perf: 提交列表等查询改走覆盖索引，避免读取 html/css/js 大行

SQLite 把 html/css/js 存在行内，而 user_id、task_id、flag 等列排在它们后面，
读这些列也要先读完整段代码；线上库也没有统计信息，规划器常挑错索引。

- 新增覆盖索引 submission_submission_slim
- 提交列表、随机打分先在索引上取 rowid，再只读这一页 / 这一行的完整内容
- 列表的 flag / zone 筛选用一元 + 避开单列索引
- 打开数据库时执行 PRAGMA optimize，之后每小时一次

线上备份实测：提交列表 844 ms → 6 ms，打分后重算分数 11.5 ms → 0.2 ms。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

server/src/db.ts

@@ -25,6 +25,9 @@ CREATE INDEX IF NOT EXISTS "submission_submission_task_id_91e9556b" ON "submissi
 CREATE INDEX IF NOT EXISTS "submission_submission_user_id_71741691" ON "submission_submission" ("user_id");
 CREATE INDEX IF NOT EXISTS "submission_submission_flag_d96bc4aa" ON "submission_submission" ("flag");
 CREATE INDEX IF NOT EXISTS "submission_submission_zone_a1bbf0a2" ON "submission_submission" ("zone");
+-- 非 Django 原有：除 html/css/js 外的常用列。user_id 等列在表里排在 html/css/js 后面，
+-- 直接读表要先读完整段代码，所以列表、打分、统计等扫描都靠这个覆盖索引
+CREATE INDEX IF NOT EXISTS "submission_submission_slim" ON "submission_submission" ("user_id", "task_id", "created", "score", "flag", "zone", "raw_score", "id");
 CREATE TABLE IF NOT EXISTS "submission_rating" ("id" integer NOT NULL PRIMARY KEY AUTOINCREMENT, "score" integer NOT NULL, "created" datetime NOT NULL, "submission_id" char(32) NOT NULL REFERENCES "submission_submission" ("id") DEFERRABLE INITIALLY DEFERRED, "user_id" bigint NOT NULL REFERENCES "account_user" ("id") DEFERRABLE INITIALLY DEFERRED);
 CREATE UNIQUE INDEX IF NOT EXISTS "submission_rating_user_id_submission_id_29be4c39_uniq" ON "submission_rating" ("user_id", "submission_id");
 CREATE INDEX IF NOT EXISTS "submission_rating_submission_id_a20dcece" ON "submission_rating" ("submission_id");
@@ -57,6 +60,8 @@ export function openDb(path: string): Database {
   db.run("PRAGMA busy_timeout = 5000")
   db.run("PRAGMA foreign_keys = ON")
   db.run(SCHEMA)
+  // SQLite 不会自动 ANALYZE，没有统计信息时规划器常挑错索引；0x10002 = 打开时只分析必要的表
+  db.run("PRAGMA optimize=0x10002")
   return db
 }
 

server/src/index.ts

@@ -1,5 +1,6 @@
 import { app } from "./app"
 import { config } from "./config"
+import { db } from "./db"
 import { upgrade, websocket, type WsData } from "./ws"
 
 const server = Bun.serve<WsData>({
@@ -16,4 +17,7 @@ const server = Bun.serve<WsData>({
   websocket,
 })
 
+// 长期运行的连接按 SQLite 文档建议定期刷新统计信息
+setInterval(() => db.run("PRAGMA optimize"), 60 * 60 * 1000)
+
 console.log(`webpreview server listening on http://localhost:${server.port}`)

server/src/routes/submission.ts

@@ -234,13 +234,14 @@ router.get("/", (c) => {
     where.push("s.user_id = $user_id")
     params.user_id = f.user_id
   }
-  if (f.flag === "any") where.push("s.flag IS NOT NULL")
+  // flag / zone 前的一元 + 让 SQLite 不用这两列的单列索引（那样要回表读大行），改走 slim 覆盖索引
+  if (f.flag === "any") where.push("+s.flag IS NOT NULL")
   else if (f.flag) {
-    where.push("s.flag = $flag")
+    where.push("+s.flag = $flag")
     params.flag = f.flag
   }
   if (f.zone) {
-    where.push("s.zone = $zone")
+    where.push("+s.zone = $zone")
     params.zone = f.zone
   }
   if (f.score_lt_threshold !== undefined) {
@@ -258,7 +259,8 @@ router.get("/", (c) => {
   }
   if (f.grouped) {
     // 分组模式：每个 (user, task) 只保留最新一条
-    where.push(`s.id = (SELECT s3.id FROM submission_submission s3
+    // 用 rowid 而不是 id，这样子查询能完全走 submission_submission_slim 覆盖索引
+    where.push(`s.rowid = (SELECT s3.rowid FROM submission_submission s3
       WHERE s3.user_id = s.user_id AND s3.task_id = s.task_id ORDER BY s3.created DESC LIMIT 1)`)
   }
 
@@ -271,12 +273,16 @@ router.get("/", (c) => {
   const whereSql = where.length ? `WHERE ${where.join(" AND ")}` : ""
   const { limit, offset } = pageWindow(f)
 
+  // 先只在索引上排序分页拿到这一页的 rowid，再读这几行的完整内容；
+  // 否则排序时要把每条提交的 html/css/js 大行都读一遍
   const items = all(
     `${SUBMISSION_SELECT},
        (SELECT COUNT(*) FROM submission_submission s2
         WHERE s2.user_id = s.user_id AND s2.task_id = s.task_id) AS submit_count
-     ${SUBMISSION_FROM} ${whereSql}
+     ${SUBMISSION_FROM}
-     ORDER BY ${orderBy} LIMIT $limit OFFSET $offset`,
+     WHERE s.rowid IN (SELECT s.rowid ${SUBMISSION_FROM} ${whereSql}
+                       ORDER BY ${orderBy} LIMIT $limit OFFSET $offset)
+     ORDER BY ${orderBy}`,
     { ...params, me: user.id, limit, offset },
   )
   const count = get<{ n: number }>(`SELECT COUNT(*) AS n ${SUBMISSION_FROM} ${whereSql}`, params)!.n
@@ -305,20 +311,25 @@ router.get("/random-for-rating", (c) => {
   }
   if (user.role === Role.NORMAL && todayRatingCount(user.id) >= 30) return c.json(null)
 
-  const base = `SELECT s.id, u.username, t.title AS task_title, t.display AS task_display,
+  // 只在索引上随机挑出 rowid，再读这一行的代码；否则每个候选都要读一遍 html/css/js
-      t.task_type, s.html, s.css, s.js
+  const base = `SELECT s.rowid AS rid FROM submission_submission s
-    ${SUBMISSION_FROM}
     WHERE s.user_id != $me
       AND NOT EXISTS (SELECT 1 FROM submission_rating r WHERE r.submission_id = s.id AND r.user_id = $me)
       ${excludeId ? "AND s.id != $exclude" : ""}`
   const params = { me: user.id, ...(excludeId ? { exclude: excludeId } : {}) }
-  const pick =
+  const picked =
-    get(
+    get<{ rid: number }>(
       `${base} AND (SELECT COUNT(*) FROM submission_rating r WHERE r.submission_id = s.id) < 5
        ORDER BY RANDOM() LIMIT 1`,
       params,
-    ) ?? get(`${base} ORDER BY RANDOM() LIMIT 1`, params)
+    ) ?? get<{ rid: number }>(`${base} ORDER BY RANDOM() LIMIT 1`, params)
-  if (!pick) return c.json(null)
+  if (!picked) return c.json(null)
+  const pick = get(
+    `SELECT s.id, u.username, t.title AS task_title, t.display AS task_display,
+       t.task_type, s.html, s.css, s.js
+     ${SUBMISSION_FROM} WHERE s.rowid = $rid`,
+    { rid: picked.rid },
+  )!
   return c.json({
     submission_id: uuidOut(pick.id),
     username: pick.username,