖 work_order_events，否则这张只增不减的表会一直拖慢历史单检索', async () => { + const src = await readJobSource() + assert.ok( + src.includes("deleteOldRowsInBatches('work_order_events'"), + 'work_order_events 未纳入保留清理（生产上 71 万行 / 269MB 且只增不减）', + ) + assert.ok( + src.includes("'task_events' | 'webhook_events' | 'admin_audit_logs' | 'work_order_events'"), + '删除助手需要显式允许 work_order_events（表名是白名单联合类型，避免拼接注入）', + ) + assert.ok(src.includes('retention.workOrderEventDays'), '保留天数必须走配置项，便于按环境调整') + assert.ok( + src.includes('workOrderEventRows: number') && src.includes('workOrderEventRows,'), + '清理结果要带上工单事件的删除行数，方便定时任务面板核对', + ) +}) + +test('保留清理默认窗口：敏感报文/事件 180 天，审计 365 天，工单事件 180 天', async () => { + const defaults = await readFile(new URL('../../config/defaults.ts', import.meta.url), 'utf8') + for (const line of [ + 'rawPayloadDays: 180', + 'taskEventDays: 180', + 'webhookEventDays: 180', + 'auditLogDays: 365', + 'workOrderEventDays: 180', + ]) { + assert.ok(defaults.includes(line), `默认保留策略缺少 ${line}`) + } + + const overrides = await readFile( + new URL('../../config/env-overrides.ts', import.meta.url), + 'utf8', + ) + assert.ok( + overrides.includes( + "integerEnv('WORK_ORDER_EVENT_RETENTION_DAYS', ['retention', 'workOrderEventDays'])", + ), + 'WORK_ORDER_EVENT_RETENTION_DAYS 需要环境变量出口，否则生产无法调整', + ) +}) diff --git a/apps/backend/src/services/scheduler/data-retention-job.ts b/apps/backend/src/services/scheduler/data-retention-job.ts index ee535c4f..7806b50c 100644 --- a/apps/backend/src/services/scheduler/data-retention-job.ts +++ b/apps/backend/src/services/scheduler/data-retention-job.ts @@ -10,6 +10,7 @@ export type DataRetentionResult = { taskEventRows: number webhookEventRows: number auditLogRows: number + workOrderEventRows: number matchLogRows: number matchLogDeletedRows: number } @@ -21,6 +22,8 @@ export async function runDataRetentionJob(): Promise { const taskEventDays = positiveDays(retention.taskEventDays, 180) const webhookEventDays = positiveDays(retention.webhookEventDays, 180) const auditLogDays = positiveDays(retention.auditLogDays, 365) + // 工单事件表以前完全没纳入保留：717k 行 / 269MB 且只增不减，会一直拖慢历史单检索。 + const workOrderEventDays = positiveDays(retention.workOrderEventDays, 180) // 1. 按配置清理超出上限或过期的打手商品匹配日志 const matchConfig = getWorkerProductMatchConfig() @@ -41,6 +44,7 @@ export async function runDataRetentionJob(): Promise { const taskEventRows = await deleteOldRowsInBatches('task_events', taskEventDays) const webhookEventRows = await deleteOldRowsInBatches('webhook_events', webhookEventDays) const auditLogRows = await deleteOldRowsInBatches('admin_audit_logs', auditLogDays) + const workOrderEventRows = await deleteOldRowsInBatches('work_order_events', workOrderEventDays) return { status: 'ok', @@ -51,6 +55,7 @@ export async function runDataRetentionJob(): Promise { taskEventRows, webhookEventRows, auditLogRows, + workOrderEventRows, } } async function scrubJsonColumnInBatches( @@ -82,7 +87,7 @@ async function scrubJsonColumnInBatches( } async function deleteOldRowsInBatches( - table: 'task_events' | 'webhook_events' | 'admin_audit_logs', + table: 'task_events' | 'webhook_events' | 'admin_audit_logs' | 'work_order_events', retentionDays: number, ) { let total = 0 diff --git a/apps/backend/src/services/worker-platform/worker-platform-service.test.ts b/apps/backend/src/services/worker-platform/worker-platform-service.test.ts index 22d7110f..6274e1ab 100644 --- a/apps/backend/src/services/worker-platform/worker-platform-service.test.ts +++ b/apps/backend/src/services/worker-platform/worker-platform-service.test.ts @@ -2533,3 +2533,59 @@ test('押金已退还或已没收时不再尝试二次冻结（台账唯一索 '这个码要被服务层识别并映射成 409 文案，不能漏成 500', ) }) + +test('筛选统计必须优先读派生列 source_order_amount_fen，不再逐行解析 material_json', async () => { + const { readFile } = await import('node:fs/promises') + const queryRepo = await readFile( + new URL('../../repositories/worker-platform/work-order-query-repo.ts', import.meta.url), + 'utf8', + ) + // 统计表达式：列 → JSON 兜底 → orders.total_amount，COALESCE 从左到右短路 + assert.ok( + queryRepo.includes('wo.source_order_amount_fen'), + '统计没有走派生列，又会退化成每次解析全表 JSON（生产实测 2.8 秒）', + ) + const paymentBlock = queryRepo.slice( + queryRepo.indexOf('const WORK_ORDER_PAYMENT_AMOUNT_SQL'), + queryRepo.indexOf('export async function getWorkOrderStatistics'), + ) + assert.ok( + paymentBlock.indexOf('wo.source_order_amount_fen') < + paymentBlock.indexOf("material_json->'source'->>'orderAmountFen'"), + '派生列必须排在 JSON 兜底之前，否则短路失效、等于没优化', + ) + assert.ok( + paymentBlock.includes('COALESCE(o.total_amount, 0)'), + '最后一级兜底必须与改造前一致（回退 orders.total_amount）', + ) + + // 两个写入点都要维护派生列，否则新单全是 NULL（虽然能兜底，但白优化） + const managementRepo = await readFile( + new URL('../../repositories/worker-platform/work-order-management-repo.ts', import.meta.url), + 'utf8', + ) + const writes = managementRepo.split('resolveSourceOrderAmountFen(').length - 1 + assert.equal(writes, 2, '派生列写入点必须正好两处：建单 INSERT 与通用 UPDATE') + assert.ok(managementRepo.includes('source_order_amount_fen = $17')) +}) + +test('会话 touch 必须同时保留鉴权语义与写入节流', async () => { + const { readFile } = await import('node:fs/promises') + const src = await readFile( + new URL('../../repositories/worker-platform/worker-session-repo.ts', import.meta.url), + 'utf8', + ) + const body = src.slice( + src.indexOf('export async function touchWorkerSession'), + src.indexOf('export async function revokeWorkerSession'), + ) + // 返回值是鉴权有效性：上层 !(await touch) 会直接 401，不能因为"不写"就返回 false + assert.ok(body.includes('EXISTS (SELECT 1 FROM target) AS valid'), '必须仍然返回会话是否有效') + assert.ok(body.includes("status = 'active'"), '有效性判定不能丢') + // 写操作必须节流：32 天里被写了 3300 万次 + assert.ok(body.includes('WORKER_SESSION_TOUCH_MIN_INTERVAL'), '缺少节流间隔常量') + assert.ok( + body.includes('ws.last_seen_at < ($2::timestamptz - INTERVAL'), + '写入必须带 last_seen_at 未刷新判定，否则等于没节流', + ) +}) diff --git a/apps/backend/src/startup/bootstrap.ts b/apps/backend/src/startup/bootstrap.ts index 45fee38e..9623ce37 100644 --- a/apps/backend/src/startup/bootstrap.ts +++ b/apps/backend/src/startup/bootstrap.ts @@ -5,6 +5,7 @@ import { migrateJsonConfigFilesToDatabase } from '../services/config/json-config import { startKuaishouIndustrySendCallbackRetryWorker } from '../services/platforms/kuaishou-industry/send-code-service.js' import { startScheduledJobs } from '../services/scheduler/scheduler-service.js' import { logError, logInfo } from '../utils/logger.js' +import { startMemoryReporter } from './memory-reporter.js' import { formatStartupError, type StartupState } from './state.js' const CORE_BOOT_RETRY_DELAY_MS = 5_000 @@ -45,6 +46,7 @@ export async function bootstrapCoreServices( startScheduledJobs() startKuaishouIndustrySendCallbackRetryWorker() + startMemoryReporter() break } catch (error) { const message = formatStartupError(error) diff --git a/apps/backend/src/startup/memory-reporter.ts b/apps/backend/src/startup/memory-reporter.ts new file mode 100644 index 00000000..c58242b2 --- /dev/null +++ b/apps/backend/src/startup/memory-reporter.ts @@ -0,0 +1,40 @@ +import v8 from 'node:v8' + +import { logInfo } from '../utils/logger.js' + +/** + * 周期性记录进程内存快照，用于判断 backend 容器内存是"正常 GC 曲线"还是"持续增长"。 + * + * 为什么需要它：生产上 backend 限额 768MB，重启后 194MB，十天后贴到 754MB(98%)， + * 之后整个服务明显变慢——但仅凭容器数字分不清是 V8 贴近 cgroup 上限导致 GC 抖动， + * 还是真有增长点。有了这份日志就能对比： + * - rssMb 长期单调上升、重启后回落明显 → 存在增长点（缓存/事件缓冲/连接堆积）； + * - rssMb 稳定、heapUsedMb 上下波动 → 正常 GC 曲线，调限额即可。 + * + * 每 10 分钟一条，logger 内部已有按天滚动与 30 天保留，不会堆积。 + */ +const MEMORY_REPORT_INTERVAL_MS = 10 * 60 * 1000 + +function toMb(bytes: number): number { + return Math.round((Number(bytes) || 0) / 1024 / 1024) +} + +export function startMemoryReporter(): NodeJS.Timeout { + const report = () => { + const usage = process.memoryUsage() + logInfo('[memory]', '进程内存快照', { + rssMb: toMb(usage.rss), + heapUsedMb: toMb(usage.heapUsed), + heapTotalMb: toMb(usage.heapTotal), + externalMb: toMb(usage.external), + arrayBuffersMb: toMb(usage.arrayBuffers || 0), + heapLimitMb: toMb(v8.getHeapStatistics().heap_size_limit), + }) + } + + report() + const timer = setInterval(report, MEMORY_REPORT_INTERVAL_MS) + // 不因为这个定时器把进程钉在事件循环里（关闭流程/测试环境都受益） + timer.unref() + return timer +} diff --git a/apps/backend/src/types/runtime-config.ts b/apps/backend/src/types/runtime-config.ts index eb08ed5b..4170d231 100644 --- a/apps/backend/src/types/runtime-config.ts +++ b/apps/backend/src/types/runtime-config.ts @@ -31,6 +31,8 @@ export type RuntimeConfig = { taskEventDays: number webhookEventDays: number auditLogDays: number + /** 工单事件（work_order_events）保留天数：这张表只增不减，必须纳入清理。 */ + workOrderEventDays: number } database: { url: string diff --git a/docker-compose.yml b/docker-compose.yml index 011034c4..20d9fc73 100644 --- a/docker-compose.yml +++ b/docker-compose.yml @@ -61,7 +61,10 @@ services: NPM_REGISTRY: ${NPM_REGISTRY:-https://registry.npmmirror.com} restart: unless-stopped # 30s 让后端在 SIGTERM 后完成优雅关闭（HTTP 连接收尾 + 数据库连接池释放）。 - mem_limit: ${BACKEND_MEM_LIMIT:-768m} + # 内存限额 768m → 1g：生产实测十天从重启后的 ~194MB 涨到 754MB(98%)， + # 贴着 cgroup 上限时 V8 会持续 GC 抖动、接口整体变慢（且未必 OOM，所以日志里看不到崩溃）。 + # 服务器 .env 里若仍写着 BACKEND_MEM_LIMIT=768m，会覆盖这里的默认值，需一并调整。 + mem_limit: ${BACKEND_MEM_LIMIT:-1g} cpus: ${BACKEND_CPUS:-1.5} stop_grace_period: 30s env_file: @@ -79,6 +82,9 @@ services: environment: TZ: ${TZ:-Asia/Shanghai} PORT: ${BACKEND_PORT:-3000} + # V8 老生代上限留出余量（限额 1g → 堆 768m）：让 GC 在撞上 cgroup 墙之前主动回收， + # 避免"贴顶后才疯狂 GC"的抖动。native/external 内存与堆外缓冲还需要剩余那部分。 + NODE_OPTIONS: ${BACKEND_NODE_OPTIONS:---max-old-space-size=768} HTTP_TRUST_PROXY: ${HTTP_TRUST_PROXY:-1} DATABASE_URL: ${DATABASE_URL:-postgres://postgres:postgres@postgres:5432/order_site} DATABASE_SSL: ${DATABASE_SSL:-false}