WonderableAI/shaguabijia-app-server
11
0
Fork 0
Code Issues Pull Requests 12 Actions Packages Projects Releases Wiki Activity

功能：新手引导视频 + 美团券首页分页索引

Browse Source 
新手引导视频:运营后台上传 MP4(上限 100MB,魔数校验只认 ISO BMFF),
App 端在领券等候浮层前 N 次以引导视频替代广告。新增 guide_video
的 model/schema/repository/router(App 侧 + 后台侧)与播放记录表迁移。

美团券:首页「销量最高 / 智能推荐」两个 tab 改游标分页,配套两条
(city_id, dedup_key, 排序键 DESC) 复合索引,让 Postgres 顺着索引流式
去重,免掉每翻一页重排整城券的开销。美团 CPS client 在 lifespan 预热
并在关闭时释放连接池。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
左辰勇
2026-07-23 22:50:44 +08:00
parent b7cfcf7495
commit a2270ee1b2
22 changed files with 1456 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files
app/api/v1/guide_video.py
+64
View File
@@ -0,0 +1,64 @@
"""新手引导视频(领券等候浮层前 N 次替代广告)。
路由前缀 `/api/v1/guide-video`(均需 Bearer):
POST /start 这次浮层放引导视频还是放广告?命中则**当场计次**并下发 play_token
POST /reward 播完 / 中途关闭都调,按 play_token 幂等发固定金币
发币额度以**服务端配置**为准(运营后台可改),客户端只报"播完/关闭",报不了金额,
所以被破解也刷不到超额金币;次数上限由 guide_video_play 行数(按账号)硬卡。
"""
from __future__ import annotations
import logging
from fastapi import APIRouter, Depends
from app.api.deps import CurrentUser, DbSession
from app.core.ratelimit import rate_limit
from app.repositories import guide_video as crud_guide
from app.schemas.guide_video import (
GuideVideoRewardIn,
GuideVideoRewardOut,
GuideVideoStartIn,
GuideVideoStartOut,
)
logger = logging.getLogger("shagua.guide_video")
router = APIRouter(prefix="/api/v1/guide-video", tags=["guide-video"])
@router.post(
"/start",
response_model=GuideVideoStartOut,
summary="领券浮层是否放新手引导视频(命中即计次)",
dependencies=[Depends(rate_limit(60, 60, "guide-video-start"))],
)
def start(payload: GuideVideoStartIn, user: CurrentUser, db: DbSession) -> GuideVideoStartOut:
"""开播即计数:返回 should_play=True 时服务端已写下这一次,客户端必须真的播。
没配视频 / 开关关 / 次数用完 → should_play=False,客户端照旧走广告链路(行为不变)。
"""
result = crud_guide.start_play(db, user.id, scene=payload.scene or "coupon")
logger.info(
"guide video start user_id=%d scene=%s should_play=%s seq=%d remaining=%d",
user.id, payload.scene, result["should_play"], result["seq"], result["remaining"],
)
return GuideVideoStartOut(**result)
@router.post(
"/reward",
response_model=GuideVideoRewardOut,
summary="引导视频发金币(播完/中途关闭都发,play_token 幂等)",
dependencies=[Depends(rate_limit(60, 60, "guide-video-reward"))],
)
def reward(payload: GuideVideoRewardIn, user: CurrentUser, db: DbSession) -> GuideVideoRewardOut:
result = crud_guide.grant_play(
db, user.id, play_token=payload.play_token, completed=payload.completed
)
logger.info(
"guide video reward user_id=%d token=%s completed=%s granted=%s coin=%d",
user.id, payload.play_token[:12], payload.completed, result["granted"], result["coin"],
)
return GuideVideoRewardOut(**result)
app/api/v1/meituan.py
+154 -75
View File
@@ -5,11 +5,12 @@
from __future__ import annotations
import logging
from concurrent.futures import ThreadPoolExecutor, as_completed
from concurrent.futures import ThreadPoolExecutor
from typing import TYPE_CHECKING, Any
from fastapi import APIRouter, Depends, HTTPException
from sqlalchemy import nullslast, select
from sqlalchemy.orm import Session, aliased
from sqlalchemy.orm import Session
from app.core.config import settings
from app.db.session import get_db
@@ -25,8 +26,14 @@ from app.schemas.meituan import (
ReferralLinkResponse,
TopSalesRequest,
)
from app.utils import mt_search_cursor
from app.utils.meituan_city import get_meituan_city
if TYPE_CHECKING: # 仅供类型标注(本模块已开 from __future__ import annotations)
from collections.abc import Callable
from sqlalchemy import ColumnElement
logger = logging.getLogger("shagua.meituan")
@@ -109,6 +116,86 @@ def _commission_pct(card: CouponCard) -> float:
return 0.0
# ────────────── 离线库分页(智能推荐 / 销量最高 共用) ──────────────
# 去重+排序阶段**只投影这几列**:够 DISTINCT ON 分组、够排序、够回表定位,且全是定长小字段。
# ⚠️ 关键性能点:`raw` 是整条美团原始返回(JSONB,每行数 KB)。原实现用 select(MeituanCoupon)
# 做子查询,等于把整城几千行连 raw 一起塞进两次排序(DISTINCT ON 一次 + 分页一次),
# 体量轻松超过 work_mem → Postgres 落盘做外部归并排序,而且**每翻一页都要重来一遍**。
# 拆成「先在小列上排出本页 id,再按 id 回表取 raw」后,排序数据量降到原来的百分之几,
# JSONB 只解析当前页 ~20 行。
_DEDUP_COLS = (
MeituanCoupon.id,
MeituanCoupon.dedup_key,
MeituanCoupon.sale_volume_num,
MeituanCoupon.commission_percent,
)
def _paged_dedup_ids(
db: Session,
*,
conds: list[ColumnElement[bool]],
dedup_order: list[ColumnElement],
page_order: Callable[[Any], list[ColumnElement]],
page: int,
page_size: int,
) -> tuple[list[int], bool]:
"""DISTINCT ON(dedup_key) 跨源去重 → 整体排序 → 分页,返回 (本页 id 列表, 是否还有下一页)。
- `dedup_order`:同一个 dedup_key 的多条里留哪条(如销量最高/佣金最高)。
- `page_order`:接收去重子查询的列集合(`sub.c`),返回去重后的整体排序。
多取 1 条用于判断 has_next。
"""
deduped = (
select(*_DEDUP_COLS)
.where(*conds)
.distinct(MeituanCoupon.dedup_key)
.order_by(MeituanCoupon.dedup_key, *dedup_order)
.subquery()
)
ids = db.execute(
select(deduped.c.id)
.order_by(*page_order(deduped.c))
.offset((page - 1) * page_size)
.limit(page_size + 1)
).scalars().all()
return list(ids[:page_size]), len(ids) > page_size
def _load_raws(db: Session, ids: list[int]) -> list[dict]:
"""按给定 id 顺序取 raw(只回表本页 ~20 行)。缺行(被 ETL 清掉)静默跳过。"""
if not ids:
return []
raw_by_id = {
row_id: raw
for row_id, raw in db.execute(
select(MeituanCoupon.id, MeituanCoupon.raw).where(MeituanCoupon.id.in_(ids))
).all()
}
return [raw_by_id[i] for i in ids if i in raw_by_id]
def _cards_from_raws(raws: list[dict], *, hide_distance: bool) -> list[CouponCard]:
"""raw → CouponCard;解析失败的单条跳过,不整页失败。
hide_distance:离线库里的距离是相对「城市默认点」算的,对用户无意义且误导 —— 智能推荐 /
销量最高两个 tab 一律置空,前端「距离 店名」那行只剩店名、自动顶到最左。
"""
cards: list[CouponCard] = []
for raw in raws:
try:
card = CouponCard.from_raw(raw or {})
except Exception: # noqa: BLE001
continue
if not card.product_view_sign:
continue
if hide_distance:
card.distance_text = None
card.distance_meters = None
cards.append(card)
return cards
@router.post("/feed", response_model=FeedResponse, summary="混合feed(外卖+到店交叉);tab=rec智能推荐/distance距离最近")
def feed(req: FeedRequest, db: Session = Depends(get_db)) -> FeedResponse:
lon, lat = req.longitude, req.latitude
@@ -133,17 +220,21 @@ def feed(req: FeedRequest, db: Session = Depends(get_db)) -> FeedResponse:
return [], True
# 距离最近:搜索召回(外卖搜"外卖" + 到店搜"美食",都 sortField=6 离我最近)一页页拉。
# 搜索翻页必须用 searchId(pageNo 翻不动),所以每个 feed 页顺序翻到第 N 页;两路并行、page 1 最快。
# 无状态、不改 APP(传页码即可);按你位置实时算距离(库里没存 POI 经纬度,只能实时)。
# 搜索翻页必须用 searchId(pageNo 翻不动),而接口是无状态的(客户端只传页码)—— 原实现因此
# 每次都从第 1 页顺序重放到第 N 页,取第 N 页要向美团发 N 次请求,越往下滑越慢。
# 现在把沿途 searchId 记进 [mt_search_cursor],稳态下每翻一页恒定 1 次请求;两路仍并行。
# 按你位置实时算距离(库里没存 POI 经纬度,只能实时)。
if tab == "distance":
lon_i, lat_i = int(lon * 1_000_000), int(lat * 1_000_000)
def _search_page_n(platform: int, biz_line: int | None, keyword: str, n: int) -> tuple[list[dict], bool, bool]:
"""顺序翻到第 n 页(搜索须 searchId 续页),返回(第 n 页 items, 是否还有下一页, 是否调用失败)。"""
sid: str | None = None
def _replay(
platform: int, biz_line: int | None, keyword: str,
key: mt_search_cursor.RouteKey, start: int, sid: str | None, n: int,
) -> tuple[list[dict], bool, bool]:
"""从第 start 页(用 sid 取)顺序翻到第 n 页。start==1 时 sid 应为 None(走 pageNo=1)。"""
data: list[dict] = []
has_next = False
for pg in range(1, n + 1):
for pg in range(start, n + 1):
body: dict = {
"platform": platform, "searchText": keyword, "sortField": 6,
"longitude": lon_i, "latitude": lat_i, "pageSize": 20,
@@ -161,10 +252,27 @@ def feed(req: FeedRequest, db: Session = Depends(get_db)) -> FeedResponse:
data = r.get("data") or []
sid = r.get("searchId")
has_next = bool(r.get("hasNext")) and bool(data)
# 记下「下一页要用哪个 searchId」;没有下一页就别记,免得存进死游标。
if sid and has_next:
mt_search_cursor.remember(key, pg + 1, sid)
if not data or (not has_next and pg < n):
return [], False, False # 没那么多页了(非错误)
return data, has_next, False
def _search_page_n(platform: int, biz_line: int | None, keyword: str, n: int) -> tuple[list[dict], bool, bool]:
"""取第 n 页,返回(第 n 页 items, 是否还有下一页, 是否调用失败)。
优先用缓存游标一发直达;缓存未命中/过期才从最近的已知页往后重放,并把沿途游标补进缓存。
"""
key = mt_search_cursor.route_key(lat, lon, platform, keyword)
start, sid = mt_search_cursor.lookup(key, n)
data, has_next, failed = _replay(platform, biz_line, keyword, key, start, sid, n)
# 用缓存游标却打不通,多半是上游 searchId 过期:作废整条路线,回到第 1 页重放一次。
if failed and start > 1:
mt_search_cursor.drop(key)
data, has_next, failed = _replay(platform, biz_line, keyword, key, 1, None, n)
return data, has_next, failed
with ThreadPoolExecutor(max_workers=2) as pool:
f_wm = pool.submit(_search_page_n, 1, None, "外卖", req.page)
f_dd = pool.submit(_search_page_n, 2, 1, "美食", req.page)
@@ -194,39 +302,25 @@ def feed(req: FeedRequest, db: Session = Depends(get_db)) -> FeedResponse:
return FeedResponse(items=[], has_next=False, page=req.page, status="degraded")
PAGE = 20
try:
base = select(MeituanCoupon).where(
MeituanCoupon.commission_percent >= 3.0,
MeituanCoupon.city_id == city_id,
)
deduped = base.distinct(MeituanCoupon.dedup_key).order_by(
MeituanCoupon.dedup_key,
MeituanCoupon.commission_percent.desc(),
).subquery()
m = aliased(MeituanCoupon, deduped)
start = (req.page - 1) * PAGE
rows = db.execute(
select(m)
ids, has_next = _paged_dedup_ids(
db,
conds=[
MeituanCoupon.commission_percent >= 3.0,
MeituanCoupon.city_id == city_id,
],
# 同一去重键留佣金最高那条
dedup_order=[MeituanCoupon.commission_percent.desc()],
# 销量高的优先(无销量档排后),同档佣金高优先,id 兜底稳定分页
.order_by(nullslast(m.sale_volume_num.desc()), m.commission_percent.desc(), m.id)
.offset(start)
.limit(PAGE + 1)
).scalars().all()
page_order=lambda c: [
nullslast(c.sale_volume_num.desc()), c.commission_percent.desc(), c.id,
],
page=req.page, page_size=PAGE,
)
raws = _load_raws(db, ids)
except Exception: # noqa: BLE001
logger.exception("[feed] rec 库查询失败,降级返空")
return FeedResponse(items=[], has_next=False, page=req.page, status="degraded")
has_next = len(rows) > PAGE
cards: list[CouponCard] = []
for row in rows[:PAGE]:
try:
card = CouponCard.from_raw(row.raw or {})
except Exception: # noqa: BLE001
continue
if card.product_view_sign:
# 智能推荐不显示距离:库里的距离是相对城市默认点的(对用户无意义、且误导)。
# 置空后前端"距离 店名"那行只剩店名、自动顶到最左(店名移到原距离的位置)。
card.distance_text = None
card.distance_meters = None
cards.append(card)
cards = _cards_from_raws(raws, hide_distance=True)
if not cards and req.page == 1:
# 命中城市却 0 券:该城确无 ≥3% 券,或 ETL 灌的 city_id 与 city_dict 口径不一致。
logger.info("[feed] rec city_id=%s 命中 0 券(该城确无券?或 ETL/city_dict 的 city_id 口径不一致)", city_id)
@@ -282,51 +376,36 @@ def top_sales(req: TopSalesRequest, db: Session = Depends(get_db)) -> CouponList
if not city_id:
return CouponListResponse(items=[], has_next=False, search_id=None, status="degraded")
# 去重 + 排序 + 分页全在 SQL 做,每页只取并解析当前页 ~20 条。
# (之前实现每翻一页都全表拉取 + 全量 from_raw 解析,翻页慢 → 客户端滑动卡顿/翻不动。)
# 去重 + 排序 + 分页全在 SQL 做,每页只回表并解析当前页 ~20 条(见 _paged_dedup_ids 的性能说明)。
# 库为空(prod 刚部署 / ETL 未跑完)时返空 + status=empty,不崩;库查询异常降级 degraded。
conds = [
MeituanCoupon.sale_volume_num.isnot(None),
MeituanCoupon.city_id == city_id,
]
if req.platform is not None:
conds.append(MeituanCoupon.platform == req.platform)
try:
# 1) DISTINCT ON (dedup_key):每个去重键(品牌|名|价)只留销量最高那条(同销量再按佣金)
base = select(MeituanCoupon).where(
MeituanCoupon.sale_volume_num.isnot(None),
MeituanCoupon.city_id == city_id,
)
if req.platform is not None:
base = base.where(MeituanCoupon.platform == req.platform)
deduped = base.distinct(MeituanCoupon.dedup_key).order_by(
MeituanCoupon.dedup_key,
MeituanCoupon.sale_volume_num.desc(),
MeituanCoupon.commission_percent.desc(),
).subquery()
# 2) 对去重结果按销量降序分页;多取 1 条判断 has_next,只对本页做 from_raw
m = aliased(MeituanCoupon, deduped)
start = (req.page - 1) * req.page_size
rows = db.execute(
select(m)
ids, has_next = _paged_dedup_ids(
db,
conds=conds,
# 每个去重键(品牌|名|价)只留销量最高那条(同销量再按佣金)
dedup_order=[
MeituanCoupon.sale_volume_num.desc(),
MeituanCoupon.commission_percent.desc(),
],
# 加 id 作稳定 tiebreaker:同销量同佣金的并列项排序确定,避免跨页重复/漏项
.order_by(m.sale_volume_num.desc(), m.commission_percent.desc(), m.id)
.offset(start)
.limit(req.page_size + 1)
).scalars().all()
page_order=lambda c: [
c.sale_volume_num.desc(), c.commission_percent.desc(), c.id,
],
page=req.page, page_size=req.page_size,
)
raws = _load_raws(db, ids)
except Exception: # noqa: BLE001
logger.exception("[top-sales] 库查询失败,降级返空")
return CouponListResponse(items=[], has_next=False, search_id=None, status="degraded")
has_next = len(rows) > req.page_size
cards: list[CouponCard] = []
for row in rows[:req.page_size]:
try:
card = CouponCard.from_raw(row.raw or {})
except Exception: # noqa: BLE001
continue
if card.product_view_sign:
# 不显示距离:库里的距离是相对城市默认点的(对用户无意义、且误导)。
# 置空后前端"距离 店名"那行只剩店名、自动顶到最左(店名移到原距离的位置)。
# 逻辑与推荐流保持一致
card.distance_text = None
card.distance_meters = None
cards.append(card)
# 不显示距离:库里的距离是相对城市默认点的(对用户无意义、且误导),与推荐流口径一致。
cards = _cards_from_raws(raws, hide_distance=True)
if not cards and req.page == 1:
# 命中城市却 0 券:可能该城确无券,也可能 ETL 灌的 city_id 与 city_dict 口径不一致(静默降级的隐患)。
logger.info("[top-sales] city_id=%s 命中 0 券(该城确无券?或 ETL/city_dict 的 city_id 口径不一致)", city_id)