feat(cache): PG kv_cache ported from MP TTLCache, TMDB request and recognize caches

Migration 0042_cache adds UNLOGGED kv_cache. PgCache mirrors MP
runtime/cache.py (region, per-key TTL, TLRU maxsize eviction, @cached as
cachedCall); only the storage moves from memory to PG. TMDB request
responses (MP TMDb.request @cached) and recognize results (MP TmdbCache:
same key, {id:0} negative cache with EMPTY_RESULT_CACHE_TTL, type-conflict
checks, backfill after manual assignment) live in kv_cache, shared by all
processes. Settings cache.tmdbTtlHours / tmdbMaxSize / emptyResultTtlSec
map to META_CACHE_EXPIRE / CONF.tmdb / EMPTY_RESULT_CACHE_TTL. TMDB
details now also store posterPath / backdropPath.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

apps/server/src/modules/index.ts

@@ -2,6 +2,7 @@ import { acquisitionOrganizeModule } from "./acquisition-organize/index.ts";
 import { agentModule } from "./agent/index.ts";
 import { authSettingsModule } from "./auth-settings/index.ts";
 import { browserModule } from "./browser/index.ts";
+import { cacheModule } from "./cache/index.ts";
 import { fileManagerModule } from "./file-manager/index.ts";
 import { mcpModule } from "./mcp/index.ts";
 import { piModule } from "./pi/index.ts";
@@ -28,5 +29,6 @@ export const SERVER_MODULES: readonly ServerModule[] = [
   searchModule,
   browserModule,
   mcpModule,
+  cacheModule,
   systemModule,
 ];

apps/server/src/modules/search/media.ts

@@ -163,6 +163,7 @@ export class TmdbSearch {
                 year: m.year,
                 overview: m.overview ? (m.overview.length > 300 ? `${m.overview.slice(0, 300)}…` : m.overview) : null,
                 popularity: m.popularity,
+                ...(m.posterPath ? { posterPath: m.posterPath } : {}),
               })),
             )
             .then((list) => {

apps/server/src/modules/sources-identity/runtime.ts

@@ -11,9 +11,11 @@ import {
   type SiteDefinition,
   type SiteRateLimiter,
   type TmdbClient,
+  type TmdbSharedCache,
 } from "@umt/adapters";
 import { DEFAULT_SETTINGS, type SitesSettings } from "@umt/contracts";
 import { pgTokenBucketFor, type SiteRow } from "@umt/db";
+import { CACHE_REGIONS, sharedCache, type TmdbCachePolicy, tmdbCachePolicy } from "../../shared-cache.ts";
 import type { ServerContext } from "../types.ts";
 
 // 模块运行时依赖:站点适配器与限流器(每站一个实例,令牌在 PG 跨进程共享)、RSS 抓取器、TMDB 客户端、时钟。
@@ -71,12 +73,14 @@ export class SourcesRuntime {
       return this.tmdbClient;
     }
     const meta = await this.ctx.settings.get("metadata").catch(() => null);
+    const policy = await tmdbCachePolicy(this.ctx);
     const env = (k: string) => Deno.env.get(k);
     const ref = meta?.tmdbApiKeySecretRef ?? null;
     const baseUrl = meta?.tmdbBaseUrl ?? env("TMDB_BASE_URL");
     const sig = JSON.stringify([
       ref,
       baseUrl ?? null,
+      policy,
       ref ? null : [!!env("TMDB_API_KEY"), !!env("TMDB_ACCESS_TOKEN")],
     ]);
     if (this.tmdbClient !== undefined && this.tmdbSig === sig) return this.tmdbClient;
@@ -103,6 +107,9 @@ export class SourcesRuntime {
             ...(accessToken ? { accessToken } : {}),
             ...(baseUrl ? { baseUrl } : {}),
             ...(this.overrides.fetch ? { fetch: this.overrides.fetch } : {}),
+            // 请求响应进 PG 缓存(region tmdb.request),三进程共享;TTL / 条数上限 / 空结果短 TTL 取设置(对照 MP)
+            cache: { ttlMs: policy.ttl * 1000, emptyTtlMs: policy.emptyTtl * 1000 },
+            sharedCache: tmdbSharedCache(this.ctx, policy),
           })
         : null;
     return this.tmdbClient;
@@ -233,6 +240,22 @@ export class SourcesRuntime {
   }
 }
 
+/** TMDB 请求响应缓存:PG kv_cache region=tmdb.request;PG 出错按未命中 / 不写处理,不让 TMDB 请求失败 */
+function tmdbSharedCache(ctx: ServerContext, policy: TmdbCachePolicy): TmdbSharedCache {
+  const cache = sharedCache(ctx);
+  const warn = (err: unknown) =>
+    ctx.logger.warn({ err: { message: (err as Error).message } }, "TMDB 缓存读写失败，按未命中处理");
+  return {
+    get: (key) =>
+      cache.get(CACHE_REGIONS.tmdbRequest, key).catch((e) => {
+        warn(e);
+        return null;
+      }),
+    set: (key, value, ttlMs) =>
+      cache.set(CACHE_REGIONS.tmdbRequest, key, value, { ttl: ttlMs / 1000, maxsize: policy.maxsize }).catch(warn),
+  };
+}
+
 function sameUrl(a: string, b: string): boolean {
   return a.replace(/\/+$/, "") === b.replace(/\/+$/, "");
 }

apps/server/src/modules/sources-identity/service/identity-admin.ts

@@ -23,6 +23,7 @@ import {
   resolveEntityId,
 } from "./entities.ts";
 import { recordAttempt } from "./resolve.ts";
+import { backfillRecognizeCache } from "./tmdb-recognize-cache.ts";
 
 // 人工纠错:指定来源资源身份(加锁)、改作品外部映射(新增修订、旧映射退役)、合并作品(旧 ID 跳转)。
 // 所有变更留 identity_revision / resolution_attempt / audit_event,不直接覆盖历史。
@@ -85,7 +86,7 @@ export async function assignReleaseIdentity(
 
   for (let attempt = 0; ; attempt++) {
     try {
-      return await ctx.db.transaction().execute(async (trx) => {
+      const done = await ctx.db.transaction().execute(async (trx) => {
         let entityId: string;
         if (localId) {
           const r = await resolveEntityId(trx, localId);
@@ -190,6 +191,8 @@ export async function assignReleaseIdentity(
         });
         return { mediaEntityId: entityId, contentUnitIds: unitIds };
       });
+      await backfillRecognizeCache(rt, parsed, done.mediaEntityId);
+      return done;
     } catch (e) {
       if (e instanceof IdentityConflictError && attempt < 2) continue;
       throw e;

apps/server/src/modules/sources-identity/service/metadata.ts

@@ -32,6 +32,9 @@ function tmdbFields(m: MetadataCandidate): Partial<Record<EntityMetadataField, u
     releaseDate: m.releaseDate,
     originalLanguage: m.originalLanguage,
     originCountry: m.originCountry,
+    // 海报 / 背景图路径(前端经图片代理显示);TMDB 没有时不写,保留旧值
+    ...(m.posterPath ? { posterPath: m.posterPath } : {}),
+    ...(m.backdropPath ? { backdropPath: m.backdropPath } : {}),
   };
 }
 

apps/server/src/modules/sources-identity/service/resolve.ts

@@ -32,6 +32,7 @@ import {
   looksLikeVariety,
   resolveEntityId,
 } from "./entities.ts";
+import { loadRecognizeCache, type RecognizeMeta, recognizeMetaOf, saveRecognizeCache } from "./tmdb-recognize-cache.ts";
 
 // resolveObservation:parseRelease → 本地召回(别名 pg_trgm + 已有外部身份 + 人工锁)→ TMDB 候选 → domain.resolveIdentity。
 // auto 才建/关联正式作品;ambiguous / no_match / 查询失败进待识别池(候选只存 JSON),有界退避复查。
@@ -155,9 +156,27 @@ async function localCandidates(
   return out;
 }
 
-async function tmdbCandidates(rt: SourcesRuntime, parsed: ParsedRelease): Promise<MetadataCandidate[]> {
+/**
+ * TMDB 召回。先查识别缓存(移植 MP TmdbCache,见 tmdb-recognize-cache.ts):命中且有标题 → 按 id 取详情作唯一远端候选;
+ * 命中负缓存 → 不再搜索(视为 TMDB 无结果)。强制复查(force)不读缓存但照常回写(MP recognize_media 的 cache=False)。
+ */
+async function tmdbCandidates(
+  rt: SourcesRuntime,
+  parsed: ParsedRelease,
+  useCache: boolean,
+): Promise<{ remote: MetadataCandidate[]; cacheHit: boolean; meta: RecognizeMeta | null }> {
   const tmdb = await rt.tmdb();
-  if (!tmdb) return [];
+  if (!tmdb) return { remote: [], cacheHit: false, meta: null };
+  const meta = recognizeMetaOf(parsed);
+  if (meta && useCache) {
+    const hit = await loadRecognizeCache(rt, meta);
+    if (hit) {
+      if (hit.id && hit.title && (hit.type === "movie" || hit.type === "tv")) {
+        return { remote: [await tmdb.details(hit.type, hit.id)], cacheHit: true, meta };
+      }
+      return { remote: [], cacheHit: true, meta };
+    }
+  }
   const queries = [parsed.titleCn, parsed.titleEn, parsed.title].filter((t): t is string => !!t);
   const seen = new Map<string, MetadataCandidate>();
   for (const q of [...new Set(queries)].slice(0, 2)) {
@@ -166,7 +185,7 @@ async function tmdbCandidates(rt: SourcesRuntime, parsed: ParsedRelease): Promis
       if (!seen.has(k)) seen.set(k, m);
     }
   }
-  return [...seen.values()].slice(0, 20);
+  return { remote: [...seen.values()].slice(0, 20), cacheHit: false, meta };
 }
 
 async function loadRelease(rt: SourcesRuntime, input: ResolveInput) {
@@ -240,14 +259,27 @@ export async function resolveRelease(rt: SourcesRuntime, input: ResolveInput): P
   const local = await localCandidates(ctx.db, parsed, release);
   let remote: MetadataCandidate[] = [];
   let lookupError: string | null = null;
+  let recognize: { cacheHit: boolean; meta: RecognizeMeta | null } = { cacheHit: false, meta: null };
   try {
-    remote = await tmdbCandidates(rt, parsed);
+    const r = await tmdbCandidates(rt, parsed, !input.force);
+    remote = r.remote;
+    recognize = r;
   } catch (e) {
     if (!(e instanceof TmdbError)) throw e;
     lookupError = `${e.kind}: ${e.message}`;
   }
   const candidates = [...local, ...remote.map(tmdbToCandidate)];
   const resolution = resolveIdentity(parsed, candidates);
+  // 识别缓存回写(MP TmdbCache.update):识别到 TMDB 条目写正常条目,确认无匹配写负缓存;TMDB 出错 / 有歧义不写
+  if (recognize.meta && !recognize.cacheHit && !lookupError) {
+    const top = resolution.status === "auto" ? resolution.best?.candidate : undefined;
+    const hit =
+      top?.provider === "tmdb"
+        ? remote.find((m) => m.externalId === top.externalId && m.entityType === top.entityType)
+        : undefined;
+    const info = hit ?? (resolution.status === "no_match" ? null : undefined);
+    await saveRecognizeCache(rt, recognize.meta, info);
+  }
   const version = evidenceVersion(parsed, resolution.ranked);
 
   // 已识别且证据未变:幂等短路(强制复查时)

apps/server/src/modules/sources-identity/service/tmdb-recognize-cache.ts

@@ -0,0 +1,154 @@
+import type { MetadataCandidate } from "@umt/adapters";
+import type { ParsedRelease } from "@umt/domain";
+import { CACHE_REGIONS, sharedCache, tmdbCachePolicy } from "../../../shared-cache.ts";
+import type { SourcesRuntime } from "../runtime.ts";
+
+// TMDB 识别缓存:移植 MP app/modules/themoviedb/cache.py 的 TmdbCache(对照表见 docs/porting/moviepilot-tmdb-cache.md)。
+// 存储从 内存 TTLCache + pickle 文件持久化 换成 PG kv_cache(region tmdb.recognize),每条独立 expires_at,
+// 三进程共享;"持久化与恢复"等价于表本身。语义照抄:
+//   - 键:[类型][语言]{tmdbid 或 名称}-{年份}-{开始季}(Python 的 None 原样写作 "None");
+//   - 值:{id, type, year, title, poster_path, backdrop_path};负缓存 {id: 0};
+//   - TTL:正常条目 = META_CACHE_EXPIRE(cache.tmdbTtlHours);负缓存 = EMPTY_RESULT_CACHE_TTL(cache.emptyResultTtlSec);
+//   - 条数上限 = CONF.tmdb(cache.tmdbMaxSize),超出删最旧;
+//   - 网络错误(None)不写;查无结果({})写负缓存;
+//   - 读 / 写时"元数据是电视剧、缓存是电影"视为类型冲突:读时丢弃该条,写时拒绝写入。
+
+/** MP MediaType 的值(键里的类型段,与 MP 一致用中文) */
+export type MpMediaType = "电影" | "电视剧" | "未知";
+
+export interface RecognizeCacheEntry {
+  /** TMDB id;0 = 负缓存(未识别) */
+  id: number;
+  type?: "movie" | "tv" | null;
+  year?: string | null;
+  title?: string | null;
+  poster_path?: string | null;
+  backdrop_path?: string | null;
+}
+
+/** 键所需的元数据(对照 MP MetaBase:type / name / year / begin_season / media_source+media_id) */
+export interface RecognizeMeta {
+  type: MpMediaType;
+  name: string;
+  year: number | null;
+  beginSeason: number | null;
+  /** 已知 TMDB id 时键用它代替名称(MP:media_source == TMDB 时取 media_id) */
+  tmdbId?: string | null;
+}
+
+const isAllChinese = (s: string) => /^[㐀-鿿豈-﫿]+$/.test(s);
+
+/** 解析结果 → 识别元数据(MP MetaBase.name:中文名全是汉字优先,否则英文名,再退回中文名) */
+export function recognizeMetaOf(p: ParsedRelease): RecognizeMeta | null {
+  const cn = p.titleCn ?? null;
+  const en = p.titleEn ?? null;
+  const name = cn && isAllChinese(cn) ? cn : (en ?? cn ?? p.title ?? "");
+  if (!name) return null;
+  // MP MetaVideo:识别出季或集即为电视剧,否则类型未知
+  const tv = p.season !== undefined || p.episodes !== undefined;
+  return { type: tv ? "电视剧" : "未知", name, year: p.year ?? null, beginSeason: p.season ?? null };
+}
+
+/** TmdbCache.__get_key */
+export function recognizeKey(meta: RecognizeMeta, locale: string): string {
+  const py = (v: number | null) => (v === null ? "None" : String(v));
+  return `[${meta.type}][${locale}]${meta.tmdbId || meta.name}-${py(meta.year)}-${py(meta.beginSeason)}`;
+}
+
+/** TmdbCache.__is_type_conflicted:只有"元数据是电视剧、结果是电影"算冲突;负缓存不带类型不算 */
+function typeConflicted(meta: RecognizeMeta, type: unknown, id: unknown): boolean {
+  if (meta.type !== "电视剧" || !id) return false;
+  return type === "movie";
+}
+
+function localeOf(tmdb: unknown): string {
+  return (tmdb as { language?: string } | null)?.language ?? "zh-CN";
+}
+
+/** TmdbCache.get:未命中返回 null;类型冲突的脏条目删除后按未命中处理 */
+export async function loadRecognizeCache(rt: SourcesRuntime, meta: RecognizeMeta): Promise<RecognizeCacheEntry | null> {
+  const tmdb = await rt.tmdb();
+  const key = recognizeKey(meta, localeOf(tmdb));
+  const cache = sharedCache(rt.ctx);
+  const v = await cache.get<RecognizeCacheEntry>(CACHE_REGIONS.tmdbRecognize, key).catch(() => null);
+  if (!v || typeof v !== "object") return null;
+  if (typeConflicted(meta, v.type, v.id)) {
+    rt.ctx.logger.warn({ key, title: v.title, type: v.type }, "识别缓存类型与元数据冲突，已丢弃并重新识别");
+    await cache.delete(CACHE_REGIONS.tmdbRecognize, key);
+    return null;
+  }
+  return v;
+}
+
+/**
+ * TmdbCache.update:info 为识别结果写正常条目;info 为 null(查无此项)写负缓存 {id: 0};
+ * info 为 undefined(网络错误 / 没有结论)不写,允许下次重新请求。
+ */
+export async function saveRecognizeCache(
+  rt: SourcesRuntime,
+  meta: RecognizeMeta,
+  info:
+    | Pick<MetadataCandidate, "externalId" | "entityType" | "title" | "year" | "posterPath" | "backdropPath">
+    | null
+    | undefined,
+): Promise<void> {
+  if (info === undefined) return;
+  const tmdb = await rt.tmdb();
+  const key = recognizeKey(meta, localeOf(tmdb));
+  const policy = await tmdbCachePolicy(rt.ctx);
+  const cache = sharedCache(rt.ctx);
+  if (info) {
+    if (typeConflicted(meta, info.entityType, info.externalId)) {
+      rt.ctx.logger.warn({ key, title: info.title, type: info.entityType }, "识别结果类型与元数据冲突，不写入识别缓存");
+      return;
+    }
+    const entry: RecognizeCacheEntry = {
+      id: Number(info.externalId),
+      type: info.entityType,
+      year: info.year === null ? null : String(info.year),
+      title: info.title,
+      poster_path: info.posterPath ?? null,
+      backdrop_path: info.backdropPath ?? null,
+    };
+    await cache.set(CACHE_REGIONS.tmdbRecognize, key, entry, { ttl: policy.ttl, maxsize: policy.maxsize });
+    return;
+  }
+  // 负识别缓存用独立短 TTL(EMPTY_RESULT_CACHE_TTL):故障期间的空结果不被固化整个有效期
+  await cache.set(CACHE_REGIONS.tmdbRecognize, key, { id: 0 }, { ttl: policy.emptyTtl, maxsize: policy.maxsize });
+}
+
+/**
+ * 人工认定后回填识别缓存(MP update_recognize_cache:覆盖名称负缓存,避免同名资源再次回查)。
+ * 作品没有已接受的 TMDB 身份时不写。失败只记日志。
+ */
+export async function backfillRecognizeCache(
+  rt: SourcesRuntime,
+  parsed: ParsedRelease,
+  entityId: string,
+): Promise<void> {
+  try {
+    const meta = recognizeMetaOf(parsed);
+    if (!meta || !(await rt.tmdb())) return;
+    const row = await rt.ctx.db
+      .selectFrom("external_identity as i")
+      .innerJoin("media_entity as e", "e.id", "i.media_entity_id")
+      .select(["i.entity_type", "i.external_id", "e.title", "e.year", "e.metadata"])
+      .where("i.media_entity_id", "=", entityId)
+      .where("i.provider", "=", "tmdb")
+      .where("i.status", "=", "accepted")
+      .executeTakeFirst();
+    if (!row || (row.entity_type !== "movie" && row.entity_type !== "tv")) return;
+    const md = (row.metadata ?? {}) as Record<string, { value?: unknown } | undefined>;
+    const str = (v: unknown) => (typeof v === "string" && v ? v : null);
+    await saveRecognizeCache(rt, meta, {
+      externalId: row.external_id,
+      entityType: row.entity_type,
+      title: row.title,
+      year: row.year,
+      posterPath: str(md.posterPath?.value),
+      backdropPath: str(md.backdropPath?.value),
+    });
+  } catch (e) {
+    rt.ctx.logger.warn({ entityId, err: { message: (e as Error).message } }, "识别缓存回填失败");
+  }
+}