be/src/cloud/cloud_tablet.h

// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

#pragma once

#include <memory>
#include <tuple>

#include "storage/partial_update_info.h"
#include "storage/rowset/rowset.h"
#include "storage/tablet/base_tablet.h"

namespace doris {

class CloudStorageEngine;

enum class WarmUpTriggerSource : int { NONE, SYNC_ROWSET, EVENT_DRIVEN, JOB };

enum class WarmUpProgress : int { NONE, DOING, DONE };

struct WarmUpState {
    WarmUpTriggerSource trigger_source {WarmUpTriggerSource::NONE};
    WarmUpProgress progress {WarmUpProgress::NONE};

    bool operator==(const WarmUpState& other) const {
        return trigger_source == other.trigger_source && progress == other.progress;
    }
};

struct SyncRowsetStats {
    int64_t get_remote_rowsets_num {0};
    int64_t get_remote_rowsets_rpc_ns {0};

    int64_t get_local_delete_bitmap_rowsets_num {0};
    int64_t get_remote_delete_bitmap_rowsets_num {0};
    int64_t get_remote_delete_bitmap_key_count {0};
    int64_t get_remote_delete_bitmap_bytes {0};
    int64_t get_remote_delete_bitmap_rpc_ns {0};

    int64_t get_remote_tablet_meta_rpc_ns {0};
    int64_t tablet_meta_cache_hit {0};
    int64_t tablet_meta_cache_miss {0};

    int64_t bthread_schedule_delay_ns {0};
    int64_t meta_lock_wait_ns {0}; // _meta_lock (BthreadSharedMutex) wait across all acquisitions
    int64_t sync_meta_lock_wait_ns {
            0}; // _sync_meta_lock (bthread::Mutex) wait across all acquisitions
};

struct SyncOptions {
    bool warmup_delta_data = false;
    bool sync_delete_bitmap = true;
    bool full_sync = false;
    bool merge_schema = false;
    int64_t query_version = -1;
};

struct RecycledRowsets {
    RowsetId rowset_id;
    int64_t num_segments;
    std::vector<int64_t> segment_ids;
    std::vector<std::string> index_file_names;
};

class CloudTablet final : public BaseTablet {
public:
    // rowset id -> [(segment id, version, serialized delete bitmap size)]
    using PreRowsetDeleteBitmapStats = std::map<
            std::string,
            std::vector<std::tuple<DeleteBitmap::SegmentId, DeleteBitmap::Version, size_t>>>;

    CloudTablet(CloudStorageEngine& engine, TabletMetaSharedPtr tablet_meta);

    ~CloudTablet() override;

    bool exceed_version_limit(int32_t limit) override;

    Result<std::unique_ptr<RowsetWriter>> create_rowset_writer(RowsetWriterContext& context,
                                                               bool vertical) override;

    Status capture_rs_readers(const Version& spec_version, std::vector<RowSetSplits>* rs_splits,
                              const CaptureRowsetOps& opts) override;

    [[nodiscard]] Result<std::vector<Version>> capture_consistent_versions_unlocked(
            const Version& version_range, const CaptureRowsetOps& options) const override;

    // Capture versions with cache preference optimization.
    // This method prioritizes using cached/warmed-up rowsets when building version paths,
    // avoiding cold data reads when possible. It uses capture_consistent_versions_prefer_cache
    // to find a consistent version path that prefers already warmed-up rowsets.
    Result<std::vector<Version>> capture_versions_prefer_cache(const Version& spec_version) const;

    // Capture versions with query freshness tolerance.
    // This method finds a consistent version path where all rowsets are warmed up,
    // but allows fallback to normal capture if there are newer rowsets that should be
    // visible (based on freshness tolerance) but haven't been warmed up yet.
    // For merge-on-write tables, uses special validation to ensure data correctness.
    //
    // IMPORTANT: The returned version may be smaller than the requested version if newer
    // data hasn't been warmed up yet. This can cause different tablets in the same query
    // to read from different versions, potentially leading to inconsistent query results.
    //
    // @param options.query_freshness_tolerance_ms: Time tolerance in milliseconds. Rowsets that
    //        became visible within this time range (after current_time - query_freshness_tolerance_ms)
    //        can be skipped if not warmed up. However, if older rowsets (before this time point)
    //        are not warmed up, the method will fallback to normal capture.
    Result<std::vector<Version>> capture_versions_with_freshness_tolerance(
            const Version& spec_version, const CaptureRowsetOps& options) const;

    size_t tablet_footprint() override {
        return _approximate_data_size.load(std::memory_order_relaxed);
    }

    std::string tablet_path() const override;

    // clang-format off
    int64_t fetch_add_approximate_num_rowsets (int64_t x) { return _approximate_num_rowsets .fetch_add(x, std::memory_order_relaxed); }
    int64_t fetch_add_approximate_num_segments(int64_t x) { return _approximate_num_segments.fetch_add(x, std::memory_order_relaxed); }
    int64_t fetch_add_approximate_num_rows    (int64_t x) { return _approximate_num_rows    .fetch_add(x, std::memory_order_relaxed); }
    int64_t fetch_add_approximate_data_size   (int64_t x) { return _approximate_data_size   .fetch_add(x, std::memory_order_relaxed); }
    int64_t fetch_add_approximate_cumu_num_rowsets (int64_t x) { return _approximate_cumu_num_rowsets.fetch_add(x, std::memory_order_relaxed); }
    int64_t fetch_add_approximate_cumu_num_deltas   (int64_t x) { return _approximate_cumu_num_deltas.fetch_add(x, std::memory_order_relaxed); }
    // clang-format on

    // meta lock must be held when calling this function
    void reset_approximate_stats(int64_t num_rowsets, int64_t num_segments, int64_t num_rows,
                                 int64_t data_size);

    // return a json string to show the compaction status of this tablet
    void get_compaction_status(std::string* json_result);

    // Synchronize the rowsets from meta service.
    // If tablet state is not `TABLET_RUNNING`, sync tablet meta and all visible rowsets.
    // If `query_version` > 0 and local max_version of the tablet >= `query_version`, do nothing.
    // If 'need_download_data_async' is true, it means that we need to download the new version
    // rowsets datum async.
    Status sync_rowsets(const SyncOptions& options = {}, SyncRowsetStats* stats = nullptr);

    // Synchronize the tablet meta from meta service.
    Status sync_meta();

    // If `version_overlap` is true, function will delete rowsets with overlapped version in this tablet.
    // If 'warmup_delta_data' is true, download the new version rowset data in background.
    // MUST hold EXCLUSIVE `_meta_lock`.
    // If 'need_download_data_async' is true, it means that we need to download the new version
    // rowsets datum async.
    void add_rowsets(std::vector<RowsetSharedPtr> to_add, bool version_overlap,
                     std::unique_lock<BthreadSharedMutex>& meta_lock,
                     bool warmup_delta_data = false);

    // MUST hold EXCLUSIVE `_meta_lock`.
    void delete_rowsets(const std::vector<RowsetSharedPtr>& to_delete,
                        std::unique_lock<BthreadSharedMutex>& meta_lock);

    // Like delete_rowsets, but also removes edges from the version graph.
    // Used by schema change to prevent the greedy capture algorithm from
    // preferring stale compaction rowsets over individual SC output rowsets.
    // MUST hold EXCLUSIVE `_meta_lock`.
    void delete_rowsets_for_schema_change(const std::vector<RowsetSharedPtr>& to_delete,
                                          std::unique_lock<BthreadSharedMutex>& meta_lock,
                                          bool recycle_deleted_rowsets = true);

    // Replace local rowsets in [2, alter_version] with schema change output rowsets.
    // Existing SC output rowsets are kept; other local/double-write/compaction rowsets
    // in this version range are removed from both _rs_version_map and version graph.
    // recycle_deleted_rowsets should only be true for the real tablet; temporary
    // schema-change delete-bitmap tablets only need to normalize their local graph.
    // MUST hold EXCLUSIVE `_meta_lock`.
    void replace_rowsets_with_schema_change_output(
            const std::vector<RowsetSharedPtr>& output_rowsets, int64_t alter_version,
            std::unique_lock<BthreadSharedMutex>& meta_lock, const char* stage,
            bool recycle_deleted_rowsets);

    // When the tablet is dropped, we need to recycle cached data:
    // 1. The data in file cache
    // 2. The memory in tablet cache
    void clear_cache() override;

    // Return number of deleted stale rowsets
    uint64_t delete_expired_stale_rowsets();

    bool has_stale_rowsets() const { return !_stale_rs_version_map.empty(); }

    int64_t get_cloud_base_compaction_score() const;
    int64_t get_cloud_cumu_compaction_score() const;

    int64_t max_version_unlocked() const override { return _max_version; }
    int64_t base_compaction_cnt() const { return _base_compaction_cnt; }
    int64_t cumulative_compaction_cnt() const { return _cumulative_compaction_cnt; }
    int64_t full_compaction_cnt() const { return _full_compaction_cnt; }
    int64_t cumulative_layer_point() const {
        return _cumulative_point.load(std::memory_order_relaxed);
    }

    void set_base_compaction_cnt(int64_t cnt) { _base_compaction_cnt = cnt; }
    void set_cumulative_compaction_cnt(int64_t cnt) { _cumulative_compaction_cnt = cnt; }
    void set_full_compaction_cnt(int64_t cnt) { _full_compaction_cnt = cnt; }
    void set_cumulative_layer_point(int64_t new_point);

    int64_t last_cumu_compaction_failure_time() { return _last_cumu_compaction_failure_millis; }
    void set_last_cumu_compaction_failure_time(int64_t millis) {
        _last_cumu_compaction_failure_millis = millis;
    }

    int64_t last_base_compaction_failure_time() { return _last_base_compaction_failure_millis; }
    void set_last_base_compaction_failure_time(int64_t millis) {
        _last_base_compaction_failure_millis = millis;
    }

    int64_t last_full_compaction_failure_time() { return _last_full_compaction_failure_millis; }
    void set_last_full_compaction_failure_time(int64_t millis) {
        _last_full_compaction_failure_millis = millis;
    }

    int64_t last_cumu_compaction_success_time() { return _last_cumu_compaction_success_millis; }
    void set_last_cumu_compaction_success_time(int64_t millis) {
        _last_cumu_compaction_success_millis = millis;
    }

    int64_t last_base_compaction_success_time() { return _last_base_compaction_success_millis; }
    void set_last_base_compaction_success_time(int64_t millis) {
        _last_base_compaction_success_millis = millis;
    }

    int64_t last_full_compaction_success_time() { return _last_full_compaction_success_millis; }
    void set_last_full_compaction_success_time(int64_t millis) {
        _last_full_compaction_success_millis = millis;
    }

    int64_t last_cumu_compaction_schedule_time() { return _last_cumu_compaction_schedule_millis; }
    void set_last_cumu_compaction_schedule_time(int64_t millis) {
        _last_cumu_compaction_schedule_millis = millis;
    }

    int64_t last_base_compaction_schedule_time() { return _last_base_compaction_schedule_millis; }
    void set_last_base_compaction_schedule_time(int64_t millis) {
        _last_base_compaction_schedule_millis = millis;
    }

    int64_t last_full_compaction_schedule_time() { return _last_full_compaction_schedule_millis; }
    void set_last_full_compaction_schedule_time(int64_t millis) {
        _last_full_compaction_schedule_millis = millis;
    }

    void set_last_cumu_compaction_status(std::string status) {
        _last_cumu_compaction_status = std::move(status);
    }

    std::string get_last_cumu_compaction_status() { return _last_cumu_compaction_status; }

    void set_last_base_compaction_status(std::string status) {
        _last_base_compaction_status = std::move(status);
    }

    std::string get_last_base_compaction_status() { return _last_base_compaction_status; }

    void set_last_full_compaction_status(std::string status) {
        _last_full_compaction_status = std::move(status);
    }

    std::string get_last_full_compaction_status() { return _last_full_compaction_status; }

    int64_t alter_version() const { return _alter_version; }
    void set_alter_version(int64_t alter_version) { _alter_version = alter_version; }

    // Last active cluster info for compaction read-write separation
    std::string last_active_cluster_id() const {
        std::shared_lock lock(_cluster_info_mutex);
        return _last_active_cluster_id;
    }
    int64_t last_active_time_ms() const {
        std::shared_lock lock(_cluster_info_mutex);
        return _last_active_time_ms;
    }
    void set_last_active_cluster_info(const std::string& cluster_id, int64_t time_ms) {
        std::unique_lock lock(_cluster_info_mutex);
        _last_active_cluster_id = cluster_id;
        _last_active_time_ms = time_ms;
    }

    // MUST hold SHARED `_meta_lock`.
    std::vector<RowsetSharedPtr> pick_candidate_rowsets_to_base_compaction_unlocked();

    inline Version max_version() const {
        std::shared_lock rdlock(_meta_lock);
        return _tablet_meta->max_version();
    }

    int64_t base_size() const { return _base_size; }

    // MUST hold SHARED `_meta_lock`.
    std::vector<RowsetSharedPtr> pick_candidate_rowsets_to_full_compaction_unlocked();
    Result<RowsetSharedPtr> pick_a_rowset_for_index_change(int schema_version,
                                                           bool& is_base_rowset);
    Status check_rowset_schema_for_build_index(std::vector<TColumn>& columns, int schema_version);

    std::mutex& get_base_compaction_lock() { return _base_compaction_lock; }
    std::mutex& get_cumulative_compaction_lock() { return _cumulative_compaction_lock; }

    Result<std::unique_ptr<RowsetWriter>> create_transient_rowset_writer(
            const Rowset& rowset, std::shared_ptr<PartialUpdateInfo> partial_update_info,
            int64_t txn_expiration = 0) override;

    CalcDeleteBitmapExecutor* calc_delete_bitmap_executor() override;

    Status save_delete_bitmap(const TabletTxnInfo* txn_info, int64_t txn_id,
                              DeleteBitmapPtr delete_bitmap, RowsetWriter* rowset_writer,
                              const RowsetIdUnorderedSet& cur_rowset_ids, int64_t lock_id = -1,
                              int64_t next_visible_version = -1) override;

    Status save_delete_bitmap_to_ms(int64_t cur_version, int64_t txn_id,
                                    DeleteBitmapPtr delete_bitmap, int64_t lock_id,
                                    int64_t next_visible_version, RowsetSharedPtr rowset);

    Status calc_delete_bitmap_for_compaction(const std::vector<RowsetSharedPtr>& input_rowsets,
                                             const RowsetSharedPtr& output_rowset,
                                             const RowIdConversion& rowid_conversion,
                                             ReaderType compaction_type, int64_t merged_rows,
                                             int64_t filtered_rows, int64_t initiator,
                                             DeleteBitmapPtr& output_rowset_delete_bitmap,
                                             bool allow_delete_in_cumu_compaction,
                                             int64_t& get_delete_bitmap_lock_start_time);

    // Find the missed versions until the spec_version.
    //
    // for example:
    //     [0-4][5-5][8-8][9-9][14-14]
    // if spec_version = 12, it will return [6-7],[10-12]
    Versions calc_missed_versions(int64_t spec_version, Versions existing_versions) const override;

    std::mutex& get_rowset_update_lock() { return _rowset_update_lock; }

    bthread::Mutex& get_sync_meta_lock() { return _sync_meta_lock; }

    const auto& rowset_map() const { return _rs_version_map; }

    // How long since this tablet's ROWSETS were pulled from MS. Only sync_rowsets() advances
    // it, and only when it actually issues the RPC.
    int64_t last_sync_rowsets_time_s = 0;
    // How long since this tablet's META was pulled from MS, which is what carries properties
    // such as the file cache TTL. Only sync_meta() advances it. Tracked separately on
    // purpose: a tablet under continuous ingest keeps the rowsets clock permanently fresh,
    // so gating meta work on that one starves the meta refresh entirely.
    int64_t last_sync_tablet_meta_time_s = 0;
    int64_t last_load_time_ms = 0;
    int64_t last_base_compaction_success_time_ms = 0;
    int64_t last_cumu_compaction_success_time_ms = 0;
    int64_t last_cumu_no_suitable_version_ms = 0;
    int64_t last_access_time_ms = 0;

    std::atomic<int64_t> local_read_time_us = 0;
    std::atomic<int64_t> remote_read_time_us = 0;
    std::atomic<int64_t> exec_compaction_time_us = 0;

    void build_tablet_report_info(TTabletInfo* tablet_info);

    // check that if the delete bitmap in delete bitmap cache has the same cardinality with the expected_delete_bitmap's
    Status check_delete_bitmap_cache(int64_t txn_id, DeleteBitmap* expected_delete_bitmap) override;

    void agg_delete_bitmap_for_compaction(
            int64_t start_version, int64_t end_version,
            const std::vector<RowsetSharedPtr>& pre_rowsets, DeleteBitmapPtr& new_delete_bitmap,
            std::map<std::string, int64_t>& pre_rowset_to_versions,
            PreRowsetDeleteBitmapStats* pre_rowset_delete_bitmap_stats);

    bool need_remove_unused_rowsets();

    void add_unused_rowsets(const std::vector<RowsetSharedPtr>& rowsets);
    void remove_unused_rowsets();

    // For each given rowset not in active use, clears its file cache and returns its
    // ID, segment count, and index file names as RecycledRowsets entries.
    static std::vector<RecycledRowsets> recycle_cached_data(
            const std::vector<RowsetSharedPtr>& rowsets);

    // Add warmup state management
    WarmUpState get_rowset_warmup_state(RowsetId rowset_id);
    bool add_rowset_warmup_state(
            const RowsetMeta& rowset, WarmUpTriggerSource source,
            std::chrono::steady_clock::time_point start_tp = std::chrono::steady_clock::now());
    bool update_rowset_warmup_state_inverted_idx_num(WarmUpTriggerSource source, RowsetId rowset_id,
                                                     int64_t delta);
    bool update_rowset_warmup_state_inverted_idx_num_unlocked(WarmUpTriggerSource source,
                                                              RowsetId rowset_id, int64_t delta);
    WarmUpState complete_rowset_segment_warmup(WarmUpTriggerSource trigger_source,
                                               RowsetId rowset_id, Status status,
                                               int64_t segment_num, int64_t inverted_idx_num);

    bool is_rowset_warmed_up(const RowsetId& rowset_id) const;

    void add_warmed_up_rowset(const RowsetId& rowset_id);
    // Test helper: add a rowset to the warmup state map with DOING progress,
    // so that is_rowset_warmed_up() returns false for it.
    void add_not_warmed_up_rowset(const RowsetId& rowset_id);

    // Try to apply visible pending rowsets to tablet meta in version order
    // This should be called after receiving FE notification or when new rowsets are added
    // @return Status::OK() if successfully applied, error otherwise
    void apply_visible_pending_rowsets();

    void try_make_committed_rs_visible(int64_t txn_id, int64_t visible_version,
                                       int64_t version_update_time_ms);
    void try_make_committed_rs_visible_for_mow(int64_t txn_id, int64_t visible_version,
                                               int64_t version_update_time_ms);

    void clear_unused_visible_pending_rowsets();

    std::string rowset_warmup_digest() const {
        std::string res;
        auto add_log = [&](const RowsetSharedPtr& rs) {
            auto tmp = fmt::format("{}{}", rs->rowset_id().to_string(), rs->version().to_string());
            if (_rowset_warm_up_states.contains(rs->rowset_id())) {
                tmp += fmt::format(
                        ", progress={}, segments_warmed_up={}/{}, inverted_idx_warmed_up={}/{}",
                        _rowset_warm_up_states.at(rs->rowset_id()).state.progress,
                        _rowset_warm_up_states.at(rs->rowset_id()).num_segments_warmed_up,
                        _rowset_warm_up_states.at(rs->rowset_id()).num_segments,
                        _rowset_warm_up_states.at(rs->rowset_id()).num_inverted_idx_warmed_up,
                        _rowset_warm_up_states.at(rs->rowset_id()).num_inverted_idx);
            }
            res += fmt::format("[{}],", tmp);
        };
        traverse_rowsets_unlocked(add_log, true);
        return res;
    }

private:
    // FIXME(plat1ko): No need to record base size if rowsets are ordered by version
    void update_base_size(const Rowset& rs);

    Status sync_if_not_running(SyncRowsetStats* stats = nullptr);

    bool add_rowset_warmup_state_unlocked(
            const RowsetMeta& rowset, WarmUpTriggerSource source,
            std::chrono::steady_clock::time_point start_tp = std::chrono::steady_clock::now());

    // used by capture_rs_reader_xxx functions
    bool rowset_is_warmed_up_unlocked(int64_t start_version, int64_t end_version) const;

    // Check if a rowset should be visible but not warmed up within freshness tolerance
    bool _check_rowset_should_be_visible_but_not_warmed_up(
            const RowsetMetaSharedPtr& rs_meta, int64_t path_max_version,
            std::chrono::system_clock::time_point freshness_limit_tp) const;

    // Submit a segment download task for warming up
    void _submit_segment_download_task(const RowsetSharedPtr& rs, io::Path segment_path,
                                       int64_t segment_file_size, int64_t expiration_time);

    // Submit an inverted index download task for warming up
    void _submit_inverted_index_download_task(const RowsetSharedPtr& rs,
                                              const StorageResource* storage_resource,
                                              const io::Path& idx_path, int64_t idx_size,
                                              int64_t expiration_time);

    // Add rowsets directly with warmup
    void _add_rowsets_directly(std::vector<RowsetSharedPtr>& rowsets, bool warmup_delta_data);

    CloudStorageEngine& _engine;

    // this mutex MUST ONLY be used when sync meta
    bthread::Mutex _sync_meta_lock;
    // ATTENTION: lock order should be: _sync_meta_lock -> _meta_lock

    std::atomic<int64_t> _cumulative_point {-1};
    std::atomic<int64_t> _approximate_num_rowsets {-1};
    std::atomic<int64_t> _approximate_num_segments {-1};
    std::atomic<int64_t> _approximate_num_rows {-1};
    std::atomic<int64_t> _approximate_data_size {-1};
    std::atomic<int64_t> _approximate_cumu_num_rowsets {-1};
    // Number of sorted arrays (e.g. for rowset with N segments, if rowset is overlapping, delta is N, otherwise 1) after cumu point
    std::atomic<int64_t> _approximate_cumu_num_deltas {-1};

    // timestamp of last cumu compaction failure
    std::atomic<int64_t> _last_cumu_compaction_failure_millis;
    // timestamp of last base compaction failure
    std::atomic<int64_t> _last_base_compaction_failure_millis;
    // timestamp of last full compaction failure
    std::atomic<int64_t> _last_full_compaction_failure_millis;
    // timestamp of last cumu compaction success
    std::atomic<int64_t> _last_cumu_compaction_success_millis;
    // timestamp of last base compaction success
    std::atomic<int64_t> _last_base_compaction_success_millis;
    // timestamp of last full compaction success
    std::atomic<int64_t> _last_full_compaction_success_millis;
    // timestamp of last cumu compaction schedule time
    std::atomic<int64_t> _last_cumu_compaction_schedule_millis;
    // timestamp of last base compaction schedule time
    std::atomic<int64_t> _last_base_compaction_schedule_millis;
    // timestamp of last full compaction schedule time
    std::atomic<int64_t> _last_full_compaction_schedule_millis;
    std::string _last_cumu_compaction_status;
    std::string _last_base_compaction_status;
    std::string _last_full_compaction_status;
    int64_t _base_compaction_cnt = 0;
    int64_t _cumulative_compaction_cnt = 0;
    int64_t _full_compaction_cnt = 0;
    int64_t _max_version = -1;
    int64_t _base_size = 0;
    int64_t _alter_version = -1;

    std::mutex _base_compaction_lock;
    std::mutex _cumulative_compaction_lock;

    // To avoid multiple calc delete bitmap tasks on same (txn_id, tablet_id) with different
    // signatures being executed concurrently, we use _rowset_update_lock to serialize them
    mutable std::mutex _rowset_update_lock;

    // unused_rowsets, [start_version, end_version]
    std::mutex _gc_mutex;
    std::unordered_map<RowsetId, RowsetSharedPtr> _unused_rowsets;
    std::vector<std::pair<std::vector<RowsetId>, DeleteBitmapKeyRanges>> _unused_delete_bitmap;

    // for warm up states management
    struct RowsetWarmUpInfo {
        WarmUpState state;
        int64_t num_segments = 0;
        int64_t num_inverted_idx = 0;
        int64_t num_segments_warmed_up = 0;
        int64_t num_inverted_idx_warmed_up = 0;
        std::chrono::steady_clock::time_point start_tp;

        void done(int64_t input_num_segments, int64_t input_num_inverted_idx) {
            num_segments_warmed_up += input_num_segments;
            num_inverted_idx_warmed_up += input_num_inverted_idx;
            update_state();
        }

        bool has_finished() const {
            return (num_segments_warmed_up >= num_segments) &&
                   (num_inverted_idx_warmed_up >= num_inverted_idx);
        }

        void update_state();
    };
    std::unordered_map<RowsetId, RowsetWarmUpInfo> _rowset_warm_up_states;

    mutable std::shared_mutex _warmed_up_rowsets_mutex;
    std::unordered_set<RowsetId> _warmed_up_rowsets;

    // Cluster info for compaction read-write separation
    mutable std::shared_mutex _cluster_info_mutex;
    std::string _last_active_cluster_id;
    int64_t _last_active_time_ms {0};

    // Map: version -> <rowset_meta, expiration_time>
    // Stores rowsets that have been notified by FE but not yet added to tablet meta
    // due to out-of-order notification or version discontinuity
    struct VisiblePendingRowset {
        const bool is_empty_rowset;
        const int64_t expiration_time; // seconds since epoch
        RowsetMetaSharedPtr rowset_meta;

        VisiblePendingRowset(RowsetMetaSharedPtr rowset_meta_, int64_t expiration_time_,
                             bool is_empty_rowset_ = false)
                : is_empty_rowset(is_empty_rowset_),
                  expiration_time(expiration_time_),
                  rowset_meta(std::move(rowset_meta_)) {}
    };
    mutable std::mutex _visible_pending_rs_lock;
    std::map<int64_t, VisiblePendingRowset> _visible_pending_rs_map;
};

using CloudTabletSPtr = std::shared_ptr<CloudTablet>;

} // namespace doris