chronix_engine/memtable/

memtable.rs

//! Concurrent memtable backed by a lock-free skip list.
//!
//! The [`Memtable`] is the primary in-memory write buffer. It accepts
//! [`Point`] inserts, stores them in a concurrent skip
//! list ordered by `(series_key_hash, timestamp)`, and supports range scans
//! for reads.
//!
//! # Threading model
//!
//! Multiple writers can insert concurrently without external locking.
//! The skip list (from [`crossbeam_skiplist`]) provides lock-free concurrent
//! access. An [`AtomicUsize`] tracks the estimated memory footprint, and
//! [`AtomicU64`] values track the WAL sequence range for flush coordination.

use std::collections::HashSet;
use std::sync::Arc;
use std::sync::atomic::{AtomicBool, AtomicU64, AtomicUsize, Ordering};

use crossbeam_skiplist::SkipMap;
use dashmap::DashMap;
use dashmap::mapref::entry::Entry;

use chronix_core::types::{Point, SeriesKey, Timestamp};

use crate::memtable::error::{MemtableError, Result};
use crate::memtable::interner::StringInterner;
use crate::memtable::key::{MemtableEntry, MemtableKey, NODE_OVERHEAD};

/// One series' tag set, shared by every point of the series in a memtable.
pub type TagSet = Arc<[(Arc<str>, Arc<str>)]>;

/// A concurrent in-memory write buffer for time-series data.
///
/// Points are stored in a [`crossbeam_skiplist::SkipMap`] keyed by
/// [`MemtableKey`] `(series_key_hash, timestamp)`. The skip list provides
/// lock-free concurrent inserts and ordered iteration.
///
/// # Deduplication
///
/// If a point with the same `(series_key_hash, timestamp)` already exists,
/// the new value overwrites the old one (last-write-wins).
pub struct Memtable {
    /// Lock-free concurrent skip list.
    data: SkipMap<MemtableKey, MemtableEntry>,
    /// String interner for deduplicating measurement/tag strings.
    interner: StringInterner,
    /// Secondary index: measurement name → set of series hashes.
    ///
    /// Enables O(S × log N) `scan_measurement()` instead of O(N) full scans,
    /// where S is the number of distinct series for the target measurement.
    ///
    /// Uses [`DashMap`] with per-shard locking instead of a
    /// global `RwLock<HashMap>` to eliminate write-lock contention on the
    /// hot insert path.
    measurement_index: DashMap<Arc<str>, HashSet<u64>>,
    /// One interned tag set per series (keyed by canonical form), shared by
    /// every point of the series in this memtable.
    series_tags: DashMap<Arc<str>, TagSet>,
    /// Estimated total memory usage in bytes, excluding `measurement_index`.
    estimated_size: AtomicUsize,
    /// Estimated `measurement_index` footprint in bytes.
    ///
    /// Accumulated at the two sites that write the index, not recomputed on
    /// read: the flush controller's capacity check calls `estimated_size()`
    /// once per point, and walking the `DashMap` there locks every shard.
    index_size: AtomicUsize,
    /// Minimum WAL sequence number covered by this memtable.
    min_wal_seq: AtomicU64,
    /// Maximum WAL sequence number covered by this memtable.
    max_wal_seq: AtomicU64,
    /// Whether this memtable has been frozen (read-only).
    frozen: AtomicBool,
}

impl std::fmt::Debug for Memtable {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("Memtable")
            .field("len", &self.data.len())
            .field(
                "estimated_size",
                &self.estimated_size.load(Ordering::Relaxed),
            )
            .field("interned_strings", &self.interner.len())
            .field("min_wal_seq", &self.min_wal_seq())
            .field("max_wal_seq", &self.max_wal_seq())
            .field("frozen", &self.frozen.load(Ordering::Relaxed))
            .finish()
    }
}

impl Memtable {
    /// Create a new empty memtable.
    #[must_use]
    pub fn new() -> Self {
        Self {
            data: SkipMap::new(),
            interner: StringInterner::new(),
            measurement_index: DashMap::new(),
            series_tags: DashMap::new(),
            estimated_size: AtomicUsize::new(0),
            index_size: AtomicUsize::new(0),
            min_wal_seq: AtomicU64::new(u64::MAX),
            max_wal_seq: AtomicU64::new(0),
            frozen: AtomicBool::new(false),
        }
    }

    /// Insert a point into the memtable.
    ///
    /// The point is stored keyed by `(series_key_hash, timestamp)`. If a
    /// point with the same key already exists, it is overwritten
    /// (last-write-wins).
    ///
    /// # Errors
    ///
    /// Returns [`MemtableError::Frozen`] if the memtable has been frozen.
    pub fn insert(&self, point: &Point) -> Result<()> {
        if self.frozen.load(Ordering::Acquire) {
            return Err(MemtableError::Frozen);
        }

        let canonical: Arc<str> = self.interner.intern(point.series_key().canonical_form());
        let key = MemtableKey::new(point.series_key().hash_fnv(), point.timestamp(), canonical);

        let measurement = self.interner.intern(point.series_key().measurement());
        let measurement_for_index = Arc::clone(&measurement);
        let tags = self.series_tag_set(&key.series_canonical, point.series_key());

        let entry = MemtableEntry {
            measurement,
            tags,
            fields: point
                .fields()
                .iter()
                .map(|(k, v)| (Arc::clone(k), v.clone()))
                .collect(),
        };

        let entry_size =
            entry.estimated_size() + std::mem::size_of::<MemtableKey>() + NODE_OVERHEAD;

        // If the key already exists, subtract the old entry size to avoid
        // monotonically inflating the estimate on overwrites.
        // Use saturating subtraction to prevent underflow when concurrent
        // overwrites race on the same key (TOCTOU between get and insert).
        // Crossbeam SkipMap::insert() doesn't return the old value,
        // so a separate get() is required. The two O(log N) traversals are
        // inherent to the SkipMap API.
        if let Some(existing) = self.data.get(&key) {
            let old_size = existing.value().estimated_size()
                + std::mem::size_of::<MemtableKey>()
                + NODE_OVERHEAD;
            let _ =
                self.estimated_size
                    .fetch_update(Ordering::Relaxed, Ordering::Relaxed, |current| {
                        Some(current.saturating_sub(old_size))
                    });
        }

        self.data.insert(key, entry);
        let new_size = self.estimated_size.fetch_add(entry_size, Ordering::Relaxed) + entry_size;

        // Emit memtable memory usage gauge so operators can monitor
        // memory pressure and flush readiness.
        metrics::gauge!("chronix_memtable_memory_bytes").set(new_size as f64);
        metrics::gauge!("chronix_memtable_entries").set(self.data.len() as f64);

        // Update secondary measurement index so scan_measurement() can
        // use targeted range scans instead of full table iteration.
        // DashMap entry API — only per-shard lock, no global contention.
        {
            let hash = point.series_key().hash_fnv();
            let delta = self.index_insert(measurement_for_index, hash);
            if delta > 0 {
                self.index_size.fetch_add(delta, Ordering::Relaxed);
            }
        }

        Ok(())
    }

    /// Insert a point with an associated WAL sequence number.
    ///
    /// Updates the min/max WAL sequence tracking in addition to inserting
    /// the point.
    ///
    /// # Errors
    ///
    /// Returns [`MemtableError::Frozen`] if the memtable has been frozen.
    pub fn insert_with_wal_seq(&self, point: &Point, wal_seq: u64) -> Result<()> {
        self.insert(point)?;
        self.update_wal_seq(wal_seq);
        Ok(())
    }

    /// Insert a batch of points in one call.
    ///
    /// More efficient than calling `insert` in a loop because:
    /// The frozen check is performed once.
    /// Strings are pre-interned and the measurement index is updated in
    ///   a single write-lock acquisition.
    /// Size adjustments are accumulated locally and flushed with a single
    ///   `fetch_add`.
    ///
    /// Returns the number of points successfully inserted.
    ///
    /// # Errors
    ///
    /// Returns [`MemtableError::Frozen`] if the memtable has been frozen.
    pub fn insert_batch(&self, points: &[Point]) -> Result<usize> {
        if self.frozen.load(Ordering::Acquire) {
            return Err(MemtableError::Frozen);
        }

        let mut total_size_delta: i64 = 0;
        // Collect (measurement, hash) pairs for batch measurement index update.
        let mut new_index_entries: Vec<(Arc<str>, u64)> = Vec::new();

        // Track previous key to skip overwrite check when
        // consecutive batch points have different keys (the common case
        // for append-only time-series ingestion).
        let mut prev_hash: u64 = 0;
        let mut prev_ts: i64 = i64::MIN;

        for point in points {
            let canonical: Arc<str> = self.interner.intern(point.series_key().canonical_form());
            let key = MemtableKey::new(point.series_key().hash_fnv(), point.timestamp(), canonical);

            let measurement = self.interner.intern(point.series_key().measurement());
            let tags = self.series_tag_set(&key.series_canonical, point.series_key());

            let entry = MemtableEntry {
                measurement: Arc::clone(&measurement),
                tags,
                fields: point
                    .fields()
                    .iter()
                    .map(|(k, v)| (Arc::clone(k), v.clone()))
                    .collect(),
            };

            let entry_size =
                entry.estimated_size() + std::mem::size_of::<MemtableKey>() + NODE_OVERHEAD;

            // Only check for overwrites when the key could collide
            // with an existing entry. For strictly increasing timestamps
            // within the same series or different series, skip the O(log N)
            // get() entirely — the insert alone is sufficient.
            let cur_hash = point.series_key().hash_fnv();
            let cur_ts = point.timestamp();
            let may_overwrite = cur_hash == prev_hash && cur_ts == prev_ts;

            if may_overwrite && let Some(existing) = self.data.get(&key) {
                let old_size = existing.value().estimated_size()
                    + std::mem::size_of::<MemtableKey>()
                    + NODE_OVERHEAD;
                total_size_delta -= old_size.min(i64::MAX as usize) as i64;
            }

            prev_hash = cur_hash;
            prev_ts = cur_ts;

            self.data.insert(key, entry);
            // Saturating cast prevents theoretical i64 overflow.
            total_size_delta += entry_size.min(i64::MAX as usize) as i64;

            new_index_entries.push((measurement, cur_hash));
        }

        // Flush accumulated size delta in one atomic operation.
        if total_size_delta >= 0 {
            self.estimated_size
                .fetch_add(total_size_delta as usize, Ordering::Relaxed);
        } else {
            let _ =
                self.estimated_size
                    .fetch_update(Ordering::Relaxed, Ordering::Relaxed, |current| {
                        Some(current.saturating_sub(total_size_delta.unsigned_abs() as usize))
                    });
        }

        // DashMap entry API — per-shard locking, no global contention.
        let mut index_delta = 0usize;
        for (measurement, hash) in new_index_entries {
            index_delta += self.index_insert(measurement, hash);
        }
        if index_delta > 0 {
            self.index_size.fetch_add(index_delta, Ordering::Relaxed);
        }

        // Emit memtable memory usage gauge after batch insert.
        metrics::gauge!("chronix_memtable_memory_bytes")
            .set(self.estimated_size.load(Ordering::Relaxed) as f64);
        metrics::gauge!("chronix_memtable_entries").set(self.data.len() as f64);

        Ok(points.len())
    }

    /// Insert a batch of points with an associated WAL sequence number.
    ///
    /// Uses the batch to set the WAL sequence range, then delegates to
    /// `insert_batch`.
    ///
    /// # Errors
    ///
    /// Returns [`MemtableError::Frozen`] if the memtable has been frozen.
    pub fn insert_batch_with_wal_seq(&self, points: &[Point], wal_seq: u64) -> Result<usize> {
        let count = self.insert_batch(points)?;
        self.update_wal_seq(wal_seq);
        Ok(count)
    }

    /// Update the WAL sequence range tracked by this memtable.
    fn update_wal_seq(&self, seq: u64) {
        // Update min: keep the smallest
        let mut current = self.min_wal_seq.load(Ordering::Relaxed);
        while seq < current {
            match self.min_wal_seq.compare_exchange_weak(
                current,
                seq,
                Ordering::Release,
                Ordering::Relaxed,
            ) {
                Ok(_) => break,
                Err(actual) => current = actual,
            }
        }

        // Update max: keep the largest
        let mut current = self.max_wal_seq.load(Ordering::Relaxed);
        while seq > current {
            match self.max_wal_seq.compare_exchange_weak(
                current,
                seq,
                Ordering::Release,
                Ordering::Relaxed,
            ) {
                Ok(_) => break,
                Err(actual) => current = actual,
            }
        }
    }

    /// Scan points for a specific series within a time range.
    ///
    /// Returns points sorted by timestamp for the series identified by
    /// `series_key`. Only points within `[min_ts, max_ts]` are returned.
    /// Uses canonical form comparison (not hash) to avoid false positives
    /// from hash collisions.
    #[must_use]
    pub fn scan(&self, series_key: &SeriesKey, min_ts: Timestamp, max_ts: Timestamp) -> Vec<Point> {
        let hash = series_key.hash_fnv();
        let canonical = series_key.canonical_form();
        // Use the canonical form for exact range boundaries so the
        // scan covers only entries for this specific series.
        let start = MemtableKey::new(hash, min_ts, Arc::from(canonical));
        let end = MemtableKey::new(hash, max_ts, Arc::from(canonical));

        self.data
            .range(start..=end)
            .filter(|entry| &*entry.key().series_canonical == canonical)
            .filter_map(|entry| Self::entry_to_point(entry.key(), entry.value()))
            .collect()
    }

    /// Scan all points across all series within a time range.
    ///
    /// Returns all points sorted by `(series_key_hash, timestamp)`.
    ///
    /// **Note:** this is intentionally O(N) — it scans all series.
    /// For measurement-scoped queries (the common case), use
    /// `scan_measurement` which leverages the measurement index for
    /// O(S × log N) per-series range scans.
    #[must_use]
    pub fn scan_range(&self, min_ts: Timestamp, max_ts: Timestamp) -> Vec<Point> {
        // Scan the full BTree but filter by timestamp. No sentinel upper
        // bound needed — just iterate all entries and filter.
        self.data
            .iter()
            .filter(|entry| {
                let ts = entry.key().timestamp;
                ts >= min_ts && ts <= max_ts
            })
            .filter_map(|entry| Self::entry_to_point(entry.key(), entry.value()))
            .collect()
    }

    /// Scan all points in the memtable, sorted by `(series_key_hash,
    /// timestamp)`.
    #[must_use]
    pub fn scan_all(&self) -> Vec<Point> {
        self.data
            .iter()
            .filter_map(|entry| Self::entry_to_point(entry.key(), entry.value()))
            .collect()
    }

    /// Scan all points belonging to a specific measurement within a time range.
    ///
    /// Uses a secondary measurement index to perform targeted range scans
    /// per series hash, achieving O(S × log N) complexity instead of O(N)
    /// full-table iteration, where S is the number of distinct series for
    /// the measurement.
    #[must_use]
    pub fn scan_measurement(
        &self,
        measurement: &str,
        min_ts: Timestamp,
        max_ts: Timestamp,
    ) -> Vec<Point> {
        let hashes = match self.measurement_index.get(measurement) {
            Some(set) => set.iter().copied().collect::<Vec<_>>(),
            None => return Vec::new(),
        };

        let mut results = Vec::new();
        for hash in hashes {
            // Build range bounds for this series hash across the timestamp window.
            let start = MemtableKey::lower_bound(hash, min_ts);
            let end = MemtableKey::upper_bound(hash, max_ts);
            results.extend(
                self.data
                    .range(start..=end)
                    .filter(|entry| {
                        let ts = entry.key().timestamp;
                        ts >= min_ts && ts <= max_ts && &*entry.value().measurement == measurement
                    })
                    .filter_map(|entry| Self::entry_to_point(entry.key(), entry.value())),
            );
        }
        results
    }

    /// The canonical form of every series this memtable holds.
    ///
    /// The cardinality budget is derived from what the database holds, and a
    /// memtable holds series no segment does yet — so a repair that consulted
    /// only the segments on disk would release a series written a second ago
    /// and then count it again on the next write.
    ///
    /// `series_tags` is already keyed by canonical form, so this walks the
    /// map rather than scanning the points.
    #[must_use]
    pub fn series_canonical_forms(&self) -> Vec<Arc<str>> {
        self.series_tags
            .iter()
            .map(|e| Arc::clone(e.key()))
            .collect()
    }

    /// Hand every `(tag key, tag value)` pair this memtable holds to `f`.
    ///
    /// Label discovery has to answer from the memtable as well as from the
    /// segments: the tag inverted index is built at flush, so a series
    /// written a second ago is in no segment and therefore in no index, and
    /// an answer drawn from the index alone omits exactly the labels a fresh
    /// database has. A visitor rather than a `Vec`, because the caller is
    /// building a set and the pairs repeat once per series.
    pub fn visit_tags(&self, mut f: impl FnMut(&str, &str)) {
        for entry in &self.series_tags {
            for (key, value) in entry.value().iter() {
                f(key, value);
            }
        }
    }

    /// Returns the number of entries in the memtable.
    #[must_use]
    pub fn len(&self) -> usize {
        self.data.len()
    }

    /// Returns `true` if the memtable contains no entries.
    #[must_use]
    pub fn is_empty(&self) -> bool {
        self.data.is_empty()
    }

    /// Returns the estimated memory usage in bytes.
    ///
    /// Heap bytes held by this memtable's string interner.
    ///
    /// Reported separately from [`estimated_size`](Self::estimated_size),
    /// which counts rows and index entries: the interner holds the tag keys,
    /// tag values and measurement names, so it grows with *cardinality* rather
    /// than with row count and is the term a high-cardinality workload sees
    /// first.
    #[must_use]
    pub fn interner_bytes(&self) -> usize {
        self.interner.memory_bytes()
    }

    /// Includes the `measurement_index` DashMap overhead.
    #[must_use]
    pub fn estimated_size(&self) -> usize {
        self.estimated_size.load(Ordering::Relaxed) + self.index_size.load(Ordering::Relaxed)
    }

    /// The interned tag set of a series, created on its first point.
    ///
    /// A new series charges its canonical form, its tag slice and the map
    /// entry to the size estimate once; every later point of the series
    /// shares the `Arc`.
    fn series_tag_set(
        &self,
        canonical: &Arc<str>,
        series_key: &SeriesKey,
    ) -> Arc<[(Arc<str>, Arc<str>)]> {
        if let Some(existing) = self.series_tags.get(canonical) {
            return Arc::clone(existing.value());
        }
        let tags: Arc<[(Arc<str>, Arc<str>)]> = series_key
            .tags()
            .iter()
            .map(|(k, v)| (self.interner.intern(k), self.interner.intern(v)))
            .collect();
        let mut charged = 0usize;
        let stored = self
            .series_tags
            .entry(Arc::clone(canonical))
            .or_insert_with(|| {
                // Canonical form bytes (interned once), the tag slice, the
                // strings the slice points at (interned, charged here on
                // first sight), and the map entry.
                charged = canonical.len()
                    + tags.len() * std::mem::size_of::<(Arc<str>, Arc<str>)>()
                    + tags.iter().map(|(k, v)| k.len() + v.len()).sum::<usize>()
                    + 2 * std::mem::size_of::<usize>()
                    + 64;
                tags
            })
            .clone();
        if charged > 0 {
            self.estimated_size.fetch_add(charged, Ordering::Relaxed);
        }
        stored
    }

    /// Add `hash` to the index under `measurement`, returning the bytes the
    /// index grew by.
    ///
    /// The accounting matches what a full walk of the map would produce: an
    /// `Arc<str>` key (pointer pair + string bytes) plus the `HashSet`'s
    /// allocation (capacity × 8) and its own overhead. Growth is charged as
    /// the set's capacity actually changes, so an amortised rehash is
    /// accounted for on the insert that triggers it.
    fn index_insert(&self, measurement: Arc<str>, hash: u64) -> usize {
        const SET_OVERHEAD: usize = 56;
        match self.measurement_index.entry(measurement) {
            Entry::Occupied(mut occupied) => {
                let before = occupied.get().capacity();
                occupied.get_mut().insert(hash);
                (occupied.get().capacity() - before) * std::mem::size_of::<u64>()
            }
            Entry::Vacant(vacant) => {
                let key_bytes = std::mem::size_of::<usize>() * 2 + vacant.key().len();
                let mut set = HashSet::new();
                set.insert(hash);
                let set_bytes = set.capacity() * std::mem::size_of::<u64>() + SET_OVERHEAD;
                vacant.insert(set);
                key_bytes + set_bytes
            }
        }
    }

    /// Returns the minimum WAL sequence number tracked.
    ///
    /// Returns `None` if no WAL sequences have been recorded.
    #[must_use]
    pub fn min_wal_seq(&self) -> Option<u64> {
        let val = self.min_wal_seq.load(Ordering::Acquire);
        if val == u64::MAX { None } else { Some(val) }
    }

    /// Returns the maximum WAL sequence number tracked.
    ///
    /// Returns `None` if no WAL sequences have been recorded.
    #[must_use]
    pub fn max_wal_seq(&self) -> Option<u64> {
        let val = self.max_wal_seq.load(Ordering::Acquire);
        if val == 0 && self.min_wal_seq.load(Ordering::Acquire) == u64::MAX {
            None
        } else {
            Some(val)
        }
    }

    /// Returns whether this memtable has been frozen.
    #[must_use]
    pub fn is_frozen(&self) -> bool {
        self.frozen.load(Ordering::Acquire)
    }

    /// Freeze this memtable, preventing any further inserts.
    ///
    /// After freezing, all `insert` calls will return
    /// [`MemtableError::Frozen`].
    pub fn freeze(&self) {
        self.frozen.store(true, Ordering::Release);
    }

    /// Collect all points as a `Vec<Point>` for flushing to a segment.
    #[must_use]
    pub fn to_points(&self) -> Vec<Point> {
        self.scan_all()
    }

    /// Convert a skip list entry back into a [`Point`].
    ///
    /// Returns `None` if reconstruction fails (e.g. invalid series key),
    /// which is logged as a warning.
    fn entry_to_point(key: &MemtableKey, entry: &MemtableEntry) -> Option<Point> {
        let tags = entry
            .tags
            .iter()
            .map(|(k, v)| (k.to_string(), v.to_string()))
            .collect();
        let series_key = SeriesKey::new(entry.measurement.to_string(), tags)
            .map_err(|e| {
                // Emit metric so operators can alert on data loss
                // during flush, not just rely on log scanning.
                tracing::error!(
                    hash = key.series_key_hash,
                    error = %e,
                    "DATA LOSS: failed to reconstruct series key during flush"
                );
                metrics::counter!("chronix_memtable_dropped_points_total").increment(1);
                e
            })
            .ok()?;
        Point::new(
            series_key,
            entry
                .fields
                .iter()
                .map(|(k, v)| (k.to_string(), v.clone()))
                .collect(),
            key.timestamp,
        )
        .map_err(|e| {
            tracing::error!(
                hash = key.series_key_hash,
                timestamp = key.timestamp,
                error = %e,
                "DATA LOSS: failed to reconstruct point during flush"
            );
            metrics::counter!("chronix_memtable_dropped_points_total").increment(1);
            e
        })
        .ok()
    }
}

impl Default for Memtable {
    fn default() -> Self {
        Self::new()
    }
}

// SAFETY: `SkipMap` is `Send + Sync`, and all other fields are atomic types.
// The `Memtable` is safe to share across threads.
// (This is automatically derived, but we document it explicitly.)

#[cfg(test)]
mod tests {
    use super::*;
    use std::collections::BTreeMap;

    use chronix_core::FieldValue;

    fn make_point(measurement: &str, host: &str, ts: i64, value: f64) -> Point {
        let tags: BTreeMap<String, String> = [("host".to_string(), host.to_string())]
            .into_iter()
            .collect();
        let series_key = SeriesKey::new(measurement.to_string(), tags).unwrap();
        let fields: BTreeMap<String, FieldValue> = [("value".to_string(), FieldValue::F64(value))]
            .into_iter()
            .collect();
        Point::new(series_key, fields, ts).unwrap()
    }

    #[test]
    fn insert_and_scan_all() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 200, 2.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 300, 3.0)).unwrap();

        let points = mt.scan_all();
        assert_eq!(points.len(), 3);
        assert_eq!(points[0].timestamp(), 100);
        assert_eq!(points[1].timestamp(), 200);
        assert_eq!(points[2].timestamp(), 300);
    }

    #[test]
    fn insert_and_scan_series() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 200, 2.0)).unwrap();
        mt.insert(&make_point("cpu", "b", 150, 9.0)).unwrap();

        let key_a = SeriesKey::new(
            "cpu".to_string(),
            [("host".to_string(), "a".to_string())]
                .into_iter()
                .collect(),
        )
        .unwrap();

        let points = mt.scan(&key_a, 0, 300);
        assert_eq!(points.len(), 2);
        assert!(
            points
                .iter()
                .all(|p| p.series_key().tag("host") == Some("a"))
        );
    }

    #[test]
    fn scan_time_range() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 200, 2.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 300, 3.0)).unwrap();

        let key = SeriesKey::new(
            "cpu".to_string(),
            [("host".to_string(), "a".to_string())]
                .into_iter()
                .collect(),
        )
        .unwrap();

        let points = mt.scan(&key, 150, 250);
        assert_eq!(points.len(), 1);
        assert_eq!(points[0].timestamp(), 200);
    }

    #[test]
    fn last_write_wins() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 100, 99.0)).unwrap();

        let points = mt.scan_all();
        assert_eq!(points.len(), 1);
        assert_eq!(points[0].field("value"), Some(&FieldValue::F64(99.0)));
    }

    #[test]
    fn empty_scan() {
        let mt = Memtable::new();
        let key = SeriesKey::new("cpu".to_string(), BTreeMap::new()).unwrap();
        let points = mt.scan(&key, 0, 1000);
        assert!(points.is_empty());
    }

    /// `estimated_size` tracks the index incrementally; this pins it to the
    /// full walk it replaced, so the two cannot drift apart silently.
    #[test]
    fn index_accounting_matches_a_full_walk() {
        let mt = Memtable::new();
        for i in 0..500u64 {
            let measurement = format!("m{}", i % 7);
            let series = format!("s{}", i % 53);
            mt.insert(&make_point(&measurement, &series, i as i64, 1.0))
                .unwrap();
        }

        // The batch path has its own index-write site; cover it too.
        let batch: Vec<_> = (500..800u64)
            .map(|i| {
                make_point(
                    &format!("m{}", i % 11),
                    &format!("s{}", i % 97),
                    i as i64,
                    2.0,
                )
            })
            .collect();
        mt.insert_batch(&batch).unwrap();

        let walked: usize = mt
            .measurement_index
            .iter()
            .map(|entry| {
                let key_bytes = std::mem::size_of::<usize>() * 2 + entry.key().len();
                let set_bytes = entry.value().capacity() * std::mem::size_of::<u64>() + 56;
                key_bytes + set_bytes
            })
            .sum();

        assert_eq!(
            mt.index_size.load(Ordering::Relaxed),
            walked,
            "incremental index accounting drifted from a full walk"
        );
    }

    #[test]
    fn estimated_size_increases() {
        let mt = Memtable::new();
        assert_eq!(mt.estimated_size(), 0);
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        assert!(mt.estimated_size() > 0);
        let size_after_one = mt.estimated_size();
        mt.insert(&make_point("cpu", "a", 200, 2.0)).unwrap();
        assert!(mt.estimated_size() > size_after_one);
    }

    #[test]
    fn wal_seq_tracking() {
        let mt = Memtable::new();
        assert_eq!(mt.min_wal_seq(), None);
        assert_eq!(mt.max_wal_seq(), None);

        mt.insert_with_wal_seq(&make_point("cpu", "a", 100, 1.0), 10)
            .unwrap();
        assert_eq!(mt.min_wal_seq(), Some(10));
        assert_eq!(mt.max_wal_seq(), Some(10));

        mt.insert_with_wal_seq(&make_point("cpu", "a", 200, 2.0), 5)
            .unwrap();
        assert_eq!(mt.min_wal_seq(), Some(5));
        assert_eq!(mt.max_wal_seq(), Some(10));

        mt.insert_with_wal_seq(&make_point("cpu", "a", 300, 3.0), 20)
            .unwrap();
        assert_eq!(mt.min_wal_seq(), Some(5));
        assert_eq!(mt.max_wal_seq(), Some(20));
    }

    #[test]
    fn freeze_prevents_inserts() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.freeze();
        assert!(mt.is_frozen());

        let result = mt.insert(&make_point("cpu", "a", 200, 2.0));
        assert!(result.is_err());

        // Existing data is still readable
        let points = mt.scan_all();
        assert_eq!(points.len(), 1);
    }

    #[test]
    fn len_and_is_empty() {
        let mt = Memtable::new();
        assert!(mt.is_empty());
        assert_eq!(mt.len(), 0);

        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        assert!(!mt.is_empty());
        assert_eq!(mt.len(), 1);
    }

    #[test]
    fn multiple_series() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "b", 100, 2.0)).unwrap();
        mt.insert(&make_point("mem", "a", 100, 3.0)).unwrap();

        assert_eq!(mt.len(), 3);

        let all = mt.scan_all();
        assert_eq!(all.len(), 3);
    }

    #[test]
    fn concurrent_inserts() {
        use std::sync::Arc;
        use std::thread;

        let mt = Arc::new(Memtable::new());
        let mut handles = Vec::new();

        for thread_id in 0..4 {
            let mt = Arc::clone(&mt);
            handles.push(thread::spawn(move || {
                for i in 0..100 {
                    let ts = i64::from(thread_id * 1000 + i);
                    let host = format!("host-{thread_id}");
                    mt.insert(&make_point("cpu", &host, ts, ts as f64)).unwrap();
                }
            }));
        }

        for h in handles {
            h.join().unwrap();
        }

        assert_eq!(mt.len(), 400);
        let all = mt.scan_all();
        assert_eq!(all.len(), 400);

        // Verify ordering: each series should be ordered by timestamp
        let mut prev_key: Option<MemtableKey> = None;
        for p in &all {
            let key = MemtableKey::new(
                p.series_key().hash_fnv(),
                p.timestamp(),
                Arc::from(p.series_key().canonical_form()),
            );
            if let Some(prev) = prev_key {
                assert!(key >= prev, "entries should be ordered");
            }
            prev_key = Some(key);
        }
    }

    #[allow(deprecated)]
    #[test]
    fn to_points_is_scan_all() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 200, 2.0)).unwrap();

        let scan = mt.scan_all();
        let points = mt.to_points();
        assert_eq!(scan.len(), points.len());
        for (s, d) in scan.iter().zip(points.iter()) {
            assert_eq!(s.timestamp(), d.timestamp());
            assert_eq!(s.fields(), d.fields());
        }
    }

    #[test]
    fn scan_range_across_series() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 50, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 100, 2.0)).unwrap();
        mt.insert(&make_point("cpu", "b", 75, 3.0)).unwrap();
        mt.insert(&make_point("cpu", "b", 200, 4.0)).unwrap();

        let points = mt.scan_range(60, 150);
        // Should get: cpu/a@100 and cpu/b@75
        assert_eq!(points.len(), 2);
        for p in &points {
            assert!(
                p.timestamp() >= 60 && p.timestamp() <= 150,
                "timestamp {} should be in [60, 150]",
                p.timestamp()
            );
        }
    }

    #[test]
    fn scan_measurement_filters_by_name() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "b", 200, 2.0)).unwrap();
        mt.insert(&make_point("mem", "a", 150, 3.0)).unwrap();
        mt.insert(&make_point("mem", "a", 300, 4.0)).unwrap();
        mt.insert(&make_point("disk", "c", 250, 5.0)).unwrap();

        // Scan only "cpu"
        let cpu = mt.scan_measurement("cpu", 0, i64::MAX);
        assert_eq!(cpu.len(), 2);
        assert!(cpu.iter().all(|p| p.series_key().measurement() == "cpu"));

        // Scan only "mem"
        let mem = mt.scan_measurement("mem", 0, i64::MAX);
        assert_eq!(mem.len(), 2);
        assert!(mem.iter().all(|p| p.series_key().measurement() == "mem"));

        // Scan "disk"
        let disk = mt.scan_measurement("disk", 0, i64::MAX);
        assert_eq!(disk.len(), 1);

        // Scan non-existent measurement
        let none = mt.scan_measurement("net", 0, i64::MAX);
        assert!(none.is_empty());
    }

    #[test]
    fn scan_measurement_respects_time_range() {
        let mt = Memtable::new();
        mt.insert(&make_point("cpu", "a", 100, 1.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 200, 2.0)).unwrap();
        mt.insert(&make_point("cpu", "a", 300, 3.0)).unwrap();
        mt.insert(&make_point("mem", "a", 200, 4.0)).unwrap();

        let result = mt.scan_measurement("cpu", 150, 250);
        assert_eq!(result.len(), 1);
        assert_eq!(result[0].timestamp(), 200);
    }

    /// Regression test: two distinct series with the same FNV-1a hash
    /// and timestamp must both be stored and retrievable (no silent
    /// overwrite).
    #[test]
    fn hash_collision_preserves_both_series() {
        let mt = Memtable::new();

        // Create two points with the same hash and timestamp but
        // different canonical forms by manually constructing keys.
        // In production, FNV-1a collisions are rare but possible.
        let p1 = make_point("cpu", "host-a", 1000, 42.0);
        let p2 = make_point("cpu", "host-b", 1000, 99.0);

        mt.insert(&p1).unwrap();
        mt.insert(&p2).unwrap();

        // Both must be present
        assert_eq!(
            mt.len(),
            2,
            "two distinct series at same timestamp must both survive"
        );

        // Each series should return its own value
        let s1 = mt.scan(p1.series_key(), 0, i64::MAX);
        assert_eq!(s1.len(), 1);
        assert_eq!(
            *s1[0].field("value").unwrap(),
            chronix_core::FieldValue::F64(42.0)
        );

        let s2 = mt.scan(p2.series_key(), 0, i64::MAX);
        assert_eq!(s2.len(), 1);
        assert_eq!(
            *s2[0].field("value").unwrap(),
            chronix_core::FieldValue::F64(99.0)
        );
    }

    #[test]
    fn insert_batch_basic() {
        let mt = Memtable::new();
        let points = vec![
            make_point("cpu", "a", 100, 1.0),
            make_point("cpu", "a", 200, 2.0),
            make_point("cpu", "b", 300, 3.0),
            make_point("mem", "a", 400, 4.0),
        ];
        let count = mt.insert_batch(&points).unwrap();
        assert_eq!(count, 4);

        let all = mt.scan_all();
        assert_eq!(all.len(), 4);

        // Measurement index covers both measurements
        let cpu = mt.scan_measurement("cpu", 0, i64::MAX);
        assert_eq!(cpu.len(), 3);
        let mem = mt.scan_measurement("mem", 0, i64::MAX);
        assert_eq!(mem.len(), 1);
    }

    #[test]
    fn insert_batch_overwrite() {
        let mt = Memtable::new();
        let initial = vec![
            make_point("cpu", "a", 100, 1.0),
            make_point("cpu", "a", 200, 2.0),
        ];
        mt.insert_batch(&initial).unwrap();

        // Overwrite same timestamps with new values
        let overwrite = vec![
            make_point("cpu", "a", 100, 10.0),
            make_point("cpu", "a", 200, 20.0),
        ];
        mt.insert_batch(&overwrite).unwrap();

        let all = mt.scan_all();
        assert_eq!(all.len(), 2);
        assert_eq!(
            *all[0].field("value").unwrap(),
            chronix_core::FieldValue::F64(10.0)
        );
        assert_eq!(
            *all[1].field("value").unwrap(),
            chronix_core::FieldValue::F64(20.0)
        );
    }

    #[test]
    fn insert_batch_frozen_fails() {
        let mt = Memtable::new();
        mt.freeze();
        let points = vec![make_point("cpu", "a", 100, 1.0)];
        assert!(mt.insert_batch(&points).is_err());
    }

    #[test]
    fn insert_batch_with_wal_seq_updates_range() {
        let mt = Memtable::new();
        let points = vec![
            make_point("cpu", "a", 100, 1.0),
            make_point("cpu", "a", 200, 2.0),
        ];
        mt.insert_batch_with_wal_seq(&points, 42).unwrap();
        assert_eq!(mt.min_wal_seq(), Some(42));
        assert_eq!(mt.max_wal_seq(), Some(42));
    }
}

/// The size estimate against the allocator's number, not against itself.
///
/// This is its own module because it installs a counting global allocator
/// for the whole test binary; that is cheap (two atomics per allocation)
/// and it is the only way to ask what a memtable *actually* costs.
#[cfg(test)]
mod calibration {
    use std::alloc::{GlobalAlloc, Layout, System};
    use std::collections::BTreeMap;
    use std::sync::atomic::{AtomicUsize, Ordering};

    use chronix_core::{FieldValue, Point, SeriesKey};

    use super::Memtable;

    struct Counting;
    static LIVE: AtomicUsize = AtomicUsize::new(0);

    // Bytes allocated minus freed *on this thread*, so the other tests
    // running in parallel in this binary do not show up in the window.
    std::thread_local! {
        static THREAD_LIVE: std::cell::Cell<isize> = const { std::cell::Cell::new(0) };
    }

    fn thread_live() -> isize {
        THREAD_LIVE.with(std::cell::Cell::get)
    }

    // SAFETY: every operation is forwarded to the system allocator; the
    // counters are the only addition. The thread-local is `const`-initialised
    // so touching it never allocates, and `try_with` tolerates teardown.
    #[allow(unsafe_code)]
    unsafe impl GlobalAlloc for Counting {
        unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
            LIVE.fetch_add(layout.size(), Ordering::Relaxed);
            let _ = THREAD_LIVE.try_with(|c| c.set(c.get() + layout.size() as isize));
            unsafe { System.alloc(layout) }
        }
        unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
            LIVE.fetch_sub(layout.size(), Ordering::Relaxed);
            let _ = THREAD_LIVE.try_with(|c| c.set(c.get() - layout.size() as isize));
            unsafe { System.dealloc(ptr, layout) }
        }
    }

    #[global_allocator]
    static ALLOC: Counting = Counting;

    /// The design partner's shape: 50 series, two fields, a few hundred
    /// points each. The estimate must land within a quarter of what the
    /// allocator says — it used to be eight times below it.
    #[test]
    fn memtable_estimate_tracks_real_allocation() {
        let keys: Vec<SeriesKey> = (0..50)
            .map(|m| {
                SeriesKey::new(
                    "power",
                    BTreeMap::from([
                        ("meter".to_string(), format!("m{m:02}")),
                        ("site".to_string(), "home".to_string()),
                    ]),
                )
                .unwrap()
            })
            .collect();
        // Build the points first so their allocation is outside the window.
        let points: Vec<Point> = (0..400i64)
            .flat_map(|s| {
                keys.iter().map(move |k| {
                    Point::new(
                        k.clone(),
                        BTreeMap::from([
                            ("w".to_string(), FieldValue::F64(230.0 + s as f64)),
                            ("kwh".to_string(), FieldValue::F64(s as f64 / 1000.0)),
                        ]),
                        s * 1_000_000_000,
                    )
                    .unwrap()
                })
            })
            .collect();

        let before = thread_live();
        let mt = Memtable::new();
        for p in &points {
            mt.insert(p).unwrap();
        }
        let measured = usize::try_from(thread_live() - before).unwrap_or(0);
        let estimated = mt.estimated_size();
        let ratio = estimated as f64 / measured as f64;
        assert!(
            (0.75..=1.25).contains(&ratio),
            "estimate {estimated} B vs allocator {measured} B for {} points: ratio {ratio:.2} \
             ({:.0} B/point measured, {:.0} B/point estimated)",
            points.len(),
            measured as f64 / points.len() as f64,
            estimated as f64 / points.len() as f64,
        );
        drop(mt);
    }
}

/// One measurement's rows, ready for the segment writer.
#[derive(Debug)]
pub struct MeasurementBatch {
    /// The measurement.
    pub measurement: String,
    /// Its rows: `timestamp`, then the tag columns, then the field columns,
    /// each group sorted by name; rows in skip-list order (series-major,
    /// then time).
    pub batch: arrow::record_batch::RecordBatch,
    /// Which of the columns are tags.
    pub tag_columns: Vec<String>,
}

/// Per-column Arrow builder for one field.
enum FieldBuilder {
    F64(arrow::array::Float64Builder),
    I64(arrow::array::Int64Builder),
    U64(arrow::array::UInt64Builder),
    Bool(arrow::array::BooleanBuilder),
    Str(arrow::array::StringBuilder),
    /// An exact decimal column, with the scale the first value carried.
    ///
    /// The scale is fixed for the column by the write path, which rescales
    /// every value to the schema's declared scale before it reaches the
    /// memtable — so the first value's scale is the column's scale.
    Decimal(arrow::array::Decimal128Builder, u8),
    /// A native histogram column.
    ///
    /// Arrow has no composite-sample type, so each value is a `postcard`
    /// blob in a `Binary` column and the query layer decodes it. That is the
    /// same shape Prometheus uses — a histogram is read through the
    /// `histogram_*` family, never through arithmetic — and it keeps the
    /// Arrow schema of a histogram column stable regardless of resolution,
    /// which a struct-of-lists encoding would not.
    Histogram(arrow::array::BinaryBuilder),
}

impl FieldBuilder {
    fn for_value(v: &chronix_core::FieldValue, rows_before: usize) -> Self {
        use chronix_core::FieldValue as V;
        let mut b = match v {
            V::F64(_) => Self::F64(arrow::array::Float64Builder::new()),
            V::I64(_) => Self::I64(arrow::array::Int64Builder::new()),
            V::U64(_) => Self::U64(arrow::array::UInt64Builder::new()),
            V::Bool(_) => Self::Bool(arrow::array::BooleanBuilder::new()),
            V::String(_) => Self::Str(arrow::array::StringBuilder::new()),
            // The Arrow type is set on the builder rather than on the
            // finished array: `with_precision_and_scale` re-validates every
            // value and can fail, and there is no honest thing to return
            // from a failure here. The scale comes from a validated
            // `Decimal`, so it is inside Arrow's range by construction.
            V::Decimal(d) => Self::Decimal(
                arrow::array::Decimal128Builder::new().with_data_type(
                    arrow::datatypes::DataType::Decimal128(
                        chronix_core::DECIMAL_PRECISION,
                        d.scale() as i8,
                    ),
                ),
                d.scale(),
            ),
            V::Histogram(_) => Self::Histogram(arrow::array::BinaryBuilder::new()),
        };
        for _ in 0..rows_before {
            b.append_null();
        }
        b
    }

    fn append_null(&mut self) {
        match self {
            Self::F64(b) => b.append_null(),
            Self::I64(b) => b.append_null(),
            Self::U64(b) => b.append_null(),
            Self::Bool(b) => b.append_null(),
            Self::Str(b) => b.append_null(),
            Self::Decimal(b, _) => b.append_null(),
            Self::Histogram(b) => b.append_null(),
        }
    }

    /// Append a value; a value of another type than the column's is a null,
    /// because the schema registry has already refused it at write time
    /// and this is the last line of defence, not the first.
    fn append(&mut self, v: &chronix_core::FieldValue) {
        use chronix_core::FieldValue as V;
        match (self, v) {
            (Self::F64(b), V::F64(x)) => b.append_value(*x),
            (Self::I64(b), V::I64(x)) => b.append_value(*x),
            (Self::U64(b), V::U64(x)) => b.append_value(*x),
            (Self::Bool(b), V::Bool(x)) => b.append_value(*x),
            (Self::Str(b), V::String(x)) => b.append_value(x),
            // A value at another scale is rescaled rather than dropped: the
            // write path has already normalised the batch, so reaching this
            // means the value came from somewhere that had not — and a
            // lossless widening is still the right answer.
            (Self::Decimal(b, scale), V::Decimal(x)) => match x.rescale(*scale) {
                Ok(v) => b.append_value(v.mantissa()),
                Err(_) => b.append_null(),
            },
            // A histogram that cannot be encoded is a null rather than a
            // panic, for the same reason a mistyped value is: the schema
            // registry refused it at write time and this is the last line of
            // defence. `postcard` on an owned struct with no borrowed data
            // has no failure mode in practice, so this arm is unreachable
            // rather than lossy.
            (Self::Histogram(b), V::Histogram(h)) => match postcard::to_allocvec(h.as_ref()) {
                Ok(bytes) => b.append_value(bytes),
                Err(_) => b.append_null(),
            },
            (this, _) => this.append_null(),
        }
    }

    fn finish(self) -> (arrow::datatypes::DataType, arrow::array::ArrayRef) {
        use arrow::datatypes::DataType;
        match self {
            Self::Histogram(mut b) => (DataType::Binary, Arc::new(b.finish())),
            Self::F64(mut b) => (DataType::Float64, Arc::new(b.finish())),
            Self::I64(mut b) => (DataType::Int64, Arc::new(b.finish())),
            Self::U64(mut b) => (DataType::UInt64, Arc::new(b.finish())),
            Self::Bool(mut b) => (DataType::Boolean, Arc::new(b.finish())),
            Self::Str(mut b) => (DataType::Utf8, Arc::new(b.finish())),
            Self::Decimal(mut b, scale) => {
                let dt = DataType::Decimal128(chronix_core::DECIMAL_PRECISION, scale as i8);
                (dt, Arc::new(b.finish()))
            }
        }
    }
}

/// Builders for one measurement, growing columns as the entries reveal them.
struct MeasurementBuilder {
    rows: usize,
    timestamps: arrow::array::Int64Builder,
    tags: std::collections::BTreeMap<Arc<str>, arrow::array::StringBuilder>,
    fields: std::collections::BTreeMap<Arc<str>, FieldBuilder>,
}

impl MeasurementBuilder {
    fn new() -> Self {
        Self {
            rows: 0,
            timestamps: arrow::array::Int64Builder::new(),
            tags: std::collections::BTreeMap::new(),
            fields: std::collections::BTreeMap::new(),
        }
    }

    fn push(&mut self, key: &MemtableKey, entry: &MemtableEntry) {
        self.timestamps.append_value(key.timestamp);
        // Tags: every known column gets a value or a null; a new column is
        // back-filled with nulls for the rows before it.
        for (k, v) in entry.tags.iter() {
            let rows = self.rows;
            self.tags.entry(Arc::clone(k)).or_insert_with(|| {
                let mut b = arrow::array::StringBuilder::new();
                for _ in 0..rows {
                    b.append_null();
                }
                b
            });
            let _ = v;
        }
        for (name, b) in &mut self.tags {
            match entry.tags.iter().find(|(k, _)| k == name) {
                Some((_, v)) => b.append_value(v),
                None => b.append_null(),
            }
        }
        for (k, v) in entry.fields.iter() {
            let rows = self.rows;
            self.fields
                .entry(Arc::clone(k))
                .or_insert_with(|| FieldBuilder::for_value(v, rows));
        }
        for (name, b) in &mut self.fields {
            match entry.fields.iter().find(|(k, _)| k == name) {
                Some((_, v)) => b.append(v),
                None => b.append_null(),
            }
        }
        self.rows += 1;
    }

    fn finish(mut self, measurement: &str) -> Option<MeasurementBatch> {
        use arrow::datatypes::{DataType, Field, Schema};
        if self.rows == 0 {
            return None;
        }
        use crate::segment::metadata::roles;
        // A tag and a string field are the same Arrow type, so the schema has
        // to carry the role or every consumer guesses at it.
        let mut fields = vec![
            Field::new(chronix_core::TIME_COLUMN, DataType::Int64, false)
                .with_metadata(roles::arrow_metadata(roles::TIMESTAMP)),
        ];
        let mut columns: Vec<arrow::array::ArrayRef> = vec![Arc::new(self.timestamps.finish())];
        let tag_columns: Vec<String> = self.tags.keys().map(ToString::to_string).collect();
        for (name, mut b) in self.tags {
            fields.push(
                Field::new(name.as_ref(), DataType::Utf8, true)
                    .with_metadata(roles::arrow_metadata(roles::TAG)),
            );
            columns.push(Arc::new(b.finish()));
        }
        for (name, b) in self.fields {
            let (dt, arr) = b.finish();
            fields.push(
                Field::new(name.as_ref(), dt, true)
                    .with_metadata(roles::arrow_metadata(roles::FIELD)),
            );
            columns.push(arr);
        }
        let batch =
            arrow::record_batch::RecordBatch::try_new(Arc::new(Schema::new(fields)), columns)
                .ok()?;
        Some(MeasurementBatch {
            measurement: measurement.to_string(),
            batch,
            tag_columns,
        })
    }
}

impl Memtable {
    /// Every row, as one Arrow batch per measurement, built straight from
    /// the skip-list entries.
    ///
    /// This is what a flush hands the segment writer. It replaced
    /// `to_points()` on that path: materialising a full memtable as
    /// `Point`s — two `BTreeMap`s each — cost a kilobyte a row and made the
    /// flush of an 8 MB memtable peak at ten times that. A batch is the
    /// columns themselves, and the writer encodes it without another copy.
    #[must_use]
    pub fn to_record_batches(&self) -> Vec<MeasurementBatch> {
        let mut builders: std::collections::BTreeMap<Arc<str>, MeasurementBuilder> =
            std::collections::BTreeMap::new();
        for entry in self.data.iter() {
            let value = entry.value();
            builders
                .entry(Arc::clone(&value.measurement))
                .or_insert_with(MeasurementBuilder::new)
                .push(entry.key(), value);
        }
        builders
            .into_iter()
            .filter_map(|(m, b)| b.finish(&m))
            .collect()
    }
}