"""usage_events 聚合读(报表側)—— 与写侧 usage.py 同层收口(2026-07-23)。 背景:`UsageEvent.units` JSONB 的 key(tokens_in/tokens_out/cache_hit_tokens) 由 usage.py 写入,此前读侧的 `cast(units[...].astext, BigInteger)` 散在 web/admin.py 与 web/app.py 各自硬编码 —— 写读跨文件隐式耦合,改 key 极易漏改 导致计费统计静默错。本模块是 units 结构**唯一的读侧出口**:web 层只调函数, 不再直接碰 JSONB cast。改 units key 时,写侧 usage.py 与本模块同文件夹同 PR 改。 口径约定(与前端展示一致,tests/test_usage_report.py 锁行为): - cost_cny:全 kind 合计(chat+image+video+vision+...)= 真实花费 - tokens_in/out + cache_hit:仅 kind='chat'。三者同源,缓存命中率 cache_hit/tokens_in 恒 ≤100%(绝不能拿 tasks.tokens_prompt 当分母 —— 那列会被「清空对话」重置而 usage_events 不重置)。 """ from __future__ import annotations from typing import Any, Optional from sqlalchemy import BigInteger, and_, cast, func, select from .models import UsageEvent, User # ── units JSONB 读侧列表达式(单一事实源;写侧 key 见 usage.py record_chat_usage)── _CHAT = UsageEvent.kind == "chat" _TIN = cast(UsageEvent.units["tokens_in"].astext, BigInteger) _TOUT = cast(UsageEvent.units["tokens_out"].astext, BigInteger) _HIT = cast(UsageEvent.units["cache_hit_tokens"].astext, BigInteger) def task_usage_aggregates(s: Any, tids: list) -> dict: """按 task_id 批量聚合:真实成本 + chat token + 缓存命中。 单查询 GROUP BY(复用列表接口 msg_counts 同款批量范式,无 N+1)。on-the-fly 现算, 不落 tasks 列 —— 对所有历史 task 即时准确,免回填。 返回 {task_id: {"cost_cny": float, "tokens_in": int, "tokens_out": int, "tokens_cache_hit": int}}。 """ if not tids: return {} rows = s.execute( select( UsageEvent.task_id, func.coalesce(func.sum(UsageEvent.cost_cny), 0), func.coalesce(func.sum(_TIN).filter(_CHAT), 0), func.coalesce(func.sum(_TOUT).filter(_CHAT), 0), func.coalesce(func.sum(_HIT).filter(_CHAT), 0), ) .where(UsageEvent.task_id.in_(tids)) .group_by(UsageEvent.task_id) ).all() return { tid: { "cost_cny": float(cost or 0), "tokens_in": int(tin or 0), "tokens_out": int(tout or 0), "tokens_cache_hit": int(hit or 0), } for tid, cost, tin, tout, hit in rows } def usage_overview(s: Any, cutoff_7d) -> dict: """全局合计(all-time)+ 近 7d 按天趋势(admin overview 的 usage section)。 按模型 / 各用户用量是独立带筛选排序的函数(models_usage / user_usage_page), 不在此 bundle。 """ # 全局合计(all-time) g = s.execute( select( func.coalesce(func.sum(UsageEvent.cost_cny), 0), func.coalesce(func.sum(_TIN).filter(_CHAT), 0), func.coalesce(func.sum(_TOUT).filter(_CHAT), 0), func.coalesce(func.sum(_HIT).filter(_CHAT), 0), func.count(), ) ).one() total = { "cost_cny": float(g[0] or 0), "tokens_in": int(g[1] or 0), "tokens_out": int(g[2] or 0), "tokens_cache_hit": int(g[3] or 0), "n_events": int(g[4] or 0), } # 近 7d 按天(date 截断;前端画成条/数字均可);按日期倒序 —— 最新一天在最上面 day = func.date(UsageEvent.created_at) by_day = [ { "date": str(d), "cost_cny": float(c or 0), "tokens_in": int(ti or 0), "tokens_out": int(to or 0), } for d, c, ti, to in s.execute( select( day, func.coalesce(func.sum(UsageEvent.cost_cny), 0), func.coalesce(func.sum(_TIN).filter(_CHAT), 0), func.coalesce(func.sum(_TOUT).filter(_CHAT), 0), ) .where(UsageEvent.created_at >= cutoff_7d) .group_by(day) .order_by(day.desc()) ).all() ] return {"total": total, "by_day_7d": by_day} def models_usage(s: Any, cutoff, sort: str) -> list: """按模型用量(支持时间筛选 + 排序)。sort: cost(按成本)/ tokens(按用量=输入+输出)。 cutoff=None 即全部;cost 全 kind 合计,token 仅 chat。模型集合从 usage_events 现取 (无"全模型"基线),故时间条件直接进 WHERE。 """ cost_sum = func.coalesce(func.sum(UsageEvent.cost_cny), 0) tin_sum = func.coalesce(func.sum(_TIN).filter(_CHAT), 0) tout_sum = func.coalesce(func.sum(_TOUT).filter(_CHAT), 0) order = (tin_sum + tout_sum).desc() if sort == "tokens" else cost_sum.desc() q = select( UsageEvent.model_profile, cost_sum, tin_sum, tout_sum, func.count(), ) if cutoff is not None: q = q.where(UsageEvent.created_at >= cutoff) q = q.group_by(UsageEvent.model_profile).order_by(order, UsageEvent.model_profile) return [ { "model_profile": mp, "cost_cny": float(c or 0), "tokens_in": int(ti or 0), "tokens_out": int(to or 0), "n_events": int(n or 0), } for mp, c, ti, to, n in s.execute(q).all() ] def user_usage_page( s: Any, page: int, page_size: int, cutoff, sort: str, ) -> dict: """分页的各用户 token 用量(时间筛选 + 排序),含零用量用户(LEFT JOIN users)。 `各用户` 取自 users 全表 LEFT JOIN usage_events,故没产生过用量的用户也出现(0); 时间筛选放 JOIN ON(非 WHERE),否则带 cutoff 时会把零用量用户挤掉。 sort: cost(按成本)/ tokens(按用量=输入+输出);+ user_id 兜底稳定分页。 cost 全 kind 合计;token/cache_hit 仅 chat。返回 {page, page_size, total_users, rows}。 """ cost_sum = func.coalesce(func.sum(UsageEvent.cost_cny), 0) tin_sum = func.coalesce(func.sum(_TIN).filter(_CHAT), 0) tout_sum = func.coalesce(func.sum(_TOUT).filter(_CHAT), 0) order = (tin_sum + tout_sum).desc() if sort == "tokens" else cost_sum.desc() join_cond = UsageEvent.user_id == User.user_id if cutoff is not None: join_cond = and_(join_cond, UsageEvent.created_at >= cutoff) # 最近使用时间:取全量(不随 range 筛选变),否则 7d/30d 会把更早的真实 last-used 藏掉。 last_used_sq = ( select(func.max(UsageEvent.created_at)) .where(UsageEvent.user_id == User.user_id) .correlate(User) .scalar_subquery() ) total_users = s.execute(select(func.count()).select_from(User)).scalar_one() rows = [ { "user_id": str(uid), "email": email or "", "name": name or "", "user_name": uname or "", "role": role or "user", "plan": plan or "", # 模型档位(空 → default 档),admin UI 内联下拉用 "cost_cny": float(c or 0), "tokens_in": int(ti or 0), "tokens_out": int(to or 0), "tokens_cache_hit": int(h or 0), "n_events": int(n or 0), "last_used_at": last_used.isoformat() if last_used else None, } for uid, email, name, uname, role, plan, c, ti, to, h, n, last_used in s.execute( select( User.user_id, User.email, User.name, User.user_name, User.role, User.plan, cost_sum, tin_sum, tout_sum, func.coalesce(func.sum(_HIT).filter(_CHAT), 0), func.count(UsageEvent.event_id), last_used_sq.label("last_used_at"), ) .join(UsageEvent, join_cond, isouter=True) .group_by(User.user_id, User.email, User.name, User.user_name, User.role, User.plan) .order_by(order, User.user_id) .limit(page_size) .offset(page * page_size) ).all() ] return {"page": page, "page_size": page_size, "total_users": total_users, "rows": rows}