feat(阶段2补课): SQL 判题链路 + 最后两个后台端点

新后端此前完全没有 SQL 判题（旧 judge/sql_runner.py 378 行 + sql_dispatcher.py 113 行无对应实现），阶段 2 纵切时漏了这条与沙箱完全不同的路径。 judge/sql/engine.ts 判题核心，移植自 sql_runner.py，判定口径逐条对齐 judge/sql/child.ts 子进程入口 judge/sql/index.ts 父进程：spawn + 硬超时 judge/run.ts language === "SQL" 时分流，不经判题沙箱 POST admin/sql-test-cases/preview 题目页展示数据预览 POST admin/sql-test-cases/generate AI 按标准答案倒推初始化脚本题目保存时重新生成 sqlDisplay（对齐旧 generate_sql_display）：取测试点 1 的初始化脚本 + 标准答案跑一遍，失败一律拦下不让保存 —— 展示数据直接决定学生看到的表结构与期望结果，宁可不保存也不能存错的。 ## 防护换了实现，逐条实测 bun:sqlite 没有 authorizer / progress_handler / setlimit，且实测 Worker.terminate() 杀不掉跑飞的查询（原生代码占着线程）。改用「WASM 引擎 + 独立子进程」： ATTACH → WASM 无宿主文件系统绑定，结构上够不到（比旧的 authorizer 更强）查询题只读 → PRAGMA query_only=1 超时 → 子进程外部 SIGKILL 单值内存 → 子进程 ulimit -d 八条提交实测：正确→Accepted；列少一个/漏过滤→Wrong Answer；语法错误→Compile Error；查询题里 INSERT→运行错误并说明；递归 CTE 死循环→CPU 超时；hex(zeroblob(2e8))→内存超限； attach '/etc/passwd'→打不开。 ## 踩到的两个坑（已写进 docs/specs/phase3-coverage.md） 1. ulimit 必须用 -d 不能用 -v。-v 限虚拟地址空间而 JS 引擎预留巨量地址，实测 -v 之下 Bun 退出时有概率 panic（SIGILL），结果早已写出但进程异常终止，父进程读到空串误判成超时 —— 6 次里坏 2 次。换 -d 后 12/12 稳定。 2. 子进程写完结果直接 SIGKILL 自己，不走 process.exit()——后者仍有清理会撞限额。至此 admin/api.ts 已无任何指向旧后端的调用。 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-07 17:03:41 -06:00
parent 007ae8619b
commit e3faa689e7
10 changed files with 913 additions and 26 deletions
--- a/apps/api/src/judge/run.ts
+++ b/apps/api/src/judge/run.ts
@@ -16,12 +16,18 @@ import {
  type JudgeStatusValue,
 } from "./status"
 import { parseProblemTemplate } from "./template"
+import { runSqlCase } from "./sql"
+import { readInfo } from "../services/test-case"
+import { readFile } from "node:fs/promises"
+import { resolve as resolvePath } from "node:path"

 interface JudgeCase {
  cpu_time: number
  memory: number
  result: number
  test_case: string
+  /** SQL 判题会带上中文原因，沙箱判题没有这个字段 */
+  error_message?: string | null
  [key: string]: unknown
 }

@@ -350,13 +356,17 @@ export async function judgeSubmission(job: JudgeJobData) {
      ? `${template.prepend}\n${row.submission.code}\n${template.append}`
      : row.submission.code

-    const response = await requestJudge(
-      row.submission.language,
-      source,
-      row.problem.timeLimit,
-      row.problem.memoryLimit,
-      row.problem.testCaseId,
-    )
+    // SQL 题不经判题沙箱：沙箱是给编译型/脚本型语言用的，SQL 判的是结果集，
+    // 走 judge/sql 的 WASM 引擎（在独立子进程里跑，见那边的说明）。
+    const response = row.submission.language === "SQL"
+      ? await judgeSqlSubmission(row.problem, row.submission.code)
+      : await requestJudge(
+          row.submission.language,
+          source,
+          row.problem.timeLimit,
+          row.problem.memoryLimit,
+          row.problem.testCaseId,
+        )

    let result: JudgeStatusValue
    let info: unknown = {}
@@ -386,6 +396,13 @@ export async function judgeSubmission(job: JudgeJobData) {
        memory_cost: Math.max(0, ...cases.map((item) => Number(item.memory) || 0)),
        score: 0,
      }
+      // SQL 判题给出的中文提示（只读拒绝/超时/内存/无结果集）只存在测试点的
+      // error_message 里，而前端只读 statistic_info.err_info。把首个失败测试点的
+      // 提示提上来，否则学生只看到一个 WA 却不知道原因。对齐旧 sql_dispatcher。
+      const failedMessage = cases.find(
+        (item) => item.result !== JudgeStatus.ACCEPTED && item.error_message,
+      )?.error_message
+      if (typeof failedMessage === "string") statisticInfo.err_info = failedMessage

      if (result === JudgeStatus.ACCEPTED) {
        const rules = astRulesForLanguage(
@@ -454,3 +471,76 @@ export async function judgeSubmission(job: JudgeJobData) {
    await markSystemError(row.submission.id, row.submission.userId, error)
  }
 }
+
+
+/**
+ * SQL 题判题：逐个测试点用各自的初始化脚本跑一遍，产出与沙箱同构的结果结构，
+ * 好让上面的状态聚合、统计、排名、WebSocket 推送逻辑完全复用。
+ * 对齐旧 `judge/sql_dispatcher.py`。
+ */
+async function judgeSqlSubmission(
+  problem: typeof schema.problem.$inferSelect,
+  studentSql: string,
+): Promise<JudgeResponse> {
+  const sqlConfig = objectValue(problem.sqlConfig)
+  const mode = sqlConfig.mode
+  if (mode !== "query" && mode !== "modify") {
+    throw new Error("题目缺少 SQL 配置（题型）")
+  }
+  const answers = Array.isArray(problem.answers) ? problem.answers : []
+  const refSql = answers
+    .map((item) => objectValue(item))
+    .find((item) => item.language === "SQL" && typeof item.code === "string" && item.code.trim())?.code
+  if (typeof refSql !== "string") throw new Error("题目缺少 SQL 标准答案")
+
+  const info = await readInfo(problem.testCaseId)
+  if (!info) throw new Error("测试点信息读取失败")
+  if (!info.sql) throw new Error("测试点不是 SQL 类型，请重新上传 SQL 测试点压缩包")
+
+  // 按 "1","2",… 的数字序遍历，保证测试点顺序稳定
+  const keys = Object.keys(info.test_cases ?? {}).sort((a, b) => Number(a) - Number(b))
+  if (keys.length === 0) throw new Error("题目没有任何测试点")
+
+  const cases: JudgeCase[] = []
+  for (const [index, key] of keys.entries()) {
+    const inputName = info.test_cases![key]!.input_name
+    const initSql = await readFile(
+      resolvePath(config.testCaseDirectory, problem.testCaseId, inputName),
+      "utf8",
+    ).catch(() => { throw new Error(`测试点脚本 ${inputName} 读取失败`) })
+
+    const outcome = await runSqlCase({
+      kind: "judge",
+      initSql,
+      refSql,
+      studentSql,
+      mode,
+      orderSensitive: sqlConfig.order_sensitive === true,
+      timeLimitMs: problem.timeLimit,
+      memoryLimitMb: problem.memoryLimit,
+    })
+    if (!outcome.ok) {
+      // 初始化/标准答案执行失败属出题配置问题，整题 SYSTEM_ERROR
+      if (outcome.result === JudgeStatus.SYSTEM_ERROR) throw new Error(outcome.message)
+      // 子进程被杀（超时/内存）也走这里，按学生错误记成一个测试点
+      cases.push({
+        test_case: String(index + 1),
+        result: outcome.result,
+        cpu_time: 0,
+        real_time: 0,
+        memory: 0,
+        error_message: outcome.message,
+      } as unknown as JudgeCase)
+      break
+    }
+    const value = outcome.value
+    value.test_case = String(index + 1)
+    // 语法错误与数据无关，首个测试点即可确认，整题按编译错误处理
+    // （ACM 不罚时，前端展示 err_info）
+    if (index === 0 && value.result === JudgeStatus.COMPILE_ERROR) {
+      return { err: "CompileError", data: value.error_message }
+    }
+    cases.push(value as unknown as JudgeCase)
+  }
+  return { err: null, data: cases }
+}
--- a/apps/api/src/judge/sql/child.ts
+++ b/apps/api/src/judge/sql/child.ts
@@ -0,0 +1,88 @@
+/**
+ * SQL 判题子进程的入口。
+ *
+ * 单独起进程的唯一理由是**能被杀掉**：SQLite 的查询跑在原生代码里，
+ * 实测 Worker.terminate() 抢占不了，只有 OS 的 SIGKILL 可靠。
+ * 父进程见 `./index.ts`。
+ *
+ * 协议：stdin 读一段 JSON 作业，stdout 写一段 JSON 结果；
+ * 阶段标记写 stderr，父进程在超时杀掉本进程后据此判断卡在哪一阶段。
+ */
+
+import { writeSync } from "node:fs"
+
+import { buildDisplay, runCase, SqlCaseError } from "./engine"
+import { JudgeStatus } from "../status"
+
+export type SqlJob =
+  | {
+      kind: "judge"
+      initSql: string
+      refSql: string
+      studentSql: string
+      mode: "query" | "modify"
+      orderSensitive: boolean
+      timeLimitMs: number
+      memoryLimitMb: number
+    }
+  | { kind: "display"; initSql: string; refSql: string; mode: "query" | "modify" }
+
+function markPhase(phase: string) {
+  process.stderr.write(`@phase:${phase}\n`)
+}
+
+/**
+ * 写完结果立刻硬退出，不走 Bun 的正常 teardown。
+ *
+ * 实测在 `ulimit -v` 之下，Bun 退出时的清理有概率撞上地址空间上限而 panic
+ * （SIGILL，exit 132）—— 结果其实已经写到 stdout 了，但进程异常终止会让父进程
+ * 读到空串，进而误判成超时。这个 panic 是不确定的，同样的输入时好时坏，
+ * 正是最难查的那种。这里跳过 teardown：子进程本来就是一次性的，没有要优雅关闭的资源。
+ */
+function finish(payload: unknown): never {
+  const bytes = new TextEncoder().encode(JSON.stringify(payload))
+  // 直接写 fd 1 并确认写完，再自杀。不用 process.exit()：实测它仍会走一段清理，
+  // 在 ulimit 之下有概率 panic（SIGILL，6 次里 2 次），而结果早已写出，
+  // 父进程却因进程异常终止读到空串、误判成超时 —— 时好时坏，最难查的那种。
+  let written = 0
+  while (written < bytes.length) {
+    written += writeSync(1, bytes, written, bytes.length - written)
+  }
+  process.kill(process.pid, "SIGKILL")
+  throw new Error("unreachable")
+}
+
+async function main() {
+  const raw = await new Response(Bun.stdin.stream()).text()
+  const job = JSON.parse(raw) as SqlJob
+  try {
+    if (job.kind === "display") {
+      markPhase("display")
+      const display = await buildDisplay(job.initSql, job.refSql, job.mode)
+      finish({ ok: true, display })
+    }
+    markPhase("judge")
+    const result = await runCase(job.initSql, job.refSql, job.studentSql, {
+      mode: job.mode,
+      orderSensitive: job.orderSensitive,
+      timeLimitMs: job.timeLimitMs,
+      memoryLimitMb: job.memoryLimitMb,
+    })
+    finish({ ok: true, case: result })
+  } catch (error) {
+    if (error instanceof SqlCaseError) {
+      finish({ ok: false, result: error.result, message: error.detail })
+    }
+    // WASM 堆触顶时 emscripten 抛的是普通 Error（"Aborted"/"out of memory"），
+    // 到这里说明连引擎自身都没撑住，按内存超限报，不当成出题人的错
+    const message = String((error as Error)?.message ?? error)
+    const memoryish = message.includes("out of memory") || message.includes("Aborted")
+    finish({
+      ok: false,
+      result: memoryish ? JudgeStatus.MEMORY_LIMIT_EXCEEDED : JudgeStatus.SYSTEM_ERROR,
+      message: memoryish ? "内存超出限制" : message.slice(0, 200),
+    })
+  }
+}
+
+await main()
--- a/apps/api/src/judge/sql/engine.ts
+++ b/apps/api/src/judge/sql/engine.ts
@@ -0,0 +1,467 @@
+/**
+ * SQL 题判题核心：在 WASM SQLite 里分别执行标准答案和学生 SQL 并比对结果。
+ * 移植自旧后端 `judge/sql_runner.py`，判定口径逐条对齐。
+ *
+ * 查询题（mode="query"）比对最后一条 SELECT 的结果集；
+ * 增删改题（mode="modify"）比对执行后所有用户表的最终状态。
+ *
+ * ## 防护为什么和旧实现不一样
+ *
+ * 旧实现用 Python sqlite3 的 set_authorizer / set_progress_handler / setlimit 三件套。
+ * bun:sqlite 一个都没有，且实测 Worker.terminate() 杀不掉跑飞的查询（原生代码占着线程）。
+ * 所以改成「WASM 引擎 + 独立子进程」，逐条替代：
+ *
+ * | 旧防护 | 新做法 |
+ * |---|---|
+ * | authorizer 禁 ATTACH（防读写服务器任意 SQLite 文件） | WASM 没有宿主文件系统绑定，ATTACH **结构上**够不到宿主，只能碰随进程消失的虚拟 FS |
+ * | authorizer 白名单让查询题只读 | `PRAGMA query_only=1`，SQLite 原生只读开关 |
+ * | progress_handler 墙钟超时 | 子进程外部 SIGKILL（OS 级，比指令计数更硬）+ 语句间 deadline 检查 |
+ * | setlimit(LIMIT_LENGTH) 防单值撑爆内存 | 子进程 `ulimit -v`，触顶时 WASM 抛可捕获错误 |
+ * | max_page_count | 原样保留 |
+ */
+
+import initSqlJs, { type Database, type SqlJsStatic } from "sql.js"
+import { readFileSync } from "node:fs"
+
+import { JudgeStatus, type JudgeStatusValue } from "../status"
+
+/** 单结果集/单表最大行数，防 CROSS JOIN 撑爆内存 */
+const ROW_LIMIT = 10_000
+/** 题目页展示的行数上限（示例数据/期望结果） */
+const DISPLAY_ROW_LIMIT = 20
+const ERROR_MESSAGE_MAX_LEN = 200
+
+/** prepare 阶段的语法类错误，映射为 COMPILE_ERROR */
+const SYNTAX_ERROR_MARKERS = ["syntax error", "unrecognized token", "incomplete input"]
+
+export class SqlCaseError extends Error {
+  constructor(readonly result: JudgeStatusValue, readonly detail: string) {
+    super(detail)
+  }
+}
+
+let cached: SqlJsStatic | null = null
+
+export async function sqlEngine() {
+  if (cached) return cached
+  const binary = readFileSync(require.resolve("sql.js/dist/sql-wasm.wasm"))
+  // @types/sql.js 把 wasmBinary 标成 ArrayBuffer，实际 emscripten 接受 TypedArray；
+  // 这里传 Uint8Array 是运行时正确的写法，类型上断言掉
+  cached = await initSqlJs({ wasmBinary: binary as unknown as ArrayBuffer })
+  return cached
+}
+
+function truncate(message: string) {
+  return message.length > ERROR_MESSAGE_MAX_LEN
+    ? `${message.slice(0, ERROR_MESSAGE_MAX_LEN)}...`
+    : message
+}
+
+// ---------------------------------------------------------------- 值归一化
+
+type Canonical = string
+
+/**
+ * 值归一化并打类型标签，防止 NULL/"NULL"、1/"1" 碰撞；数值统一比对
+ * （1 == 1.0，浮点保留 6 位有效数字）。与旧 `_canonical_value` 同口径。
+ */
+function canonicalValue(value: unknown): Canonical {
+  if (value === null || value === undefined) return "null"
+  if (value instanceof Uint8Array) return `blob:${Buffer.from(value).toString("hex")}`
+  if (typeof value === "number") {
+    if (Number.isInteger(value) && Math.abs(value) < 2 ** 53) return `num:${value}`
+    // Python 的 format(v, ".6g")
+    return `num:${formatG6(value)}`
+  }
+  if (typeof value === "bigint") return `num:${value}`
+  return `str:${String(value)}`
+}
+
+/** 等价于 Python 的 format(v, ".6g") */
+function formatG6(value: number) {
+  const exponent = value === 0 ? 0 : Math.floor(Math.log10(Math.abs(value)))
+  if (exponent < -4 || exponent >= 6) {
+    return value.toExponential(5).replace(/\.?0+e/, "e").replace(/e([+-])(\d)$/, "e$10$2")
+  }
+  const text = value.toPrecision(6)
+  return text.includes(".") ? text.replace(/\.?0+$/, "") : text
+}
+
+function canonicalRow(row: unknown[]) {
+  return row.map(canonicalValue).join("")
+}
+
+// ---------------------------------------------------------------- 执行
+
+interface ResultSet {
+  columns: number
+  rows: string[]
+}
+
+function newDatabase(SQL: SqlJsStatic, memoryLimitMb: number) {
+  const db = new SQL.Database()
+  const limit = Math.max(Math.trunc(memoryLimitMb), 1)
+  db.run("PRAGMA page_size=4096")
+  // 4096B/页 × 256 页/MB，超限报 "database or disk is full"
+  db.run(`PRAGMA max_page_count=${limit * 256}`)
+  return db
+}
+
+/**
+ * 逐条执行，返回最后一条产生结果集的语句的 (列数, 行)；无结果集返回 null。
+ *
+ * 用 sql.js 的 iterateStatements（底层是 sqlite3_prepare_v2 逐条推进），
+ * 比旧实现手写的分号切分更准 —— 字符串和注释里的分号天然不会误切。
+ */
+function executeStatements(db: Database, script: string, deadline: number): ResultSet | null {
+  let last: ResultSet | null = null
+  for (const statement of (db as unknown as {
+    iterateStatements(sql: string): Iterable<{
+      step(): boolean
+      get(): unknown[]
+      getColumnNames(): string[]
+      free(): void
+    }>
+  }).iterateStatements(script)) {
+    if (Date.now() > deadline) {
+      statement.free()
+      throw new Error("interrupted")
+    }
+    try {
+      const names = statement.getColumnNames()
+      if (names.length > 0) {
+        const rows: string[] = []
+        while (statement.step()) {
+          rows.push(canonicalRow(statement.get()))
+          if (rows.length > ROW_LIMIT) {
+            throw new SqlCaseError(JudgeStatus.MEMORY_LIMIT_EXCEEDED, `查询结果超过 ${ROW_LIMIT} 行`)
+          }
+        }
+        last = { columns: names.length, rows }
+      } else {
+        while (statement.step()) { /* 无结果集语句，推进到结束 */ }
+      }
+    } finally {
+      statement.free()
+    }
+  }
+  return last
+}
+
+/** dump 所有用户表：{表名: 列数 + 已排序的行}，表状态天然无序 */
+function dumpTables(db: Database) {
+  const names = queryColumn(db, "SELECT name FROM sqlite_master WHERE type='table' AND name NOT LIKE 'sqlite_%' ORDER BY name")
+  const state: Record<string, { columns: number; rows: string[] }> = {}
+  for (const table of names) {
+    const quoted = String(table).replaceAll('"', '""')
+    const result = db.exec(`SELECT * FROM "${quoted}"`)
+    const first = result[0]
+    const rows = (first?.values ?? []).map((row) => canonicalRow(row as unknown[]))
+    if (rows.length > ROW_LIMIT) {
+      throw new SqlCaseError(JudgeStatus.MEMORY_LIMIT_EXCEEDED, `表 ${table} 超过 ${ROW_LIMIT} 行`)
+    }
+    state[String(table)] = {
+      // 空表 exec 不返回结果，列数用 table_info 兜底
+      columns: first?.columns.length ?? tableColumnCount(db, quoted),
+      rows: rows.sort(),
+    }
+  }
+  return state
+}
+
+function tableColumnCount(db: Database, quotedTable: string) {
+  return db.exec(`PRAGMA table_info("${quotedTable}")`)[0]?.values.length ?? 0
+}
+
+function queryColumn(db: Database, sql: string) {
+  return (db.exec(sql)[0]?.values ?? []).map((row) => row[0])
+}
+
+function trustedErrorText(message: string) {
+  if (message.includes("interrupted")) return "超时"
+  return truncate(message)
+}
+
+/** 执行受信脚本（初始化/标准答案），任何失败都是出题问题 → SYSTEM_ERROR */
+function executeTrusted(db: Database, script: string, deadline: number, prefix: string) {
+  try {
+    return executeStatements(db, script, deadline)
+  } catch (error) {
+    if (error instanceof SqlCaseError) {
+      throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, `${prefix}: ${error.detail}`)
+    }
+    throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, `${prefix}: ${trustedErrorText(String((error as Error).message))}`)
+  }
+}
+
+/** 带防护执行学生 SQL，异常映射为学生级 JudgeStatus */
+function runStudent(db: Database, script: string, mode: string, deadline: number) {
+  // 查询题只读：PRAGMA query_only 是 SQLite 原生开关，替代旧实现的 authorizer 白名单
+  if (mode === "query") db.run("PRAGMA query_only=1")
+  try {
+    const last = executeStatements(db, script, deadline)
+    if (mode === "query") return last
+    return dumpTables(db)
+  } catch (error) {
+    if (error instanceof SqlCaseError) throw error
+    const message = String((error as Error).message)
+    if (message.includes("interrupted")) {
+      throw new SqlCaseError(JudgeStatus.CPU_TIME_LIMIT_EXCEEDED, "SQL 执行超时")
+    }
+    if (message.includes("database or disk is full")) {
+      throw new SqlCaseError(JudgeStatus.MEMORY_LIMIT_EXCEEDED, "数据量超出内存限制")
+    }
+    // WASM 堆触顶（zeroblob/group_concat 构造出的超大单值）或 SQLite 自身的长度上限
+    if (message.includes("too big") || message.includes("out of memory") || message.includes("Aborted")) {
+      throw new SqlCaseError(JudgeStatus.MEMORY_LIMIT_EXCEEDED, "单个数据值超出内存限制")
+    }
+    if (message.includes("readonly database")) {
+      throw new SqlCaseError(JudgeStatus.RUNTIME_ERROR, "本题为查询题，禁止修改数据或表结构（INSERT/UPDATE/DELETE/CREATE 等）")
+    }
+    if (SYNTAX_ERROR_MARKERS.some((marker) => message.includes(marker))) {
+      throw new SqlCaseError(JudgeStatus.COMPILE_ERROR, truncate(message))
+    }
+    throw new SqlCaseError(JudgeStatus.RUNTIME_ERROR, truncate(message))
+  } finally {
+    if (mode === "query") {
+      try { db.run("PRAGMA query_only=0") } catch { /* 连接可能已不可用 */ }
+    }
+  }
+}
+
+function compare(
+  expected: unknown,
+  actual: unknown,
+  mode: string,
+  orderSensitive: boolean,
+) {
+  if (mode === "query") {
+    const exp = expected as ResultSet
+    const act = actual as ResultSet
+    if (exp.columns !== act.columns) return false
+    if (orderSensitive) return exp.rows.join("") === act.rows.join("")
+    return [...exp.rows].sort().join("") === [...act.rows].sort().join("")
+  }
+  return JSON.stringify(expected) === JSON.stringify(actual)
+}
+
+export interface RunCaseOptions {
+  mode: "query" | "modify"
+  orderSensitive: boolean
+  timeLimitMs: number
+  memoryLimitMb: number
+}
+
+export interface CaseResult {
+  test_case: string
+  result: JudgeStatusValue
+  cpu_time: number
+  real_time: number
+  memory: number
+  signal: number
+  exit_code: number
+  error: number
+  output_md5: string
+  error_message: string | null
+}
+
+/**
+ * 判一个测试点，返回与外部 judger 单测试点同构的结构。
+ * 学生错误（CE/WA/TLE/MLE/RE）体现在返回值里；出题配置错误抛 SqlCaseError(SYSTEM_ERROR)。
+ */
+export async function runCase(
+  initSql: string,
+  refSql: string,
+  studentSql: string,
+  options: RunCaseOptions,
+): Promise<CaseResult> {
+  const SQL = await sqlEngine()
+  // 受信脚本的运行上限放宽，避免出题数据较大时误报；仍防子进程永久阻塞
+  const trustedLimitMs = Math.max(options.timeLimitMs * 5, 10_000)
+
+  let expected: unknown
+  const refDb = newDatabase(SQL, options.memoryLimitMb)
+  try {
+    executeTrusted(refDb, initSql, Date.now() + trustedLimitMs, "初始化脚本执行失败")
+    const last = executeTrusted(refDb, refSql, Date.now() + trustedLimitMs, "标准答案执行失败")
+    if (options.mode === "query") {
+      expected = last
+    } else {
+      try {
+        expected = dumpTables(refDb)
+      } catch (error) {
+        throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, `标准答案结果超出限制: ${(error as SqlCaseError).detail}`)
+      }
+    }
+  } finally {
+    refDb.close()
+  }
+  if (options.mode === "query" && expected === null) {
+    throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, "标准答案未产生查询结果集")
+  }
+
+  const result: CaseResult = {
+    test_case: "",
+    result: JudgeStatus.ACCEPTED,
+    cpu_time: 0,
+    real_time: 0,
+    memory: 0,
+    signal: 0,
+    exit_code: 0,
+    error: 0,
+    output_md5: "",
+    error_message: null,
+  }
+
+  const studentDb = newDatabase(SQL, options.memoryLimitMb)
+  let actual: unknown
+  let elapsed = 0
+  try {
+    executeTrusted(studentDb, initSql, Date.now() + trustedLimitMs, "初始化脚本执行失败")
+    const start = Date.now()
+    try {
+      actual = runStudent(studentDb, studentSql, options.mode, start + options.timeLimitMs)
+    } catch (error) {
+      elapsed = Date.now() - start
+      const failure = error as SqlCaseError
+      return { ...result, result: failure.result, error_message: failure.detail, cpu_time: elapsed, real_time: elapsed }
+    }
+    elapsed = Date.now() - start
+  } finally {
+    studentDb.close()
+  }
+
+  result.cpu_time = elapsed
+  result.real_time = elapsed
+  if (options.mode === "query" && (actual === null || actual === undefined)) {
+    result.result = JudgeStatus.WRONG_ANSWER
+    result.error_message = "提交的 SQL 未产生查询结果集"
+  } else if (!compare(expected, actual, options.mode, options.orderSensitive)) {
+    result.result = JudgeStatus.WRONG_ANSWER
+  }
+  return result
+}
+
+// ---------------------------------------------------------------- 题目页展示数据
+
+function displayValue(value: unknown) {
+  if (value instanceof Uint8Array) return Buffer.from(value).toString("hex")
+  return value as string | number | null
+}
+
+interface DisplayTable {
+  name: string
+  columns: { name: string; type: string }[]
+  rows: (string | number | null)[][]
+  total_rows: number
+  truncated: boolean
+  dropped?: boolean
+}
+
+/** 按建表顺序 dump 用户表的原始行用于展示（区别于 dumpTables 的归一化判题态） */
+function dumpDisplayTables(db: Database, only?: Set<string>): DisplayTable[] {
+  const names = queryColumn(db, "SELECT name FROM sqlite_master WHERE type='table' AND name NOT LIKE 'sqlite_%'")
+  const tables: DisplayTable[] = []
+  for (const raw of names) {
+    const name = String(raw)
+    if (only && !only.has(name)) continue
+    const quoted = name.replaceAll('"', '""')
+    const columns = (db.exec(`PRAGMA table_info("${quoted}")`)[0]?.values ?? []).map((row) => ({
+      name: String(row[1]),
+      type: String(row[2] ?? ""),
+    }))
+    const total = Number(db.exec(`SELECT COUNT(*) FROM "${quoted}"`)[0]?.values[0]?.[0] ?? 0)
+    const rows = (db.exec(`SELECT * FROM "${quoted}" LIMIT ${DISPLAY_ROW_LIMIT}`)[0]?.values ?? [])
+      .map((row) => (row as unknown[]).map(displayValue))
+    tables.push({ name, columns, rows, total_rows: total, truncated: total > DISPLAY_ROW_LIMIT })
+  }
+  return tables
+}
+
+/**
+ * 给查询结果的列名标上类型：按列名回查数据表的声明类型，与数据表展示同源（如 VARCHAR(20)）。
+ * 表达式/聚合列（COUNT(*)、别名等）在数据表里无同名列，类型留空（前端隐藏）。
+ */
+function queryResultColumns(names: string[], tables: DisplayTable[]) {
+  const types = new Map<string, string>()
+  for (const table of tables) {
+    for (const column of table.columns) types.set(column.name, column.type)
+  }
+  return names.map((name) => ({ name, type: types.get(name) ?? "" }))
+}
+
+/** 生成题目页展示数据：初始数据表 + 期望结果。失败一律抛 SqlCaseError（出题配置问题） */
+export async function buildDisplay(
+  initSql: string,
+  refSql: string,
+  mode: "query" | "modify",
+  memoryLimitMb = 64,
+) {
+  const SQL = await sqlEngine()
+  const db = newDatabase(SQL, memoryLimitMb)
+  const deadline = Date.now() + 10_000
+  try {
+    executeTrusted(db, initSql, deadline, "初始化脚本执行失败")
+    const tables = dumpDisplayTables(db)
+
+    if (mode === "query") {
+      let expected: unknown = null
+      try {
+        for (const statement of (db as unknown as {
+          iterateStatements(sql: string): Iterable<{
+            step(): boolean; get(): unknown[]; getColumnNames(): string[]; free(): void
+          }>
+        }).iterateStatements(refSql)) {
+          try {
+            const names = statement.getColumnNames()
+            if (names.length === 0) { while (statement.step()) { /* 无结果集 */ } ; continue }
+            const rows: unknown[][] = []
+            while (statement.step()) {
+              rows.push(statement.get())
+              if (rows.length > ROW_LIMIT) {
+                throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, `标准答案结果超过 ${ROW_LIMIT} 行`)
+              }
+            }
+            expected = {
+              columns: queryResultColumns(names, tables),
+              rows: rows.slice(0, DISPLAY_ROW_LIMIT).map((row) => row.map(displayValue)),
+              total_rows: rows.length,
+              truncated: rows.length > DISPLAY_ROW_LIMIT,
+            }
+          } finally {
+            statement.free()
+          }
+        }
+      } catch (error) {
+        if (error instanceof SqlCaseError) throw error
+        throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, `标准答案执行失败: ${trustedErrorText(String((error as Error).message))}`)
+      }
+      if (expected === null) {
+        throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, "标准答案未产生查询结果集")
+      }
+      return { tables, expected }
+    }
+
+    const before = dumpTables(db)
+    executeTrusted(db, refSql, Date.now() + 10_000, "标准答案执行失败")
+    const after = dumpTables(db)
+    const changed = new Set<string>()
+    for (const name of new Set([...Object.keys(before), ...Object.keys(after)])) {
+      if (JSON.stringify(before[name]) !== JSON.stringify(after[name])) changed.add(name)
+    }
+    if (changed.size === 0) {
+      throw new SqlCaseError(JudgeStatus.SYSTEM_ERROR, "标准答案未修改任何表数据，请检查题目配置")
+    }
+    const changedTables = dumpDisplayTables(db, changed)
+    // 被标准答案 DROP 的表已不在库中，用初始展示数据补齐条目（前端据 dropped 提示「表已删除」）
+    const existing = new Set(changedTables.map((table) => table.name))
+    for (const table of tables) {
+      if (changed.has(table.name) && !existing.has(table.name)) {
+        changedTables.push({ ...table, rows: [], total_rows: 0, truncated: false, dropped: true })
+      }
+    }
+    return { tables, expected: { changed_tables: changedTables } }
+  } finally {
+    db.close()
+  }
+}
--- a/apps/api/src/judge/sql/index.ts
+++ b/apps/api/src/judge/sql/index.ts
@@ -0,0 +1,88 @@
+import { resolve } from "node:path"
+
+import { JudgeStatus, type JudgeStatusValue } from "../status"
+import type { CaseResult } from "./engine"
+import type { SqlJob } from "./child"
+
+/**
+ * SQL 作业的父进程侧：起一个短命子进程跑，到点 SIGKILL。
+ *
+ * 为什么非要子进程：SQLite 查询在原生代码里执行，实测 Worker.terminate() 抢占不了
+ * （递归 CTE 死循环能把整个 worker 卡死），只有 OS 的 SIGKILL 可靠。
+ *
+ * 内存交给 OS：`ulimit -d`（数据段）给子进程封顶，触顶时 WASM 抛可捕获错误，
+ * 子进程照常回报 MEMORY_LIMIT_EXCEEDED。
+ *
+ * **必须用 -d 不能用 -v。** -v 限的是虚拟地址空间，而 JS 引擎会预留巨量地址，
+ * 实测 -v 之下 Bun 退出时有概率 panic（SIGILL）—— 结果早已写出但进程异常终止，
+ * 父进程读到空串误判成超时，时好时坏。-d 限的是实际提交的内存（Linux 4.7 起
+ * 也覆盖匿名 mmap），512MB 下正常判题稳定、`hex(zeroblob(2e8))` 被拦。
+ */
+
+/** 子进程数据段上限（KB）。低于 512MB Bun 自己起不来 */
+const CHILD_DATA_LIMIT_KB = 512 * 1024
+/** 父进程的兜底墙钟。比作业自报的时限宽裕，只负责杀掉真正跑飞的进程 */
+const HARD_TIMEOUT_SLACK_MS = 15_000
+
+export interface SqlJobFailure {
+  ok: false
+  result: JudgeStatusValue
+  message: string
+}
+
+type SqlJobOutcome<T> = { ok: true; value: T } | SqlJobFailure
+
+async function runJob<T>(job: SqlJob, budgetMs: number): Promise<SqlJobOutcome<T>> {
+  const entry = resolve(import.meta.dir, "child.ts")
+  // 经 sh 起是为了用 ulimit —— Bun.spawn 没有直接设 rlimit 的接口
+  const child = Bun.spawn(
+    ["sh", "-c", `ulimit -d ${CHILD_DATA_LIMIT_KB}; exec "$0" run "$1"`, process.execPath, entry],
+    { stdin: "pipe", stdout: "pipe", stderr: "pipe" },
+  )
+  child.stdin.write(JSON.stringify(job))
+  await child.stdin.end()
+
+  const timer = setTimeout(() => child.kill("SIGKILL"), budgetMs + HARD_TIMEOUT_SLACK_MS)
+  let stdout = ""
+  let stderr = ""
+  try {
+    ;[stdout, stderr] = await Promise.all([
+      new Response(child.stdout).text(),
+      new Response(child.stderr).text(),
+    ])
+    await child.exited
+  } finally {
+    clearTimeout(timer)
+  }
+
+  if (!stdout.trim()) {
+    // 子进程没来得及写结果就没了 —— 要么被我们 SIGKILL，要么被内核 OOM 掉。
+    // 用 stderr 里的阶段标记区分：卡在受信脚本是出题问题，卡在学生 SQL 是超时。
+    const phase = stderr.includes("@phase:") ? stderr.split("@phase:")[1]?.split("\n")[0] : null
+    if (phase === "display") {
+      return { ok: false, result: JudgeStatus.SYSTEM_ERROR, message: "生成展示数据超时或内存超限，请检查初始化脚本与标准答案" }
+    }
+    return { ok: false, result: JudgeStatus.CPU_TIME_LIMIT_EXCEEDED, message: "SQL 执行超时" }
+  }
+
+  try {
+    const parsed = JSON.parse(stdout) as
+      | { ok: true; case?: CaseResult; display?: unknown }
+      | SqlJobFailure
+    if (!parsed.ok) return parsed
+    return { ok: true, value: (parsed.case ?? parsed.display) as T }
+  } catch {
+    return { ok: false, result: JudgeStatus.SYSTEM_ERROR, message: "SQL 判题子进程返回了无法解析的结果" }
+  }
+}
+
+export function runSqlCase(job: Extract<SqlJob, { kind: "judge" }>) {
+  return runJob<CaseResult>(job, Math.max(job.timeLimitMs * 5, 10_000))
+}
+
+export function buildSqlDisplay(initSql: string, refSql: string, mode: "query" | "modify") {
+  return runJob<{ tables: unknown[]; expected: unknown }>(
+    { kind: "display", initSql, refSql, mode },
+    10_000,
+  )
+}
--- a/apps/api/src/routes/admin/problem.ts
+++ b/apps/api/src/routes/admin/problem.ts
@@ -6,6 +6,9 @@ import {
  createProblemRequestSchema,
  makeProblemPublicRequestSchema,
  updateProblemRequestSchema,
+  generateSqlTestCaseRequestSchema,
+  generateSqlTestCaseResponseSchema,
+  sqlPreviewRequestSchema,
  sqlTestCaseScriptSchema,
  uploadTestCaseResponseSchema,
 } from "@oj2/contract"
@@ -16,8 +19,13 @@ import { requireProblemPermission, type AppEnv } from "../../auth/middleware"
 import type { AuthUser } from "../../auth/session"
 import { db, schema } from "../../db"
 import { failure, success } from "../../http"
+import { buildSqlDisplay } from "../../judge/sql"
+import { completeChat } from "../../services/ai"
 import { contestStatus } from "../../services/contest"
 import { packTestCaseZip, processTestCaseZip, readInfo, readSqlScripts, TestCaseError } from "../../services/test-case"
+import { config } from "../../config"
+import { readFile } from "node:fs/promises"
+import { resolve } from "node:path"
 import { objectValue, queryInteger, sampleUser, stringArray } from "../helpers"

 export const adminProblemRoutes = new Hono<AppEnv>()
@@ -123,13 +131,7 @@ async function serialize(row: ProblemRow) {
  })
 }

-/**
- * 非 SQL 题的公共校验，对齐旧 `ProblemBase.common_checks` 的 else 分支。
- *
- * SQL 分支在这里**只做前置校验、不生成 sqlDisplay** —— 生成需要跑 SQLite，
- * 而新后端还没有 SQL 判题链路（旧 judge/sql_runner.py 378 行无对应实现）。
- * 因此新建 SQL 题会被拒绝，编辑 SQL 题则保留原有 sqlDisplay 不动，不会把已有数据弄坏。
- */
+/** 公共校验，对齐旧 `ProblemBase.common_checks` */
 function commonChecks(data: {
  languages: string[]
  inputDescription: string
@@ -153,8 +155,38 @@ function commonChecks(data: {
  return { sql: false }
 }

-const SQL_NOT_SUPPORTED =
-  "新后端尚未实现 SQL 判题链路（题目页展示数据需要跑 SQLite 生成），暂不能新建 SQL 题。已有 SQL 题可以正常编辑，其展示数据保持不变。"
+/**
+ * SQL 题保存时生成题目页展示数据（数据表 + 期望结果）。
+ * 对齐旧 `problem/utils.py:generate_sql_display`：取**测试点 1** 的初始化脚本 + 标准答案跑一遍。
+ * 失败一律拦下不让保存 —— 展示数据直接决定学生看到的表结构与期望结果，宁可不保存也不能存错的。
+ */
+async function generateSqlDisplay(
+  testCaseId: string,
+  answers: Record<string, unknown>[],
+  sqlConfig: Record<string, unknown>,
+): Promise<{ error: string } | { display: unknown }> {
+  const info = await readInfo(testCaseId)
+  if (!info) return { error: "测试点信息读取失败，请重新上传测试点" }
+  if (!info.sql) return { error: "测试点不是 SQL 类型，请重新上传 SQL 测试点压缩包" }
+  const keys = Object.keys(info.test_cases ?? {}).sort((a, b) => Number(a) - Number(b))
+  if (keys.length === 0) return { error: "题目没有任何测试点" }
+  const inputName = info.test_cases![keys[0]!]?.input_name
+  if (!inputName) return { error: "测试点信息损坏，请重新上传测试点" }
+  let initSql: string
+  try {
+    initSql = await readFile(resolve(config.testCaseDirectory, testCaseId, inputName), "utf8")
+  } catch {
+    return { error: `测试点脚本 ${inputName} 读取失败` }
+  }
+  const refSql = answers.find(
+    (item) => item.language === "SQL" && typeof item.code === "string" && item.code.trim(),
+  )?.code
+  if (typeof refSql !== "string") return { error: "题目缺少 SQL 标准答案" }
+  const mode = sqlConfig.mode === "modify" ? "modify" as const : "query" as const
+  const outcome = await buildSqlDisplay(initSql, refSql, mode)
+  if (!outcome.ok) return { error: `SQL 展示数据生成失败: ${outcome.message}` }
+  return { display: outcome.value }
+}

 function problemValues(data: ReturnType<typeof createProblemRequestSchema.parse>, isSql: boolean) {
  return {
@@ -257,7 +289,12 @@ adminProblemRoutes.post("/problems", requireProblemPermission, async (c) => {
  }
  const checked = commonChecks(parsed.data)
  if ("error" in checked) return failure(c, 400, "invalid-problem", checked.error)
-  if (checked.sql) return failure(c, 501, "sql-not-supported", SQL_NOT_SUPPORTED)
+  let sqlDisplay: unknown = null
+  if (checked.sql) {
+    const built = await generateSqlDisplay(parsed.data.testCaseId, parsed.data.answers, parsed.data.sqlConfig!)
+    if ("error" in built) return failure(c, 400, "invalid-problem", built.error)
+    sqlDisplay = built.display
+  }

  const [duplicate] = await db.select({ id: schema.problem.id }).from(schema.problem)
    .where(and(eq(schema.problem.displayId, parsed.data._id), isNull(schema.problem.contestId))).limit(1)
@@ -266,7 +303,7 @@ adminProblemRoutes.post("/problems", requireProblemPermission, async (c) => {
  const now = new Date().toISOString()
  const created = await db.transaction(async (tx) => {
    const [row] = await tx.insert(schema.problem).values({
-      ...problemValues(parsed.data, false),
+      ...problemValues(parsed.data, checked.sql),
      contestId: null,
      createdById: c.get("user")!.id,
      createTime: now,
@@ -275,7 +312,7 @@ adminProblemRoutes.post("/problems", requireProblemPermission, async (c) => {
      acceptedNumber: 0,
      statisticInfo: {},
      isPublic: false,
-      sqlDisplay: null,
+      sqlDisplay,
    }).returning()
    await setTags(tx as unknown as typeof db, row!.id, parsed.data.tags)
    return row!
@@ -308,12 +345,19 @@ adminProblemRoutes.put("/problems/:id", requireProblemPermission, async (c) => {
    )).limit(1)
  if (duplicate) return failure(c, 409, "display-id-exists", "Display ID already exists")

+  // SQL 题每次保存都重算展示数据：测试点或标准答案可能刚改过，留着旧的就会和判题结果对不上
+  let sqlDisplay: unknown = null
+  if (checked.sql) {
+    const built = await generateSqlDisplay(parsed.data.testCaseId, parsed.data.answers, parsed.data.sqlConfig!)
+    if ("error" in built) return failure(c, 400, "invalid-problem", built.error)
+    sqlDisplay = built.display
+  }
+
  const updated = await db.transaction(async (tx) => {
    const [row] = await tx.update(schema.problem).set({
      ...problemValues(parsed.data, checked.sql),
+      sqlDisplay,
      lastUpdateTime: new Date().toISOString(),
-      // SQL 题的 sqlDisplay 原样保留：重新生成需要跑 SQLite，新后端还没有那条链路。
-      // 不动它比生成一个错的更安全 —— 它直接决定题目页给学生看的表结构与期望结果。
    }).where(eq(schema.problem.id, id)).returning()
    await setTags(tx as unknown as typeof db, id, parsed.data.tags)
    return row!
@@ -411,7 +455,12 @@ adminProblemRoutes.post("/contests/:contestId/problems", requireProblemPermissio
  }
  const checked = commonChecks(parsed.data)
  if ("error" in checked) return failure(c, 400, "invalid-problem", checked.error)
-  if (checked.sql) return failure(c, 501, "sql-not-supported", SQL_NOT_SUPPORTED)
+  let sqlDisplay: unknown = null
+  if (checked.sql) {
+    const built = await generateSqlDisplay(parsed.data.testCaseId, parsed.data.answers, parsed.data.sqlConfig!)
+    if ("error" in built) return failure(c, 400, "invalid-problem", built.error)
+    sqlDisplay = built.display
+  }

  const [duplicate] = await db.select({ id: schema.problem.id }).from(schema.problem)
    .where(and(eq(schema.problem.displayId, parsed.data._id), eq(schema.problem.contestId, contestId))).limit(1)
@@ -420,7 +469,7 @@ adminProblemRoutes.post("/contests/:contestId/problems", requireProblemPermissio
  const now = new Date().toISOString()
  const created = await db.transaction(async (tx) => {
    const [row] = await tx.insert(schema.problem).values({
-      ...problemValues(parsed.data, false),
+      ...problemValues(parsed.data, checked.sql),
      contestId,
      createdById: user.id,
      createTime: now,
@@ -594,3 +643,37 @@ adminProblemRoutes.get("/problems/:id/sql-scripts", requireProblemPermission, as
    return failure(c, 500, "test-case-error", "测试点脚本读取失败")
  }
 })
+
+/** SQL 题测试点预览：跑一遍初始化脚本 + 标准答案，返回题目页要展示的数据表与期望结果 */
+adminProblemRoutes.post("/sql-test-cases/preview", requireProblemPermission, async (c) => {
+  const parsed = sqlPreviewRequestSchema.safeParse(await c.req.json().catch(() => null))
+  if (!parsed.success) {
+    return failure(c, 400, "invalid-request", parsed.error.issues[0]?.message ?? "参数错误")
+  }
+  const outcome = await buildSqlDisplay(parsed.data.initSql, parsed.data.refSql, parsed.data.mode)
+  if (!outcome.ok) return failure(c, 400, "sql-preview-failed", outcome.message)
+  return success(c, outcome.value)
+})
+
+/** AI 按标准答案倒推表结构、生成一份自洽的初始化脚本 */
+adminProblemRoutes.post("/sql-test-cases/generate", requireProblemPermission, async (c) => {
+  const parsed = generateSqlTestCaseRequestSchema.safeParse(await c.req.json().catch(() => null))
+  if (!parsed.success) {
+    return failure(c, 400, "invalid-request", parsed.error.issues[0]?.message ?? "参数错误")
+  }
+  try {
+    const sql = await completeChat(
+      `你是一个 SQL 出题助手。用户会给你一道 SQL 题的标准答案（查询题的
+SELECT 语句，或增删改题的 UPDATE/DELETE/INSERT 语句）和题型。
+请你推断出该标准答案所需要的表结构，生成一份自洽的 SQLite 兼容初始化脚本，
+包含 CREATE TABLE 和若干条 INSERT 语句，插入的数据要足够让标准答案跑出有意义的结果
+（比如查询题要有能被筛选出来和被过滤掉的行；增删改题要有能被改动和不受影响的行）。
+请只返回 SQL 脚本本身，连 \`\`\` 都不需要，不要任何解释文字。`,
+      `题型：${parsed.data.mode}\n标准答案：\n${parsed.data.refSql}`,
+    )
+    return success(c, generateSqlTestCaseResponseSchema.parse({ sql }))
+  } catch (error) {
+    console.error("SQL test case generation failed", error)
+    return failure(c, 502, "ai-unavailable", "生成失败，请稍后再试")
+  }
+})