21 KiB
Admin Flux Grant Batch
- Date: 2026-05-07 (initial), 2026-05-08 (rename + docs move)
- Status: Implemented (v1) — pending end-to-end live verification
- Surface:
apps/server
Naming note: an earlier draft used the table name
campaign. Renamed toflux_grant_batchbefore merge becausecampaignis a generic marketing term that overpromises (suggests support for coupons / discounts / referral rewards) while underdescribing the actual capability (batch FLUX grants only).
1. 背景
目前 BillingService.creditFlux 提供了"给单个 userId 加余额"的事务安全实现(billing-service.ts:162),但没有任何上层入口(路由 / 脚本 / CLI)调用它。所有 promo / 客诉补偿 / 活动赠送类的余额发放,目前只能:
- 直接写裸 SQL 修改
user_flux+flux_transaction+ 手动失效 Redis(绕过BillingService,违反apps/server/CLAUDE.md"balance 写入只走 BillingService" 的架构约束) - 临时跑脚本,每次活动都要 ad-hoc
随着 FLUX 商业化推进,预计批量发放会成为常规运营动作(每月 1+ 次,每次 1k–10k 用户)。需要一个稳定、可观测、可重试的 admin 接口。
2. 目标 / 非目标
目标
- Dev 通过 admin API 给指定邮箱列表批量发放 FLUX
- 发放过程异步执行,可查询进度(待发 / 已发 / 跳过 / 失败 计数)
- 失败项可手动重试
- ledger 留下"活动赠送"语义标记(区别于 Stripe 充值、初始赠送)
- 创建前可 dry-run 预览(避免错 amount / 错人群)
- 通过 better-auth session + env allowlist 限制访问
非目标(v1)
- 回滚 / 反向操作:罕见(年 0–2 次),用 ad-hoc SQL 处理;建固定接口 ROI 不成立
- email → userId 解析以外的输入解析:不支持 CSV 上传、不支持电话号码、不支持用户分群条件
- 运营自助 UI:dashboard 提供只读查看 + 可以发起 API 调用,但不是为非技术用户设计
- 完整 RBAC role 系统:env-var allowlist 够用,未来加 role 列再升级
- 细粒度 amount 上限 / 审批流:dev 自律,必要时加 daily 总额限制即可
- 优惠码 / 折扣券 / 邀请奖励等其它优惠形式:本 schema 不包含 code 字段、
max_redemptions、用户行为触发器等机制;未来要做这些功能是新的 schema(coupon/referral_reward),不在本 batch 表上加列
3. 数据模型
3.1 新增表
flux_grant_batch
export const fluxGrantBatch = pgTable('flux_grant_batch', {
id: text('id').primaryKey().$defaultFn(() => nanoid()),
name: text('name').notNull(), // 操作者填写,e.g. "Spring 2026 Promo"
type: text('type').notNull(), // 'promo'(v1 只支持这一种)
amount: bigint('amount', { mode: 'number' }).notNull(), // 每人发多少 FLUX
description: text('description'), // 写到每条 flux_transaction.description
status: text('status').notNull(), // 'created' | 'running' | 'completed' | 'failed_partial'
createdByUserId: text('created_by_user_id').notNull(), // 操作者 userId(来自 session)
createdAt: timestamp('created_at').defaultNow().notNull(),
startedAt: timestamp('started_at'), // worker 第一次拣起 recipient 的时间
completedAt: timestamp('completed_at'), // 所有 recipient 终态后的时间
}, table => [
index('flux_grant_batch_status_idx').on(table.status),
index('flux_grant_batch_created_by_idx').on(table.createdByUserId),
])
status 状态机:
draft (dryRun, 不入库) → 不存在
created → running → completed (全部 granted)
created → running → failed_partial (部分 failed/skipped 后无法继续)
draft 仅在 dryRun 响应中出现,不持久化。创建即 created,worker 首次拣起改 running。
flux_grant_batch_recipient
export const fluxGrantBatchRecipient = pgTable('flux_grant_batch_recipient', {
id: text('id').primaryKey().$defaultFn(() => nanoid()),
batchId: text('batch_id').notNull(), // FK 概念上指向 flux_grant_batch.id(不强制 FK 约束以避免级联删除)
inputEmail: text('input_email').notNull(), // 操作者输入快照(审计用),保留原始大小写
userId: text('user_id'), // 解析成功才有;NULL = email 没匹配上
status: text('status').notNull(), // 'pending' | 'granted' | 'skipped' | 'failed'
errorReason: text('error_reason'), // 'not_found' | 'user_deleted' | 'duplicate_in_input' | 'user_deleted_after_resolution' | DB error message
fluxTransactionId: text('flux_transaction_id'), // granted 后回填,指向 flux_transaction.id
attemptCount: integer('attempt_count').notNull().default(0),
lastAttemptedAt: timestamp('last_attempted_at'),
createdAt: timestamp('created_at').defaultNow().notNull(),
}, table => [
index('flux_grant_batch_recipient_batch_status_idx').on(table.batchId, table.status),
index('flux_grant_batch_recipient_pending_idx').on(table.status, table.lastAttemptedAt).where(sql`status = 'pending'`),
uniqueIndex('flux_grant_batch_recipient_batch_email_uniq').on(table.batchId, table.inputEmail),
])
status 状态机:
pending → granted (creditFlux 成功)
pending → failed (重试上限后)
pending → skipped (worker 执行时 user_flux.deleted_at 已置位)
created at insert:
- pending (email 解析成功,userId 已写入)
- skipped (email 没匹配 / 用户已注销 / input 重复)
(batch_id, input_email) 唯一索引:同一 batch 内邮箱不会重复处理。
Partial index WHERE status = 'pending':worker 轮询效率,避免扫整表。
3.2 现有表扩展
flux_transaction.type 枚举注释扩展
flux-transaction.ts:13 现在的注释:
type: text('type').notNull(), // 'credit' | 'debit' | 'initial'
改为:
type: text('type').notNull(), // 'credit' | 'debit' | 'initial' | 'promo'
注:type 是 text + 注释伪枚举,DB 不强制。新增值无需 migration,但需要 grep 现有代码里是否有硬编码 type === 'credit' 的地方。已知影响面:
BillingService.creditFlux写入处(billing-service.ts:194):硬编码'credit',需要扩展为可入参BillingService.creditFluxFromStripeCheckout(billing-service.ts:285):仍写'credit',不动BillingService.creditFluxFromInvoice(billing-service.ts:397):仍写'credit',不动flux-transaction.ts的TransactionEntry.type类型 union 加'promo'getStats的 capacity 算法把'promo'计入inArray列表(promo 入账后用户进度条 capacity 应跟着涨)
为什么不拆 type + category 两列:当前 credit 子类型只会增加 1 个(promo)。等出现第 2 个不同来源(refund / staff_grant / beta)时再做这次重构。YAGNI。
3.3 BillingService.creditFlux 签名扩展
async creditFlux(input: {
userId: string
amount: number
requestId?: string
description: string
source: string
type?: 'credit' | 'promo' // 新增,默认 'credit'
auditMetadata?: Record<string, unknown>
}): Promise<{ balanceBefore: number, balanceAfter: number, fluxTransactionId: string }>
写入 ledger 时使用 input.type ?? 'credit'。其他逻辑(事务、Redis 失效、flux.credited 事件)不变。
返回值新增 fluxTransactionId:worker 拿到后回填到 flux_grant_batch_recipient.flux_transaction_id,为后续报表查询省掉一次 join。
4. API 表面
所有路由挂在 /api/admin/*,前置 sessionMiddleware + 新增 adminGuard middleware。
4.1 POST /api/admin/flux-grant-batches
创建 batch。支持 dry-run。
Request:
POST /api/admin/flux-grant-batches?dryRun=false
Content-Type: application/json
Authorization: Bearer <session-jwt>
{
"name": "Spring 2026 Promo",
"amount": 200,
"description": "Spring promo grant",
"emails": ["alice@example.com", "bob@example.com"]
}
Query params:
dryRun(true|false,默认false):true 时只解析、不写库、不调度
Body 字段:
name: 1–100 charsamount: integer ≥ 1, ≤MAX_GRANT_AMOUNT_PER_USER(默认 10000)description: 0–500 charsemails: array of email strings, length 1–10000
Response (dryRun=true):
{
"preview": {
"totalEmails": 1000,
"willGrant": 947,
"willSkip": {
"notFound": 30,
"userDeleted": 18,
"duplicateInInput": 5
},
"totalFluxToIssue": 189400,
"estimatedDurationSec": 19,
"samples": {
"willGrant": ["alice@example.com", "bob@example.com", "..."],
"notFound": ["typo@old-domain.com", "..."],
"userDeleted": ["deleted-user@example.com", "..."]
}
}
}
samples 每类截取前 5 条用于人工 sanity check。
Response (dryRun=false):
{
"batch": {
"id": "fgb_abc123",
"name": "Spring 2026 Promo",
"status": "created",
"createdAt": "2026-05-07T10:00:00.000Z",
"createdByUserId": "uid_xyz"
},
"summary": {
"totalEmails": 1000,
"pending": 947,
"skipped": 53,
"totalFluxToIssue": 189400
}
}
写库后立刻返回 202 Accepted,worker 异步执行。
Errors:
400validation 失败 / amount 超上限 / emails 列表为空401未登录403已登录但不在 admin 白名单
4.2 GET /api/admin/flux-grant-batches/:id
查询单个 batch 详情和进度。
Response:
{
"batch": {
"id": "fgb_abc123",
"name": "Spring 2026 Promo",
"type": "promo",
"amount": 200,
"status": "running",
"createdByUserId": "uid_xyz",
"createdAt": "2026-05-07T10:00:00.000Z",
"startedAt": "2026-05-07T10:00:05.000Z",
"completedAt": null
},
"progress": {
"total": 1000,
"pending": 200,
"granted": 745,
"skipped": 53,
"failed": 2
},
"recentFailures": [
{
"id": "fgr_xxx",
"inputEmail": "alice@example.com",
"userId": "uid_alice",
"errorReason": "DB connection timeout",
"attemptCount": 3,
"lastAttemptedAt": "2026-05-07T10:05:00.000Z"
}
]
}
recentFailures 限 20 条,最近优先。
4.3 GET /api/admin/flux-grant-batches
列表分页查询。
Query:
limit: default 20, max 100cursor: opaque cursor (created_at + id, base64-encoded)status: optional filter
Response:
{
"batches": [{ "id": "...", "name": "...", "status": "...", ... }],
"nextCursor": "..."
}
4.4 POST /api/admin/flux-grant-batches/:id/retry
把所有 status='failed' 的 recipient 改回 status='pending',重置 attemptCount 和 lastAttemptedAt。worker 下次轮询自动拣起。如果 batch 已经在 completed / failed_partial 终态,会被改回 running。
Response:
{ "retriedCount": 2 }
Idempotency: 重复调用安全(没有 failed 项时返回 retriedCount: 0)。
5. 调度与执行
5.1 Worker 部署
复用 billing-consumer Railway role(apps/server/CLAUDE.md "Railway role 越少越好维护"),不开新 role。
billing-consumer 当前只跑 Redis Stream 消费循环(src/bin/run-billing-consumer.ts)。增加一个并行的 flux grant batch 轮询循环,独立 lifecycle。两个 loop 共享 DB pool 和 Redis 连接,通过 Promise.all 一起跑,任一抛错 + abortSignal 共享让另一个也退出。
5.2 调度循环
loop forever:
1. findActiveBatches: 找出有 pending recipient 的 batch(status ∈ {created, running})
2. 对每个 batch,开事务取 N 条:
SELECT recipient.* FROM flux_grant_batch_recipient
INNER JOIN flux_grant_batch ON ...
WHERE batch_id = $1
AND status = 'pending'
AND (last_attempted_at IS NULL OR last_attempted_at + backoff <= NOW())
ORDER BY created_at
LIMIT N
FOR UPDATE SKIP LOCKED -- 多 worker 实例安全
3. 对每条 recipient:
a. re-check user_flux.deleted_at;如果已注销 → status='skipped' + reason='user_deleted_after_resolution'
b. 否则调 BillingService.creditFlux({
userId, amount: batch.amount, type: 'promo',
requestId: `flux-grant-batch-${batchId}-${recipientId}`,
description: batch.description ?? `Flux grant batch: ${batch.name}`,
source: 'admin_promo',
auditMetadata: { batchId, batchName, recipientId },
})
c. 成功 → status='granted', flux_transaction_id 回填, attempt_count++
d. 失败 (attempt < MAX_ATTEMPTS) → 保持 pending,attempt_count++,下轮重试
e. 失败 (attempt >= MAX_ATTEMPTS) → status='failed', error_reason=...
4. 当 batch 没有 pending 项时 (finalizeBatchIfDone):
UPDATE flux_grant_batch
SET status = (any failed → 'failed_partial' else 'completed'),
completed_at = NOW()
5. throttle: 每发一条 sleep (1000 / throttlePerSec) ms
6. 没拿到行时 sleep idleSleepMs (默认 1000)
关键约束:
FOR UPDATE SKIP LOCKED:多billing-consumer实例同时跑安全,不会重复处理同一行(user_id, request_id)唯一索引(flux-transaction.ts:24-26):crash-recovery 时即使creditFlux重入也不会双发requestId = flux-grant-batch-${batchId}-${recipientId}:用 recipientId 而非 userId,因为同一 user 在不同 batch 可以被发多次(不同 recipient 行 → 不同 requestId)
5.3 Backoff 策略
backoffMs(attempt) =
attempt == 0 → 0s
attempt == 1 → 30s
attempt >= 2 → 5min
MAX_ATTEMPTS = 3,可调(worker option)。
5.4 节流
throttlePerSec 默认 50。理由:
creditFlux单条约 10–20ms(DB tx + redis + stream xadd)- 50/s 已经足够 1k 用户在 20s 内完成
- 不抢线上 LLM 流量(
/api/v1/openai高峰能到 100+ req/s)
6. Auth
6.1 新增 env var
// libs/env.ts EnvSchema 中追加一个字段:
const EnvSchema = object({
// ...其它字段...
ADMIN_EMAILS: optional(string(), ''), // 逗号分隔邮箱,e.g. "alice@example.com,bob@example.com"
})
空字符串 = 没人是 admin(生产默认安全)。
为什么用 email 不用 user.id:
- 运营心智模型 = 邮箱,没人记得
uid_aaa - env 配置可读性 + grep 可读性都更好
- 离职 / 接任时直接编辑邮箱列表,不需要先去 DB 查 userId
为什么必须 email_verified = true:
- 否则攻击者可以注册一个新账号,用 admin 邮箱(在 admin 注销 / 邮箱被释放后),不验证就尝试访问 admin 接口
- 走 OAuth (Google/GitHub) 的用户邮箱默认 verified;email/password 注册要求显式验证
- 这是零成本的额外防线,admin 用户都会有 verified 邮箱
6.2 adminGuard middleware
新建 apps/server/src/middlewares/admin-guard.ts:
export function adminGuard(env: Env): MiddlewareHandler<HonoEnv> {
const adminEmails = parseAdminEmails(env.ADMIN_EMAILS) // lowercase + trim
return async (c, next) => {
const user = c.get('user')
if (!user) throw createUnauthorizedError(...)
if (!user.emailVerified) throw createForbiddenError('Admin access requires a verified email')
if (!adminEmails.has(user.email.toLowerCase())) throw createForbiddenError(...)
await next()
}
}
挂载顺序:sessionMiddleware → authGuard → adminGuard → 路由。
6.3 审计
每个 batch 写 created_by_user_id。每条 ledger 的 metadata 含 {batchId, batchName, recipientId},description 字段写 batch 名或 description。事后查"谁在什么时候发了什么活动"通过:
SELECT b.id, b.name, b.amount,
COUNT(r.id) FILTER (WHERE r.status='granted') AS granted_count,
u.email AS operator_email
FROM flux_grant_batch b
JOIN "user" u ON u.id = b.created_by_user_id
LEFT JOIN flux_grant_batch_recipient r ON r.batch_id = b.id
GROUP BY b.id, u.email
ORDER BY b.created_at DESC;
未来加 RBAC 时 adminGuard 改成 user.role === 'admin',env-var 删除,其他保持。
7. Failure Modes & Recovery
| 故障 | 检测 | 恢复 |
|---|---|---|
| worker crash | Railway 自动重启 | pending 行被新 worker 拣起,(user_id, request_id) 唯一索引防双发 |
| Postgres 短暂故障 | creditFlux 抛错 |
recipient attempt_count++,30s 后重试 |
| Redis 短暂故障 | creditFlux 在 update cache 阶段抛错 |
同上。已在事务中提交的余额不会回滚 → 下次 retry 会被唯一索引挡住但 cache 仍未刷新。已知缺口:v1 接受这个边界,下次 getFlux 触发 cache miss 会自动从 DB 读最新值。 |
| batch 创建后立刻 worker 还没拣起就有用户注销 | worker 执行前 re-check user_flux.deleted_at |
标 skipped + user_deleted_after_resolution |
| 单个 recipient 永久失败(非瞬时) | attempt >= 3 后标 failed |
操作者 POST /retry 或人工查 error_reason |
| 整个 batch 卡死(worker 不工作) | progress.pending > 0 && lastUpdated > 5min |
报警 → 看 billing-consumer 日志;手动重启 |
8. Configuration
8.1 env vars (libs/env.ts)
ADMIN_USER_IDS 逗号分隔 userId allowlist;空 = 没人是 admin
8.2 Worker options(可调,目前在代码里写死默认)
batchSize 默认 50 每次轮询拣多少 recipient
throttlePerSec 默认 50 每秒最多发多少条
maxAttempts 默认 3 超过即标 failed
idleSleepMs 默认 1000 无 work 时 sleep 多久
未来可以接 configKV 实时调整。
8.3 Route-level constants
MAX_GRANT_AMOUNT_PER_USER 10_000 单人 amount 上限
MAX_EMAILS_PER_BATCH 10_000 单 batch emails 列表上限
9. 测试策略
9.1 Unit / Integration
parseAdminUserIds+adminGuardmiddleware:allowlist 命中 / 不命中 / 空 allowlist / 无 sessionresolveEmails:case insensitive、找不到、用户注销、input 重复、samples 截断、estimatedDurationSec 边界createFluxGrantBatchService.create:persist batch + recipient rows,pending/skipped 状态正确createFluxGrantBatchService.preview:dry-run 不写库createFluxGrantBatchService.retryFailed:failed → pending 重置,completed batch 改回 running,无 failed 时 idempotentbackoffMs:3 个分支边界
9.2 文件位置
apps/server/src/middlewares/tests/admin-guard.test.tsapps/server/src/services/admin-flux-grant-batch/tests/flux-grant-batch-service.test.tsapps/server/src/services/admin-flux-grant-batch/tests/flux-grant-batch-worker.test.ts
10. 实施顺序
- Schema + migration:
flux_grant_batch+flux_grant_batch_recipient表,drizzle-kit generate BillingService.creditFlux扩展:加type入参,返回fluxTransactionId- Service 层:
createFluxGrantBatchService(创建、查询、retry)+ 共享 worker helper - Worker:
runFluxGrantBatchWorker轮询循环 - Middleware:
adminGuard+parseAdminUserIds - Routes:
/api/admin/flux-grant-batches/* - Worker 接入:在
src/bin/run-billing-consumer.ts用Promise.all并行启动 - 配置:env var
- Verification:
docs/ai/context/verifications/admin-flux-grant-batch.md(按 CLAUDE.md verification 格式手动跑一次端到端)
11. 已知风险 / Open Questions
- Redis cache 失效缺口:worker 在 DB 事务 commit 后、cache update 前 crash → 下次
getFluxcache miss 触发自动修复。能接受。 type='promo'软枚举:DB 不强制;下游若有人写type IN ('credit', 'initial')没改的话会漏掉 promo 记录。实施时已 grep 全量过一遍(命中flux-transaction.ts:60的 capacity 计算,已修复)。- 跨 batch 同 user 多次发:允许(不同 recipientId → 不同 requestId)。同一 batch 内通过
(batch_id, input_email)唯一索引防多发。 - 没有 daily total cap:理论上 dev 可以一次发 10k × 10000 FLUX = 100M FLUX。当前没硬限。靠 dry-run + code review。
- dashboard 联调:本设计假设 dashboard 后续会接入。spec 里只规范 API;UI 设计不在范围内。
12. 参考
apps/server/CLAUDE.md— 服务架构总览apps/server/docs/ai-context/billing-architecture.md— Flux/Stripe/outbox/Streamsbilling-service.ts— 现有creditFlux实现flux-transaction.ts— ledger schemaflux.ts—user_fluxschemaredis-keys.ts— Redis 键规范