mito2/region/

opener.rs

// Copyright 2023 Greptime Team
//
// Licensed 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.

//! Region opener.

use std::any::TypeId;
use std::collections::HashMap;
use std::sync::atomic::{AtomicI64, AtomicU64};
use std::sync::{Arc, LazyLock};
use std::time::Instant;

use arrow_schema::extension::ExtensionType;
use common_telemetry::{debug, error, info, warn};
use common_wal::options::WalOptions;
use datatypes::extension::json::{Json2ExtensionType, JsonMetadata};
use futures::StreamExt;
use futures::future::BoxFuture;
use log_store::kafka::log_store::KafkaLogStore;
use log_store::noop::log_store::NoopLogStore;
use log_store::raft_engine::log_store::RaftEngineLogStore;
use object_store::ObjectStore;
use object_store::manager::ObjectStoreManagerRef;
use object_store::util::{is_object_storage, normalize_dir};
use parquet::file::metadata::PageIndexPolicy;
use snafu::{OptionExt, ResultExt, ensure};
use store_api::codec::PrimaryKeyEncoding;
use store_api::logstore::LogStore;
use store_api::logstore::provider::Provider;
use store_api::metadata::{
    ColumnMetadata, RegionMetadata, RegionMetadataBuilder, RegionMetadataRef,
};
use store_api::region_engine::RegionRole;
use store_api::region_request::{PathType, RegionRequirements};
use store_api::storage::{ColumnId, RegionId};
use tokio::sync::Semaphore;

use crate::access_layer::AccessLayer;
use crate::cache::file_cache::{FileCache, FileType, IndexKey};
use crate::cache::{CacheManagerRef, SstMetaPreparation, prepare_sst_meta};
use crate::config::MitoConfig;
use crate::engine::region_hook::RegionHookRef;
use crate::error;
use crate::error::{
    DataTypeMismatchSnafu, EmptyRegionDirSnafu, InvalidMetadataSnafu, InvalidRegionOptionsSnafu,
    ObjectStoreNotFoundSnafu, RegionCorruptedSnafu, Result, StaleLogEntrySnafu,
};
use crate::manifest::action::RegionManifest;
use crate::manifest::manager::{RegionManifestManager, RegionManifestOptions};
use crate::memtable::bulk::part::BulkPart;
use crate::memtable::time_partition::TimePartitions;
use crate::memtable::{MemtableBuilderProvider, ensure_json2_not_use_time_series_memtable};
use crate::metrics::{CACHE_FILL_DOWNLOADED_FILES, CACHE_FILL_PENDING_FILES};
use crate::read::series_candidate::is_sparse_metric_metadata;
use crate::region::options::RegionOptions;
use crate::region::version::{Version, VersionBuilder, VersionControl, VersionControlRef};
use crate::region::{
    ManifestContext, ManifestStats, MitoRegion, MitoRegionRef, RegionLeaderState, RegionRoleState,
    RegionStats,
};
use crate::region_write_ctx::{RegionWriteCtx, WriteSource};
use crate::request::OptionOutputTx;
use crate::schedule::scheduler::SchedulerRef;
use crate::series_index::{IndexFilePurger, load_version_control};
use crate::sst::FormatType;
use crate::sst::file::{FileHandle, RegionFileId, RegionIndexId};
use crate::sst::file_purger::{FilePurgerRef, create_file_purger};
use crate::sst::file_ref::FileReferenceManagerRef;
use crate::sst::index::intermediate::IntermediateManager;
use crate::sst::index::puffin_manager::PuffinManagerFactory;
use crate::sst::location::{self, region_dir_from_table_dir};
use crate::sst::parquet::metadata::{MetadataLoader, extract_primary_key_range};
use crate::sst::parquet::reader::MetadataCacheMetrics;
use crate::sst::range_index::RangeIndexDeleter;
use crate::time_provider::TimeProviderRef;
use crate::wal::entry_reader::WalEntryReader;
use crate::wal::{EntryId, Wal};

const PARQUET_META_PRELOAD_CONCURRENCY: usize = 8;

static PARQUET_META_PRELOAD_SEMAPHORE: LazyLock<Semaphore> =
    LazyLock::new(|| Semaphore::new(PARQUET_META_PRELOAD_CONCURRENCY));

fn initial_pruned_entry_id(wal_options: &WalOptions) -> EntryId {
    match wal_options {
        WalOptions::Kafka(options) => options.initial_pruned_entry_id.unwrap_or(0),
        WalOptions::RaftEngine | WalOptions::Noop | WalOptions::ObjectStore(_) => 0,
    }
}

fn maybe_upgrade_json2_layout(metadata: RegionMetadataRef) -> Result<RegionMetadataRef> {
    let mut upgrades = Vec::new();
    for (index, column) in metadata.column_metadatas.iter().enumerate() {
        if !column.column_schema.data_type.is_json2() {
            continue;
        }
        let Some(extension) = column
            .column_schema
            .extension_type::<Json2ExtensionType>()
            .context(DataTypeMismatchSnafu)?
        else {
            continue;
        };
        if extension.metadata().is_version_2() {
            continue;
        }
        upgrades.push((index, extension.metadata().json_settings().clone()));
    }

    if upgrades.is_empty() {
        return Ok(metadata);
    }

    let mut upgraded = metadata.as_ref().clone();
    for (index, settings) in upgrades {
        let extension = Json2ExtensionType::new(Arc::new(JsonMetadata::new(settings)));
        upgraded.column_metadatas[index]
            .column_schema
            .with_extension_type(&extension);
    }

    let builder = RegionMetadataBuilder::from_existing(upgraded);
    Ok(Arc::new(builder.build().context(InvalidMetadataSnafu)?))
}

/// A fetcher to retrieve partition expr for a region.
///
/// Compatibility: older regions didn't persist `partition_expr` in engine metadata,
/// while newer ones do. On open, we backfill it via this fetcher and persist it
/// to the manifest so future opens don't need refetching.
#[async_trait::async_trait]
pub trait PartitionExprFetcher {
    async fn fetch_expr(&self, region_id: RegionId) -> Option<String>;
}

pub type PartitionExprFetcherRef = Arc<dyn PartitionExprFetcher + Send + Sync>;

/// Builder to create a new [MitoRegion] or open an existing one.
pub(crate) struct RegionOpener {
    region_id: RegionId,
    metadata_builder: Option<RegionMetadataBuilder>,
    memtable_builder_provider: MemtableBuilderProvider,
    object_store_manager: ObjectStoreManagerRef,
    table_dir: String,
    path_type: PathType,
    purge_scheduler: SchedulerRef,
    options: Option<RegionOptions>,
    cache_manager: Option<CacheManagerRef>,
    skip_wal_replay: bool,
    puffin_manager_factory: PuffinManagerFactory,
    intermediate_manager: IntermediateManager,
    time_provider: TimeProviderRef,
    stats: ManifestStats,
    wal_entry_reader: Option<Box<dyn WalEntryReader>>,
    replay_checkpoint: Option<u64>,
    file_ref_manager: FileReferenceManagerRef,
    partition_expr_fetcher: PartitionExprFetcherRef,
    hook: Option<RegionHookRef>,
    series_index_store: Option<ObjectStore>,
    series_index_purger: Option<IndexFilePurger>,
}

impl RegionOpener {
    /// Returns a new opener.
    // TODO(LFC): Reduce the number of arguments.
    #[allow(clippy::too_many_arguments)]
    pub(crate) fn new(
        region_id: RegionId,
        table_dir: &str,
        path_type: PathType,
        memtable_builder_provider: MemtableBuilderProvider,
        object_store_manager: ObjectStoreManagerRef,
        purge_scheduler: SchedulerRef,
        puffin_manager_factory: PuffinManagerFactory,
        intermediate_manager: IntermediateManager,
        time_provider: TimeProviderRef,
        file_ref_manager: FileReferenceManagerRef,
        partition_expr_fetcher: PartitionExprFetcherRef,
    ) -> RegionOpener {
        RegionOpener {
            region_id,
            metadata_builder: None,
            memtable_builder_provider,
            object_store_manager,
            table_dir: normalize_dir(table_dir),
            path_type,
            purge_scheduler,
            options: None,
            cache_manager: None,
            skip_wal_replay: false,
            puffin_manager_factory,
            intermediate_manager,
            time_provider,
            stats: Default::default(),
            wal_entry_reader: None,
            replay_checkpoint: None,
            file_ref_manager,
            partition_expr_fetcher,
            hook: None,
            series_index_store: None,
            series_index_purger: None,
        }
    }

    /// Sets the store for companion range indexes.
    pub(crate) fn series_index_store(mut self, store: Option<ObjectStore>) -> Self {
        self.series_index_store = store;
        self
    }

    /// Sets the purger shared with the worker's series-index maintenance task.
    pub(crate) fn series_index_purger(mut self, purger: Option<IndexFilePurger>) -> Self {
        self.series_index_purger = purger;
        self
    }

    /// Sets the region hook for observing manifest mutations.
    pub(crate) fn hook(mut self, hook: Option<RegionHookRef>) -> Self {
        self.hook = hook;
        self
    }

    /// Sets metadata builder of the region to create.
    pub(crate) fn metadata_builder(mut self, builder: RegionMetadataBuilder) -> Self {
        self.metadata_builder = Some(builder);
        self
    }

    /// Computes the region directory from table_dir and region_id.
    fn region_dir(&self) -> String {
        region_dir_from_table_dir(&self.table_dir, self.region_id, self.path_type)
    }

    /// Builds the region metadata.
    ///
    /// # Panics
    /// - Panics if `options` is not set.
    /// - Panics if `metadata_builder` is not set.
    fn build_metadata(&mut self) -> Result<RegionMetadata> {
        let options = self.options.as_ref().unwrap();
        let mut metadata_builder = self.metadata_builder.take().unwrap();
        metadata_builder.primary_key_encoding(options.primary_key_encoding());
        metadata_builder.build().context(InvalidMetadataSnafu)
    }

    /// Parses and sets options for the region.
    pub(crate) fn parse_options(self, options: HashMap<String, String>) -> Result<Self> {
        let region_id = self.region_id;
        let options = self
            .memtable_builder_provider
            .parse_options(region_id, &options)?;
        self.options(options)
    }

    /// Sets the replay checkpoint for the region.
    pub(crate) fn replay_checkpoint(mut self, replay_checkpoint: Option<u64>) -> Self {
        self.replay_checkpoint = replay_checkpoint;
        self
    }

    /// If a [WalEntryReader] is set, the [RegionOpener] will use [WalEntryReader] instead of
    /// constructing a new one from scratch.
    pub(crate) fn wal_entry_reader(
        mut self,
        wal_entry_reader: Option<Box<dyn WalEntryReader>>,
    ) -> Self {
        self.wal_entry_reader = wal_entry_reader;
        self
    }

    /// Sets options for the region.
    pub(crate) fn options(mut self, options: RegionOptions) -> Result<Self> {
        options.validate()?;
        self.options = Some(options);
        Ok(self)
    }

    /// Ensures the current region request satisfies its requirements.
    pub(crate) fn ensure_region_requirements(
        &self,
        requirements: RegionRequirements,
    ) -> Result<()> {
        if !requirements.object_storage {
            return Ok(());
        }

        let options = self.options.as_ref().context(InvalidRegionOptionsSnafu {
            reason: "missing region options before requirement check".to_string(),
        })?;
        let object_store = get_object_store(&options.storage, &self.object_store_manager)?;

        ensure!(
            supports_open_region_object_storage_requirement(&object_store),
            error::RegionRequirementSnafu {
                region_id: self.region_id,
                requirement: "object storage",
                reason: "region data must be accessible from another datanode",
            }
        );

        Ok(())
    }

    /// Sets the cache manager for the region.
    pub(crate) fn cache(mut self, cache_manager: Option<CacheManagerRef>) -> Self {
        self.cache_manager = cache_manager;
        self
    }

    /// Sets the `skip_wal_replay`.
    pub(crate) fn skip_wal_replay(mut self, skip: bool) -> Self {
        self.skip_wal_replay = skip;
        self
    }

    /// Writes region manifest and creates a new region if it does not exist.
    /// Opens the region if it already exists.
    ///
    /// # Panics
    /// - Panics if `metadata_builder` is not set.
    /// - Panics if `options` is not set.
    pub(crate) async fn create_or_open<S: LogStore>(
        mut self,
        config: &MitoConfig,
        wal: &Wal<S>,
    ) -> Result<MitoRegionRef> {
        let region_id = self.region_id;
        let region_dir = self.region_dir();
        let metadata = self.build_metadata()?;
        // Tries to open the region.
        match self.maybe_open(config, wal).await {
            Ok(Some(region)) => {
                let recovered = region.metadata();
                // Checks the schema of the region.
                let expect = &metadata;
                check_recovered_region(
                    &recovered,
                    expect.region_id,
                    &expect.column_metadatas,
                    &expect.primary_key,
                )?;
                // To keep consistency with Create behavior, set the opened Region to RegionRole::Leader.
                region.set_role(RegionRole::Leader);

                return Ok(region);
            }
            Ok(None) => {
                debug!(
                    "No data under directory {}, region_id: {}",
                    region_dir, self.region_id
                );
            }
            // The region exists but the WAL does not hold its history:
            // creating it again would replace its manifest.
            Err(e @ error::Error::CheckWalHistory { .. }) => return Err(e),
            Err(e) => {
                warn!(e;
                    "Failed to open region {} before creating it, region_dir: {}",
                    self.region_id, region_dir
                );
            }
        }
        // Safety: must be set before calling this method.
        let mut options = self.options.take().unwrap();
        let object_store = get_object_store(&options.storage, &self.object_store_manager)?;
        let provider = self.provider::<S>(&options.wal_options)?;
        let metadata = Arc::new(metadata);
        // Sets the sst_format based on options or flat_format flag
        let sst_format = if let Some(format) = options.sst_format {
            format
        } else if config.default_flat_format {
            options.sst_format = Some(FormatType::Flat);
            FormatType::Flat
        } else {
            // Default to PrimaryKeyParquet for newly created regions
            options.sst_format = Some(FormatType::PrimaryKey);
            FormatType::PrimaryKey
        };
        ensure_json2_not_use_time_series_memtable(&metadata, &options)?;
        options.validate_column_compression(&metadata)?;
        // The region's WAL history starts in this log store, so reopening
        // the region on it passes the history check without any entry. It is
        // recorded before the manifest exists: a log store that gave up its
        // history meanwhile fails the create before it leaves a manifest.
        wal.start_region_history(region_id, &provider)?;
        // Create a manifest manager for this region and writes regions to the manifest file.
        let mut region_manifest_options =
            RegionManifestOptions::new(config, &region_dir, &object_store);
        // Set manifest cache if available
        region_manifest_options.manifest_cache = self
            .cache_manager
            .as_ref()
            .and_then(|cm| cm.write_cache())
            .and_then(|wc| wc.manifest_cache());
        // For remote WAL, we need to set flushed_entry_id to current topic's latest entry id.
        // Kafka WAL allocation also carries the topic's pruned entry id as a create-time hint.
        let flushed_entry_id = provider
            .initial_flushed_entry_id::<S>(wal.store())
            .max(initial_pruned_entry_id(&options.wal_options));
        let manifest_manager = RegionManifestManager::new(
            metadata.clone(),
            flushed_entry_id,
            region_manifest_options,
            sst_format,
            &self.stats,
        )
        .await?;

        let memtable_builder = self.memtable_builder_provider.builder_for_options(&options);
        let part_duration = options.compaction.time_window();
        // Initial memtable id is 0.
        let mutable =
            TimePartitions::new(metadata.clone(), memtable_builder.clone(), 0, part_duration);

        debug!(
            "Create region {} with options: {:?}, default_flat_format: {}",
            region_id, options, config.default_flat_format
        );

        let version = VersionBuilder::new(metadata, mutable)
            .options(options)
            .flushed_entry_id(flushed_entry_id)
            .build();
        let version_control = Arc::new(VersionControl::new(version));
        let access_layer = Arc::new(AccessLayer::new(
            self.table_dir.clone(),
            self.path_type,
            object_store,
            self.puffin_manager_factory,
            self.intermediate_manager,
        ));
        let now = self.time_provider.current_time_millis();

        Ok(Arc::new(MitoRegion {
            region_id,
            version_control,
            series_index_version_control: Default::default(),
            series_index_store: self.series_index_store.clone(),
            access_layer: access_layer.clone(),
            // Region is writable after it is created.
            manifest_ctx: Arc::new(ManifestContext::new(
                manifest_manager,
                RegionRoleState::Leader(RegionLeaderState::Writable),
                self.hook.clone(),
            )),
            file_purger: create_file_purger(
                config.gc.enable,
                self.path_type,
                self.purge_scheduler,
                access_layer,
                self.cache_manager,
                self.file_ref_manager.clone(),
                self.series_index_store
                    .map(|store| RangeIndexDeleter::new(store, region_id)),
            ),
            provider,
            last_flush_millis: AtomicI64::new(now),
            last_schedule_compaction_millis: AtomicI64::new(now),
            time_provider: self.time_provider.clone(),
            topic_latest_entry_id: AtomicU64::new(flushed_entry_id),
            region_stats: RegionStats::new(),
            stats: self.stats,
        }))
    }

    /// Opens an existing region in read only mode.
    ///
    /// Returns error if the region doesn't exist.
    pub(crate) async fn open<S: LogStore>(
        mut self,
        config: &MitoConfig,
        wal: &Wal<S>,
    ) -> Result<MitoRegionRef> {
        let region_id = self.region_id;
        let region_dir = self.region_dir();
        let region = self
            .maybe_open(config, wal)
            .await?
            .with_context(|| EmptyRegionDirSnafu {
                region_id,
                region_dir: &region_dir,
            })?;

        ensure!(
            region.region_id == self.region_id,
            RegionCorruptedSnafu {
                region_id: self.region_id,
                reason: format!(
                    "recovered region has different region id {}",
                    region.region_id
                ),
            }
        );

        Ok(region)
    }

    fn provider<S: LogStore>(&self, wal_options: &WalOptions) -> Result<Provider> {
        provider_from_wal_options::<S>(self.region_id, wal_options)
    }

    /// Tries to open the region and returns `None` if the region directory is empty.
    /// A region it replays must find its WAL history in the log store, see
    /// [`check_wal_history`].
    async fn maybe_open<S: LogStore>(
        &mut self,
        config: &MitoConfig,
        wal: &Wal<S>,
    ) -> Result<Option<MitoRegionRef>> {
        let now = Instant::now();
        let mut region_options = self.options.as_ref().unwrap().clone();
        let object_storage = get_object_store(&region_options.storage, &self.object_store_manager)?;
        let mut region_manifest_options =
            RegionManifestOptions::new(config, &self.region_dir(), &object_storage);
        // Set manifest cache if available
        region_manifest_options.manifest_cache = self
            .cache_manager
            .as_ref()
            .and_then(|cm| cm.write_cache())
            .and_then(|wc| wc.manifest_cache());
        let Some(manifest_manager) =
            RegionManifestManager::open(region_manifest_options, &self.stats).await?
        else {
            return Ok(None);
        };

        // Backfill `partition_expr` if missing. Use the backfilled metadata in-memory during this open.
        let manifest = manifest_manager.manifest();
        let metadata = if manifest.metadata.partition_expr.is_none()
            && let Some(expr_json) = self.partition_expr_fetcher.fetch_expr(self.region_id).await
        {
            let metadata = manifest.metadata.as_ref().clone();
            let mut builder = RegionMetadataBuilder::from_existing(metadata);
            builder.partition_expr_json(Some(expr_json));
            Arc::new(builder.build().context(InvalidMetadataSnafu)?)
        } else {
            manifest.metadata.clone()
        };
        let metadata = maybe_upgrade_json2_layout(metadata)?;
        // Updates the region options with the manifest.
        sanitize_region_options(&manifest, &mut region_options);
        ensure_json2_not_use_time_series_memtable(&metadata, &region_options)?;
        region_options.validate_column_compression(&metadata)?;

        let region_id = self.region_id;
        let provider = self.provider::<S>(&region_options.wal_options)?;
        let wal_entry_reader = self
            .wal_entry_reader
            .take()
            .unwrap_or_else(|| wal.wal_entry_reader(&provider, region_id));
        let on_region_opened = wal.on_region_opened();
        let object_store = get_object_store(&region_options.storage, &self.object_store_manager)?;

        debug!(
            "Open region {} at {} with options: {:?}",
            region_id, self.table_dir, self.options
        );

        let access_layer = Arc::new(AccessLayer::new(
            self.table_dir.clone(),
            self.path_type,
            object_store,
            self.puffin_manager_factory.clone(),
            self.intermediate_manager.clone(),
        ));
        let file_purger = create_file_purger(
            config.gc.enable,
            self.path_type,
            self.purge_scheduler.clone(),
            access_layer.clone(),
            self.cache_manager.clone(),
            self.file_ref_manager.clone(),
            self.series_index_store
                .clone()
                .map(|store| RangeIndexDeleter::new(store, region_id)),
        );
        // We should sanitize the region options before creating a new memtable.
        let memtable_builder = self
            .memtable_builder_provider
            .builder_for_options(&region_options);
        // Use compaction time window in the manifest if region doesn't provide
        // the time window option.
        let part_duration = region_options
            .compaction
            .time_window()
            .or(manifest.compaction_time_window);
        // Initial memtable id is 0.
        let mutable =
            TimePartitions::new(metadata.clone(), memtable_builder.clone(), 0, part_duration);

        // Updates region options by manifest before creating version.
        let version_builder = version_builder_from_manifest(
            &manifest,
            metadata,
            file_purger.clone(),
            mutable,
            region_options,
        );
        let version = version_builder.build();
        let flushed_entry_id = version.flushed_entry_id;
        let version_control = Arc::new(VersionControl::new(version));

        let replay_from_entry_id = self
            .replay_checkpoint
            .unwrap_or_default()
            .max(flushed_entry_id);
        let topic_latest_entry_id = if !self.skip_wal_replay {
            check_wal_history(wal, &provider, &version_control.current().version).await?;
            info!(
                "Start replaying memtable at replay_from_entry_id: {} for region {}, manifest version: {}, flushed entry id: {}, elapsed: {:?}",
                replay_from_entry_id,
                region_id,
                manifest.manifest_version,
                flushed_entry_id,
                now.elapsed()
            );
            replay_memtable(
                &provider,
                wal_entry_reader,
                region_id,
                replay_from_entry_id,
                &version_control,
                config.allow_stale_entries,
                on_region_opened,
            )
            .await?;
            // For remote WAL, we need to set topic_latest_entry_id to current topic's latest entry id.
            // Only set after the WAL replay is completed.

            if provider.is_remote_wal() && version_control.current().version.memtables.is_empty() {
                wal.store()
                    .latest_entry_id(&provider)
                    .unwrap_or(replay_from_entry_id)
            } else if provider.is_remote_wal() {
                replay_from_entry_id
            } else {
                0
            }
        } else {
            info!(
                "Skip the WAL replay for region: {}, manifest version: {}, flushed_entry_id: {}, elapsed: {:?}",
                region_id,
                manifest.manifest_version,
                flushed_entry_id,
                now.elapsed()
            );

            if provider.is_remote_wal() {
                replay_from_entry_id
            } else {
                0
            }
        };

        if let Some(committed_in_manifest) = manifest.committed_sequence {
            let committed_after_replay = version_control.committed_sequence();
            if committed_in_manifest > committed_after_replay {
                info!(
                    "Overriding committed sequence, region: {}, flushed_sequence: {}, committed_sequence: {} -> {}",
                    self.region_id,
                    version_control.current().version.flushed_sequence,
                    version_control.committed_sequence(),
                    committed_in_manifest
                );
                version_control.set_committed_sequence(committed_in_manifest);
            }
        }

        let now = self.time_provider.current_time_millis();

        let series_index_version_control =
            match (&self.series_index_store, &self.series_index_purger) {
                (Some(store), Some(purger))
                    if is_sparse_metric_metadata(&version_control.current().version.metadata) =>
                {
                    load_version_control(store, self.region_id, purger).await
                }
                _ => Default::default(),
            };

        let region = MitoRegion {
            region_id: self.region_id,
            version_control: version_control.clone(),
            series_index_version_control,
            series_index_store: self.series_index_store.clone(),
            access_layer: access_layer.clone(),
            // Region is always opened in read only mode.
            manifest_ctx: Arc::new(ManifestContext::new(
                manifest_manager,
                RegionRoleState::Follower,
                self.hook.clone(),
            )),
            file_purger,
            provider: provider.clone(),
            last_flush_millis: AtomicI64::new(now),
            last_schedule_compaction_millis: AtomicI64::new(now),
            time_provider: self.time_provider.clone(),
            topic_latest_entry_id: AtomicU64::new(topic_latest_entry_id),
            region_stats: RegionStats::new(),
            stats: self.stats.clone(),
        };

        let region = Arc::new(region);

        maybe_load_cache(&region, config, &self.cache_manager);
        maybe_preload_parquet_meta_cache(&region, config, &self.cache_manager);

        Ok(Some(region))
    }
}

pub(crate) fn provider_from_wal_options<S: LogStore>(
    region_id: RegionId,
    wal_options: &WalOptions,
) -> Result<Provider> {
    match wal_options {
        WalOptions::RaftEngine => {
            ensure!(
                TypeId::of::<RaftEngineLogStore>() == TypeId::of::<S>()
                    || TypeId::of::<NoopLogStore>() == TypeId::of::<S>(),
                error::IncompatibleWalProviderChangeSnafu {
                    global: "`kafka`",
                    region: "`raft_engine`",
                }
            );
            Ok(Provider::raft_engine_provider(region_id.as_u64()))
        }
        WalOptions::Kafka(options) => {
            ensure!(
                TypeId::of::<KafkaLogStore>() == TypeId::of::<S>()
                    || TypeId::of::<NoopLogStore>() == TypeId::of::<S>(),
                error::IncompatibleWalProviderChangeSnafu {
                    global: "`raft_engine`",
                    region: "`kafka`",
                }
            );
            Ok(Provider::kafka_provider(options.topic.clone()))
        }
        WalOptions::ObjectStore(options) => {
            ensure!(
                TypeId::of::<RaftEngineLogStore>() != TypeId::of::<S>(),
                error::IncompatibleWalProviderChangeSnafu {
                    global: "`raft_engine`",
                    region: "`object_store`",
                }
            );
            ensure!(
                TypeId::of::<KafkaLogStore>() != TypeId::of::<S>(),
                error::IncompatibleWalProviderChangeSnafu {
                    global: "`kafka`",
                    region: "`object_store`",
                }
            );
            Ok(Provider::object_store_provider(
                region_id,
                options.prefix.clone(),
            ))
        }
        WalOptions::Noop => Ok(Provider::noop_provider()),
    }
}

#[cfg(not(feature = "test-shared-fs-region-migration"))]
fn supports_open_region_object_storage_requirement(object_store: &ObjectStore) -> bool {
    is_object_storage(object_store)
}

#[cfg(feature = "test-shared-fs-region-migration")]
fn supports_open_region_object_storage_requirement(object_store: &ObjectStore) -> bool {
    // Integration tests can configure multiple datanodes to share the same
    // temporary home dir. That makes file storage accessible to all test
    // datanodes, but production file storage still does not satisfy this
    // requirement.
    is_object_storage(object_store)
        || object_store.info().scheme() == object_store::services::FS_SCHEME
}

/// Creates a version builder from a region manifest.
pub(crate) fn version_builder_from_manifest(
    manifest: &RegionManifest,
    metadata: RegionMetadataRef,
    file_purger: FilePurgerRef,
    mutable: TimePartitions,
    region_options: RegionOptions,
) -> VersionBuilder {
    VersionBuilder::new(metadata, mutable)
        .add_files(file_purger, manifest.files.values().cloned())
        .flushed_entry_id(manifest.flushed_entry_id)
        .flushed_sequence(manifest.flushed_sequence)
        .truncated_entry_id(manifest.truncated_entry_id)
        .compaction_time_window(manifest.compaction_time_window)
        .options(region_options)
}

/// Updates region options by persistent options.
pub(crate) fn sanitize_region_options(manifest: &RegionManifest, options: &mut RegionOptions) {
    // The caller-supplied options win when they specify an explicit `sst_format`,
    // so re-parsed options (e.g. bulk memtable forcing flat format) take effect on
    // the running region instead of being clobbered by the persisted manifest value.
    // Fall back to the manifest only when the caller did not provide a value.
    match options.sst_format {
        Some(format) if format != manifest.sst_format => {
            common_telemetry::warn!(
                "Overriding SST format from {:?} (manifest) to {:?} (options) for region {}",
                manifest.sst_format,
                format,
                manifest.metadata.region_id,
            );
        }
        Some(_) => {}
        None => {
            options.sst_format = Some(manifest.sst_format);
        }
    }
    if let Some(manifest_append_mode) = manifest.append_mode
        && options.append_mode != manifest_append_mode
    {
        common_telemetry::warn!(
            "Overriding append_mode from {} to {} for region {}",
            options.append_mode,
            manifest_append_mode,
            manifest.metadata.region_id,
        );
        options.append_mode = manifest_append_mode;
    }
    if options.append_mode && options.merge_mode.take().is_some() {
        common_telemetry::warn!(
            "Ignoring merge_mode because append_mode is enabled for region {}",
            manifest.metadata.region_id,
        );
    }
}

/// Sanitizes open request options before parsing.
pub(crate) fn sanitize_open_request_options(options: &mut HashMap<String, String>) {
    let append_mode_enabled = options
        .get("append_mode")
        .is_some_and(|v| matches!(v.trim().to_ascii_lowercase().as_str(), "true" | "1"));

    if append_mode_enabled && options.remove("merge_mode").is_some() {
        common_telemetry::warn!(
            "Ignoring merge_mode in open request options because append_mode is enabled"
        );
    }
}

/// Returns an object store corresponding to `name`. If `name` is `None`, this method returns the default object store.
pub fn get_object_store(
    name: &Option<String>,
    object_store_manager: &ObjectStoreManagerRef,
) -> Result<object_store::ObjectStore> {
    if let Some(name) = name {
        Ok(object_store_manager
            .find(name)
            .with_context(|| ObjectStoreNotFoundSnafu {
                object_store: name.clone(),
            })?
            .clone())
    } else {
        Ok(object_store_manager.default_object_store().clone())
    }
}

/// Checks whether the recovered region has the same schema as region to create.
pub(crate) fn check_recovered_region(
    recovered: &RegionMetadata,
    region_id: RegionId,
    column_metadatas: &[ColumnMetadata],
    primary_key: &[ColumnId],
) -> Result<()> {
    if recovered.region_id != region_id {
        error!(
            "Recovered region {}, expect region {}",
            recovered.region_id, region_id
        );
        return RegionCorruptedSnafu {
            region_id,
            reason: format!(
                "recovered metadata has different region id {}",
                recovered.region_id
            ),
        }
        .fail();
    }
    if recovered.column_metadatas != column_metadatas {
        error!(
            "Unexpected schema in recovered region {}, recovered: {:?}, expect: {:?}",
            recovered.region_id, recovered.column_metadatas, column_metadatas
        );

        return RegionCorruptedSnafu {
            region_id,
            reason: "recovered metadata has different schema",
        }
        .fail();
    }
    if recovered.primary_key != primary_key {
        error!(
            "Unexpected primary key in recovered region {}, recovered: {:?}, expect: {:?}",
            recovered.region_id, recovered.primary_key, primary_key
        );

        return RegionCorruptedSnafu {
            region_id,
            reason: "recovered metadata has different primary key",
        }
        .fail();
    }

    Ok(())
}

/// Checks that the log store holds the WAL history of a region that writes
/// to the WAL through the largest replay watermark `version` persisted,
/// before the region is replayed. A region that skips the WAL keeps no
/// history there.
pub(crate) async fn check_wal_history<S: LogStore>(
    wal: &Wal<S>,
    provider: &Provider,
    version: &Version,
) -> Result<()> {
    if version.options.skip_wal {
        return Ok(());
    }
    let watermark = version
        .flushed_entry_id
        .max(version.truncated_entry_id.unwrap_or_default());
    wal.check_region_history(version.metadata.region_id, provider, watermark)
        .await
}

/// Replays the mutations from WAL and inserts mutations to memtable of given region.
pub(crate) async fn replay_memtable<F>(
    provider: &Provider,
    mut wal_entry_reader: Box<dyn WalEntryReader>,
    region_id: RegionId,
    flushed_entry_id: EntryId,
    version_control: &VersionControlRef,
    allow_stale_entries: bool,
    on_region_opened: F,
) -> Result<EntryId>
where
    F: FnOnce(RegionId, EntryId, &Provider) -> BoxFuture<Result<()>> + Send,
{
    let now = Instant::now();
    let mut rows_replayed = 0;
    // Last entry id should start from flushed entry id since there might be no
    // data in the WAL.
    let mut last_entry_id = flushed_entry_id;
    let replay_from_entry_id = flushed_entry_id + 1;
    let region_metadata = version_control.current().version.metadata.clone();

    let mut wal_stream = wal_entry_reader.read(provider, replay_from_entry_id)?;
    while let Some(res) = wal_stream.next().await {
        let (entry_id, entry) = res?;
        if entry_id <= flushed_entry_id {
            warn!(
                "Stale WAL entries read during replay, region id: {}, flushed entry id: {}, entry id read: {}",
                region_id, flushed_entry_id, entry_id
            );
            ensure!(
                allow_stale_entries,
                StaleLogEntrySnafu {
                    region_id,
                    flushed_entry_id,
                    unexpected_entry_id: entry_id,
                }
            );
        }
        last_entry_id = last_entry_id.max(entry_id);

        let mut region_write_ctx = RegionWriteCtx::new(
            region_id,
            version_control,
            provider.clone(),
            // For WAL replay, we don't need to track the write bytes rate.
            None,
            WriteSource::WalReplay,
        );
        for mutation in entry.mutations {
            rows_replayed += mutation
                .rows
                .as_ref()
                .map(|rows| rows.rows.len())
                .unwrap_or(0);
            region_write_ctx.push_mutation(
                mutation.op_type,
                mutation.rows,
                mutation.write_hint,
                OptionOutputTx::none(),
                // We should respect the sequence in WAL during replay.
                Some(mutation.sequence),
                false,
            );
        }

        for bulk_entry in entry.bulk_entries {
            let mut part = BulkPart::try_from(bulk_entry)?;
            // The entry may miss columns added by a concurrent alter if it was
            // written by a writer with a stale schema (older versions kept the
            // stale raw data when filling missing columns). Fills missing
            // columns like the write path, otherwise the memtable rejects the
            // batch. Skips sparse batches as they don't carry tag columns by
            // design.
            if region_metadata.primary_key_encoding != PrimaryKeyEncoding::Sparse {
                part.fill_missing_columns(&region_metadata)?;
            }
            rows_replayed += part.num_rows();
            // During replay, we should adopt the sequence from WAL.
            let bulk_sequence_from_wal = part.sequence;
            ensure!(
                region_write_ctx.push_bulk(
                    OptionOutputTx::none(),
                    part,
                    Some(bulk_sequence_from_wal),
                    false
                ),
                RegionCorruptedSnafu {
                    region_id,
                    reason: "unable to replay memtable with bulk entries",
                }
            );
        }

        // set next_entry_id and write to memtable.
        region_write_ctx.set_next_entry_id(last_entry_id + 1);
        // Publish the replayed sequences only after all rows (including bulk
        // parts) are installed, matching the write path ordering.
        region_write_ctx.write_memtables().await;
    }

    // TODO(weny): We need to update `flushed_entry_id` in the region manifest
    // to avoid reading potentially incomplete entries in the future.
    (on_region_opened)(region_id, flushed_entry_id, provider).await?;

    let series_count = version_control.current().series_count();
    info!(
        "Replay WAL for region: {}, provider: {:?}, rows recovered: {}, replay from entry id: {}, last entry id: {}, total timeseries replayed: {}, elapsed: {:?}",
        region_id,
        provider,
        rows_replayed,
        replay_from_entry_id,
        last_entry_id,
        series_count,
        now.elapsed()
    );
    Ok(last_entry_id)
}

/// A task to load and fill the region file cache.
pub(crate) struct RegionLoadCacheTask {
    region: MitoRegionRef,
}

impl RegionLoadCacheTask {
    pub(crate) fn new(region: MitoRegionRef) -> Self {
        Self { region }
    }

    /// Fills the file cache with index files from the region.
    pub(crate) async fn fill_cache(&self, file_cache: &FileCache) {
        let region_id = self.region.region_id;
        let table_dir = self.region.access_layer.table_dir();
        let path_type = self.region.access_layer.path_type();
        let object_store = self.region.access_layer.object_store();
        let version_control = &self.region.version_control;

        // Collects IndexKeys, file sizes, and max timestamps for files that need to be downloaded
        let mut files_to_download = Vec::new();
        let mut files_already_cached = 0;

        {
            let version = version_control.current().version;
            for level in version.ssts.levels() {
                for file_handle in level.files.values() {
                    let file_meta = file_handle.meta_ref();
                    if file_meta.exists_index() {
                        let puffin_key = IndexKey::new(
                            file_meta.region_id,
                            file_meta.file_id,
                            FileType::Puffin(file_meta.index_version),
                        );

                        if !file_cache.contains_key(&puffin_key) {
                            files_to_download.push((
                                puffin_key,
                                file_meta.index_file_size,
                                file_meta.time_range.1, // max timestamp
                            ));
                        } else {
                            files_already_cached += 1;
                        }
                    }
                }
            }
            // Releases the Version after the scope to avoid holding the memtables and file handles
            // for a long time.
        }

        // Sorts files by max timestamp in descending order to loads latest files first
        files_to_download.sort_by_key(|b| std::cmp::Reverse(b.2));

        let total_files = files_to_download.len() as i64;

        info!(
            "Starting background index cache preload for region {}, total_files_to_download: {}, files_already_cached: {}",
            region_id, total_files, files_already_cached
        );

        CACHE_FILL_PENDING_FILES.add(total_files);

        let mut files_downloaded = 0;
        let mut files_skipped = 0;

        for (puffin_key, file_size, max_timestamp) in files_to_download {
            let current_size = file_cache.puffin_cache_size();
            let capacity = file_cache.puffin_cache_capacity();
            let region_state = self.region.state();
            if !can_load_cache(region_state) {
                info!(
                    "Stopping index cache by state: {:?}, region: {}, current_size: {}, capacity: {}",
                    region_state, region_id, current_size, capacity
                );
                break;
            }

            // Checks if adding this file would exceed capacity
            if current_size + file_size > capacity {
                info!(
                    "Stopping index cache preload due to capacity limit, region: {}, file_id: {}, current_size: {}, file_size: {}, capacity: {}, file_timestamp: {:?}",
                    region_id, puffin_key.file_id, current_size, file_size, capacity, max_timestamp
                );
                files_skipped = (total_files - files_downloaded) as usize;
                CACHE_FILL_PENDING_FILES.sub(total_files - files_downloaded);
                break;
            }

            let index_version = if let FileType::Puffin(version) = puffin_key.file_type {
                version
            } else {
                unreachable!("`files_to_download` should only contains Puffin files");
            };
            let index_id = RegionIndexId::new(
                RegionFileId::new(puffin_key.region_id, puffin_key.file_id),
                index_version,
            );

            let index_remote_path = location::index_file_path(table_dir, index_id, path_type);

            match file_cache
                .download(puffin_key, &index_remote_path, object_store, file_size)
                .await
            {
                Ok(_) => {
                    debug!(
                        "Downloaded index file to write cache, region: {}, file_id: {}",
                        region_id, puffin_key.file_id
                    );
                    files_downloaded += 1;
                    CACHE_FILL_DOWNLOADED_FILES.inc_by(1);
                    CACHE_FILL_PENDING_FILES.dec();
                }
                Err(e) => {
                    warn!(
                        e; "Failed to download index file to write cache, region: {}, file_id: {}",
                        region_id, puffin_key.file_id
                    );
                    CACHE_FILL_PENDING_FILES.dec();
                }
            }
        }

        info!(
            "Completed background cache fill task for region {}, total_files: {}, files_downloaded: {}, files_already_cached: {}, files_skipped: {}",
            region_id, total_files, files_downloaded, files_already_cached, files_skipped
        );
    }
}

/// Loads all index (Puffin) files from the version into the write cache.
fn maybe_load_cache(
    region: &MitoRegionRef,
    config: &MitoConfig,
    cache_manager: &Option<CacheManagerRef>,
) {
    let Some(cache_manager) = cache_manager else {
        return;
    };
    let Some(write_cache) = cache_manager.write_cache() else {
        return;
    };

    let preload_enabled = config.preload_index_cache;
    if !preload_enabled {
        return;
    }

    let task = RegionLoadCacheTask::new(region.clone());
    write_cache.load_region_cache(task);
}

/// Preloads Parquet metadata into the in-memory SST meta cache on region open.
///
/// This improves the latency of the first query after server start by avoiding large Parquet
/// metadata reads on demand.
///
/// The preload is best-effort:
/// - Always tries to warm from the local write cache (file cache) first.
/// - If the region storage backend is local filesystem (`Scheme::Fs`), it may also load metadata
///   directly from the local store.
/// - It will not fetch metadata from remote object stores (S3/GCS/OSS/...).
#[allow(clippy::too_many_arguments)]
async fn preload_parquet_meta_cache_for_files(
    region_id: RegionId,
    cache_manager: CacheManagerRef,
    sst_meta_cache_capacity: u64,
    table_dir: String,
    path_type: PathType,
    object_store: object_store::ObjectStore,
    region_metadata: RegionMetadataRef,
    mut files: Vec<FileHandle>,
) -> usize {
    if !cache_manager.sst_meta_cache_enabled()
        || sst_meta_cache_capacity == 0
        || cache_manager.sst_meta_cache_is_full()
    {
        return 0;
    }

    let allow_direct_load = object_store.info().scheme() == object_store::services::FS_SCHEME;

    // Sort by time range so we can prefer preloading newer files first.
    files.sort_by_key(|b| std::cmp::Reverse(b.meta_ref().time_range.1));

    let mut loaded = 0usize;
    for file_handle in files {
        // Stop when the protected authoritative tier is full.
        if cache_manager.sst_meta_cache_is_full() {
            break;
        }

        let file_id = file_handle.file_id();
        let mut cache_metrics = MetadataCacheMetrics::default();
        if let Some(metadata) = cache_manager
            .get_compact_sst_meta_data(file_id, PageIndexPolicy::Optional)
            .await
        {
            if file_handle.raw_primary_key_range().is_none()
                && let Some(primary_key_range) = extract_primary_key_range(
                    metadata.parquet_metadata().as_ref(),
                    &region_metadata,
                )
            {
                file_handle.set_primary_key_range(primary_key_range);
            }
            continue;
        }

        if let Some(write_cache) = cache_manager.write_cache() {
            let key = IndexKey::new(file_id.region_id(), file_id.file_id(), FileType::Parquet);
            if let Some(metadata) = write_cache
                .file_cache()
                .get_sst_meta_data(
                    key,
                    &mut cache_metrics,
                    PageIndexPolicy::Optional,
                    &common_runtime::global_runtime(),
                )
                .await
            {
                let decoded = metadata.decoded();
                if file_handle.raw_primary_key_range().is_none()
                    && let Some(primary_key_range) = extract_primary_key_range(
                        decoded.parquet_metadata().as_ref(),
                        &region_metadata,
                    )
                {
                    file_handle.set_primary_key_range(primary_key_range);
                }
                match metadata {
                    SstMetaPreparation::Prepared(metadata) => {
                        cache_manager.put_prepared_sst_meta(file_id, metadata, false);
                        loaded += 1;
                    }
                    SstMetaPreparation::DecodedOnly { encoding_error, .. } => warn!(
                        encoding_error;
                        "Failed to encode file-cached SST metadata during preload, region: {}, file: {}",
                        region_id,
                        file_id.file_id()
                    ),
                }
                continue;
            }
        }

        if !allow_direct_load {
            continue;
        }

        let file_size = file_handle.meta_ref().file_size;
        let file_path = file_handle.file_path(&table_dir, path_type);
        let mut loader = MetadataLoader::new(object_store.clone(), &file_path, file_size);
        loader.with_page_index_policy(PageIndexPolicy::Optional);
        match loader.load(&mut cache_metrics).await {
            Ok(metadata) => {
                if let Some(primary_key_range) =
                    extract_primary_key_range(&metadata, &region_metadata)
                {
                    file_handle.set_primary_key_range(primary_key_range);
                }
                match prepare_sst_meta(
                    &file_path,
                    metadata,
                    // The SST can predate schema changes, so decode its embedded region metadata
                    // instead of substituting the region's current schema.
                    None,
                    PageIndexPolicy::Optional,
                    &common_runtime::global_runtime(),
                )
                .await
                {
                    Ok(SstMetaPreparation::Prepared(metadata)) => {
                        // Preload the compact canonical entry only. Query traffic promotes decoded
                        // entries into the separate acceleration tier according to actual demand.
                        cache_manager.put_prepared_sst_meta(file_id, metadata, false);
                        loaded += 1;
                    }
                    Ok(SstMetaPreparation::DecodedOnly { encoding_error, .. }) => warn!(
                        encoding_error;
                        "Failed to encode preloaded parquet metadata, region: {}, file: {}",
                        region_id,
                        file_path
                    ),
                    Err(err) => {
                        warn!(
                            err; "Failed to prepare preloaded parquet metadata, region: {}, file: {}",
                            region_id, file_path
                        );
                    }
                }
            }
            Err(err) => {
                // Preloading is best-effort. Failure shouldn't affect region open.
                warn!(
                    err; "Failed to preload parquet metadata from local store, region: {}, file: {}",
                    region_id, file_path
                );
            }
        }
    }

    loaded
}

fn maybe_preload_parquet_meta_cache(
    region: &MitoRegionRef,
    config: &MitoConfig,
    cache_manager: &Option<CacheManagerRef>,
) {
    let Some(cache_manager) = cache_manager else {
        return;
    };
    if !cache_manager.sst_meta_cache_enabled() {
        return;
    }

    // Skip if SST meta cache is disabled.
    if config.sst_meta_cache_size.as_bytes() == 0 {
        return;
    }
    if !config.preload_index_cache {
        return;
    }

    let region = region.clone();
    let cache_manager = cache_manager.clone();
    let sst_meta_cache_capacity = config.sst_meta_cache_size.as_bytes();

    tokio::spawn(async move {
        // Safety: semaphore must exist.
        let _permit = PARQUET_META_PRELOAD_SEMAPHORE.acquire().await.unwrap();

        let region_id = region.region_id;
        let table_dir = region.access_layer.table_dir().to_string();
        let path_type = region.access_layer.path_type();
        let object_store = region.access_layer.object_store().clone();
        let region_metadata = region.version_control.current().version.metadata.clone();

        // Collect SST files. Do not hold the version longer than needed.
        let mut files = Vec::new();
        {
            let version = region.version_control.current().version;
            for level in version.ssts.levels() {
                for file_handle in level.files.values() {
                    files.push(file_handle.clone());
                }
            }
        }
        let preloading_start = Instant::now();
        let loaded = preload_parquet_meta_cache_for_files(
            region_id,
            cache_manager,
            sst_meta_cache_capacity,
            table_dir,
            path_type,
            object_store,
            region_metadata,
            files,
        )
        .await;
        let preloading_cost = preloading_start.elapsed();

        if loaded > 0 {
            info!(
                "Preloaded parquet metadata for region {}, loaded_files: {}, elapsed_ms: {}",
                region_id,
                loaded,
                preloading_cost.as_millis()
            );
        }
    });
}

fn can_load_cache(state: RegionRoleState) -> bool {
    match state {
        RegionRoleState::Leader(RegionLeaderState::Writable)
        | RegionRoleState::Leader(RegionLeaderState::Staging)
        | RegionRoleState::Leader(RegionLeaderState::Altering)
        | RegionRoleState::Leader(RegionLeaderState::EnteringStaging)
        | RegionRoleState::Leader(RegionLeaderState::Editing)
        | RegionRoleState::Follower => true,
        // The region will be closed soon if it is downgrading.
        RegionRoleState::Leader(RegionLeaderState::Downgrading)
        | RegionRoleState::Leader(RegionLeaderState::Dropping)
        | RegionRoleState::Leader(RegionLeaderState::Truncating) => false,
    }
}

#[cfg(test)]
mod tests {
    use std::collections::HashMap;
    use std::sync::Arc;

    use arrow_schema::extension::ExtensionType;
    use common_base::readable_size::ReadableSize;
    use common_error::ext::WhateverResult;
    use common_test_util::temp_dir::create_temp_dir;
    use common_time::Timestamp;
    use common_wal::options::{KafkaWalOptions, ObjectStoreWalOptions, WalOptions};
    use datatypes::arrow::array::{ArrayRef, BinaryArray, Int64Array};
    use datatypes::arrow::record_batch::RecordBatch;
    use datatypes::extension::json::{Json2ExtensionType, JsonMetadata};
    use datatypes::json::JsonSettings;
    use datatypes::prelude::ConcreteDataType;
    use datatypes::schema::ColumnSchema;
    use datatypes::types::json_type::{JsonNativeType, JsonObjectType};
    use log_store::kafka::log_store::KafkaLogStore;
    use log_store::noop::log_store::NoopLogStore;
    use log_store::raft_engine::log_store::RaftEngineLogStore;
    use object_store::ObjectStore;
    use object_store::services::{Fs, Memory, S3};
    use parquet::arrow::ArrowWriter;
    use parquet::file::metadata::{KeyValue, PageIndexPolicy};
    use parquet::file::properties::WriterProperties;
    use store_api::logstore::provider::Provider;
    use store_api::metadata::RegionMetadataBuilder;
    use store_api::region_request::PathType;
    use store_api::storage::{FileId, RegionId};

    use super::{
        initial_pruned_entry_id, maybe_upgrade_json2_layout, preload_parquet_meta_cache_for_files,
        provider_from_wal_options, sanitize_region_options,
        supports_open_region_object_storage_requirement,
    };
    use crate::cache::CacheManager;
    use crate::cache::file_cache::{FileType, IndexKey};
    use crate::error;
    use crate::manifest::action::{RegionManifest, RemovedFilesRecord};
    use crate::region::options::RegionOptions;
    use crate::sst::FormatType;
    use crate::sst::file::{FileHandle, FileMeta};
    use crate::sst::file_purger::NoopFilePurger;
    use crate::sst::parquet::PARQUET_METADATA_KEY;
    use crate::test_util::TestEnv;
    use crate::test_util::sst_util::sst_region_metadata;

    fn build_test_manifest(sst_format: FormatType) -> RegionManifest {
        RegionManifest {
            metadata: Arc::new(sst_region_metadata()),
            files: HashMap::new(),
            removed_files: RemovedFilesRecord::default(),
            flushed_entry_id: 0,
            flushed_sequence: 0,
            committed_sequence: None,
            manifest_version: 0,
            truncated_entry_id: None,
            compaction_time_window: None,
            sst_format,
            append_mode: None,
        }
    }

    fn build_fs_object_store() -> ObjectStore {
        ObjectStore::new(Fs::default().root("/tmp")).unwrap()
    }

    #[test]
    fn test_provider_from_object_store_wal_options() {
        let region_id = RegionId::new(1, 2);
        let wal_options = WalOptions::ObjectStore(ObjectStoreWalOptions::new("wal".to_string()));

        let provider = provider_from_wal_options::<NoopLogStore>(region_id, &wal_options).unwrap();
        assert_eq!(
            Provider::object_store_provider(region_id, "wal".to_string()),
            provider
        );

        let err =
            provider_from_wal_options::<RaftEngineLogStore>(region_id, &wal_options).unwrap_err();
        assert!(matches!(
            err,
            error::Error::IncompatibleWalProviderChange { .. }
        ));
        let err = provider_from_wal_options::<KafkaLogStore>(region_id, &wal_options).unwrap_err();
        assert!(matches!(
            err,
            error::Error::IncompatibleWalProviderChange { .. }
        ));
    }

    #[test]
    fn test_initial_pruned_entry_id() {
        assert_eq!(0, initial_pruned_entry_id(&WalOptions::RaftEngine));
        assert_eq!(0, initial_pruned_entry_id(&WalOptions::Noop));
        assert_eq!(
            0,
            initial_pruned_entry_id(&WalOptions::ObjectStore(ObjectStoreWalOptions::new(
                "wal".to_string()
            )))
        );
        assert_eq!(
            0,
            initial_pruned_entry_id(&WalOptions::Kafka(KafkaWalOptions::new(
                "test_topic".to_string()
            )))
        );
        assert_eq!(
            42,
            initial_pruned_entry_id(&WalOptions::Kafka(KafkaWalOptions {
                topic: "test_topic".to_string(),
                initial_pruned_entry_id: Some(42),
            }))
        );
    }

    #[test]
    fn test_upgrade_json2_layout() -> WhateverResult<()> {
        let settings = JsonSettings::try_new(vec![], Some(3))?;
        let extension = Json2ExtensionType::new(Arc::new(JsonMetadata::new_v1(settings.clone())));
        let mut column = ColumnSchema::new(
            "field_0",
            ConcreteDataType::json2(JsonNativeType::Object(JsonObjectType::new())),
            true,
        );
        column.with_extension_type(&extension);

        let mut metadata = sst_region_metadata();
        metadata.column_metadatas[2].column_schema = column;
        let builder = RegionMetadataBuilder::from_existing(metadata);
        let metadata = Arc::new(builder.build()?);

        let upgraded = maybe_upgrade_json2_layout(metadata)?;
        let column = &upgraded.column_metadatas[2].column_schema;
        let extension = column.extension_type::<Json2ExtensionType>()?.unwrap();
        assert!(extension.metadata().is_version_2());
        assert_eq!(&settings, extension.metadata().json_settings());

        let arrow_schema = upgraded.schema.arrow_schema();
        let field = arrow_schema.field_with_name("field_0").unwrap();
        let extension = field.try_extension_type::<Json2ExtensionType>().unwrap();
        assert!(extension.metadata().is_version_2());

        let unchanged = maybe_upgrade_json2_layout(upgraded.clone())?;
        assert!(Arc::ptr_eq(&upgraded, &unchanged));
        Ok(())
    }

    #[test]
    #[cfg(not(feature = "test-shared-fs-region-migration"))]
    fn test_open_requirement_rejects_fs_object_store() {
        let object_store = build_fs_object_store();

        assert!(!supports_open_region_object_storage_requirement(
            &object_store
        ));
    }

    #[test]
    #[cfg(feature = "test-shared-fs-region-migration")]
    fn test_open_requirement_accepts_shared_fs_object_store_for_tests() {
        let object_store = build_fs_object_store();

        assert!(supports_open_region_object_storage_requirement(
            &object_store
        ));
    }

    #[test]
    fn test_open_requirement_accepts_s3_object_store() {
        let object_store = ObjectStore::new(
            S3::default()
                .bucket("test-bucket")
                .region("us-east-1")
                .disable_ec2_metadata(),
        )
        .unwrap();

        assert!(supports_open_region_object_storage_requirement(
            &object_store
        ));
    }

    #[test]
    fn test_sanitize_region_options_options_format_wins() {
        // Manifest persisted PrimaryKey, but the re-parsed options now request Flat
        // (e.g., bulk memtable). The options' value must win.
        let manifest = build_test_manifest(FormatType::PrimaryKey);
        let mut options = RegionOptions {
            sst_format: Some(FormatType::Flat),
            ..Default::default()
        };
        sanitize_region_options(&manifest, &mut options);
        assert_eq!(options.sst_format, Some(FormatType::Flat));
    }

    #[test]
    fn test_sanitize_region_options_fills_from_manifest_when_unset() {
        // When the caller didn't specify sst_format, fall back to the manifest value
        // so downstream code always sees a concrete format.
        let manifest = build_test_manifest(FormatType::Flat);
        let mut options = RegionOptions {
            sst_format: None,
            ..Default::default()
        };
        sanitize_region_options(&manifest, &mut options);
        assert_eq!(options.sst_format, Some(FormatType::Flat));
    }

    fn sst_parquet_bytes(batch: &RecordBatch) -> Vec<u8> {
        let key_value_meta = KeyValue::new(
            PARQUET_METADATA_KEY.to_string(),
            sst_region_metadata().to_json().unwrap(),
        );
        let props = WriterProperties::builder()
            .set_key_value_metadata(Some(vec![key_value_meta]))
            .build();

        let mut parquet_bytes = Vec::new();
        let mut writer =
            ArrowWriter::try_new(&mut parquet_bytes, batch.schema(), Some(props)).unwrap();
        writer.write(batch).unwrap();
        writer.close().unwrap();

        parquet_bytes
    }

    #[tokio::test]
    async fn test_preload_parquet_meta_cache_uses_file_cache() {
        let env = TestEnv::new().await;

        let local_store = ObjectStore::new(Memory::default()).unwrap();
        let write_cache = env
            .create_write_cache(local_store, ReadableSize::mb(1024))
            .await;
        let cache_manager = Arc::new(
            CacheManager::builder()
                .sst_meta_cache_size(1024 * 1024)
                .write_cache(Some(write_cache.clone()))
                .build(),
        );

        let region_id = RegionId::new(1, 1);
        let file_id = FileId::random();

        let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
        let keys =
            ["a", "b", "c"].map(|tag| crate::test_util::sst_util::new_primary_key(&[tag, ""]));
        let primary_key = Arc::new(BinaryArray::from_iter_values(&keys)) as ArrayRef;
        let batch = RecordBatch::try_from_iter([
            ("col", col),
            (
                store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME,
                primary_key,
            ),
        ])
        .unwrap();
        let parquet_bytes = sst_parquet_bytes(&batch);
        let file_size = parquet_bytes.len() as u64;

        let file_meta = FileMeta {
            region_id,
            file_id,
            time_range: (Timestamp::new_millisecond(0), Timestamp::new_millisecond(1)),
            level: 0,
            file_size,
            max_row_group_uncompressed_size: 0,
            available_indexes: Default::default(),
            indexes: vec![],
            index_file_size: 0,
            index_version: 0,
            num_rows: 3,
            num_row_groups: 1,
            sequence: None,
            partition_expr: None,
            num_series: 0,
            ..Default::default()
        };
        let file_handle = FileHandle::new(file_meta, Arc::new(NoopFilePurger));

        let table_dir = "test_table";
        let path_type = PathType::Bare;
        let remote_path = file_handle.file_path(table_dir, path_type);

        let source_store = ObjectStore::new(Memory::default()).unwrap();
        source_store
            .write(&remote_path, parquet_bytes)
            .await
            .unwrap();

        // Put the parquet file into the write cache, so file cache contains metadata.
        let index_key = IndexKey::new(region_id, file_id, FileType::Parquet);
        write_cache
            .file_cache()
            .download(index_key, &remote_path, &source_store, file_size)
            .await
            .unwrap();

        let region_file_id = file_handle.file_id();
        assert!(
            cache_manager
                .get_parquet_meta_data_from_mem_cache(region_file_id)
                .is_none()
        );

        let loaded = preload_parquet_meta_cache_for_files(
            region_id,
            cache_manager.clone(),
            1024 * 1024,
            table_dir.to_string(),
            path_type,
            source_store.clone(),
            Arc::new(sst_region_metadata()),
            vec![file_handle.clone()],
        )
        .await;

        // Should warm the in-memory cache from the local file cache.
        assert_eq!(loaded, 1);
        assert!(
            cache_manager
                .get_compact_sst_meta_data(
                    region_file_id,
                    parquet::file::metadata::PageIndexPolicy::Optional,
                )
                .await
                .is_some()
        );
        assert!(file_handle.raw_primary_key_range().is_some());
    }

    #[tokio::test]
    async fn test_preload_parquet_meta_cache_skips_files_not_in_file_cache() {
        let cache_manager = Arc::new(
            CacheManager::builder()
                .sst_meta_cache_size(1024 * 1024)
                .build(),
        );

        let region_id = RegionId::new(1, 1);
        let file_id = FileId::random();

        // Without a local file cache entry, preloading should skip the file.
        let file_meta = FileMeta {
            region_id,
            file_id,
            time_range: (Timestamp::new_millisecond(0), Timestamp::new_millisecond(1)),
            level: 0,
            file_size: 0,
            max_row_group_uncompressed_size: 0,
            available_indexes: Default::default(),
            indexes: vec![],
            index_file_size: 0,
            index_version: 0,
            num_rows: 3,
            num_row_groups: 1,
            sequence: None,
            partition_expr: None,
            num_series: 0,
            ..Default::default()
        };
        let file_handle = FileHandle::new(file_meta, Arc::new(NoopFilePurger));

        let table_dir = "test_table";
        let path_type = PathType::Bare;
        let remote_path = file_handle.file_path(table_dir, path_type);

        // Even if the remote object store has the file, we should not preload from it.
        let object_store = ObjectStore::new(Memory::default()).unwrap();
        object_store
            .write(&remote_path, b"noop".as_slice())
            .await
            .unwrap();

        let region_file_id = file_handle.file_id();
        assert!(
            cache_manager
                .get_parquet_meta_data_from_mem_cache(region_file_id)
                .is_none()
        );

        let loaded = preload_parquet_meta_cache_for_files(
            region_id,
            cache_manager.clone(),
            1024 * 1024,
            table_dir.to_string(),
            path_type,
            object_store,
            Arc::new(sst_region_metadata()),
            vec![file_handle],
        )
        .await;

        assert_eq!(loaded, 0);
        assert!(
            cache_manager
                .get_parquet_meta_data_from_mem_cache(region_file_id)
                .is_none()
        );
    }

    #[tokio::test]
    async fn test_preload_parquet_meta_cache_loads_from_local_fs() {
        let cache_manager = Arc::new(
            CacheManager::builder()
                .sst_meta_cache_size(1024 * 1024)
                .build(),
        );

        let region_id = RegionId::new(1, 1);
        let file_id = FileId::random();

        let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
        let batch = RecordBatch::try_from_iter([("col", col)]).unwrap();
        let parquet_bytes = sst_parquet_bytes(&batch);

        // file_size is 0 when it's missing/defaulted in manifests; MetadataLoader::load will stat
        // the local filesystem to retrieve it.
        let file_meta = FileMeta {
            region_id,
            file_id,
            time_range: (Timestamp::new_millisecond(0), Timestamp::new_millisecond(1)),
            level: 0,
            file_size: 0,
            max_row_group_uncompressed_size: 0,
            available_indexes: Default::default(),
            indexes: vec![],
            index_file_size: 0,
            index_version: 0,
            num_rows: 3,
            num_row_groups: 1,
            sequence: None,
            partition_expr: None,
            num_series: 0,
            ..Default::default()
        };
        let file_handle = FileHandle::new(file_meta, Arc::new(NoopFilePurger));

        let table_dir = "test_table";
        let path_type = PathType::Bare;
        let file_path = file_handle.file_path(table_dir, path_type);

        let root = create_temp_dir("parquet-meta-preload");
        let object_store =
            ObjectStore::new(Fs::default().root(root.path().to_str().unwrap())).unwrap();
        object_store.write(&file_path, parquet_bytes).await.unwrap();

        let region_file_id = file_handle.file_id();
        assert!(
            cache_manager
                .get_parquet_meta_data_from_mem_cache(region_file_id)
                .is_none()
        );

        let loaded = preload_parquet_meta_cache_for_files(
            region_id,
            cache_manager.clone(),
            1024 * 1024,
            table_dir.to_string(),
            path_type,
            object_store,
            Arc::new(sst_region_metadata()),
            vec![file_handle],
        )
        .await;

        assert_eq!(loaded, 1);
        assert!(
            cache_manager
                .get_compact_sst_meta_data(region_file_id, PageIndexPolicy::Optional)
                .await
                .is_some()
        );
    }
}