mito2/

flush.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.

//! Flush related utilities and structs.

use std::collections::HashMap;
use std::num::NonZeroU64;
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::time::Instant;

use common_telemetry::{debug, error, info, trace};
use datatypes::arrow::datatypes::SchemaRef;
use either::Either;
use partition::expr::PartitionExpr;
use smallvec::{SmallVec, smallvec};
use snafu::ResultExt;
use store_api::storage::RegionId;
use strum::IntoStaticStr;
use tokio::sync::{Semaphore, mpsc, watch};

use crate::access_layer::{
    AccessLayerRef, Metrics, OperationType, SstInfoArray, SstWriteRequest, WriteType,
};
use crate::cache::CacheManagerRef;
use crate::config::MitoConfig;
use crate::error::{
    Error, FlushRegionSnafu, InvalidPartitionExprSnafu, JoinSnafu, RegionClosedSnafu,
    RegionDroppedSnafu, RegionTruncatedSnafu, Result,
};
use crate::manifest::action::{RegionEdit, RegionMetaAction, RegionMetaActionList};
use crate::memtable::{
    BoxedRecordBatchIterator, EncodedRange, IterBuilder, MemtableRanges, RangesOptions,
};
use crate::metrics::{
    FLUSH_BYTES_TOTAL, FLUSH_ELAPSED, FLUSH_FAILURE_TOTAL, FLUSH_REQUESTS_TOTAL,
    INFLIGHT_FLUSH_COUNT,
};
use crate::read::dedup::{DedupReader, LastNonNull, LastRow};
use crate::read::flat_dedup::{FlatDedupIterator, FlatLastNonNull, FlatLastRow};
use crate::read::flat_merge::FlatMergeIterator;
use crate::read::merge::MergeReaderBuilder;
use crate::read::{FlatSource, Source};
use crate::region::options::{IndexOptions, MergeMode, RegionOptions};
use crate::region::version::{VersionControlData, VersionControlRef, VersionRef};
use crate::region::{ManifestContextRef, RegionLeaderState, RegionRoleState};
use crate::request::{
    BackgroundNotify, FlushFailed, FlushFinished, OptionOutputTx, OutputTx, SenderBulkRequest,
    SenderDdlRequest, SenderWriteRequest, WorkerRequest, WorkerRequestWithTime,
};
use crate::schedule::scheduler::{Job, SchedulerRef};
use crate::sst::file::FileMeta;
use crate::sst::parquet::{DEFAULT_READ_BATCH_SIZE, DEFAULT_ROW_GROUP_SIZE, SstInfo, WriteOptions};
use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
use crate::worker::WorkerListener;

/// Global write buffer (memtable) manager.
///
/// Tracks write buffer (memtable) usages and decide whether the engine needs to flush.
pub trait WriteBufferManager: Send + Sync + std::fmt::Debug {
    /// Returns whether to trigger the engine.
    fn should_flush_engine(&self) -> bool;

    /// Returns whether to stall write requests.
    fn should_stall(&self) -> bool;

    /// Reserves `mem` bytes.
    fn reserve_mem(&self, mem: usize);

    /// Tells the manager we are freeing `mem` bytes.
    ///
    /// We are in the process of freeing `mem` bytes, so it is not considered
    /// when checking the soft limit.
    fn schedule_free_mem(&self, mem: usize);

    /// We have freed `mem` bytes.
    fn free_mem(&self, mem: usize);

    /// Returns the total memory used by memtables.
    fn memory_usage(&self) -> usize;
}

pub type WriteBufferManagerRef = Arc<dyn WriteBufferManager>;

/// Default [WriteBufferManager] implementation.
///
/// Inspired by RocksDB's WriteBufferManager.
/// <https://github.com/facebook/rocksdb/blob/main/include/rocksdb/write_buffer_manager.h>
#[derive(Debug)]
pub struct WriteBufferManagerImpl {
    /// Write buffer size for the engine.
    global_write_buffer_size: usize,
    /// Mutable memtable memory size limit.
    mutable_limit: usize,
    /// Memory in used (e.g. used by mutable and immutable memtables).
    memory_used: AtomicUsize,
    /// Memory that hasn't been scheduled to free (e.g. used by mutable memtables).
    memory_active: AtomicUsize,
    /// Optional notifier.
    /// The manager can wake up the worker once we free the write buffer.
    notifier: Option<watch::Sender<()>>,
}

impl WriteBufferManagerImpl {
    /// Returns a new manager with specific `global_write_buffer_size`.
    pub fn new(global_write_buffer_size: usize) -> Self {
        Self {
            global_write_buffer_size,
            mutable_limit: Self::get_mutable_limit(global_write_buffer_size),
            memory_used: AtomicUsize::new(0),
            memory_active: AtomicUsize::new(0),
            notifier: None,
        }
    }

    /// Attaches a notifier to the manager.
    pub fn with_notifier(mut self, notifier: watch::Sender<()>) -> Self {
        self.notifier = Some(notifier);
        self
    }

    /// Returns memory usage of mutable memtables.
    pub fn mutable_usage(&self) -> usize {
        self.memory_active.load(Ordering::Relaxed)
    }

    /// Returns the size limit for mutable memtables.
    fn get_mutable_limit(global_write_buffer_size: usize) -> usize {
        // Reserves half of the write buffer for mutable memtable.
        global_write_buffer_size / 2
    }
}

impl WriteBufferManager for WriteBufferManagerImpl {
    fn should_flush_engine(&self) -> bool {
        let mutable_memtable_memory_usage = self.memory_active.load(Ordering::Relaxed);
        if mutable_memtable_memory_usage > self.mutable_limit {
            debug!(
                "Engine should flush (over mutable limit), mutable_usage: {}, memory_usage: {}, mutable_limit: {}, global_limit: {}",
                mutable_memtable_memory_usage,
                self.memory_usage(),
                self.mutable_limit,
                self.global_write_buffer_size,
            );
            return true;
        }

        let memory_usage = self.memory_used.load(Ordering::Relaxed);
        // If the memory exceeds the buffer size, we trigger more aggressive
        // flush. But if already more than half memory is being flushed,
        // triggering more flush may not help. We will hold it instead.
        if memory_usage >= self.global_write_buffer_size {
            if mutable_memtable_memory_usage >= self.global_write_buffer_size / 2 {
                debug!(
                    "Engine should flush (over total limit), memory_usage: {}, global_write_buffer_size: {}, \
                 mutable_usage: {}.",
                    memory_usage, self.global_write_buffer_size, mutable_memtable_memory_usage
                );
                return true;
            } else {
                trace!(
                    "Engine won't flush, memory_usage: {}, global_write_buffer_size: {}, mutable_usage: {}.",
                    memory_usage, self.global_write_buffer_size, mutable_memtable_memory_usage
                );
            }
        }

        false
    }

    fn should_stall(&self) -> bool {
        self.memory_usage() >= self.global_write_buffer_size
    }

    fn reserve_mem(&self, mem: usize) {
        self.memory_used.fetch_add(mem, Ordering::Relaxed);
        self.memory_active.fetch_add(mem, Ordering::Relaxed);
    }

    fn schedule_free_mem(&self, mem: usize) {
        self.memory_active.fetch_sub(mem, Ordering::Relaxed);
    }

    fn free_mem(&self, mem: usize) {
        self.memory_used.fetch_sub(mem, Ordering::Relaxed);
        if let Some(notifier) = &self.notifier {
            // Notifies the worker after the memory usage is decreased. When we drop the memtable
            // outside of the worker, the worker may still stall requests because the memory usage
            // is not updated. So we need to notify the worker to handle stalled requests again.
            let _ = notifier.send(());
        }
    }

    fn memory_usage(&self) -> usize {
        self.memory_used.load(Ordering::Relaxed)
    }
}

/// Reason of a flush task.
#[derive(Debug, IntoStaticStr)]
pub enum FlushReason {
    /// Other reasons.
    Others,
    /// Engine reaches flush threshold.
    EngineFull,
    /// Manual flush.
    Manual,
    /// Flush to alter table.
    Alter,
    /// Flush periodically.
    Periodically,
    /// Flush memtable during downgrading state.
    Downgrading,
}

impl FlushReason {
    /// Get flush reason as static str.
    fn as_str(&self) -> &'static str {
        self.into()
    }
}

/// Task to flush a region.
pub(crate) struct RegionFlushTask {
    /// Region to flush.
    pub(crate) region_id: RegionId,
    /// Reason to flush.
    pub(crate) reason: FlushReason,
    /// Flush result senders.
    pub(crate) senders: Vec<OutputTx>,
    /// Request sender to notify the worker.
    pub(crate) request_sender: mpsc::Sender<WorkerRequestWithTime>,

    pub(crate) access_layer: AccessLayerRef,
    pub(crate) listener: WorkerListener,
    pub(crate) engine_config: Arc<MitoConfig>,
    pub(crate) row_group_size: Option<usize>,
    pub(crate) cache_manager: CacheManagerRef,
    pub(crate) manifest_ctx: ManifestContextRef,

    /// Index options for the region.
    pub(crate) index_options: IndexOptions,
    /// Semaphore to control flush concurrency.
    pub(crate) flush_semaphore: Arc<Semaphore>,
}

impl RegionFlushTask {
    /// Push the sender if it is not none.
    pub(crate) fn push_sender(&mut self, mut sender: OptionOutputTx) {
        if let Some(sender) = sender.take_inner() {
            self.senders.push(sender);
        }
    }

    /// Consumes the task and notify the sender the job is success.
    fn on_success(self) {
        for sender in self.senders {
            sender.send(Ok(0));
        }
    }

    /// Send flush error to waiter.
    fn on_failure(&mut self, err: Arc<Error>) {
        for sender in self.senders.drain(..) {
            sender.send(Err(err.clone()).context(FlushRegionSnafu {
                region_id: self.region_id,
            }));
        }
    }

    /// Converts the flush task into a background job.
    ///
    /// We must call this in the region worker.
    fn into_flush_job(mut self, version_control: &VersionControlRef) -> Job {
        // Get a version of this region before creating a job to get current
        // wal entry id, sequence and immutable memtables.
        let version_data = version_control.current();

        Box::pin(async move {
            INFLIGHT_FLUSH_COUNT.inc();
            self.do_flush(version_data).await;
            INFLIGHT_FLUSH_COUNT.dec();
        })
    }

    /// Runs the flush task.
    async fn do_flush(&mut self, version_data: VersionControlData) {
        let timer = FLUSH_ELAPSED.with_label_values(&["total"]).start_timer();
        self.listener.on_flush_begin(self.region_id).await;

        let worker_request = match self.flush_memtables(&version_data).await {
            Ok(edit) => {
                let memtables_to_remove = version_data
                    .version
                    .memtables
                    .immutables()
                    .iter()
                    .map(|m| m.id())
                    .collect();
                let flush_finished = FlushFinished {
                    region_id: self.region_id,
                    // The last entry has been flushed.
                    flushed_entry_id: version_data.last_entry_id,
                    senders: std::mem::take(&mut self.senders),
                    _timer: timer,
                    edit,
                    memtables_to_remove,
                };
                WorkerRequest::Background {
                    region_id: self.region_id,
                    notify: BackgroundNotify::FlushFinished(flush_finished),
                }
            }
            Err(e) => {
                error!(e; "Failed to flush region {}", self.region_id);
                // Discard the timer.
                timer.stop_and_discard();

                let err = Arc::new(e);
                self.on_failure(err.clone());
                WorkerRequest::Background {
                    region_id: self.region_id,
                    notify: BackgroundNotify::FlushFailed(FlushFailed { err }),
                }
            }
        };
        self.send_worker_request(worker_request).await;
    }

    /// Flushes memtables to level 0 SSTs and updates the manifest.
    /// Returns the [RegionEdit] to apply.
    async fn flush_memtables(&self, version_data: &VersionControlData) -> Result<RegionEdit> {
        // We must use the immutable memtables list and entry ids from the `version_data`
        // for consistency as others might already modify the version in the `version_control`.
        let version = &version_data.version;
        let timer = FLUSH_ELAPSED
            .with_label_values(&["flush_memtables"])
            .start_timer();

        let mut write_opts = WriteOptions {
            write_buffer_size: self.engine_config.sst_write_buffer_size,
            ..Default::default()
        };
        if let Some(row_group_size) = self.row_group_size {
            write_opts.row_group_size = row_group_size;
        }

        let DoFlushMemtablesResult {
            file_metas,
            flushed_bytes,
            series_count,
            flush_metrics,
        } = self.do_flush_memtables(version, write_opts).await?;

        if !file_metas.is_empty() {
            FLUSH_BYTES_TOTAL.inc_by(flushed_bytes);
        }

        let mut file_ids = Vec::with_capacity(file_metas.len());
        let mut total_rows = 0;
        let mut total_bytes = 0;
        for meta in &file_metas {
            file_ids.push(meta.file_id);
            total_rows += meta.num_rows;
            total_bytes += meta.file_size;
        }
        info!(
            "Successfully flush memtables, region: {}, reason: {}, files: {:?}, series count: {}, total_rows: {}, total_bytes: {}, cost: {:?}, metrics: {:?}",
            self.region_id,
            self.reason.as_str(),
            file_ids,
            series_count,
            total_rows,
            total_bytes,
            timer.stop_and_record(),
            flush_metrics,
        );
        flush_metrics.observe();

        let edit = RegionEdit {
            files_to_add: file_metas,
            files_to_remove: Vec::new(),
            timestamp_ms: Some(chrono::Utc::now().timestamp_millis()),
            compaction_time_window: None,
            // The last entry has been flushed.
            flushed_entry_id: Some(version_data.last_entry_id),
            flushed_sequence: Some(version_data.committed_sequence),
            committed_sequence: None,
        };
        info!("Applying {edit:?} to region {}", self.region_id);

        let action_list = RegionMetaActionList::with_action(RegionMetaAction::Edit(edit.clone()));

        let expected_state = if matches!(self.reason, FlushReason::Downgrading) {
            RegionLeaderState::Downgrading
        } else {
            // Check if region is in staging mode
            let current_state = self.manifest_ctx.current_state();
            if current_state == RegionRoleState::Leader(RegionLeaderState::Staging) {
                RegionLeaderState::Staging
            } else {
                RegionLeaderState::Writable
            }
        };
        // We will leak files if the manifest update fails, but we ignore them for simplicity. We can
        // add a cleanup job to remove them later.
        let version = self
            .manifest_ctx
            .update_manifest(expected_state, action_list)
            .await?;
        info!(
            "Successfully update manifest version to {version}, region: {}, reason: {}",
            self.region_id,
            self.reason.as_str()
        );

        Ok(edit)
    }

    async fn do_flush_memtables(
        &self,
        version: &VersionRef,
        write_opts: WriteOptions,
    ) -> Result<DoFlushMemtablesResult> {
        let memtables = version.memtables.immutables();
        let mut file_metas = Vec::with_capacity(memtables.len());
        let mut flushed_bytes = 0;
        let mut series_count = 0;
        // Convert partition expression once outside the map
        let partition_expr = match &version.metadata.partition_expr {
            None => None,
            Some(json_expr) if json_expr.is_empty() => None,
            Some(json_str) => partition::expr::PartitionExpr::from_json_str(json_str)
                .with_context(|_| InvalidPartitionExprSnafu { expr: json_str })?,
        };
        let mut flush_metrics = Metrics::new(WriteType::Flush);
        for mem in memtables {
            if mem.is_empty() {
                // Skip empty memtables.
                continue;
            }

            // Compact the memtable first, this waits the background compaction to finish.
            let compact_start = std::time::Instant::now();
            if let Err(e) = mem.compact(true) {
                common_telemetry::error!(e; "Failed to compact memtable before flush");
            }
            let compact_cost = compact_start.elapsed();
            flush_metrics.compact_memtable += compact_cost;

            // Sets `for_flush` flag to true.
            let mem_ranges = mem.ranges(None, RangesOptions::for_flush())?;
            let num_mem_ranges = mem_ranges.ranges.len();
            let num_mem_rows = mem_ranges.stats.num_rows();
            let memtable_id = mem.id();
            // Increases series count for each mem range. We consider each mem range has different series so
            // the counter may have more series than the actual series count.
            series_count += mem_ranges.stats.series_count();

            if mem_ranges.is_record_batch() {
                let flush_start = Instant::now();
                let FlushFlatMemResult {
                    num_encoded,
                    max_sequence,
                    num_sources,
                    results,
                } = self
                    .flush_flat_mem_ranges(version, &write_opts, mem_ranges)
                    .await?;
                for (source_idx, result) in results.into_iter().enumerate() {
                    let (ssts_written, metrics) = result?;
                    if ssts_written.is_empty() {
                        // No data written.
                        continue;
                    }

                    common_telemetry::debug!(
                        "Region {} flush one memtable {} {}/{}, metrics: {:?}",
                        self.region_id,
                        memtable_id,
                        source_idx,
                        num_sources,
                        metrics
                    );

                    flush_metrics = flush_metrics.merge(metrics);

                    file_metas.extend(ssts_written.into_iter().map(|sst_info| {
                        flushed_bytes += sst_info.file_size;
                        Self::new_file_meta(
                            self.region_id,
                            max_sequence,
                            sst_info,
                            partition_expr.clone(),
                        )
                    }));
                }

                common_telemetry::debug!(
                    "Region {} flush {} memtables for {}, num_mem_ranges: {}, num_encoded: {}, num_rows: {}, flush_cost: {:?}, compact_cost: {:?}",
                    self.region_id,
                    num_sources,
                    memtable_id,
                    num_mem_ranges,
                    num_encoded,
                    num_mem_rows,
                    flush_start.elapsed(),
                    compact_cost,
                );
            } else {
                let max_sequence = mem_ranges.stats.max_sequence();
                let source = memtable_source(mem_ranges, &version.options).await?;

                // Flush to level 0.
                let source = Either::Left(source);
                let write_request = self.new_write_request(version, max_sequence, source);

                let mut metrics = Metrics::new(WriteType::Flush);
                let ssts_written = self
                    .access_layer
                    .write_sst(write_request, &write_opts, &mut metrics)
                    .await?;
                if ssts_written.is_empty() {
                    // No data written.
                    continue;
                }

                debug!(
                    "Region {} flush one memtable, num_mem_ranges: {}, num_rows: {}, metrics: {:?}",
                    self.region_id, num_mem_ranges, num_mem_rows, metrics
                );

                flush_metrics = flush_metrics.merge(metrics);

                file_metas.extend(ssts_written.into_iter().map(|sst_info| {
                    flushed_bytes += sst_info.file_size;
                    Self::new_file_meta(
                        self.region_id,
                        max_sequence,
                        sst_info,
                        partition_expr.clone(),
                    )
                }));
            };
        }

        Ok(DoFlushMemtablesResult {
            file_metas,
            flushed_bytes,
            series_count,
            flush_metrics,
        })
    }

    async fn flush_flat_mem_ranges(
        &self,
        version: &VersionRef,
        write_opts: &WriteOptions,
        mem_ranges: MemtableRanges,
    ) -> Result<FlushFlatMemResult> {
        let batch_schema = to_flat_sst_arrow_schema(
            &version.metadata,
            &FlatSchemaOptions::from_encoding(version.metadata.primary_key_encoding),
        );
        let flat_sources = memtable_flat_sources(
            batch_schema,
            mem_ranges,
            &version.options,
            version.metadata.primary_key.len(),
        )?;
        let mut tasks = Vec::with_capacity(flat_sources.encoded.len() + flat_sources.sources.len());
        let num_encoded = flat_sources.encoded.len();
        let max_sequence = flat_sources.max_sequence;
        for source in flat_sources.sources {
            let source = Either::Right(source);
            let write_request = self.new_write_request(version, max_sequence, source);
            let access_layer = self.access_layer.clone();
            let write_opts = write_opts.clone();
            let semaphore = self.flush_semaphore.clone();
            let task = common_runtime::spawn_global(async move {
                let _permit = semaphore.acquire().await.unwrap();
                let mut metrics = Metrics::new(WriteType::Flush);
                let ssts = access_layer
                    .write_sst(write_request, &write_opts, &mut metrics)
                    .await?;
                Ok((ssts, metrics))
            });
            tasks.push(task);
        }
        for encoded in flat_sources.encoded {
            let access_layer = self.access_layer.clone();
            let cache_manager = self.cache_manager.clone();
            let region_id = version.metadata.region_id;
            let semaphore = self.flush_semaphore.clone();
            let task = common_runtime::spawn_global(async move {
                let _permit = semaphore.acquire().await.unwrap();
                let metrics = access_layer
                    .put_sst(&encoded.data, region_id, &encoded.sst_info, &cache_manager)
                    .await?;
                Ok((smallvec![encoded.sst_info], metrics))
            });
            tasks.push(task);
        }
        let num_sources = tasks.len();
        let results = futures::future::try_join_all(tasks)
            .await
            .context(JoinSnafu)?;
        Ok(FlushFlatMemResult {
            num_encoded,
            max_sequence,
            num_sources,
            results,
        })
    }

    fn new_file_meta(
        region_id: RegionId,
        max_sequence: u64,
        sst_info: SstInfo,
        partition_expr: Option<PartitionExpr>,
    ) -> FileMeta {
        FileMeta {
            region_id,
            file_id: sst_info.file_id,
            time_range: sst_info.time_range,
            level: 0,
            file_size: sst_info.file_size,
            available_indexes: sst_info.index_metadata.build_available_indexes(),
            index_file_size: sst_info.index_metadata.file_size,
            index_file_id: None,
            num_rows: sst_info.num_rows as u64,
            num_row_groups: sst_info.num_row_groups,
            sequence: NonZeroU64::new(max_sequence),
            partition_expr,
            num_series: sst_info.num_series,
        }
    }

    fn new_write_request(
        &self,
        version: &VersionRef,
        max_sequence: u64,
        source: Either<Source, FlatSource>,
    ) -> SstWriteRequest {
        SstWriteRequest {
            op_type: OperationType::Flush,
            metadata: version.metadata.clone(),
            source,
            cache_manager: self.cache_manager.clone(),
            storage: version.options.storage.clone(),
            max_sequence: Some(max_sequence),
            index_options: self.index_options.clone(),
            index_config: self.engine_config.index.clone(),
            inverted_index_config: self.engine_config.inverted_index.clone(),
            fulltext_index_config: self.engine_config.fulltext_index.clone(),
            bloom_filter_index_config: self.engine_config.bloom_filter_index.clone(),
        }
    }

    /// Notify flush job status.
    pub(crate) async fn send_worker_request(&self, request: WorkerRequest) {
        if let Err(e) = self
            .request_sender
            .send(WorkerRequestWithTime::new(request))
            .await
        {
            error!(
                "Failed to notify flush job status for region {}, request: {:?}",
                self.region_id, e.0
            );
        }
    }

    /// Merge two flush tasks.
    fn merge(&mut self, mut other: RegionFlushTask) {
        assert_eq!(self.region_id, other.region_id);
        // Now we only merge senders. They share the same flush reason.
        self.senders.append(&mut other.senders);
    }
}

struct FlushFlatMemResult {
    num_encoded: usize,
    max_sequence: u64,
    num_sources: usize,
    results: Vec<Result<(SstInfoArray, Metrics)>>,
}

struct DoFlushMemtablesResult {
    file_metas: Vec<FileMeta>,
    flushed_bytes: u64,
    series_count: usize,
    flush_metrics: Metrics,
}

/// Returns a [Source] for the given memtable.
async fn memtable_source(mem_ranges: MemtableRanges, options: &RegionOptions) -> Result<Source> {
    let source = if mem_ranges.ranges.len() == 1 {
        let only_range = mem_ranges.ranges.into_values().next().unwrap();
        let iter = only_range.build_iter()?;
        Source::Iter(iter)
    } else {
        // todo(hl): a workaround since sync version of MergeReader is wip.
        let sources = mem_ranges
            .ranges
            .into_values()
            .map(|r| r.build_iter().map(Source::Iter))
            .collect::<Result<Vec<_>>>()?;
        let merge_reader = MergeReaderBuilder::from_sources(sources).build().await?;
        let maybe_dedup = if options.append_mode {
            // no dedup in append mode
            Box::new(merge_reader) as _
        } else {
            // dedup according to merge mode
            match options.merge_mode.unwrap_or(MergeMode::LastRow) {
                MergeMode::LastRow => {
                    Box::new(DedupReader::new(merge_reader, LastRow::new(false))) as _
                }
                MergeMode::LastNonNull => {
                    Box::new(DedupReader::new(merge_reader, LastNonNull::new(false))) as _
                }
            }
        };
        Source::Reader(maybe_dedup)
    };
    Ok(source)
}

struct FlatSources {
    max_sequence: u64,
    sources: SmallVec<[FlatSource; 4]>,
    encoded: SmallVec<[EncodedRange; 4]>,
}

/// Returns the max sequence and [FlatSource] for the given memtable.
fn memtable_flat_sources(
    schema: SchemaRef,
    mem_ranges: MemtableRanges,
    options: &RegionOptions,
    field_column_start: usize,
) -> Result<FlatSources> {
    let MemtableRanges { ranges, stats } = mem_ranges;
    let max_sequence = stats.max_sequence();
    let mut flat_sources = FlatSources {
        max_sequence,
        sources: SmallVec::new(),
        encoded: SmallVec::new(),
    };

    if ranges.len() == 1 {
        let only_range = ranges.into_values().next().unwrap();
        if let Some(encoded) = only_range.encoded() {
            flat_sources.encoded.push(encoded);
        } else {
            let iter = only_range.build_record_batch_iter(None)?;
            // Dedup according to append mode and merge mode.
            // Even single range may have duplicate rows.
            let iter = maybe_dedup_one(options, field_column_start, iter);
            flat_sources.sources.push(FlatSource::Iter(iter));
        };
    } else {
        let min_flush_rows = stats.num_rows / 8;
        let min_flush_rows = min_flush_rows.max(DEFAULT_ROW_GROUP_SIZE);
        let mut last_iter_rows = 0;
        let num_ranges = ranges.len();
        let mut input_iters = Vec::with_capacity(num_ranges);
        for (_range_id, range) in ranges {
            if let Some(encoded) = range.encoded() {
                flat_sources.encoded.push(encoded);
                continue;
            }

            let iter = range.build_record_batch_iter(None)?;
            input_iters.push(iter);
            last_iter_rows += range.num_rows();

            if last_iter_rows > min_flush_rows {
                let maybe_dedup = merge_and_dedup(
                    &schema,
                    options,
                    field_column_start,
                    std::mem::replace(&mut input_iters, Vec::with_capacity(num_ranges)),
                )?;

                flat_sources.sources.push(FlatSource::Iter(maybe_dedup));
                last_iter_rows = 0;
            }
        }

        // Handle remaining iters.
        if !input_iters.is_empty() {
            let maybe_dedup = merge_and_dedup(&schema, options, field_column_start, input_iters)?;

            flat_sources.sources.push(FlatSource::Iter(maybe_dedup));
        }
    }

    Ok(flat_sources)
}

fn merge_and_dedup(
    schema: &SchemaRef,
    options: &RegionOptions,
    field_column_start: usize,
    input_iters: Vec<BoxedRecordBatchIterator>,
) -> Result<BoxedRecordBatchIterator> {
    let merge_iter = FlatMergeIterator::new(schema.clone(), input_iters, DEFAULT_READ_BATCH_SIZE)?;
    let maybe_dedup = if options.append_mode {
        // No dedup in append mode
        Box::new(merge_iter) as _
    } else {
        // Dedup according to merge mode.
        match options.merge_mode() {
            MergeMode::LastRow => {
                Box::new(FlatDedupIterator::new(merge_iter, FlatLastRow::new(false))) as _
            }
            MergeMode::LastNonNull => Box::new(FlatDedupIterator::new(
                merge_iter,
                FlatLastNonNull::new(field_column_start, false),
            )) as _,
        }
    };
    Ok(maybe_dedup)
}

fn maybe_dedup_one(
    options: &RegionOptions,
    field_column_start: usize,
    input_iter: BoxedRecordBatchIterator,
) -> BoxedRecordBatchIterator {
    if options.append_mode {
        // No dedup in append mode
        input_iter
    } else {
        // Dedup according to merge mode.
        match options.merge_mode() {
            MergeMode::LastRow => {
                Box::new(FlatDedupIterator::new(input_iter, FlatLastRow::new(false)))
            }
            MergeMode::LastNonNull => Box::new(FlatDedupIterator::new(
                input_iter,
                FlatLastNonNull::new(field_column_start, false),
            )),
        }
    }
}

/// Manages background flushes of a worker.
pub(crate) struct FlushScheduler {
    /// Tracks regions need to flush.
    region_status: HashMap<RegionId, FlushStatus>,
    /// Background job scheduler.
    scheduler: SchedulerRef,
}

impl FlushScheduler {
    /// Creates a new flush scheduler.
    pub(crate) fn new(scheduler: SchedulerRef) -> FlushScheduler {
        FlushScheduler {
            region_status: HashMap::new(),
            scheduler,
        }
    }

    /// Returns true if the region already requested flush.
    pub(crate) fn is_flush_requested(&self, region_id: RegionId) -> bool {
        self.region_status.contains_key(&region_id)
    }

    /// Schedules a flush `task` for specific `region`.
    pub(crate) fn schedule_flush(
        &mut self,
        region_id: RegionId,
        version_control: &VersionControlRef,
        task: RegionFlushTask,
    ) -> Result<()> {
        debug_assert_eq!(region_id, task.region_id);

        let version = version_control.current().version;
        if version.memtables.is_empty() {
            debug_assert!(!self.region_status.contains_key(&region_id));
            // The region has nothing to flush.
            task.on_success();
            return Ok(());
        }

        // Don't increase the counter if a region has nothing to flush.
        FLUSH_REQUESTS_TOTAL
            .with_label_values(&[task.reason.as_str()])
            .inc();

        // Add this region to status map.
        let flush_status = self
            .region_status
            .entry(region_id)
            .or_insert_with(|| FlushStatus::new(region_id, version_control.clone()));
        // Checks whether we can flush the region now.
        if flush_status.flushing {
            // There is already a flush job running.
            flush_status.merge_task(task);
            return Ok(());
        }

        // TODO(yingwen): We can merge with pending and execute directly.
        // If there are pending tasks, then we should push it to pending list.
        if flush_status.pending_task.is_some() {
            flush_status.merge_task(task);
            return Ok(());
        }

        // Now we can flush the region directly.
        if let Err(e) = version_control.freeze_mutable() {
            error!(e; "Failed to freeze the mutable memtable for region {}", region_id);

            // Remove from region status if we can't freeze the mutable memtable.
            self.region_status.remove(&region_id);
            return Err(e);
        }
        // Submit a flush job.
        let job = task.into_flush_job(version_control);
        if let Err(e) = self.scheduler.schedule(job) {
            // If scheduler returns error, senders in the job will be dropped and waiters
            // can get recv errors.
            error!(e; "Failed to schedule flush job for region {}", region_id);

            // Remove from region status if we can't submit the task.
            self.region_status.remove(&region_id);
            return Err(e);
        }

        flush_status.flushing = true;

        Ok(())
    }

    /// Notifies the scheduler that the flush job is finished.
    ///
    /// Returns all pending requests if the region doesn't need to flush again.
    pub(crate) fn on_flush_success(
        &mut self,
        region_id: RegionId,
    ) -> Option<(
        Vec<SenderDdlRequest>,
        Vec<SenderWriteRequest>,
        Vec<SenderBulkRequest>,
    )> {
        let flush_status = self.region_status.get_mut(&region_id)?;

        // This region doesn't have running flush job.
        flush_status.flushing = false;

        let pending_requests = if flush_status.pending_task.is_none() {
            // The region doesn't have any pending flush task.
            // Safety: The flush status must exist.
            let flush_status = self.region_status.remove(&region_id).unwrap();
            Some((
                flush_status.pending_ddls,
                flush_status.pending_writes,
                flush_status.pending_bulk_writes,
            ))
        } else {
            let version_data = flush_status.version_control.current();
            if version_data.version.memtables.is_empty() {
                // The region has nothing to flush, we also need to remove it from the status.
                // Safety: The pending task is not None.
                let task = flush_status.pending_task.take().unwrap();
                // The region has nothing to flush. We can notify pending task.
                task.on_success();
                // `schedule_next_flush()` may pick up the same region to flush, so we must remove
                // it from the status to avoid leaking pending requests.
                // Safety: The flush status must exist.
                let flush_status = self.region_status.remove(&region_id).unwrap();
                Some((
                    flush_status.pending_ddls,
                    flush_status.pending_writes,
                    flush_status.pending_bulk_writes,
                ))
            } else {
                // We can flush the region again, keep it in the region status.
                None
            }
        };

        // Schedule next flush job.
        if let Err(e) = self.schedule_next_flush() {
            error!(e; "Flush of region {} is successful, but failed to schedule next flush", region_id);
        }

        pending_requests
    }

    /// Notifies the scheduler that the flush job is failed.
    pub(crate) fn on_flush_failed(&mut self, region_id: RegionId, err: Arc<Error>) {
        error!(err; "Region {} failed to flush, cancel all pending tasks", region_id);

        FLUSH_FAILURE_TOTAL.inc();

        // Remove this region.
        let Some(flush_status) = self.region_status.remove(&region_id) else {
            return;
        };

        // Fast fail: cancels all pending tasks and sends error to their waiters.
        flush_status.on_failure(err);

        // Still tries to schedule a new flush.
        if let Err(e) = self.schedule_next_flush() {
            error!(e; "Failed to schedule next flush after region {} flush is failed", region_id);
        }
    }

    /// Notifies the scheduler that the region is dropped.
    pub(crate) fn on_region_dropped(&mut self, region_id: RegionId) {
        self.remove_region_on_failure(
            region_id,
            Arc::new(RegionDroppedSnafu { region_id }.build()),
        );
    }

    /// Notifies the scheduler that the region is closed.
    pub(crate) fn on_region_closed(&mut self, region_id: RegionId) {
        self.remove_region_on_failure(region_id, Arc::new(RegionClosedSnafu { region_id }.build()));
    }

    /// Notifies the scheduler that the region is truncated.
    pub(crate) fn on_region_truncated(&mut self, region_id: RegionId) {
        self.remove_region_on_failure(
            region_id,
            Arc::new(RegionTruncatedSnafu { region_id }.build()),
        );
    }

    fn remove_region_on_failure(&mut self, region_id: RegionId, err: Arc<Error>) {
        // Remove this region.
        let Some(flush_status) = self.region_status.remove(&region_id) else {
            return;
        };

        // Notifies all pending tasks.
        flush_status.on_failure(err);
    }

    /// Add ddl request to pending queue.
    ///
    /// # Panics
    /// Panics if region didn't request flush.
    pub(crate) fn add_ddl_request_to_pending(&mut self, request: SenderDdlRequest) {
        let status = self.region_status.get_mut(&request.region_id).unwrap();
        status.pending_ddls.push(request);
    }

    /// Add write request to pending queue.
    ///
    /// # Panics
    /// Panics if region didn't request flush.
    pub(crate) fn add_write_request_to_pending(&mut self, request: SenderWriteRequest) {
        let status = self
            .region_status
            .get_mut(&request.request.region_id)
            .unwrap();
        status.pending_writes.push(request);
    }

    /// Add bulk write request to pending queue.
    ///
    /// # Panics
    /// Panics if region didn't request flush.
    pub(crate) fn add_bulk_request_to_pending(&mut self, request: SenderBulkRequest) {
        let status = self.region_status.get_mut(&request.region_id).unwrap();
        status.pending_bulk_writes.push(request);
    }

    /// Returns true if the region has pending DDLs.
    pub(crate) fn has_pending_ddls(&self, region_id: RegionId) -> bool {
        self.region_status
            .get(&region_id)
            .map(|status| !status.pending_ddls.is_empty())
            .unwrap_or(false)
    }

    /// Schedules a new flush task when the scheduler can submit next task.
    pub(crate) fn schedule_next_flush(&mut self) -> Result<()> {
        debug_assert!(
            self.region_status
                .values()
                .all(|status| status.flushing || status.pending_task.is_some())
        );

        // Get the first region from status map.
        let Some(flush_status) = self
            .region_status
            .values_mut()
            .find(|status| status.pending_task.is_some())
        else {
            return Ok(());
        };
        debug_assert!(!flush_status.flushing);
        let task = flush_status.pending_task.take().unwrap();
        let region_id = flush_status.region_id;
        let version_control = flush_status.version_control.clone();

        self.schedule_flush(region_id, &version_control, task)
    }
}

impl Drop for FlushScheduler {
    fn drop(&mut self) {
        for (region_id, flush_status) in self.region_status.drain() {
            // We are shutting down so notify all pending tasks.
            flush_status.on_failure(Arc::new(RegionClosedSnafu { region_id }.build()));
        }
    }
}

/// Flush status of a region scheduled by the [FlushScheduler].
///
/// Tracks running and pending flush tasks and all pending requests of a region.
struct FlushStatus {
    /// Current region.
    region_id: RegionId,
    /// Version control of the region.
    version_control: VersionControlRef,
    /// There is a flush task running.
    ///
    /// It is possible that a region is not flushing but has pending task if the scheduler
    /// doesn't schedules this region.
    flushing: bool,
    /// Task waiting for next flush.
    pending_task: Option<RegionFlushTask>,
    /// Pending ddl requests.
    pending_ddls: Vec<SenderDdlRequest>,
    /// Requests waiting to write after altering the region.
    pending_writes: Vec<SenderWriteRequest>,
    /// Bulk requests waiting to write after altering the region.
    pending_bulk_writes: Vec<SenderBulkRequest>,
}

impl FlushStatus {
    fn new(region_id: RegionId, version_control: VersionControlRef) -> FlushStatus {
        FlushStatus {
            region_id,
            version_control,
            flushing: false,
            pending_task: None,
            pending_ddls: Vec::new(),
            pending_writes: Vec::new(),
            pending_bulk_writes: Vec::new(),
        }
    }

    /// Merges the task to pending task.
    fn merge_task(&mut self, task: RegionFlushTask) {
        if let Some(pending) = &mut self.pending_task {
            pending.merge(task);
        } else {
            self.pending_task = Some(task);
        }
    }

    fn on_failure(self, err: Arc<Error>) {
        if let Some(mut task) = self.pending_task {
            task.on_failure(err.clone());
        }
        for ddl in self.pending_ddls {
            ddl.sender.send(Err(err.clone()).context(FlushRegionSnafu {
                region_id: self.region_id,
            }));
        }
        for write_req in self.pending_writes {
            write_req
                .sender
                .send(Err(err.clone()).context(FlushRegionSnafu {
                    region_id: self.region_id,
                }));
        }
    }
}

#[cfg(test)]
mod tests {
    use mito_codec::row_converter::build_primary_key_codec;
    use tokio::sync::oneshot;

    use super::*;
    use crate::cache::CacheManager;
    use crate::memtable::bulk::part::BulkPartConverter;
    use crate::memtable::time_series::TimeSeriesMemtableBuilder;
    use crate::memtable::{Memtable, RangesOptions};
    use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
    use crate::test_util::memtable_util::{build_key_values_with_ts_seq_values, metadata_for_test};
    use crate::test_util::scheduler_util::{SchedulerEnv, VecScheduler};
    use crate::test_util::version_util::{VersionControlBuilder, write_rows_to_version};

    #[test]
    fn test_get_mutable_limit() {
        assert_eq!(4, WriteBufferManagerImpl::get_mutable_limit(8));
        assert_eq!(5, WriteBufferManagerImpl::get_mutable_limit(10));
        assert_eq!(32, WriteBufferManagerImpl::get_mutable_limit(64));
        assert_eq!(0, WriteBufferManagerImpl::get_mutable_limit(0));
    }

    #[test]
    fn test_over_mutable_limit() {
        // Mutable limit is 500.
        let manager = WriteBufferManagerImpl::new(1000);
        manager.reserve_mem(400);
        assert!(!manager.should_flush_engine());
        assert!(!manager.should_stall());

        // More than mutable limit.
        manager.reserve_mem(400);
        assert!(manager.should_flush_engine());

        // Freezes mutable.
        manager.schedule_free_mem(400);
        assert!(!manager.should_flush_engine());
        assert_eq!(800, manager.memory_used.load(Ordering::Relaxed));
        assert_eq!(400, manager.memory_active.load(Ordering::Relaxed));

        // Releases immutable.
        manager.free_mem(400);
        assert_eq!(400, manager.memory_used.load(Ordering::Relaxed));
        assert_eq!(400, manager.memory_active.load(Ordering::Relaxed));
    }

    #[test]
    fn test_over_global() {
        // Mutable limit is 500.
        let manager = WriteBufferManagerImpl::new(1000);
        manager.reserve_mem(1100);
        assert!(manager.should_stall());
        // Global usage is still 1100.
        manager.schedule_free_mem(200);
        assert!(manager.should_flush_engine());

        // More than global limit, but mutable (1100-200-450=450) is not enough (< 500).
        manager.schedule_free_mem(450);
        assert!(!manager.should_flush_engine());

        // Now mutable is enough.
        manager.reserve_mem(50);
        assert!(manager.should_flush_engine());
        manager.reserve_mem(100);
        assert!(manager.should_flush_engine());
    }

    #[test]
    fn test_manager_notify() {
        let (sender, receiver) = watch::channel(());
        let manager = WriteBufferManagerImpl::new(1000).with_notifier(sender);
        manager.reserve_mem(500);
        assert!(!receiver.has_changed().unwrap());
        manager.schedule_free_mem(500);
        assert!(!receiver.has_changed().unwrap());
        manager.free_mem(500);
        assert!(receiver.has_changed().unwrap());
    }

    #[tokio::test]
    async fn test_schedule_empty() {
        let env = SchedulerEnv::new().await;
        let (tx, _rx) = mpsc::channel(4);
        let mut scheduler = env.mock_flush_scheduler();
        let builder = VersionControlBuilder::new();

        let version_control = Arc::new(builder.build());
        let (output_tx, output_rx) = oneshot::channel();
        let mut task = RegionFlushTask {
            region_id: builder.region_id(),
            reason: FlushReason::Others,
            senders: Vec::new(),
            request_sender: tx,
            access_layer: env.access_layer.clone(),
            listener: WorkerListener::default(),
            engine_config: Arc::new(MitoConfig::default()),
            row_group_size: None,
            cache_manager: Arc::new(CacheManager::default()),
            manifest_ctx: env
                .mock_manifest_context(version_control.current().version.metadata.clone())
                .await,
            index_options: IndexOptions::default(),
            flush_semaphore: Arc::new(Semaphore::new(2)),
        };
        task.push_sender(OptionOutputTx::from(output_tx));
        scheduler
            .schedule_flush(builder.region_id(), &version_control, task)
            .unwrap();
        assert!(scheduler.region_status.is_empty());
        let output = output_rx.await.unwrap().unwrap();
        assert_eq!(output, 0);
        assert!(scheduler.region_status.is_empty());
    }

    #[tokio::test]
    async fn test_schedule_pending_request() {
        let job_scheduler = Arc::new(VecScheduler::default());
        let env = SchedulerEnv::new().await.scheduler(job_scheduler.clone());
        let (tx, _rx) = mpsc::channel(4);
        let mut scheduler = env.mock_flush_scheduler();
        let mut builder = VersionControlBuilder::new();
        // Overwrites the empty memtable builder.
        builder.set_memtable_builder(Arc::new(TimeSeriesMemtableBuilder::default()));
        let version_control = Arc::new(builder.build());
        // Writes data to the memtable so it is not empty.
        let version_data = version_control.current();
        write_rows_to_version(&version_data.version, "host0", 0, 10);
        let manifest_ctx = env
            .mock_manifest_context(version_data.version.metadata.clone())
            .await;
        // Creates 3 tasks.
        let mut tasks: Vec<_> = (0..3)
            .map(|_| RegionFlushTask {
                region_id: builder.region_id(),
                reason: FlushReason::Others,
                senders: Vec::new(),
                request_sender: tx.clone(),
                access_layer: env.access_layer.clone(),
                listener: WorkerListener::default(),
                engine_config: Arc::new(MitoConfig::default()),
                row_group_size: None,
                cache_manager: Arc::new(CacheManager::default()),
                manifest_ctx: manifest_ctx.clone(),
                index_options: IndexOptions::default(),
                flush_semaphore: Arc::new(Semaphore::new(2)),
            })
            .collect();
        // Schedule first task.
        let task = tasks.pop().unwrap();
        scheduler
            .schedule_flush(builder.region_id(), &version_control, task)
            .unwrap();
        // Should schedule 1 flush.
        assert_eq!(1, scheduler.region_status.len());
        assert_eq!(1, job_scheduler.num_jobs());
        // Check the new version.
        let version_data = version_control.current();
        assert_eq!(0, version_data.version.memtables.immutables()[0].id());
        // Schedule remaining tasks.
        let output_rxs: Vec<_> = tasks
            .into_iter()
            .map(|mut task| {
                let (output_tx, output_rx) = oneshot::channel();
                task.push_sender(OptionOutputTx::from(output_tx));
                scheduler
                    .schedule_flush(builder.region_id(), &version_control, task)
                    .unwrap();
                output_rx
            })
            .collect();
        // Assumes the flush job is finished.
        version_control.apply_edit(
            Some(RegionEdit {
                files_to_add: Vec::new(),
                files_to_remove: Vec::new(),
                timestamp_ms: None,
                compaction_time_window: None,
                flushed_entry_id: None,
                flushed_sequence: None,
                committed_sequence: None,
            }),
            &[0],
            builder.file_purger(),
        );
        scheduler.on_flush_success(builder.region_id());
        // No new flush task.
        assert_eq!(1, job_scheduler.num_jobs());
        // The flush status is cleared.
        assert!(scheduler.region_status.is_empty());
        for output_rx in output_rxs {
            let output = output_rx.await.unwrap().unwrap();
            assert_eq!(output, 0);
        }
    }

    // Verifies single-range flat flush path respects append_mode (no dedup) vs dedup when disabled.
    #[test]
    fn test_memtable_flat_sources_single_range_append_mode_behavior() {
        // Build test metadata and flat schema
        let metadata = metadata_for_test();
        let schema = to_flat_sst_arrow_schema(
            &metadata,
            &FlatSchemaOptions::from_encoding(metadata.primary_key_encoding),
        );

        // Prepare a bulk part containing duplicate rows for the same PK and timestamp
        // Two rows with identical keys and timestamps (ts = 1000), different field values
        let capacity = 16;
        let pk_codec = build_primary_key_codec(&metadata);
        let mut converter =
            BulkPartConverter::new(&metadata, schema.clone(), capacity, pk_codec, true);
        let kvs = build_key_values_with_ts_seq_values(
            &metadata,
            "dup_key".to_string(),
            1,
            vec![1000i64, 1000i64].into_iter(),
            vec![Some(1.0f64), Some(2.0f64)].into_iter(),
            1,
        );
        converter.append_key_values(&kvs).unwrap();
        let part = converter.convert().unwrap();

        // Helper to build MemtableRanges with a single range from one bulk part.
        // We use BulkMemtable directly because it produces record batch iterators.
        let build_ranges = |append_mode: bool| -> MemtableRanges {
            let memtable = crate::memtable::bulk::BulkMemtable::new(
                1,
                metadata.clone(),
                None,
                None,
                append_mode,
                MergeMode::LastRow,
            );
            memtable.write_bulk(part.clone()).unwrap();
            memtable.ranges(None, RangesOptions::for_flush()).unwrap()
        };

        // Case 1: append_mode = false => dedup happens, total rows should be 1
        {
            let mem_ranges = build_ranges(false);
            assert_eq!(1, mem_ranges.ranges.len());

            let options = RegionOptions {
                append_mode: false,
                merge_mode: Some(MergeMode::LastRow),
                ..Default::default()
            };

            let flat_sources = memtable_flat_sources(
                schema.clone(),
                mem_ranges,
                &options,
                metadata.primary_key.len(),
            )
            .unwrap();
            assert!(flat_sources.encoded.is_empty());
            assert_eq!(1, flat_sources.sources.len());

            // Consume the iterator and count rows
            let mut total_rows = 0usize;
            for source in flat_sources.sources {
                match source {
                    crate::read::FlatSource::Iter(iter) => {
                        for rb in iter {
                            total_rows += rb.unwrap().num_rows();
                        }
                    }
                    crate::read::FlatSource::Stream(_) => unreachable!(),
                }
            }
            assert_eq!(1, total_rows, "dedup should keep a single row");
        }

        // Case 2: append_mode = true => no dedup, total rows should be 2
        {
            let mem_ranges = build_ranges(true);
            assert_eq!(1, mem_ranges.ranges.len());

            let options = RegionOptions {
                append_mode: true,
                ..Default::default()
            };

            let flat_sources =
                memtable_flat_sources(schema, mem_ranges, &options, metadata.primary_key.len())
                    .unwrap();
            assert!(flat_sources.encoded.is_empty());
            assert_eq!(1, flat_sources.sources.len());

            let mut total_rows = 0usize;
            for source in flat_sources.sources {
                match source {
                    crate::read::FlatSource::Iter(iter) => {
                        for rb in iter {
                            total_rows += rb.unwrap().num_rows();
                        }
                    }
                    crate::read::FlatSource::Stream(_) => unreachable!(),
                }
            }
            assert_eq!(2, total_rows, "append_mode should preserve duplicates");
        }
    }
}