scylladb/replica/compaction_group.hh

/*
 * Copyright (C) 2022-present ScyllaDB
 */

/*
 * SPDX-License-Identifier: LicenseRef-ScyllaDB-Source-Available-1.1
 */

#include <seastar/core/condition-variable.hh>
#include <seastar/core/gate.hh>
#include <seastar/core/rwlock.hh>

#include "database_fwd.hh"
#include "compaction/compaction_descriptor.hh"
#include "compaction/compaction_backlog_manager.hh"
#include "compaction/compaction_strategy_state.hh"
// FIXME: un-nest compaction_reenabler, so we can forward declare it and remove this include.
#include "compaction/compaction_manager.hh"
#include "locator/tablets.hh"
#include "replica/logstor/compaction.hh"
#include "replica/logstor/segment_manager.hh"
#include "sstables/sstable_set.hh"
#include "utils/chunked_vector.hh"
#include <absl/container/flat_hash_map.h>

#pragma once

namespace compaction {
class compaction_manager;
}

namespace locator {
class effective_replication_map;
}

namespace replica {

namespace logstor {
class primary_index;
}

using enable_backlog_tracker = bool_class<class enable_backlog_tracker_tag>;

enum class repair_sstable_classification {
    unrepaired,
    repairing,
    repaired,
};

using repair_classifier_func = std::function<repair_sstable_classification(const sstables::shared_sstable&, int64_t sstables_repaired_at)>;

// Compaction group is a set of SSTables which are eligible to be compacted together.
// By this definition, we can say:
//      - A group contains SSTables that are owned by the same shard.
//      - Also, a group will be owned by a single table. Different tables own different groups.
//      - Each group can be thought of an isolated LSM tree, where Memtable(s) and SSTable(s) are
//          isolated from other groups.
class compaction_group {
    table& _t;
    // The compaction group views are the logical compaction groups, each having its own logical
    // set of sstables. Even though they share the same instance of sstable_set, compaction will
    // only see the sstables belonging to a particular view, with the help of the classifier.
    // This way, we guarantee that sstables falling under different groups cannot be compacted
    // together.
    class compaction_group_view;
    // This is held throughout group lifetime, in order to have compaction disabled on non-compacting views.
    std::vector<compaction::compaction_reenabler> _compaction_disabler_for_views;
    // Logical compaction group representing the unrepaired sstables.
    std::unique_ptr<compaction_group_view> _unrepaired_view;
    // Logical compaction group representing the repairing sstables. Compaction disabled altogether on it.
    std::unique_ptr<compaction_group_view> _repairing_view;
    // Logical compaction group representing the repaired sstables.
    std::unique_ptr<compaction_group_view> _repaired_view;
    size_t _group_id;
    // Tokens included in this compaction_groups
    dht::token_range _token_range;
    compaction::compaction_strategy_state _compaction_strategy_state;
    // Holds list of memtables for this group
    lw_shared_ptr<memtable_list> _memtables;
    // SSTable set which contains all non-maintenance sstables
    lw_shared_ptr<sstables::sstable_set> _main_sstables;
    // Holds SSTables created by maintenance operations, which need reshaping before integration into the main set
    lw_shared_ptr<sstables::sstable_set> _maintenance_sstables;
    // sstables that have been compacted (so don't look up in query) but
    // have not been deleted yet, so must not GC any tombstones in other sstables
    // that may delete data in these sstables:
    std::vector<sstables::shared_sstable> _sstables_compacted_but_not_deleted;
    seastar::condition_variable _staging_done_condition;
    // Gates async operations confined to a single group.
    seastar::named_gate _async_gate;
    // Gates flushes.
    seastar::named_gate _flush_gate;
    // Gates sstable being added to the group.
    // This prevents the group from being considered empty when sstables are being added.
    // Crucial for tablet split which ACKs split for a table when all pre-split groups are empty.
    seastar::named_gate _sstable_add_gate;
    bool _tombstone_gc_enabled = true;
    std::optional<compaction::compaction_backlog_tracker> _backlog_tracker;
    repair_classifier_func _repair_sstable_classifier;

    counter_id _counter_id;

    lw_shared_ptr<logstor::segment_set> _logstor_segments;
    std::optional<logstor::separator_buffer> _logstor_separator;
    std::vector<future<>> _separator_flushes;
    seastar::semaphore _separator_flush_sem{1};

private:
    std::unique_ptr<compaction_group_view> make_compacting_view();
    std::unique_ptr<compaction_group_view> make_non_compacting_view();

    // Adds new sstable to the set of sstables
    // Doesn't update the cache. The cache must be synchronized in order for reads to see
    // the writes contained in this sstable.
    // Cache must be synchronized atomically with this, otherwise write atomicity may not be respected.
    // Doesn't trigger compaction.
    // Strong exception guarantees.
    lw_shared_ptr<sstables::sstable_set>
    do_add_sstable(lw_shared_ptr<sstables::sstable_set> sstables, sstables::shared_sstable sstable,
                   enable_backlog_tracker backlog_tracker);
    // Update compaction backlog tracker with the same changes applied to the underlying sstable set.
    void backlog_tracker_adjust_charges(const std::vector<sstables::shared_sstable>& old_sstables, const std::vector<sstables::shared_sstable>& new_sstables);

    // Input SSTables that weren't added to any SSTable set, are considered unused and can be unlinked.
    // An input SSTable remains linked if it wasn't actually compacted, yet compaction manager wants
    // it to be moved from its original sstable set (e.g. maintenance) into a new one (e.g. main).
    std::vector<sstables::shared_sstable> unused_sstables_for_deletion(compaction::compaction_completion_desc desc) const;
    // Tracks the maximum timestamp observed across all SSTables in this group.
    // This is used by the compacting reader to determine if a memtable contains entries
    // with timestamps that overlap with those in the SSTables of the compaction group.
    // For this purpose, tracking the maximum seen timestamp is sufficient rather than the
    // actual maximum across all SSTables. So, the variable is updated only when a new SSTable
    // is added to the group. While `set_main_sstables` and `set_maintenance_sstables` can
    // replace entire sstable sets, they are still called only by compaction, so the maximum
    // seen timestamp remains the same and there is no need to update the variable in those cases.
    api::timestamp_type _max_seen_timestamp = api::missing_timestamp;
public:
    compaction_group(table& t, size_t gid, dht::token_range token_range, repair_classifier_func repair_classifier);
    ~compaction_group();

    // Create a group with same metadata of base like range, id, but with empty data (sstable & memtable).
    static lw_shared_ptr<compaction_group> make_empty_group(const compaction_group& base);

    void update_id(size_t id) {
        _group_id = id;
    }
    void update_id_and_range(size_t id, dht::token_range token_range) {
        _group_id = id;
        _token_range = std::move(token_range);
    }

    size_t group_id() const noexcept {
        return _group_id;
    }

    const schema_ptr& schema() const;

    // Stops all activity in the group, synchronizes with in-flight writes, before
    // flushing memtable(s), so all data can be found in the SSTable set.
    future<> stop(sstring reason) noexcept;

    bool stopped() const noexcept;

    bool empty() const noexcept;

    // This removes all the storage belonging to the group. In order to avoid data
    // resurrection, makes sure that all data is flushed into SSTables before
    // proceeding with atomic deletion on them.
    future<> cleanup();

    // Clear sstable sets
    void clear_sstables();

    // Clear memtable(s) content
    future<> clear_memtables();

    future<> flush() noexcept;
    bool can_flush() const;
    bool needs_flush() const;

    const dht::token_range& token_range() const noexcept {
        return _token_range;
    }

    void set_tombstone_gc_enabled(bool tombstone_gc_enabled) noexcept {
        _tombstone_gc_enabled = tombstone_gc_enabled;
    }

    bool tombstone_gc_enabled() const noexcept {
        return _tombstone_gc_enabled;
    }

    int64_t get_sstables_repaired_at() const noexcept;

    future<> update_repaired_at_for_merge();

    void set_counter_id(counter_id cid) noexcept {
        _counter_id = cid;
    }

    counter_id get_counter_id() const noexcept {
        return _counter_id;
    }

    void set_compaction_strategy_state(compaction::compaction_strategy_state compaction_strategy_state) noexcept;

    lw_shared_ptr<memtable_list>& memtables() noexcept;
    size_t memtable_count() const noexcept;
    // Returns minimum timestamp from memtable list
    api::timestamp_type min_memtable_timestamp() const;
    // Returns maximum timestamp from memtable list
    api::timestamp_type max_memtable_timestamp() const;
    // Returns minimum timestamp of live data from memtable list
    api::timestamp_type min_memtable_live_timestamp() const;
    // Returns minimum timestamp of live row markers from memtable list
    api::timestamp_type min_memtable_live_row_marker_timestamp() const;
    // Returns true if memtable(s) contains key.
    bool memtable_has_key(const dht::decorated_key& key) const;
    // Add sstable to main set
    void add_sstable(sstables::shared_sstable sstable);
    // Add sstable to maintenance set
    void add_maintenance_sstable(sstables::shared_sstable sst);
    api::timestamp_type max_seen_timestamp() const { return _max_seen_timestamp; }

    // Update main and/or maintenance sstable sets based in info in completion descriptor,
    // where input sstables will be replaced by output ones, row cache ranges are possibly
    // invalidated and statistics are updated.
    future<> update_sstable_sets_on_compaction_completion(compaction::compaction_completion_desc desc);

    // Merges all sstables from another group into this one.
    future<> merge_sstables_from(compaction_group& group);

    // Merges all logstor segments from another group into this one.
    future<> merge_logstor_segments_from(compaction_group& group);

    const lw_shared_ptr<sstables::sstable_set>& main_sstables() const noexcept;
    sstables::sstable_set make_main_sstable_set() const;
    void set_main_sstables(lw_shared_ptr<sstables::sstable_set> new_main_sstables);

    const lw_shared_ptr<sstables::sstable_set>& maintenance_sstables() const noexcept;
    lw_shared_ptr<sstables::sstable_set> make_maintenance_sstable_set() const;
    void set_maintenance_sstables(lw_shared_ptr<sstables::sstable_set> new_maintenance_sstables);

    // Makes a sstable set, which includes all sstables managed by this group
    lw_shared_ptr<sstables::sstable_set> make_sstable_set() const;
    std::vector<sstables::shared_sstable> all_sstables() const;

    const std::vector<sstables::shared_sstable>& compacted_undeleted_sstables() const noexcept;
    // Triggers regular compaction.
    void trigger_compaction();
    void trigger_logstor_compaction();
    bool compaction_disabled() const;
    future<unsigned> estimate_pending_compactions() const;

    compaction::compaction_backlog_tracker& get_backlog_tracker();
    void register_backlog_tracker(compaction::compaction_backlog_tracker new_backlog_tracker);

    size_t live_sstable_count() const noexcept;
    uint64_t live_disk_space_used() const noexcept;
    size_t logstor_disk_space_used() const noexcept;
    sstables::file_size_stats live_disk_space_used_full_stats() const noexcept;
    uint64_t total_disk_space_used() const noexcept;
    sstables::file_size_stats total_disk_space_used_full_stats() const noexcept;

    // With static sharding, i.e. vnodes, there will be only one active view.
    compaction::compaction_group_view& as_view_for_static_sharding() const;
    // Default view to be used on newly created sstables, e.g. those produced by repair or memtable.
    compaction::compaction_group_view& view_for_unrepaired_data() const;
    // Gets the view a sstable currently belongs to.
    compaction::compaction_group_view& view_for_sstable(const sstables::shared_sstable& sst) const;
    utils::small_vector<compaction::compaction_group_view*, 3> all_views() const;
    // Returns true iff v is the repaired view of this compaction group.
    bool is_repaired_view(const compaction::compaction_group_view* v) const noexcept;
    // Returns an sstable set containing only repaired sstables (those classified as repaired).
    lw_shared_ptr<sstables::sstable_set> make_repaired_sstable_set() const;

    seastar::condition_variable& get_staging_done_condition() noexcept {
        return _staging_done_condition;
    }

    seastar::named_gate& async_gate() noexcept {
        return _async_gate;
    }

    seastar::named_gate& flush_gate() noexcept {
        return _flush_gate;
    }

    seastar::named_gate& sstable_add_gate() noexcept {
        return _sstable_add_gate;
    }

    compaction::compaction_manager& get_compaction_manager() noexcept;
    const compaction::compaction_manager& get_compaction_manager() const noexcept;

    logstor::segment_manager& get_logstor_segment_manager() noexcept;
    const logstor::segment_manager& get_logstor_segment_manager() const noexcept;

    logstor::compaction_manager& get_logstor_compaction_manager() noexcept;
    const logstor::compaction_manager& get_logstor_compaction_manager() const noexcept;

    logstor::primary_index& get_logstor_index() noexcept;

    future<> split(compaction::compaction_type_options::split opt, tasks::task_info tablet_split_task_info);

    void set_repair_sstable_classifier(repair_classifier_func repair_sstable_classifier) {
        _repair_sstable_classifier = std::move(repair_sstable_classifier);
    }

    void add_logstor_segment(logstor::segment_descriptor& desc) {
        _logstor_segments->add_segment(desc);
    }

    future<> discard_logstor_segments();

    future<> flush_separator(std::optional<size_t> seq_num = std::nullopt);
    logstor::separator_buffer& get_separator_buffer(size_t write_size);

    logstor::segment_set& logstor_segments() noexcept {
        return *_logstor_segments;
    }

    const logstor::segment_set& logstor_segments() const noexcept {
        return *_logstor_segments;
    }

    future<utils::chunked_vector<logstor::segment_snapshot>> take_logstor_snapshot();

    friend class storage_group;
};

using compaction_group_ptr = lw_shared_ptr<compaction_group>;
using const_compaction_group_ptr = lw_shared_ptr<const compaction_group>;

// Storage group is responsible for storage that belongs to a single tablet.
// A storage group can manage 1 or more compaction groups, each of which can be compacted independently.
// If a tablet needs splitting, the storage group can be put in splitting mode, allowing the storage
// in main compaction groups to be split into two new compaction groups, all of which will be managed
// by the same storage group.
//
// With vnodes, a table instance in a given shard will have a single group. With tablets, a table in a
// shard will have as many groups as there are tablet replicas owned by that shard.
class storage_group {
    compaction_group_ptr _main_cg;
    // Holds compaction groups that now belongs to same tablet after merge. Compaction groups here will
    // eventually have all their data moved into main group.
    std::vector<compaction_group_ptr> _merging_groups;
    std::vector<compaction_group_ptr> _split_ready_groups;
    seastar::named_gate _async_gate;
private:
    bool splitting_mode() const {
        return !_split_ready_groups.empty();
    }
    size_t to_idx(locator::tablet_range_side) const;
public:
    storage_group(compaction_group_ptr cg);

    seastar::named_gate& async_gate() {
        return _async_gate;
    }

    // Closes storage group without stopping its compaction groups that might be referenced elsewhere.
    future<> close() noexcept {
        return _async_gate.close();
    }

    const dht::token_range& token_range() const noexcept;

    size_t memtable_count() const;

    const compaction_group_ptr& main_compaction_group() const noexcept;
    const std::vector<compaction_group_ptr>& split_ready_compaction_groups() const;
    // Selects the compaction group for the given token. Computes the range side
    // from the token only when in splitting mode. This avoids the cost of computing
    // range side on the hot path when it's not needed.
    compaction_group_ptr& select_compaction_group(dht::token, const locator::tablet_map&) noexcept;
    // Selects the compaction group for an sstable spanning a token range.
    // If the first and last tokens fall on different sides of the split point,
    // the sstable belongs to the main compaction group.
    compaction_group_ptr& select_compaction_group(dht::token first, dht::token last, const locator::tablet_map&) noexcept;

    uint64_t live_disk_space_used() const;

    void for_each_compaction_group(std::function<void(const compaction_group_ptr&)> action) const;
    utils::small_vector<compaction_group_ptr, 3> compaction_groups_immediate();
    utils::small_vector<const_compaction_group_ptr, 3> compaction_groups_immediate() const;

    utils::small_vector<compaction_group_ptr, 3> split_unready_groups() const;
    bool split_unready_groups_are_empty() const;

    void add_merging_group(compaction_group_ptr);
    const std::vector<compaction_group_ptr>& merging_groups() const;
    future<> remove_empty_merging_groups();

    // Puts the storage group in split mode, in which it internally segregates data
    // into two sstable sets and two memtable sets corresponding to the two adjacent
    // tablets post-split.
    // Preexisting sstables and memtables are not split yet.
    // Returns true if post-conditions for split() are met.
    bool set_split_mode();

    // Like set_split_mode() but triggers splitting for old sstables and memtables and waits
    // for it:
    //  1) Flushes all memtables which were created in non-split mode, and waits for that to complete.
    //  2) Compacts all sstables which overlap with the split point
    // Returns a future which resolves when this process is complete.
    future<> split(compaction::compaction_type_options::split opt, tasks::task_info tablet_split_task_info);

    // Make an sstable set spanning all sstables in the storage_group
    lw_shared_ptr<const sstables::sstable_set> make_sstable_set() const;
    // Like make_sstable_set(), but restricted to repaired sstables only across all compaction groups.
    lw_shared_ptr<const sstables::sstable_set> make_repaired_sstable_set() const;

    future<utils::chunked_vector<logstor::segment_snapshot>> take_logstor_snapshot() const;

    // Flush all memtables.
    future<> flush() noexcept;
    future<> flush_separator() noexcept;
    bool can_flush() const;
    bool needs_flush() const;
    api::timestamp_type min_memtable_timestamp() const;
    api::timestamp_type min_memtable_live_timestamp() const;
    api::timestamp_type min_memtable_live_row_marker_timestamp() const;

    bool compaction_disabled() const;
    // Returns true when all compacted sstables were already deleted.
    bool no_compacted_sstable_undeleted() const;

    future<> stop(sstring reason = "table removal") noexcept;

    // Clear sstable sets
    void clear_sstables();
};

using storage_group_ptr = lw_shared_ptr<storage_group>;
using storage_group_map = absl::flat_hash_map<size_t, storage_group_ptr, absl::Hash<size_t>>;

class storage_group_manager {
protected:
    storage_group_map _storage_groups;
protected:
    virtual future<> stop() = 0;
public:
    virtual ~storage_group_manager();

    //    How concurrent loop and updates on the group map works without a lock:
    //
    //    Firstly, all yielding loops will work on a copy of map, to prevent a
    //    concurrent update to the map from interfering with it.
    //
    //    scenario 1:
    //    T
    //    1   loop on the map
    //    2                               storage group X is stopped on cleanup
    //    3   loop reaches X
    //
    //    Here, X is stopped before it is reached. This is handled by teaching
    //    iteration to skip groups that were stopped by cleanup (implemented
    //    using gate).
    //    X survives its removal from the map since it is a lw_shared_ptr.
    //
    //
    //    scenario 2:
    //    T
    //    1   loop on the map
    //    2   loop reaches X
    //    3                               storage group X is stopped on cleanup
    //
    //    Here, X is stopped while being used, but that also happens during shutdown.
    //    When X is stopped, flush happens and compactions are all stopped (exception
    //    is not propagated upwards) and new ones cannot start afterward.
    //
    //
    //    scenario 3:
    //    T
    //    1   loop on the map
    //    2                               storage groups are split
    //    3   loop reaches old groups
    //
    //    Here, the loop continues post storage group split, which rebuilds the old
    //    map into a new one. This is handled by allowing the old map to still access
    //    the compaction groups that were reassigned according to the new tablet count.
    //    We don't move the compaction groups, but rather they're still visible by old
    //    and new storage groups.

    // Important to not return storage_group_id in yielding variants, since ids can be
    // invalidated when storage group count changes (e.g. split or merge).
    future<> parallel_foreach_storage_group(std::function<future<>(storage_group&)> f);
    future<> for_each_storage_group_gently(std::function<future<>(storage_group&)> f);
    void for_each_storage_group(std::function<void(size_t, storage_group&)> f) const;
    const storage_group_map& storage_groups() const;

    future<> stop_storage_groups() noexcept;
    void clear_storage_groups();
    void remove_storage_group(size_t id);
    storage_group& storage_group_for_id(const schema_ptr&, size_t i) const;
    storage_group* maybe_storage_group_for_id(const schema_ptr&, size_t i) const;

    // Caller must keep the current effective_replication_map_ptr valid
    // until the storage_group_manager finishes update_effective_replication_map
    //
    // refresh_mutation_source must be called when there are changes to data source
    // structures but logical state of data is not changed (e.g. when state for a
    // new tablet replica is allocated).
    virtual void update_effective_replication_map(const locator::effective_replication_map_ptr& old_erm,
                                                  const locator::effective_replication_map& erm,
                                                  noncopyable_function<void()> refresh_mutation_source) = 0;

    virtual compaction_group& compaction_group_for_token(dht::token token) const = 0;
    virtual compaction_group& compaction_group_for_key(partition_key_view key, const schema_ptr& s) const = 0;
    virtual compaction_group& compaction_group_for_sstable(const sstables::shared_sstable& sst) const = 0;
    virtual compaction_group& compaction_group_for_logstor_segment(logstor::log_segment_id seg_id, dht::token first_token, dht::token last_token) const = 0;

    virtual storage_group& storage_group_for_token(dht::token) const = 0;
    virtual utils::chunked_vector<storage_group_ptr> storage_groups_for_token_range(dht::token_range tr) const = 0;

    virtual locator::combined_load_stats table_load_stats() const = 0;
    virtual bool all_storage_groups_split() = 0;
    virtual future<> split_all_storage_groups(tasks::task_info tablet_split_task_info) = 0;
    virtual future<> maybe_split_compaction_group_of(size_t idx) = 0;
    virtual future<std::vector<sstables::shared_sstable>> maybe_split_new_sstable(const sstables::shared_sstable& sst) = 0;
    virtual dht::token_range get_token_range_after_split(const dht::token&) const noexcept = 0;
    virtual future<> wait_for_background_tablet_resize_work() = 0;

    virtual lw_shared_ptr<sstables::sstable_set> make_sstable_set() const = 0;
};

}