ailake_query/

compaction.rs

// SPDX-License-Identifier: MIT OR Apache-2.0
use std::sync::Arc;
use tracing::{debug, error, info, warn};

use ailake_catalog::{
    make_data_file_entry, make_data_file_entry_indexing, CatalogProvider, DataFileEntry,
    NewSnapshot, SnapshotOperation, TableIdent, VectorIndexInfo,
};
use ailake_core::{AilakeResult, RowId, VectorStoragePolicy};
use ailake_file::{AilakeFileReader, AilakeFileWriter};
use ailake_store::Store;
use ailake_vec::compute_centroid_and_radius;
use arrow_array::RecordBatch;
use arrow_schema::SchemaRef;
use bytes::Bytes;
use futures::future::try_join_all;

use crate::writer::build_and_patch_index;

/// Index strategy for the merged file produced by compaction.
#[derive(Debug, Clone, Default)]
pub enum CompactionIndexStrategy {
    /// Detect GPU / CPU cores at compaction time and pick the best index.
    /// IVF-PQ on GPU/many-core machines; HNSW elsewhere. (default)
    #[default]
    Auto,
    /// Always rebuild with HNSW — highest recall, larger index.
    ForceHnsw,
    /// Always rebuild with IVF-PQ — smaller index, better S3 throughput.
    ///
    /// Recommended for large compactions (N > 100 000) on CPU-only machines
    /// where HNSW rebuild cost becomes prohibitive.
    ForceIvfPq,
}

#[derive(Debug, Clone)]
pub struct CompactionConfig {
    /// Trigger compaction only if at least this many files are eligible.
    pub min_files_to_compact: usize,
    /// Target output file size in bytes. Files below this are merged.
    pub target_file_size_bytes: u64,
    /// Index algorithm for the merged output file.
    pub index_strategy: CompactionIndexStrategy,
    /// Maximum files merged in a single compaction pass.
    ///
    /// Candidates are sorted smallest-first; only the first `max_files_per_pass`
    /// are compacted each run. This bounds peak RAM and HNSW rebuild CPU cost —
    /// O(N log N) stays proportional to this limit rather than table size.
    /// Default: 20. Set to `usize::MAX` to compact all eligible files at once.
    pub max_files_per_pass: usize,
}

impl Default for CompactionConfig {
    fn default() -> Self {
        Self {
            min_files_to_compact: 4,
            target_file_size_bytes: 128 * 1024 * 1024, // 128 MB
            index_strategy: CompactionIndexStrategy::Auto,
            max_files_per_pass: 20,
        }
    }
}

#[derive(Debug, Clone, Copy)]
pub enum CompactionMode {
    Full,    // compact all files below target size
    Partial, // compact the smallest N files
}

pub struct CompactionPlanner {
    config: CompactionConfig,
}

impl CompactionPlanner {
    pub fn new(config: CompactionConfig) -> Self {
        Self { config }
    }

    /// Select files to compact.
    ///
    /// Picks files smaller than `target_file_size_bytes`, sorts them smallest-first
    /// (cheapest to read), and caps the selection at `max_files_per_pass`. This
    /// tiered approach prevents a single pass from compacting the entire table into
    /// memory when thousands of small files exist.
    ///
    /// Foreign-write detection: every file written by the AI-Lake SDK always carries a
    /// `centroid_b64` (see `make_data_file_entry`/`make_data_file_entry_indexing` —
    /// centroid is computed and stored even for `IndexStatus::Indexing` files, before
    /// the HNSW build itself completes). A manifest entry with no centroid was never
    /// produced by AI-Lake — almost certainly a generic Iceberg engine (Spark/Trino
    /// `OPTIMIZE` / `rewrite_data_files`, DuckDB) rewrote the file with no knowledge of
    /// AI-Lake. Every query against it silently degrades to an O(N) flat scan (see the
    /// `flat_scan_unexpected` warning in `scanner.rs::search`). These files bypass both
    /// the size filter and `min_files_to_compact` — a single such file is worth
    /// repairing on its own; it shouldn't wait for enough small files to accumulate.
    pub fn plan(&self, files: &[DataFileEntry]) -> Vec<DataFileEntry> {
        // Single pass: split into foreign (no AI-Lake index) vs. native candidates.
        let (mut foreign, natives): (Vec<DataFileEntry>, Vec<DataFileEntry>) =
            files.iter().cloned().partition(DataFileEntry::is_foreign);
        if !foreign.is_empty() {
            warn!(
                "ailake: compaction plan — {} file(s) with no AI-Lake index detected \
                 (likely a foreign/external rewrite) — prioritizing for reindex: {:?}",
                foreign.len(),
                foreign.iter().map(|f| f.path.as_str()).collect::<Vec<_>>()
            );
        }

        let mut size_candidates: Vec<DataFileEntry> = natives
            .into_iter()
            .filter(|f| f.file_size_bytes < self.config.target_file_size_bytes)
            .collect();

        if size_candidates.len() < self.config.min_files_to_compact {
            if foreign.is_empty() {
                debug!(
                    "ailake: compaction skipped — {} eligible files < min_files_to_compact={}",
                    size_candidates.len(),
                    self.config.min_files_to_compact
                );
                return vec![];
            }
            // Not enough small files to justify a size-based pass on their own, but
            // foreign files still need repair regardless of batching thresholds.
            size_candidates.clear();
        }

        // Sort both smallest-first so each pass handles the cheapest files first —
        // this bounds peak RAM to max_files_per_pass * avg_selected_file_size. Foreign
        // files are NOT size-filtered (repair takes priority over the size threshold —
        // a foreign file may itself be large, e.g. after an external OPTIMIZE), but they
        // ARE size-sorted so that if there are more foreign files than max_files_per_pass,
        // this pass still prefers the cheapest ones to merge rather than an arbitrary subset.
        foreign.sort_unstable_by_key(|f| f.file_size_bytes);
        size_candidates.sort_unstable_by_key(|f| f.file_size_bytes);
        // Foreign files take priority; size-based candidates fill the rest of the pass.
        foreign.extend(size_candidates);
        foreign.truncate(self.config.max_files_per_pass);
        let candidates = foreign;

        let total_bytes: u64 = candidates.iter().map(|f| f.file_size_bytes).sum();
        info!(
            "ailake: compaction plan — {} files ({} bytes) → 1 merged file",
            candidates.len(),
            total_bytes
        );
        candidates
    }
}

/// Executes compaction plans: reads N small files, merges them into a single
/// AI-Lake file with a rebuilt index, and commits to the catalog.
///
/// The index algorithm is chosen via `CompactionIndexStrategy` (default: `Auto`,
/// which detects GPU / CPU cores at compaction time — the same heuristic used
/// by `write_batch_auto`).
///
/// For large tables use `compact_deferred` / `run_deferred`: the merged Parquet
/// is persisted immediately and the HNSW build runs in a background Tokio task,
/// decoupling I/O cost from CPU cost.
#[derive(Clone)]
pub struct CompactionExecutor {
    store: Arc<dyn Store>,
    policy: VectorStoragePolicy,
    index_strategy: CompactionIndexStrategy,
    /// When set, rebuilds and embeds a Tantivy FTS index in the compacted output file.
    fts_config: Option<ailake_fts::FtsConfig>,
}

impl CompactionExecutor {
    pub fn new(store: Arc<dyn Store>, policy: VectorStoragePolicy) -> Self {
        Self {
            store,
            policy,
            index_strategy: CompactionIndexStrategy::Auto,
            fts_config: None,
        }
    }

    /// Override the default (Auto) index strategy for this executor.
    pub fn with_index_strategy(mut self, strategy: CompactionIndexStrategy) -> Self {
        self.index_strategy = strategy;
        self
    }

    /// Rebuild and embed a Tantivy FTS index in the compacted output file.
    pub fn with_fts_config(mut self, cfg: ailake_fts::FtsConfig) -> Self {
        self.fts_config = Some(cfg);
        self
    }

    /// Return a clone of this executor whose `fts_config` is filled from table properties
    /// when the caller did not explicitly call `with_fts_config`.
    ///
    /// This is the auto-detect path: if `ailake.fts.enabled=true` is present in the table
    /// metadata but the operator forgot (or never knew to) set an FTS config, compaction
    /// would silently drop the FTS index. This method prevents that loss.
    fn with_effective_fts(
        &self,
        table_props: &std::collections::HashMap<String, String>,
    ) -> std::borrow::Cow<'_, Self> {
        if self.fts_config.is_some() {
            return std::borrow::Cow::Borrowed(self);
        }
        match ailake_fts::FtsConfig::from_table_props(table_props) {
            Some(cfg) => {
                let mut cloned = self.clone();
                cloned.fts_config = Some(cfg);
                std::borrow::Cow::Owned(cloned)
            }
            None => std::borrow::Cow::Borrowed(self),
        }
    }

    /// Read all input files in parallel, returning ordered (batch, embeddings) pairs.
    ///
    /// `read_parquet()` decodes the vector column straight from Parquet and never
    /// touches the AILK footer, so a file missing its index (external-engine rewrite,
    /// or still `IndexStatus::Indexing`) still holds real data — it must be read like
    /// any other input, not dropped. The merged output rebuilds a fresh index for
    /// every row regardless, so a missing per-file index costs nothing here.
    async fn read_files_parallel(
        &self,
        files: &[DataFileEntry],
    ) -> AilakeResult<Vec<(RecordBatch, Vec<Vec<f32>>)>> {
        let futs = files.iter().map(|entry| {
            let store = self.store.clone();
            let path = entry.path.clone();
            let column = self.policy.column_name.clone();
            let dim = self.policy.dim;
            let dv = entry.deletion_vector.clone();
            async move {
                let bytes: Bytes = store.get(&path).await?;
                let reader = AilakeFileReader::new(bytes, &column, dim);
                if !reader.is_ailake_file() {
                    debug!(
                        "ailake: compaction reading {} without an AI-Lake index \
                         (external write or indexing in progress) — data still merged",
                        path
                    );
                }
                let (batch, embeddings) = reader.read_parquet()?;
                // Drop DV-masked rows before they're merged into a new physical file —
                // row positions are about to change, so the old bitmap can't just be
                // carried forward (see dv::filter_deleted_rows).
                let pair = if let Some(dv) = dv {
                    let bitmap = crate::dv::load_deletion_vector(&store, &dv).await?;
                    crate::dv::filter_deleted_rows(batch, embeddings, &bitmap)?
                } else {
                    (batch, embeddings)
                };
                Ok::<(RecordBatch, Vec<Vec<f32>>), ailake_core::AilakeError>(pair)
            }
        });
        try_join_all(futs).await
    }

    /// Merge `files` into a single new file at `output_path`.
    ///
    /// Reads all input files **in parallel** to minimise S3 latency, then
    /// rebuilds the HNSW / IVF-PQ index synchronously. For very large merges
    /// (N > 100 000 vectors) prefer `compact_deferred`, which offloads the
    /// index build to a background Tokio task.
    ///
    /// Returns the DataFileEntry for the merged file.
    pub async fn compact(
        &self,
        files: &[DataFileEntry],
        output_path: &str,
    ) -> AilakeResult<DataFileEntry> {
        if files.is_empty() {
            return Err(ailake_core::AilakeError::Catalog(
                "compact: no files provided".into(),
            ));
        }

        // `read_files_parallel` always returns exactly one pair per input file (never
        // filters), so `pairs.len() == files.len() > 0` given the `files.is_empty()` guard
        // above — no separate emptiness check needed here.
        let pairs = self.read_files_parallel(files).await?;

        let schema: SchemaRef = pairs[0].0.schema();
        let (all_batches, all_embeddings): (Vec<_>, Vec<_>) = pairs.into_iter().unzip();
        let all_embeddings: Vec<Vec<f32>> = all_embeddings.into_iter().flatten().collect();

        // Concatenate all row groups into one batch
        let merged_batch = concat_batches(schema, &all_batches)?;
        let record_count = merged_batch.num_rows() as u64;

        // Write merged file with adaptive index selection.
        let writer = {
            let base = AilakeFileWriter::new(self.policy.clone());
            let base = match &self.index_strategy {
                CompactionIndexStrategy::Auto => base.with_auto_index(),
                CompactionIndexStrategy::ForceHnsw => base,
                CompactionIndexStrategy::ForceIvfPq => {
                    let cfg = ailake_index::IvfPqConfig::for_dataset(
                        self.policy.dim as usize,
                        all_embeddings.len(),
                    );
                    base.with_ivf_pq(cfg)
                }
            };
            if let Some(ref fts_cfg) = self.fts_config {
                match ailake_fts::merge_fts_blobs(fts_cfg, &merged_batch) {
                    Ok(blob) => base.with_prebuilt_fts_blob(blob),
                    Err(e) => {
                        warn!("ailake: FTS re-index during compaction failed: {e}");
                        base
                    }
                }
            } else {
                base
            }
        };
        let file_bytes = writer.write(&merged_batch, &all_embeddings)?;
        let file_size = file_bytes.len() as u64;
        let column_stats =
            ailake_catalog::extract_column_stats(&file_bytes, &[self.policy.column_name.as_str()])
                .and_then(|m| serde_json::to_string(&m).ok());
        self.store.put(output_path, file_bytes.clone()).await?;

        // Compute centroid and HNSW offsets for catalog entry
        let centroid = compute_centroid_and_radius(&all_embeddings, self.policy.metric);
        let reader = AilakeFileReader::new(file_bytes, &self.policy.column_name, self.policy.dim);
        let header = reader.read_header()?;
        let ailk_start = reader.ailk_offset()?;

        // Positional invariant check: parquet_count == hnsw_node_count == header.record_count.
        // Catches a mismatched merge before it's committed to the catalog rather than
        // surfacing as a wrong search result (or the invariant-violated error path in
        // scanner.rs) later.
        reader.verify_integrity()?;

        // Preserve row-ID continuity: merged file inherits the minimum first_row_id of
        // its sources so commit_snapshot doesn't allocate fresh IDs and grow next_row_id.
        let source_first_row_id = files.iter().filter_map(|f| f.first_row_id).min();

        let mut entry = make_data_file_entry(
            output_path,
            record_count,
            file_size,
            &centroid,
            VectorIndexInfo {
                column: &self.policy.column_name,
                dim: self.policy.dim,
                hnsw_offset: ailk_start + header.hnsw_offset,
                hnsw_len: header.hnsw_len,
            },
        );
        entry.first_row_id = source_first_row_id;
        // Preserve idempotency keys: a retry of a source write must still see itself
        // as already-committed after this merge (see `DataFileEntry::merge_batch_ids`
        // and `write_batch_idempotent`).
        entry.batch_id = DataFileEntry::merge_batch_ids(files);
        entry.column_stats = column_stats;
        Ok(entry)
    }

    /// Merge `files` into a single new file using incremental HNSW insertion.
    ///
    /// Identifies the **dominant file** — the file holding >= 40 % of the total
    /// row count — loads its existing HNSW graph from the AILK section, then
    /// calls `HnswIndex::insert_node` for every vector from the remaining files.
    ///
    /// **Complexity vs `compact`**:
    /// - Full rebuild: O(N log N), N = total rows.
    /// - Incremental (this method): O(N_dom) deserialization + O(N_small × log N_dom).
    ///   For a 90 / 10 split (N = 1 M, N_dom = 900 k) the speedup is ~7×.
    ///
    /// **Fallbacks** (all degrade gracefully to `compact`):
    /// - No file holds >= 40 % of rows.
    /// - Dominant file's HNSW cannot be loaded (IVF-PQ, `IndexStatus::Indexing`, corrupt).
    ///
    /// **RowId contract**: dominant file's vectors are placed first in the merged
    /// Parquet (positions 0..N_dom-1); other files follow. The existing RowIds from
    /// the dominant HNSW remain valid; new nodes receive RowIds N_dom..N-1.
    pub async fn compact_incremental(
        &self,
        files: &[DataFileEntry],
        output_path: &str,
    ) -> AilakeResult<DataFileEntry> {
        const DOMINANT_RATIO: f64 = 0.40;

        if files.is_empty() {
            return Err(ailake_core::AilakeError::Catalog(
                "compact_incremental: no files provided".into(),
            ));
        }

        // This method only ever produces an HNSW index (dominant-file graph extended via
        // `insert_node`) — it never builds IVF-PQ. `ForceIvfPq` is an explicit caller
        // request for a specific index type, so it must never be silently satisfied with
        // HNSW instead; fall back to `compact()`, which does respect it. `Auto`/`ForceHnsw`
        // are unaffected: `Auto` treats "the dominant file already has a valid HNSW"
        // as a legitimate reason to keep reusing HNSW rather than pay for a fresh
        // hardware-picked rebuild, and `ForceHnsw` is satisfied by this method by
        // construction either way.
        if matches!(self.index_strategy, CompactionIndexStrategy::ForceIvfPq) {
            debug!(
                "ailake: compact_incremental — index_strategy=ForceIvfPq, which this method \
                 can never produce (HNSW graph extension only); falling back to full rebuild"
            );
            return self.compact(files, output_path).await;
        }

        // Find the dominant file by record_count.
        let total_rows: u64 = files.iter().map(|f| f.record_count).sum();
        let dom_idx = files
            .iter()
            .enumerate()
            .max_by_key(|(_, f)| f.record_count)
            .map(|(i, _)| i)
            .unwrap_or(0);
        let dom_rows = files[dom_idx].record_count;

        // The dominant file's existing HNSW graph is reused as-is (only non-dominant
        // vectors are inserted into it) — its node IDs are tied to the dominant file's
        // current row positions. If it has DV-masked rows, filtering them out of the
        // Parquet data would desync the reused graph's node-to-row mapping (and HNSW
        // doesn't support cheap node removal). Fall back to a full rebuild, which
        // filters correctly via `read_files_parallel`.
        if files[dom_idx].deletion_vector.is_some() {
            debug!(
                "ailake: compact_incremental — dominant file {} has DV-masked rows, \
                 falling back to full rebuild (graph reuse would desync row positions)",
                files[dom_idx].path
            );
            return self.compact(files, output_path).await;
        }

        if (dom_rows as f64 / total_rows as f64) < DOMINANT_RATIO {
            debug!(
                "ailake: compact_incremental — no dominant file ({}/{} rows < {:.0}% threshold), \
                 falling back to full rebuild",
                dom_rows,
                total_rows,
                DOMINANT_RATIO * 100.0
            );
            return self.compact(files, output_path).await;
        }

        let column = self.policy.column_name.clone();
        let dim = self.policy.dim;
        let dom_path = files[dom_idx].path.clone();

        // Read all files in parallel. `read_parquet()` decodes the vector column
        // directly from Parquet and never touches the AILK footer, so every file's
        // rows are read regardless of whether it has an AI-Lake index — a file
        // missing its footer (external-engine rewrite, or still `IndexStatus::Indexing`)
        // still holds real data and must not be dropped from the merge. Raw bytes are
        // retained only for the dominant file, and only when it actually has an index
        // to reuse (needed to load its HNSW without a second round-trip); otherwise the
        // dominant-file match below falls back to a full rebuild via `compact()`.
        let futs: Vec<_> =
            files
                .iter()
                .map(|entry| {
                    let store = self.store.clone();
                    let path = entry.path.clone();
                    let col = column.clone();
                    let is_dom = path == dom_path;
                    let dv = entry.deletion_vector.clone();
                    async move {
                        let bytes: Bytes = store.get(&path).await?;
                        let reader = AilakeFileReader::new(bytes.clone(), &col, dim);
                        let has_index = reader.is_ailake_file();
                        if is_dom && !has_index {
                            debug!(
                                "ailake: compact_incremental — dominant candidate {} has no \
                             AI-Lake index; will fall back to full rebuild if no HNSW to reuse",
                                path
                            );
                        }
                        let (raw_batch, raw_vecs) = reader.read_parquet()?;
                        // Non-dominant rows are freshly inserted into the graph below, so
                        // DV-masked rows must be dropped here (dominant is guaranteed
                        // DV-free by the caller's earlier fallback check).
                        let (batch, vecs) = if let Some(dv) = dv {
                            let bitmap = crate::dv::load_deletion_vector(&store, &dv).await?;
                            crate::dv::filter_deleted_rows(raw_batch, raw_vecs, &bitmap)?
                        } else {
                            (raw_batch, raw_vecs)
                        };
                        let retained = if is_dom && has_index {
                            Some(bytes)
                        } else {
                            None
                        };
                        Ok::<
                            (RecordBatch, Vec<Vec<f32>>, bool, Option<Bytes>),
                            ailake_core::AilakeError,
                        >((batch, vecs, is_dom, retained))
                    }
                })
                .collect();

        // Every future above always resolves to one tuple per input file (or propagates an
        // Err via `?`) — no per-file filtering, so no emptiness check is needed given the
        // `files.is_empty()` guard at the top of this function.
        #[allow(clippy::type_complexity)]
        let raw: Vec<(RecordBatch, Vec<Vec<f32>>, bool, Option<Bytes>)> =
            try_join_all(futs).await?;

        // Separate dominant from others; dominant goes first in the merged file.
        let mut dom_batch: Option<RecordBatch> = None;
        let mut dom_vecs: Vec<Vec<f32>> = Vec::new();
        let mut dom_bytes_found: Option<Bytes> = None;
        let mut other_batches: Vec<RecordBatch> = Vec::new();
        let mut other_vecs: Vec<Vec<f32>> = Vec::new();

        for (batch, vecs, is_dom, retained) in raw {
            if is_dom {
                dom_batch = Some(batch);
                dom_vecs = vecs;
                dom_bytes_found = retained;
            } else {
                other_batches.push(batch);
                other_vecs.extend(vecs);
            }
        }

        let (dom_batch, dom_bytes) = match (dom_batch, dom_bytes_found) {
            (Some(b), Some(byt)) => (b, byt),
            _ => {
                debug!(
                    "ailake: compact_incremental — dominant file missing from read results, \
                     falling back to full rebuild"
                );
                return self.compact(files, output_path).await;
            }
        };

        // Load the dominant file's existing HNSW graph.
        let dom_reader = AilakeFileReader::new(dom_bytes, &column, dim);
        let mut hnsw = match dom_reader.load_index() {
            Ok(idx) => idx,
            Err(e) => {
                debug!(
                    "ailake: compact_incremental — cannot load dominant HNSW ({}), \
                     falling back to full rebuild",
                    e
                );
                return self.compact(files, output_path).await;
            }
        };

        let dom_count = dom_batch.num_rows() as u64;

        // Insert vectors from non-dominant files into the loaded graph.
        // RowIds are assigned starting at dom_count to match positions in the merged Parquet.
        for (j, vec) in other_vecs.iter().enumerate() {
            hnsw.insert_node(RowId::new(dom_count + j as u64), vec.clone());
        }
        hnsw.quantize_to_f16();

        // Assemble merged batch (dominant rows first) and all embeddings.
        let schema: SchemaRef = dom_batch.schema();
        let mut all_batches = vec![dom_batch];
        all_batches.extend(other_batches);
        let merged_batch = concat_batches(schema, &all_batches)?;
        let record_count = merged_batch.num_rows() as u64;

        let mut all_embeddings = dom_vecs;
        all_embeddings.extend(other_vecs);

        // Write the merged file using the pre-built index (no rebuild).
        // Attach FTS blob when configured — data is already in merged_batch so cost is tokenization only.
        let writer = {
            let base = AilakeFileWriter::new(self.policy.clone());
            if let Some(ref fts_cfg) = self.fts_config {
                match ailake_fts::merge_fts_blobs(fts_cfg, &merged_batch) {
                    Ok(blob) => base.with_prebuilt_fts_blob(blob),
                    Err(e) => {
                        warn!("ailake: FTS re-index during incremental compaction failed: {e}");
                        base
                    }
                }
            } else {
                base
            }
        };
        let file_bytes = writer.write_with_prebuilt_hnsw(&merged_batch, &all_embeddings, &hnsw)?;
        let file_size = file_bytes.len() as u64;
        let column_stats =
            ailake_catalog::extract_column_stats(&file_bytes, &[self.policy.column_name.as_str()])
                .and_then(|m| serde_json::to_string(&m).ok());
        self.store.put(output_path, file_bytes.clone()).await?;

        let centroid = compute_centroid_and_radius(&all_embeddings, self.policy.metric);
        let reader = AilakeFileReader::new(file_bytes, &self.policy.column_name, self.policy.dim);
        let header = reader.read_header()?;
        let ailk_start = reader.ailk_offset()?;

        // Positional invariant check — see `compact()` for rationale. Especially relevant
        // here since the index is grown incrementally (insert_node) rather than rebuilt.
        reader.verify_integrity()?;

        // Dominant file goes first in the merged output, so the merged file's first
        // logical row was the dominant file's first row.  Use its first_row_id so
        // commit_snapshot doesn't grow next_row_id unnecessarily.
        let source_first_row_id = files[dom_idx].first_row_id;

        let mut entry = make_data_file_entry(
            output_path,
            record_count,
            file_size,
            &centroid,
            VectorIndexInfo {
                column: &self.policy.column_name,
                dim: self.policy.dim,
                hnsw_offset: ailk_start + header.hnsw_offset,
                hnsw_len: header.hnsw_len,
            },
        );
        entry.first_row_id = source_first_row_id;
        entry.batch_id = DataFileEntry::merge_batch_ids(files);
        entry.column_stats = column_stats;

        info!(
            "ailake: compact_incremental — merged {} files into {} \
             ({} rows from dominant + {} inserted incrementally)",
            files.len(),
            output_path,
            dom_count,
            record_count - dom_count
        );

        Ok(entry)
    }

    /// Merge `files` into a single new file at `output_path`, writing Parquet
    /// immediately and building the HNSW / IVF-PQ index in a background Tokio task.
    ///
    /// The merged file appears in the catalog as `IndexStatus::Indexing` until
    /// the background task completes; queries fall back to flat scan during that
    /// window (same behaviour as `write_batch_deferred`).
    ///
    /// Returns the `DataFileEntry` with `IndexStatus::Indexing`. The entry
    /// transitions to `Ready` automatically when the background build finishes.
    pub async fn compact_deferred(
        &self,
        files: &[DataFileEntry],
        output_path: &str,
        catalog: Arc<dyn CatalogProvider>,
        table: &TableIdent,
    ) -> AilakeResult<DataFileEntry> {
        if files.is_empty() {
            return Err(ailake_core::AilakeError::Catalog(
                "compact_deferred: no files provided".into(),
            ));
        }

        // See `compact()` above: read_files_parallel never filters, so no emptiness
        // check is needed given the `files.is_empty()` guard above.
        let pairs = self.read_files_parallel(files).await?;

        let schema: SchemaRef = pairs[0].0.schema();
        let (all_batches, all_embeddings): (Vec<_>, Vec<_>) = pairs.into_iter().unzip();
        let all_embeddings: Vec<Vec<f32>> = all_embeddings.into_iter().flatten().collect();

        let merged_batch = concat_batches(schema, &all_batches)?;
        let record_count = merged_batch.num_rows() as u64;

        // Write Parquet-only immediately — fast path, no HNSW build.
        let file_writer = AilakeFileWriter::new(self.policy.clone());
        let parquet_bytes = file_writer.write_parquet_only(&merged_batch, &all_embeddings)?;
        let file_size = parquet_bytes.len() as u64;
        let column_stats = ailake_catalog::extract_column_stats(
            &parquet_bytes,
            &[self.policy.column_name.as_str()],
        )
        .and_then(|m| serde_json::to_string(&m).ok());
        self.store.put(output_path, parquet_bytes).await?;

        // Centroid available for geometric pruning during the build window.
        let centroid = compute_centroid_and_radius(&all_embeddings, self.policy.metric);
        let source_first_row_id = files.iter().filter_map(|f| f.first_row_id).min();
        let mut entry = make_data_file_entry_indexing(
            output_path,
            record_count,
            file_size,
            &centroid,
            &self.policy.column_name,
            self.policy.dim,
        );
        entry.first_row_id = source_first_row_id;
        entry.batch_id = DataFileEntry::merge_batch_ids(files);
        entry.column_stats = column_stats;

        // Spawn background index build; errors are logged, not propagated.
        let store = self.store.clone();
        let policy = self.policy.clone();
        let table_id = table.clone();
        let fp = output_path.to_string();
        tokio::spawn(async move {
            if let Err(e) = build_and_patch_index(store, catalog, policy, table_id, fp).await {
                error!(
                    "ailake: compaction deferred HNSW build failed — file indexed as \
                     Parquet-only until next compaction rebuilds the index: {}",
                    e
                );
            }
        });

        Ok(entry)
    }

    /// Full compaction workflow: plan, compact (synchronous HNSW rebuild),
    /// drop old files from catalog, commit.
    pub async fn run(
        &self,
        planner: &CompactionPlanner,
        table: &TableIdent,
        catalog: Arc<dyn CatalogProvider>,
        output_prefix: &str,
    ) -> AilakeResult<Option<DataFileEntry>> {
        let all_files = catalog.list_files(table, None).await?;
        let to_compact = planner.plan(&all_files);
        if to_compact.is_empty() {
            return Ok(None);
        }

        // Auto-detect FTS from table metadata so compaction never silently drops an FTS index
        // that was present in the source files. Uses ailake.fts.* properties written at write time.
        let meta_props = catalog
            .load_table(table)
            .await
            .map(|m| m.properties)
            .unwrap_or_default();
        let executor = self.with_effective_fts(&meta_props);

        let ts = std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap_or_else(|e| e.duration())
            .as_millis();
        let output_path = format!("{output_prefix}/compacted-{ts}.parquet");

        // Use incremental merge when a dominant file exists (falls back to full rebuild automatically).
        let merged = executor
            .compact_incremental(&to_compact, &output_path)
            .await?;

        // Commit: add merged file, remove input files (via Replace snapshot).
        let files = build_replace_file_list(&catalog, table, &to_compact, merged.clone()).await?;
        // Fetched fresh, right before the commit — same freshness rationale as the
        // file-list re-list in `build_replace_file_list` above.
        let parent_snapshot_id = catalog
            .load_table(table)
            .await
            .ok()
            .and_then(|m| m.current_snapshot_id);
        let snapshot = NewSnapshot {
            snapshot_id: ailake_catalog::new_snapshot_id(),
            parent_snapshot_id,
            files,
            operation: SnapshotOperation::Replace,
            iceberg_schema: None,
            extra_properties: std::collections::HashMap::new(),
            bloom_filters: vec![],
            equality_delete_files: vec![],
        };
        catalog.commit_snapshot(table, snapshot).await?;

        info!(
            "ailake: compaction committed — merged {} files into {}",
            to_compact.len(),
            output_path
        );

        if catalog.retires_files_physically() {
            delete_old_files(&self.store, &to_compact).await;
        } else {
            info!(
                "ailake: compaction — leaving {} retired file(s) in place; catalog backend \
                 manages physical reclamation itself (see docs/guides/DUCKLAKE_CATALOG.md)",
                to_compact.len()
            );
        }

        Ok(Some(merged))
    }

    /// Full compaction workflow with deferred HNSW build: plan, write merged
    /// Parquet immediately, commit as `Indexing`, spawn background index build.
    ///
    /// Use for large tables where inline HNSW rebuild blocks too long.
    ///
    /// Note: FTS index is **not** rebuilt in deferred mode — `compact_deferred` writes
    /// Parquet-only immediately and the background task (`build_and_patch_index`) only
    /// builds the HNSW/IVF-PQ index. Use `run` (synchronous) when FTS preservation
    /// on compaction is required.
    pub async fn run_deferred(
        &self,
        planner: &CompactionPlanner,
        table: &TableIdent,
        catalog: Arc<dyn CatalogProvider>,
        output_prefix: &str,
    ) -> AilakeResult<Option<DataFileEntry>> {
        // The background index build patches the merged file in place at its
        // committed path — refuse on backends where committed bytes are
        // immutable (see TableWriter::ensure_deferred_supported for why a
        // commit-time guard alone is too late: the physical put comes first).
        if !catalog.supports_in_place_rewrite() {
            return Err(ailake_core::AilakeError::Catalog(
                "deferred compaction is not supported with this catalog backend: the \
                 background index build patches the merged file in place at its committed \
                 path, which this catalog cannot re-register — run a blocking compact"
                    .into(),
            ));
        }
        let all_files = catalog.list_files(table, None).await?;
        let to_compact = planner.plan(&all_files);
        if to_compact.is_empty() {
            return Ok(None);
        }

        let ts = std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap_or_else(|e| e.duration())
            .as_millis();
        let output_path = format!("{output_prefix}/compacted-{ts}.parquet");

        let merged = self
            .compact_deferred(&to_compact, &output_path, catalog.clone(), table)
            .await?;

        // Commit immediately: merged file in Indexing state replaces input files.
        let files = build_replace_file_list(&catalog, table, &to_compact, merged.clone()).await?;
        // Fetched fresh, right before the commit — same freshness rationale as the
        // file-list re-list in `build_replace_file_list` above.
        let parent_snapshot_id = catalog
            .load_table(table)
            .await
            .ok()
            .and_then(|m| m.current_snapshot_id);
        let snapshot = NewSnapshot {
            snapshot_id: ailake_catalog::new_snapshot_id(),
            parent_snapshot_id,
            files,
            operation: SnapshotOperation::Replace,
            iceberg_schema: None,
            extra_properties: std::collections::HashMap::new(),
            bloom_filters: vec![],
            equality_delete_files: vec![],
        };
        catalog.commit_snapshot(table, snapshot).await?;

        info!(
            "ailake: compaction committed (deferred) — merged {} files into {} \
             (index building in background)",
            to_compact.len(),
            output_path
        );

        if catalog.retires_files_physically() {
            delete_old_files(&self.store, &to_compact).await;
        } else {
            info!(
                "ailake: compaction — leaving {} retired file(s) in place; catalog backend \
                 manages physical reclamation itself (see docs/guides/DUCKLAKE_CATALOG.md)",
                to_compact.len()
            );
        }

        Ok(Some(merged))
    }
}

/// Builds the file list for a post-compaction `Replace` snapshot: the merged output
/// plus every current file that wasn't part of this compaction pass.
///
/// `Replace` does not inherit the previous manifest (see `HadoopCatalog::commit_snapshot`),
/// so untouched files (above `target_file_size_bytes`, or beyond `max_files_per_pass`) must
/// be carried forward explicitly or they vanish from the table. Re-lists via `list_files()`
/// right before this call returns (rather than reusing a pre-merge snapshot) to narrow —
/// though not eliminate, `HadoopCatalog` has no optimistic-concurrency check on
/// `commit_snapshot` — the window for a concurrent writer's commit to be silently dropped
/// by the `Replace` this list feeds into.
async fn build_replace_file_list(
    catalog: &Arc<dyn CatalogProvider>,
    table: &TableIdent,
    to_compact: &[DataFileEntry],
    merged: DataFileEntry,
) -> AilakeResult<Vec<DataFileEntry>> {
    let current_files = catalog.list_files(table, None).await?;
    let compacted_paths: std::collections::HashSet<&str> =
        to_compact.iter().map(|f| f.path.as_str()).collect();
    let mut files: Vec<DataFileEntry> = current_files
        .into_iter()
        .filter(|f| !compacted_paths.contains(f.path.as_str()))
        .collect();
    files.push(merged);
    Ok(files)
}

async fn delete_old_files(store: &Arc<dyn Store>, files: &[DataFileEntry]) {
    for entry in files {
        if let Err(e) = store.delete(&entry.path).await {
            error!(
                "ailake: compaction cleanup failed — could not delete {}: {} \
                 (orphan file in object store after successful catalog commit; \
                 delete manually to reclaim storage)",
                entry.path, e
            );
        }
    }
}

fn concat_batches(schema: SchemaRef, batches: &[RecordBatch]) -> AilakeResult<RecordBatch> {
    arrow_select::concat::concat_batches(&schema, batches)
        .map_err(|e| ailake_core::AilakeError::Arrow(e.to_string()))
}

#[cfg(test)]
mod tests {
    use super::*;
    use ailake_catalog::IndexStatus;

    #[test]
    fn plan_returns_empty_if_too_few_files() {
        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 4,
            target_file_size_bytes: 1024 * 1024,
            ..Default::default()
        });
        let files: Vec<DataFileEntry> = (0..3)
            .map(|i| DataFileEntry {
                path: format!("file-{i}.parquet"),
                record_count: 10,
                file_size_bytes: 100,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            })
            .collect();
        assert!(planner.plan(&files).is_empty());
    }

    #[test]
    fn plan_selects_small_files() {
        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 2,
            target_file_size_bytes: 1000,
            ..Default::default()
        });
        let files = vec![
            DataFileEntry {
                path: "small.parquet".into(),
                record_count: 5,
                file_size_bytes: 500,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "large.parquet".into(),
                record_count: 5000,
                file_size_bytes: 200_000_000,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "also-small.parquet".into(),
                record_count: 5,
                file_size_bytes: 800,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
        ];
        let selected = planner.plan(&files);
        assert_eq!(selected.len(), 2);
        assert!(selected.iter().any(|f| f.path == "small.parquet"));
        assert!(selected.iter().any(|f| f.path == "also-small.parquet"));
    }

    #[test]
    fn plan_respects_max_files_per_pass() {
        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 2,
            target_file_size_bytes: 1_000_000,
            max_files_per_pass: 3,
            ..Default::default()
        });
        let files: Vec<DataFileEntry> = (0..5)
            .map(|i| DataFileEntry {
                path: format!("f{i}.parquet"),
                record_count: 10,
                file_size_bytes: 100 + i as u64 * 100,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            })
            .collect();
        let selected = planner.plan(&files);
        assert_eq!(selected.len(), 3);
        assert_eq!(selected[0].file_size_bytes, 100);
        assert_eq!(selected[1].file_size_bytes, 200);
        assert_eq!(selected[2].file_size_bytes, 300);
    }

    #[test]
    fn plan_sorts_smallest_first() {
        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 2,
            target_file_size_bytes: 10_000,
            max_files_per_pass: 4,
            ..Default::default()
        });
        let files = vec![
            DataFileEntry {
                path: "c.parquet".into(),
                record_count: 1,
                file_size_bytes: 300,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "a.parquet".into(),
                record_count: 1,
                file_size_bytes: 100,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "b.parquet".into(),
                record_count: 1,
                file_size_bytes: 200,
                centroid_b64: Some("AAAA".into()),
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
        ];
        let selected = planner.plan(&files);
        assert_eq!(selected[0].file_size_bytes, 100);
        assert_eq!(selected[1].file_size_bytes, 200);
        assert_eq!(selected[2].file_size_bytes, 300);
    }

    fn make_plan_entry(path: &str, size: u64, centroid_b64: Option<String>) -> DataFileEntry {
        DataFileEntry {
            path: path.to_string(),
            record_count: 10,
            file_size_bytes: size,
            centroid_b64,
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: None,
            vector_dim: None,
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: None,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        }
    }

    #[test]
    fn plan_prioritizes_foreign_written_files_regardless_of_size() {
        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 4, // way more than the 1 native small file below
            target_file_size_bytes: 1000,
            max_files_per_pass: 20,
            ..Default::default()
        });
        let files = vec![
            // Native file, below target size, but alone — not enough to hit
            // min_files_to_compact on its own.
            make_plan_entry("native_small.parquet", 500, Some("AAAA".into())),
            // Foreign file — no centroid — large (would never pass the size filter),
            // but must still be selected because it has no AI-Lake index at all.
            make_plan_entry("foreign_big.parquet", 200_000_000, None),
        ];
        let selected = planner.plan(&files);
        assert_eq!(
            selected.len(),
            1,
            "foreign file alone should trigger a pass even below min_files_to_compact"
        );
        assert_eq!(selected[0].path, "foreign_big.parquet");
    }

    #[test]
    fn plan_merges_foreign_and_size_candidates_together() {
        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 2,
            target_file_size_bytes: 1000,
            max_files_per_pass: 20,
            ..Default::default()
        });
        let files = vec![
            make_plan_entry("native_a.parquet", 300, Some("AAAA".into())),
            make_plan_entry("native_b.parquet", 400, Some("AAAA".into())),
            make_plan_entry("foreign.parquet", 200_000_000, None),
        ];
        let selected = planner.plan(&files);
        let paths: Vec<&str> = selected.iter().map(|f| f.path.as_str()).collect();
        assert_eq!(selected.len(), 3, "paths={paths:?}");
        // Foreign file must be first (repair priority).
        assert_eq!(selected[0].path, "foreign.parquet", "paths={paths:?}");
    }

    #[test]
    fn plan_sorts_foreign_files_smallest_first_too() {
        // Regression: foreign files used to keep list_files()'s arbitrary order, so a
        // pass could select several large foreign files instead of the cheapest ones —
        // violating the documented "bounds peak RAM" invariant of max_files_per_pass.
        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 1,
            target_file_size_bytes: 1000,
            max_files_per_pass: 2,
            ..Default::default()
        });
        let files = vec![
            make_plan_entry("foreign_huge.parquet", 500_000_000, None),
            make_plan_entry("foreign_small.parquet", 100, None),
            make_plan_entry("foreign_medium.parquet", 10_000_000, None),
        ];
        let selected = planner.plan(&files);
        let paths: Vec<&str> = selected.iter().map(|f| f.path.as_str()).collect();
        assert_eq!(selected.len(), 2, "max_files_per_pass=2, paths={paths:?}");
        assert_eq!(
            paths,
            vec!["foreign_small.parquet", "foreign_medium.parquet"],
            "foreign files must be size-sorted before truncation, cheapest first"
        );
    }

    #[tokio::test]
    async fn compact_merges_two_files() {
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
        let embs_a: Vec<Vec<f32>> = vec![vec![1.0, 0.0, 0.0, 0.0], vec![0.0, 1.0, 0.0, 0.0]];
        let embs_b: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 1.0, 0.0], vec![0.0, 0.0, 0.0, 1.0]];

        let batch_a = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![0i32, 1]))],
        )
        .unwrap();
        let batch_b = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![2i32, 3]))],
        )
        .unwrap();

        let writer_a = AilakeFileWriter::new(policy.clone());
        let bytes_a = writer_a.write(&batch_a, &embs_a).unwrap();
        let writer_b = AilakeFileWriter::new(policy.clone());
        let bytes_b = writer_b.write(&batch_b, &embs_b).unwrap();

        store.put("data/a.parquet", bytes_a.clone()).await.unwrap();
        store.put("data/b.parquet", bytes_b.clone()).await.unwrap();

        let entries = vec![
            DataFileEntry {
                path: "data/a.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_a.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "data/b.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_b.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
        ];

        let executor = CompactionExecutor::new(store.clone(), policy.clone());
        let merged = executor
            .compact(&entries, "data/merged.parquet")
            .await
            .unwrap();

        assert_eq!(merged.record_count, 4);
        assert_eq!(merged.path, "data/merged.parquet");

        let merged_bytes = store.get("data/merged.parquet").await.unwrap();
        let reader = AilakeFileReader::new(merged_bytes, "embedding", 4);
        reader.verify_integrity().unwrap();
        let (batch, embs) = reader.read_parquet().unwrap();
        assert_eq!(batch.num_rows(), 4);
        assert_eq!(embs.len(), 4);
    }

    /// Regression test: `compact()`/`read_files_parallel` used to read every input file's
    /// rows unconditionally, ignoring any `deletion_vector` on the entry — so a row deleted
    /// via `delete_rows()` before compaction reappeared in the merged output (new physical
    /// file, fresh row positions, no DV carried over — the row was never actually removed).
    #[tokio::test]
    async fn compact_drops_deletion_vector_masked_rows() {
        use ailake_catalog::provider::DeletionVector;
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use roaring::RoaringBitmap;
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
        // File A: row 0 (id=0) will be marked deleted; row 1 (id=1) survives.
        let embs_a: Vec<Vec<f32>> = vec![vec![1.0, 0.0, 0.0, 0.0], vec![0.0, 1.0, 0.0, 0.0]];
        let embs_b: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 1.0, 0.0]];

        let batch_a = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![0i32, 1]))],
        )
        .unwrap();
        let batch_b =
            RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(vec![2i32]))])
                .unwrap();

        let bytes_a = AilakeFileWriter::new(policy.clone())
            .write(&batch_a, &embs_a)
            .unwrap();
        let bytes_b = AilakeFileWriter::new(policy.clone())
            .write(&batch_b, &embs_b)
            .unwrap();
        store.put("data/a.parquet", bytes_a.clone()).await.unwrap();
        store.put("data/b.parquet", bytes_b.clone()).await.unwrap();

        // Puffin DV marking row 0 of file A as deleted.
        let mut bitmap = RoaringBitmap::new();
        bitmap.insert(0);
        let (puffin_bytes, offset, length) =
            crate::delete::PuffinWriter::write_single_dv(&bitmap, 1).unwrap();
        store.put("metadata/dv-1.dvd", puffin_bytes).await.unwrap();

        let entry_a = DataFileEntry {
            path: "data/a.parquet".into(),
            record_count: 2,
            file_size_bytes: bytes_a.len() as u64,
            centroid_b64: None,
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: None,
            vector_dim: None,
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: Some(DeletionVector {
                path: "metadata/dv-1.dvd".into(),
                offset,
                length,
                cardinality: 1,
            }),
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };
        let entry_b = DataFileEntry {
            path: "data/b.parquet".into(),
            record_count: 1,
            file_size_bytes: bytes_b.len() as u64,
            centroid_b64: None,
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: None,
            vector_dim: None,
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: None,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };

        let executor = CompactionExecutor::new(store.clone(), policy.clone());
        let merged = executor
            .compact(&[entry_a, entry_b], "data/merged.parquet")
            .await
            .unwrap();

        // 2 rows in A + 1 in B, minus 1 deleted from A = 2 surviving rows.
        assert_eq!(merged.record_count, 2);
        let merged_bytes = store.get("data/merged.parquet").await.unwrap();
        let reader = AilakeFileReader::new(merged_bytes, "embedding", 4);
        let (batch, embs) = reader.read_parquet().unwrap();
        assert_eq!(batch.num_rows(), 2);
        assert_eq!(embs.len(), 2);
        let ids: Vec<i32> = batch
            .column_by_name("id")
            .unwrap()
            .as_any()
            .downcast_ref::<Int32Array>()
            .unwrap()
            .values()
            .to_vec();
        assert_eq!(
            ids,
            vec![1, 2],
            "id=0 (deleted) must not survive compaction"
        );
    }

    /// Regression test: `verify_integrity()` used to always call `load_index()`, which
    /// unconditionally deserializes as `HnswIndex` regardless of the IVF-PQ flag —
    /// so any compaction using `ForceIvfPq` (or `Auto` selecting IVF-PQ) would write a
    /// correct merged file and then fail on this call, aborting the whole compaction.
    #[tokio::test]
    async fn compact_with_ivf_pq_strategy_does_not_crash_on_verify_integrity() {
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let dim = 8;
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
        let n_per_file = 30usize;
        let make_file = |path: &str, offset: i32| {
            let ids: Vec<i32> = (offset..offset + n_per_file as i32).collect();
            let embs: Vec<Vec<f32>> = ids
                .iter()
                .map(|&i| {
                    (0..dim as i32)
                        .map(|j| ((i * 31 + j * 7) % 97) as f32 / 97.0)
                        .collect()
                })
                .collect();
            let batch = RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(ids))])
                .unwrap();
            let bytes = AilakeFileWriter::new(policy.clone())
                .write(&batch, &embs)
                .unwrap();
            (path.to_string(), bytes)
        };

        let (path_a, bytes_a) = make_file("data/ivfpq_a.parquet", 0);
        let (path_b, bytes_b) = make_file("data/ivfpq_b.parquet", n_per_file as i32);
        for (path, bytes) in [(&path_a, &bytes_a), (&path_b, &bytes_b)] {
            store.put(path, bytes.clone()).await.unwrap();
        }

        let make_entry = |path: &str, size: u64| DataFileEntry {
            path: path.to_string(),
            record_count: n_per_file as u64,
            file_size_bytes: size,
            centroid_b64: None,
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: None,
            vector_dim: None,
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: None,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };
        let entries = vec![
            make_entry(&path_a, bytes_a.len() as u64),
            make_entry(&path_b, bytes_b.len() as u64),
        ];

        let executor = CompactionExecutor::new(store.clone(), policy.clone())
            .with_index_strategy(CompactionIndexStrategy::ForceIvfPq);
        let merged = executor
            .compact(&entries, "data/ivfpq_merged.parquet")
            .await
            .expect("compact() with ForceIvfPq must not fail in verify_integrity()");

        assert_eq!(merged.record_count, 2 * n_per_file as u64);

        // The bug: this used to panic/error trying to load IVF-PQ bytes as HNSW.
        let merged_bytes = store.get("data/ivfpq_merged.parquet").await.unwrap();
        let reader = AilakeFileReader::new(merged_bytes, "embedding", dim);
        reader
            .verify_integrity()
            .expect("verify_integrity() must handle an IVF-PQ-indexed file");
    }

    #[tokio::test]
    async fn compact_incremental_merges_dominant_plus_small() {
        use ailake_core::{RowId, VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));

        // Dominant file: 6 rows (75% of total 8 rows — above 40% threshold).
        let embs_dom: Vec<Vec<f32>> = vec![
            vec![1.0, 0.0, 0.0, 0.0],
            vec![0.0, 1.0, 0.0, 0.0],
            vec![0.0, 0.0, 1.0, 0.0],
            vec![0.7, 0.7, 0.0, 0.0],
            vec![0.0, 0.7, 0.7, 0.0],
            vec![0.0, 0.0, 0.7, 0.7],
        ];
        let batch_dom = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![0i32, 1, 2, 3, 4, 5]))],
        )
        .unwrap();

        // Small file: 2 rows.
        let embs_small: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 0.0, 1.0], vec![0.5, 0.5, 0.5, 0.5]];
        let batch_small = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![6i32, 7]))],
        )
        .unwrap();

        let bytes_dom = AilakeFileWriter::new(policy.clone())
            .write(&batch_dom, &embs_dom)
            .unwrap();
        let bytes_small = AilakeFileWriter::new(policy.clone())
            .write(&batch_small, &embs_small)
            .unwrap();

        store
            .put("data/dominant.parquet", bytes_dom.clone())
            .await
            .unwrap();
        store
            .put("data/small.parquet", bytes_small.clone())
            .await
            .unwrap();

        let entries = vec![
            DataFileEntry {
                path: "data/dominant.parquet".into(),
                record_count: 6,
                file_size_bytes: bytes_dom.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "data/small.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_small.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
        ];

        let executor = CompactionExecutor::new(store.clone(), policy.clone());
        let merged = executor
            .compact_incremental(&entries, "data/merged.parquet")
            .await
            .unwrap();

        // Structural checks.
        assert_eq!(merged.record_count, 8);
        assert_eq!(merged.path, "data/merged.parquet");

        // Load merged file and verify it's a valid AI-Lake file.
        let merged_bytes = store.get("data/merged.parquet").await.unwrap();
        let reader = AilakeFileReader::new(merged_bytes, "embedding", 4);
        reader.verify_integrity().unwrap();

        let (batch, embs) = reader.read_parquet().unwrap();
        assert_eq!(batch.num_rows(), 8);
        assert_eq!(embs.len(), 8);

        // Dominant rows must come first (positions 0..5).
        for f in &embs[..6] {
            assert_eq!(f.len(), 4);
        }

        // HNSW must be searchable and return the nearest neighbor for a known query.
        let hnsw = reader.load_index().unwrap();
        assert_eq!(hnsw.node_count(), 8);

        // Query [1, 0, 0, 0] → nearest should be RowId 0 (embs_dom[0]).
        let results = hnsw.search(&[1.0, 0.0, 0.0, 0.0], 1, 50);
        assert_eq!(results[0].0, RowId::new(0));

        // Query [0, 0, 0, 1] → nearest should be RowId 6 (first row of small file,
        // inserted at position 6 in the merged file).
        let results = hnsw.search(&[0.0, 0.0, 0.0, 1.0], 1, 50);
        assert_eq!(results[0].0, RowId::new(6));
    }

    /// Regression: `compact_incremental()` only ever produces HNSW (it extends the
    /// dominant file's existing graph) — with a dominant file present, it never checked
    /// `self.index_strategy` before taking that path, so an explicit `ForceIvfPq` request
    /// was silently satisfied with HNSW instead. Reachable from `CompactionExecutor::run()`/
    /// `run_deferred()` (used by the CLI `compact` command and every JNI/Spark/Trino/Flink/
    /// DuckDB compact call), which always try `compact_incremental()` first — including on
    /// a GPU machine where `Auto` would pick IVF-PQ for a fresh build, or where a caller
    /// explicitly forces IVF-PQ. Same dominant/small file shape as
    /// `compact_incremental_merges_dominant_plus_small`, but with `ForceIvfPq` set.
    #[tokio::test]
    async fn compact_incremental_respects_force_ivf_pq_even_with_dominant_file() {
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let dim = 8;
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
        let make_file = |path: &str, offset: i32, n: usize| {
            let ids: Vec<i32> = (offset..offset + n as i32).collect();
            let embs: Vec<Vec<f32>> = ids
                .iter()
                .map(|&i| {
                    (0..dim as i32)
                        .map(|j| ((i * 31 + j * 7) % 97) as f32 / 97.0)
                        .collect()
                })
                .collect();
            let batch = RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(ids))])
                .unwrap();
            // Dominant file already has a real, loadable HNSW index — the exact
            // condition that used to make compact_incremental() bypass ForceIvfPq.
            let bytes = AilakeFileWriter::new(policy.clone())
                .write(&batch, &embs)
                .unwrap();
            (path.to_string(), bytes, n as u64)
        };

        // Dominant: 90 rows (75% of 120 total — comfortably above the dominant-file
        // threshold), small: 30 rows.
        let (path_dom, bytes_dom, n_dom) = make_file("data/dom.parquet", 0, 90);
        let (path_small, bytes_small, n_small) = make_file("data/small.parquet", 999, 30);

        for (path, bytes) in [(&path_dom, &bytes_dom), (&path_small, &bytes_small)] {
            store.put(path, bytes.clone()).await.unwrap();
        }

        let make_entry = |path: &str, record_count: u64, size: u64| DataFileEntry {
            path: path.to_string(),
            record_count,
            file_size_bytes: size,
            centroid_b64: None,
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: None,
            vector_dim: None,
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: None,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };
        let entries = vec![
            make_entry(&path_dom, n_dom, bytes_dom.len() as u64),
            make_entry(&path_small, n_small, bytes_small.len() as u64),
        ];

        let executor = CompactionExecutor::new(store.clone(), policy.clone())
            .with_index_strategy(CompactionIndexStrategy::ForceIvfPq);
        let merged = executor
            .compact_incremental(&entries, "data/merged_force_ivfpq.parquet")
            .await
            .expect("compact_incremental() with ForceIvfPq must fall back to compact() cleanly");

        let merged_bytes = store.get("data/merged_force_ivfpq.parquet").await.unwrap();
        let reader = AilakeFileReader::new(merged_bytes, "embedding", dim);
        reader.verify_integrity().unwrap();

        match reader.load_any_index().unwrap() {
            ailake_index::AnyIndex::IvfPq(_) => {}
            ailake_index::AnyIndex::Hnsw(_) => panic!(
                "ForceIvfPq was silently satisfied with HNSW instead — merged.record_count={}",
                merged.record_count
            ),
        }
    }

    #[tokio::test]
    async fn compact_incremental_falls_back_when_no_dominant() {
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));

        // Two equal-sized files (50/50 split — no dominant, both below 40% threshold).
        let make_batch = |ids: Vec<i32>, embs: Vec<Vec<f32>>| {
            let batch = RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(ids))])
                .unwrap();
            AilakeFileWriter::new(policy.clone())
                .write(&batch, &embs)
                .unwrap()
        };

        let embs_a: Vec<Vec<f32>> = vec![vec![1.0, 0.0, 0.0, 0.0], vec![0.0, 1.0, 0.0, 0.0]];
        let embs_b: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 1.0, 0.0], vec![0.0, 0.0, 0.0, 1.0]];
        let bytes_a = make_batch(vec![0, 1], embs_a);
        let bytes_b = make_batch(vec![2, 3], embs_b);

        store.put("data/a.parquet", bytes_a.clone()).await.unwrap();
        store.put("data/b.parquet", bytes_b.clone()).await.unwrap();

        let entries = vec![
            DataFileEntry {
                path: "data/a.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_a.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "data/b.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_b.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
        ];

        let executor = CompactionExecutor::new(store.clone(), policy.clone());
        // Should fall back to full rebuild without error.
        let merged = executor
            .compact_incremental(&entries, "data/merged.parquet")
            .await
            .unwrap();

        assert_eq!(merged.record_count, 4);

        let merged_bytes = store.get("data/merged.parquet").await.unwrap();
        let reader = AilakeFileReader::new(merged_bytes, "embedding", 4);
        reader.verify_integrity().unwrap();
    }

    #[tokio::test]
    async fn compact_deferred_produces_parquet_only_file() {
        use ailake_catalog::HadoopCatalog;
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let catalog_dir = TempDir::new().unwrap();
        let catalog_store = Arc::new(LocalStore::new(catalog_dir.path()));
        let catalog = Arc::new(HadoopCatalog::new(catalog_store, ""));
        let table = TableIdent {
            namespace: "ns".into(),
            name: "tbl".into(),
        };

        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        use ailake_catalog::TableProperties;
        catalog
            .create_table(
                &table,
                &TableProperties {
                    policy: policy.clone(),
                    extra: std::collections::HashMap::new(),
                    format_version: 2,
                    partition_column_type: None,
                },
            )
            .await
            .unwrap();

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
        let embs_a: Vec<Vec<f32>> = vec![vec![1.0, 0.0, 0.0, 0.0], vec![0.0, 1.0, 0.0, 0.0]];
        let batch_a = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![0i32, 1]))],
        )
        .unwrap();
        let bytes_a = AilakeFileWriter::new(policy.clone())
            .write(&batch_a, &embs_a)
            .unwrap();
        store.put("data/a.parquet", bytes_a.clone()).await.unwrap();

        let embs_b: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 1.0, 0.0], vec![0.0, 0.0, 0.0, 1.0]];
        let batch_b = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![2i32, 3]))],
        )
        .unwrap();
        let bytes_b = AilakeFileWriter::new(policy.clone())
            .write(&batch_b, &embs_b)
            .unwrap();
        store.put("data/b.parquet", bytes_b.clone()).await.unwrap();

        let entries = vec![
            DataFileEntry {
                path: "data/a.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_a.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "data/b.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_b.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
        ];

        let executor = CompactionExecutor::new(store.clone(), policy.clone());
        let entry = executor
            .compact_deferred(&entries, "data/merged.parquet", catalog.clone(), &table)
            .await
            .unwrap();

        // Entry is Indexing — HNSW build pending in background
        assert_eq!(entry.index_status, IndexStatus::Indexing);
        assert_eq!(entry.record_count, 4);

        // The written file must be valid Parquet (readable) even without HNSW
        let merged_bytes = store.get("data/merged.parquet").await.unwrap();
        let pq_reader = ailake_parquet::ParquetVectorReader::new(merged_bytes, "embedding");
        let count = pq_reader.record_count().unwrap();
        assert_eq!(count, 4);
    }

    /// Regression test (ADR-018 / CLAUDE.md Fase 5 "Idempotência batch_id sobrevivendo
    /// a compaction"): before `DataFileEntry::merge_batch_ids` existed, `compact()`
    /// always produced `batch_id: None` on the merged file — a retry of either source
    /// write, dispatched after compaction swept up its file, would find no existing
    /// entry carrying its key and silently re-insert. All three merge entry-points
    /// (`compact`, `compact_incremental`, `compact_deferred`) must aggregate source
    /// `batch_id`s; this covers the plain (non-deferred) `compact()` path.
    #[tokio::test]
    async fn compact_aggregates_batch_ids_from_sources() {
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
        let embs_a: Vec<Vec<f32>> = vec![vec![1.0, 0.0, 0.0, 0.0]];
        let batch_a =
            RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(vec![0i32]))])
                .unwrap();
        let bytes_a = AilakeFileWriter::new(policy.clone())
            .write(&batch_a, &embs_a)
            .unwrap();
        store.put("data/a.parquet", bytes_a.clone()).await.unwrap();

        let embs_b: Vec<Vec<f32>> = vec![vec![0.0, 1.0, 0.0, 0.0]];
        let batch_b =
            RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(vec![1i32]))])
                .unwrap();
        let bytes_b = AilakeFileWriter::new(policy.clone())
            .write(&batch_b, &embs_b)
            .unwrap();
        store.put("data/b.parquet", bytes_b.clone()).await.unwrap();

        // A file compacted with no batch_id at all (e.g. written via plain
        // `write_batch`, never idempotently) — must not inject a spurious key.
        let embs_c: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 1.0, 0.0]];
        let batch_c =
            RecordBatch::try_new(schema, vec![Arc::new(Int32Array::from(vec![2i32]))]).unwrap();
        let bytes_c = AilakeFileWriter::new(policy.clone())
            .write(&batch_c, &embs_c)
            .unwrap();
        store.put("data/c.parquet", bytes_c.clone()).await.unwrap();

        fn entry(path: &str, size: u64, batch_id: Option<&str>) -> DataFileEntry {
            DataFileEntry {
                path: path.into(),
                record_count: 1,
                file_size_bytes: size,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: batch_id.map(String::from),
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            }
        }
        let entries = vec![
            entry("data/a.parquet", bytes_a.len() as u64, Some("k-a")),
            entry("data/b.parquet", bytes_b.len() as u64, Some("k-b")),
            entry("data/c.parquet", bytes_c.len() as u64, None),
        ];

        let executor = CompactionExecutor::new(store.clone(), policy);
        let merged = executor
            .compact(&entries, "data/merged.parquet")
            .await
            .unwrap();

        assert_eq!(merged.record_count, 3);
        assert_eq!(
            merged.batch_ids(),
            vec!["k-a".to_string(), "k-b".to_string()],
            "merged file must carry every source's idempotency key, none invented"
        );
    }

    /// Regression test: `CompactionExecutor::run()` must not drop files that fall
    /// outside the compaction pass (too large for `target_file_size_bytes`, or beyond
    /// `max_files_per_pass`). `Replace` snapshots don't inherit the previous manifest
    /// (see `HadoopCatalog::commit_snapshot`), so `run()` must explicitly carry forward
    /// every untouched file alongside the merged output — mirroring the pattern already
    /// used by the CLI `compact` command and `MemoryDecayJob::run`.
    #[tokio::test]
    async fn run_preserves_untouched_files_outside_compaction_pass() {
        use ailake_catalog::HadoopCatalog;
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let catalog_dir = TempDir::new().unwrap();
        let catalog_store = Arc::new(LocalStore::new(catalog_dir.path()));
        let catalog = Arc::new(HadoopCatalog::new(catalog_store, ""));
        let table = TableIdent {
            namespace: "ns".into(),
            name: "tbl".into(),
        };

        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        use ailake_catalog::TableProperties;
        catalog
            .create_table(
                &table,
                &TableProperties {
                    policy: policy.clone(),
                    extra: std::collections::HashMap::new(),
                    format_version: 2,
                    partition_column_type: None,
                },
            )
            .await
            .unwrap();

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));

        // Two small files — eligible for compaction.
        let write_file = |path: &str, ids: Vec<i32>, embs: Vec<Vec<f32>>| {
            let batch = RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(ids))])
                .unwrap();
            let bytes = AilakeFileWriter::new(policy.clone())
                .write(&batch, &embs)
                .unwrap();
            (path.to_string(), bytes)
        };

        let (path_a, bytes_a) = write_file(
            "data/small_a.parquet",
            vec![0, 1],
            vec![vec![1.0, 0.0, 0.0, 0.0], vec![0.0, 1.0, 0.0, 0.0]],
        );
        let (path_b, bytes_b) = write_file(
            "data/small_b.parquet",
            vec![2, 3],
            vec![vec![0.0, 0.0, 1.0, 0.0], vec![0.0, 0.0, 0.0, 1.0]],
        );
        // "Big" file — same tiny payload in this test, but its DataFileEntry reports a
        // size above target_file_size_bytes so the planner must never select it.
        let (path_big, bytes_big) = write_file(
            "data/big.parquet",
            vec![4, 5],
            vec![vec![1.0, 1.0, 0.0, 0.0], vec![0.0, 1.0, 1.0, 0.0]],
        );

        for (path, bytes) in [
            (&path_a, &bytes_a),
            (&path_b, &bytes_b),
            (&path_big, &bytes_big),
        ] {
            store.put(path, bytes.clone()).await.unwrap();
        }

        let make_entry = |path: &str, size: u64| DataFileEntry {
            path: path.to_string(),
            record_count: 2,
            file_size_bytes: size,
            // Non-None: all three files here are meant to represent normal,
            // already-indexed AI-Lake files — only size should decide eligibility.
            // `plan_prioritizes_foreign_written_files_regardless_of_size` covers
            // the `centroid_b64: None` (foreign-write) case separately.
            centroid_b64: Some("AAAA".into()),
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: None,
            vector_dim: None,
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: None,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };

        let initial_snap_id = ailake_catalog::new_snapshot_id();
        let initial_snapshot = NewSnapshot {
            snapshot_id: initial_snap_id,
            parent_snapshot_id: None,
            files: vec![
                make_entry(&path_a, 500),
                make_entry(&path_b, 500),
                make_entry(&path_big, 200_000_000), // far above target_file_size_bytes below
            ],
            operation: SnapshotOperation::Append,
            iceberg_schema: None,
            extra_properties: std::collections::HashMap::new(),
            bloom_filters: vec![],
            equality_delete_files: vec![],
        };
        catalog
            .commit_snapshot(&table, initial_snapshot)
            .await
            .unwrap();

        let planner = CompactionPlanner::new(CompactionConfig {
            min_files_to_compact: 2,
            target_file_size_bytes: 1000,
            index_strategy: CompactionIndexStrategy::ForceHnsw,
            max_files_per_pass: 20,
        });
        let executor = CompactionExecutor::new(store.clone(), policy.clone());

        let merged = executor
            .run(&planner, &table, catalog.clone(), "data")
            .await
            .unwrap()
            .expect("compaction should have run — 2 eligible small files");

        let files_after = catalog.list_files(&table, None).await.unwrap();
        let paths_after: Vec<&str> = files_after.iter().map(|f| f.path.as_str()).collect();

        assert!(
            paths_after.contains(&path_big.as_str()),
            "BUG: untouched 'big.parquet' vanished after run() — files_after={paths_after:?}"
        );
        assert!(
            paths_after.contains(&merged.path.as_str()),
            "merged output file must be present — files_after={paths_after:?}"
        );

        // Regression: `run()` used to hardcode `parent_snapshot_id: None` on the
        // post-compaction Replace snapshot even though a current snapshot always exists
        // here, breaking Iceberg snapshot lineage (`expire_snapshots`/`rollback_to_snapshot`)
        // for any compacted table. Read the committed metadata.json directly (no public
        // CatalogProvider method exposes snapshot lineage) and confirm the new snapshot's
        // parent points at the snapshot committed before this compaction ran.
        let meta_dir = catalog_dir.path().join("ns/tbl/metadata");
        let latest_metadata = std::fs::read_dir(&meta_dir)
            .unwrap()
            .filter_map(|e| e.ok())
            .filter(|e| e.path().extension().is_some_and(|ext| ext == "json"))
            .max_by_key(|e| e.metadata().unwrap().modified().unwrap())
            .expect("metadata.json must exist after commit");
        let json: serde_json::Value =
            serde_json::from_slice(&std::fs::read(latest_metadata.path()).unwrap()).unwrap();
        let last_snapshot = json["snapshots"].as_array().unwrap().last().unwrap();
        assert_eq!(
            last_snapshot["parent-snapshot-id"].as_i64(),
            Some(initial_snap_id),
            "compaction's Replace snapshot must chain to the pre-compaction snapshot, not be orphaned"
        );
        assert!(
            !paths_after.contains(&path_a.as_str()) && !paths_after.contains(&path_b.as_str()),
            "compacted input files must no longer be listed — files_after={paths_after:?}"
        );
        assert_eq!(
            files_after.len(),
            2,
            "expected exactly [big.parquet, merged] — files_after={paths_after:?}"
        );
    }

    /// Regression test: a file with no AILK footer (e.g. rewritten by a generic
    /// Iceberg engine — Spark/Trino `OPTIMIZE` — with no knowledge of AI-Lake) still
    /// holds valid Parquet data. `read_parquet()` decodes the vector column directly
    /// from Parquet and never touches the footer, so `compact()`/`compact_incremental()`
    /// must include such a file's rows in the merge, not silently drop them.
    #[tokio::test]
    async fn compact_preserves_rows_from_footerless_file() {
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;

        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));

        // Normal AI-Lake file, written the usual way (has an AILK footer).
        let embs_native: Vec<Vec<f32>> = vec![vec![1.0, 0.0, 0.0, 0.0], vec![0.0, 1.0, 0.0, 0.0]];
        let batch_native = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![0i32, 1]))],
        )
        .unwrap();
        let bytes_native = AilakeFileWriter::new(policy.clone())
            .write(&batch_native, &embs_native)
            .unwrap();
        store
            .put("data/native.parquet", bytes_native.clone())
            .await
            .unwrap();

        // "Foreign" file — plain Parquet, no AILK footer, same shape a generic
        // Iceberg engine's rewrite would produce (`write_parquet_only` is the exact
        // primitive `compact_deferred` uses for its Parquet-only fast path).
        let embs_foreign: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 1.0, 0.0], vec![0.0, 0.0, 0.0, 1.0]];
        let batch_foreign = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![2i32, 3]))],
        )
        .unwrap();
        let bytes_foreign = AilakeFileWriter::new(policy.clone())
            .write_parquet_only(&batch_foreign, &embs_foreign)
            .unwrap();
        store
            .put("data/foreign.parquet", bytes_foreign.clone())
            .await
            .unwrap();

        let reader_foreign = AilakeFileReader::new(bytes_foreign.clone(), "embedding", 4);
        assert!(
            !reader_foreign.is_ailake_file(),
            "sanity: write_parquet_only must not embed an AILK footer"
        );

        let entries = vec![
            DataFileEntry {
                path: "data/native.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_native.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
            DataFileEntry {
                path: "data/foreign.parquet".into(),
                record_count: 2,
                file_size_bytes: bytes_foreign.len() as u64,
                centroid_b64: None,
                radius: None,
                hnsw_offset: None,
                hnsw_len: None,
                vector_column: None,
                vector_dim: None,
                extra_vector_indexes: vec![],
                index_status: IndexStatus::Ready,
                index_error: None,
                batch_id: None,
                embedding_model: None,
                partition_value: None,
                deletion_vector: None,
                first_row_id: None,
                column_stats: None,
                sequence_number: 0,
            },
        ];

        // Full rebuild path (`compact`).
        let executor = CompactionExecutor::new(store.clone(), policy.clone());
        let merged = executor
            .compact(&entries, "data/merged_full.parquet")
            .await
            .unwrap();
        assert_eq!(
            merged.record_count, 4,
            "compact() must include all 4 rows — 2 native + 2 from the footerless file"
        );
        let merged_bytes = store.get("data/merged_full.parquet").await.unwrap();
        let reader = AilakeFileReader::new(merged_bytes, "embedding", 4);
        reader.verify_integrity().unwrap();

        // Incremental path (`compact_incremental` — native file becomes dominant at 50/50,
        // exercising the non-dominant-file read path where the bug lived).
        let merged_inc = executor
            .compact_incremental(&entries, "data/merged_inc.parquet")
            .await
            .unwrap();
        assert_eq!(
            merged_inc.record_count, 4,
            "compact_incremental() must include all 4 rows — 2 native + 2 from the footerless file"
        );
    }

    /// End-to-end: search returns the same top-K results before and after compaction.
    /// This tests the actual recall — not just verify_integrity().
    #[tokio::test]
    async fn compact_preserves_search_results() {
        use crate::scanner::SearchConfig;
        use ailake_catalog::HadoopCatalog;
        use ailake_core::{VectorMetric, VectorPrecision};
        use ailake_store::LocalStore;
        use arrow_array::{Int32Array, RecordBatch};
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;
        use tempfile::TempDir;
        let dir = TempDir::new().unwrap();
        let store = Arc::new(LocalStore::new(dir.path()));
        let policy = VectorStoragePolicy {
            column_name: "embedding".into(),
            dim: 4,
            metric: VectorMetric::Cosine,
            precision: VectorPrecision::F16,
            pq: None,
            keep_raw_for_reranking: true,
            pre_normalize: false,
            hnsw_m: None,
            hnsw_ef_construction: None,
            ivf_residual: false,
            embedding_model: None,
            modality: None,
            partition_by: None,
            partition_value: None,
            partition_column_type: None,
            partition_fields: vec![],
        };

        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));

        // File A: 3 rows — unit basis vectors along dimensions 0, 1, 2
        let embs_a: Vec<Vec<f32>> = vec![
            vec![1.0, 0.0, 0.0, 0.0],
            vec![0.0, 1.0, 0.0, 0.0],
            vec![0.0, 0.0, 1.0, 0.0],
        ];
        let batch_a = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![0i32, 1, 2]))],
        )
        .unwrap();

        // File B: 2 rows — unit basis vectors along dimensions 3 and mixed
        let embs_b: Vec<Vec<f32>> = vec![vec![0.0, 0.0, 0.0, 1.0], vec![0.7, 0.7, 0.0, 0.0]];
        let batch_b = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(Int32Array::from(vec![3i32, 4]))],
        )
        .unwrap();

        let bytes_a = AilakeFileWriter::new(policy.clone())
            .write(&batch_a, &embs_a)
            .unwrap();
        let bytes_b = AilakeFileWriter::new(policy.clone())
            .write(&batch_b, &embs_b)
            .unwrap();

        store.put("data/a.parquet", bytes_a).await.unwrap();
        store.put("data/b.parquet", bytes_b).await.unwrap();
        let file_size_a = store.file_size("data/a.parquet").await.unwrap();
        let file_size_b = store.file_size("data/b.parquet").await.unwrap();

        // We need a CatalogProvider to create/snapshot the table.
        // Use the catalog_provider helper from ailake-query.
        let catalog: Arc<dyn CatalogProvider> =
            Arc::new(HadoopCatalog::new(store.clone(), "warehouse"));
        let table = TableIdent::new("default", "compact_search_test");

        // Bootstrap the table via create_table, then commit the two files
        catalog
            .create_table(
                &table,
                &ailake_catalog::TableProperties {
                    policy: policy.clone(),
                    extra: std::collections::HashMap::new(),
                    format_version: 2,
                    partition_column_type: None,
                },
            )
            .await
            .unwrap();

        // Load current iceberg schema so commit_snapshot can point at it
        let meta = catalog.load_table(&table).await.unwrap();
        let parent_snapshot_id = meta.current_snapshot_id;

        let entry_a = DataFileEntry {
            path: "data/a.parquet".into(),
            record_count: 3,
            file_size_bytes: file_size_a,
            centroid_b64: None,
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: Some("embedding".into()),
            vector_dim: Some(4),
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: None,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };
        let entry_b = DataFileEntry {
            path: "data/b.parquet".into(),
            record_count: 2,
            file_size_bytes: file_size_b,
            centroid_b64: None,
            radius: None,
            hnsw_offset: None,
            hnsw_len: None,
            vector_column: Some("embedding".into()),
            vector_dim: Some(4),
            extra_vector_indexes: vec![],
            index_status: IndexStatus::Ready,
            index_error: None,
            batch_id: None,
            embedding_model: None,
            partition_value: None,
            deletion_vector: None,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };

        catalog
            .commit_snapshot(
                &table,
                NewSnapshot {
                    snapshot_id: 200,
                    parent_snapshot_id,
                    files: vec![entry_a, entry_b],
                    operation: SnapshotOperation::Append,
                    iceberg_schema: None,
                    extra_properties: std::collections::HashMap::new(),
                    bloom_filters: vec![],
                    equality_delete_files: vec![],
                },
            )
            .await
            .unwrap();

        // ── Search before compaction ──
        let query_before = vec![1.0f32, 0.0, 0.0, 0.0];
        let config = SearchConfig {
            top_k: 5,
            ef_search: 100,
            pruning_threshold: f32::INFINITY,
            ..Default::default()
        };
        let before = crate::scanner::search(
            &table,
            &query_before,
            config.clone(),
            "embedding",
            4,
            catalog.clone(),
            store.clone(),
        )
        .await
        .unwrap();
        assert!(
            !before.is_empty(),
            "search before compaction must return results"
        );
        let before_dists: Vec<f32> = before.iter().map(|r| r.distance).collect();

        // ── Compact the two files ──
        let executor = CompactionExecutor::new(store.clone(), policy.clone());
        let files_to_compact = catalog.list_files(&table, None).await.unwrap();
        let merged = executor
            .compact(&files_to_compact, "data/merged.parquet")
            .await
            .unwrap();

        // Commit the merge: Replace old entries with the merged one
        let meta2 = catalog.load_table(&table).await.unwrap();
        let merged_entry = DataFileEntry {
            path: merged.path.clone(),
            record_count: merged.record_count,
            file_size_bytes: merged.file_size_bytes,
            centroid_b64: merged.centroid_b64,
            radius: merged.radius,
            hnsw_offset: merged.hnsw_offset,
            hnsw_len: merged.hnsw_len,
            vector_column: merged.vector_column,
            vector_dim: merged.vector_dim,
            extra_vector_indexes: merged.extra_vector_indexes,
            index_status: merged.index_status,
            index_error: merged.index_error,
            batch_id: merged.batch_id,
            embedding_model: merged.embedding_model,
            partition_value: merged.partition_value,
            deletion_vector: merged.deletion_vector,
            first_row_id: None,
            column_stats: None,
            sequence_number: 0,
        };
        catalog
            .commit_snapshot(
                &table,
                NewSnapshot {
                    snapshot_id: 201,
                    parent_snapshot_id: meta2.current_snapshot_id,
                    files: vec![merged_entry],
                    operation: SnapshotOperation::Replace,
                    iceberg_schema: None,
                    extra_properties: std::collections::HashMap::new(),
                    bloom_filters: vec![],
                    equality_delete_files: vec![],
                },
            )
            .await
            .unwrap();

        // ── Search after compaction ──
        let after = crate::scanner::search(
            &table,
            &query_before,
            config.clone(),
            "embedding",
            4,
            catalog,
            store,
        )
        .await
        .unwrap();
        assert!(
            !after.is_empty(),
            "search after compaction must return results"
        );
        let after_dists: Vec<f32> = after.iter().map(|r| r.distance).collect();

        // ── Assert recall parity ──
        assert_eq!(
            before.len(),
            after.len(),
            "number of results should match before ({}) and after ({}) compaction",
            before.len(),
            after.len(),
        );
        // Row-ids are NOT expected to match — compaction reassigns row positions.
        // What matters: distances should be approximately the same for each rank,
        // meaning the same items are returned with similar relevance scores.
        // Mean Average Distance Error (MADE) should be small.
        let mut total_diff = 0.0f32;
        for (i, (bd, ad)) in before_dists.iter().zip(after_dists.iter()).enumerate() {
            let diff = (bd - ad).abs();
            total_diff += diff;
            assert!(
                diff < 0.1,
                "distance mismatch at position {i}: before={bd}, after={ad}, diff={diff}"
            );
        }
        let avg_diff = total_diff / before_dists.len() as f32;
        assert!(
            avg_diff < 0.05,
            "average distance error across all results too high: {avg_diff}"
        );
    }
}