[Performance](clould) reduce tablet cache read in topn query (#50803)

Topn lazy read only read the cache in cloud mode to reduce tablet cache
read.

Author: HappenLee

be/src/cloud/cloud_storage_engine.cpp

@@ -249,10 +249,10 @@ bool CloudStorageEngine::stopped() {
 }
 
 Result<BaseTabletSPtr> CloudStorageEngine::get_tablet(int64_t tablet_id,
-                                                      SyncRowsetStats* sync_stats) {
-    return _tablet_mgr->get_tablet(tablet_id, false, true, sync_stats).transform([](auto&& t) {
-        return static_pointer_cast<BaseTablet>(std::move(t));
-    });
+                                                      SyncRowsetStats* sync_stats,
+                                                      bool force_use_cache) {
+    return _tablet_mgr->get_tablet(tablet_id, false, true, sync_stats, force_use_cache)
+            .transform([](auto&& t) { return static_pointer_cast<BaseTablet>(std::move(t)); });
 }
 
 Status CloudStorageEngine::start_bg_threads() {

be/src/cloud/cloud_storage_engine.h

@@ -58,8 +58,8 @@ class CloudStorageEngine final : public BaseStorageEngine {
     void stop() override;
     bool stopped() override;
 
-    Result<BaseTabletSPtr> get_tablet(int64_t tablet_id,
-                                      SyncRowsetStats* sync_stats = nullptr) override;
+    Result<BaseTabletSPtr> get_tablet(int64_t tablet_id, SyncRowsetStats* sync_stats = nullptr,
+                                      bool force_use_cache = false) override;
 
     Status start_bg_threads() override;
 

be/src/cloud/cloud_tablet_mgr.cpp

@@ -153,7 +153,8 @@ void set_tablet_access_time_ms(CloudTablet* tablet) {
 
 Result<std::shared_ptr<CloudTablet>> CloudTabletMgr::get_tablet(int64_t tablet_id, bool warmup_data,
                                                                 bool sync_delete_bitmap,
-                                                                SyncRowsetStats* sync_stats) {
+                                                                SyncRowsetStats* sync_stats,
+                                                                bool force_use_cache) {
     // LRU value type. `Value`'s lifetime MUST NOT be longer than `CloudTabletMgr`
     class Value : public LRUCacheValueBase {
     public:
@@ -170,6 +171,12 @@ Result<std::shared_ptr<CloudTablet>> CloudTabletMgr::get_tablet(int64_t tablet_i
     auto tablet_id_str = std::to_string(tablet_id);
     CacheKey key(tablet_id_str);
     auto* handle = _cache->lookup(key);
+
+    if (handle == nullptr && force_use_cache) {
+        return ResultError(
+                Status::InternalError("failed to get cloud tablet from cache {}", tablet_id));
+    }
+
     if (handle == nullptr) {
         if (sync_stats) {
             ++sync_stats->tablet_meta_cache_miss;

be/src/cloud/cloud_tablet_mgr.h

@@ -46,7 +46,8 @@ class CloudTabletMgr {
     // sync rowsets after, and download segment data in background if `warmup_data` is true.
     Result<std::shared_ptr<CloudTablet>> get_tablet(int64_t tablet_id, bool warmup_data = false,
                                                     bool sync_delete_bitmap = true,
-                                                    SyncRowsetStats* sync_stats = nullptr);
+                                                    SyncRowsetStats* sync_stats = nullptr,
+                                                    bool force_use_cache = false);
 
     void erase_tablet(int64_t tablet_id);
 

be/src/exec/rowid_fetcher.cpp

@@ -359,7 +359,7 @@ Status RowIdStorageReader::read_by_rowids(const PMultiGetRequest& request,
         watch.start();
         BaseTabletSPtr tablet = scope_timer_run(
                 [&]() {
-                    auto res = ExecEnv::get_tablet(row_loc.tablet_id());
+                    auto res = ExecEnv::get_tablet(row_loc.tablet_id(), nullptr, true);
                     return !res.has_value() ? nullptr
                                             : std::dynamic_pointer_cast<BaseTablet>(res.value());
                 },

be/src/olap/storage_engine.cpp

@@ -1266,7 +1266,8 @@ Status StorageEngine::create_tablet(const TCreateTabletReq& request, RuntimeProf
     return _tablet_manager->create_tablet(request, stores, profile);
 }
 
-Result<BaseTabletSPtr> StorageEngine::get_tablet(int64_t tablet_id, SyncRowsetStats* sync_stats) {
+Result<BaseTabletSPtr> StorageEngine::get_tablet(int64_t tablet_id, SyncRowsetStats* sync_stats,
+                                                 bool force_use_cache) {
     BaseTabletSPtr tablet;
     std::string err;
     tablet = _tablet_manager->get_tablet(tablet_id, true, &err);

be/src/olap/storage_engine.h

@@ -109,7 +109,8 @@ class BaseStorageEngine {
     virtual Status start_bg_threads() = 0;
 
     virtual Result<BaseTabletSPtr> get_tablet(int64_t tablet_id,
-                                              SyncRowsetStats* sync_stats = nullptr) = 0;
+                                              SyncRowsetStats* sync_stats = nullptr,
+                                              bool force_use_cache = false) = 0;
 
     void register_report_listener(ReportWorker* listener);
     void deregister_report_listener(ReportWorker* listener);
@@ -227,8 +228,8 @@ class StorageEngine final : public BaseStorageEngine {
 
     Status create_tablet(const TCreateTabletReq& request, RuntimeProfile* profile);
 
-    Result<BaseTabletSPtr> get_tablet(int64_t tablet_id,
-                                      SyncRowsetStats* sync_stats = nullptr) override;
+    Result<BaseTabletSPtr> get_tablet(int64_t tablet_id, SyncRowsetStats* sync_stats = nullptr,
+                                      bool force_use_cache = false) override;
 
     void clear_transaction_task(const TTransactionId transaction_id);
     void clear_transaction_task(const TTransactionId transaction_id,

be/src/runtime/exec_env.cpp

@@ -57,7 +57,8 @@ void ExecEnv::set_write_cooldown_meta_executors() {
 }
 #endif // BE_TEST
 
-Result<BaseTabletSPtr> ExecEnv::get_tablet(int64_t tablet_id, SyncRowsetStats* sync_stats) {
+Result<BaseTabletSPtr> ExecEnv::get_tablet(int64_t tablet_id, SyncRowsetStats* sync_stats,
+                                           bool force_use_cache) {
     auto storage_engine = GetInstance()->_storage_engine.get();
     return storage_engine != nullptr
                    ? storage_engine->get_tablet(tablet_id, sync_stats)

be/src/runtime/exec_env.h

@@ -148,7 +148,8 @@ class ExecEnv {
 
     // Requires ExenEnv ready
     static Result<BaseTabletSPtr> get_tablet(int64_t tablet_id,
-                                             SyncRowsetStats* sync_stats = nullptr);
+                                             SyncRowsetStats* sync_stats = nullptr,
+                                             bool force_use_cache = false);
 
     static bool ready() { return _s_ready.load(std::memory_order_acquire); }
     static bool tracking_memory() { return _s_tracking_memory.load(std::memory_order_acquire); }