Speed up range-query intersections via seek_danger on RangeDocSet (up to ~50x faster)

A regular seek on RangeDocSet is costly: on a miss it fetches blocks and
scans the column forward to materialize the next matching doc. As a
non-leading docset in an intersection that work is wasted — the driver only
asks "does this candidate match?". seek_danger answers that with a cheap
point lookup via Column::values_for_doc, returning a lower bound on a miss
and leaving forward progress to the caller.

Forward seek_danger through ConstScorer.

Benchmarks (bool_queries_with_range, _all_results / DocSetCollector):

```
dense and 0.1% a
a_AND_num_rand:[0_TO_9]_all_results                                 Avg: 0.0827ms (-4.60%)     Median: 0.0825ms (-4.82%)     [0.0809ms .. 0.0891ms]    Output: 43
a_AND_num_asc:[0_TO_9]_all_results                                  Avg: 0.1937ms (-3.70%)     Median: 0.1930ms (-3.59%)     [0.1806ms .. 0.2044ms]    Output: 100
a_AND_num_rand_fast:[0_TO_9]_all_results                            Avg: 0.0367ms (-92.67%)    Median: 0.0365ms (-92.65%)    [0.0340ms .. 0.0398ms]    Output: 43
a_AND_num_asc_fast:[0_TO_9]_all_results                             Avg: 0.1052ms (-98.05%)    Median: 0.1050ms (-97.98%)    [0.1009ms .. 0.1117ms]    Output: 100
num_rand_fast:[0_TO_9]_AND_num_asc_fast:[0_TO_9]_all_results        Avg: 2.7147ms (-51.42%)    Median: 2.7075ms (-49.58%)    [2.6806ms .. 2.7799ms]    Output: 968
dense and 1% a
a_AND_num_rand:[0_TO_9]_all_results                                 Avg: 0.4373ms (-9.71%)     Median: 0.4357ms (-10.12%)    [0.4117ms .. 0.4711ms]    Output: 463
a_AND_num_asc:[0_TO_9]_all_results                                  Avg: 0.2342ms (-2.50%)     Median: 0.2338ms (-2.56%)     [0.2247ms .. 0.2452ms]    Output: 1_054
a_AND_num_rand_fast:[0_TO_9]_all_results                            Avg: 0.3956ms (-82.86%)    Median: 0.3943ms (-82.90%)    [0.3815ms .. 0.4119ms]    Output: 463
a_AND_num_asc_fast:[0_TO_9]_all_results                             Avg: 0.4896ms (-91.16%)    Median: 0.4862ms (-90.81%)    [0.4797ms .. 0.5084ms]    Output: 1_054
num_rand_fast:[0_TO_9]_AND_num_asc_fast:[0_TO_9]_all_results        Avg: 2.7108ms (-50.81%)    Median: 2.6925ms (-49.51%)    [2.6688ms .. 2.7868ms]    Output: 968
dense and 10% a
a_AND_num_rand:[0_TO_9]_all_results                                 Avg: 0.9869ms (-3.71%)     Median: 0.9833ms (-3.83%)     [0.9518ms .. 1.1218ms]    Output: 4_914
a_AND_num_asc:[0_TO_9]_all_results                                  Avg: 0.6352ms (-3.74%)     Median: 0.6363ms (-3.32%)     [0.6158ms .. 0.6488ms]    Output: 10_152
a_AND_num_rand_fast:[0_TO_9]_all_results                            Avg: 3.1264ms (+0.39%)     Median: 3.1466ms (+1.34%)     [3.0261ms .. 3.2051ms]    Output: 4_914
a_AND_num_asc_fast:[0_TO_9]_all_results                             Avg: 4.1547ms (-31.12%)    Median: 4.0933ms (-28.55%)    [3.7648ms .. 4.7600ms]    Output: 10_152
num_rand_fast:[0_TO_9]_AND_num_asc_fast:[0_TO_9]_all_results        Avg: 2.6973ms (-52.30%)    Median: 2.6901ms (-49.86%)    [2.6689ms .. 2.7677ms]    Output: 968
```

Gains are largest when the range query is the non-leading docset of a low-cardinality intersection.

.github/workflows/coverage.yml

@@ -20,7 +20,7 @@ jobs:
       contents: read
 
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
       - name: Install Rust
         run: rustup toolchain install nightly-2025-12-01 --profile minimal --component llvm-tools-preview
       - uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
@@ -28,7 +28,7 @@ jobs:
       - name: Generate code coverage
         run: cargo +nightly-2025-12-01 llvm-cov --all-features --workspace --doctests --lcov --output-path lcov.info
       - name: Upload coverage to Codecov
-        uses: codecov/codecov-action@57e3a136b779b570ffcdbf80b3bdc90e7fab3de2 # v6.0.0
+        uses: codecov/codecov-action@fb8b3582c8e4def4969c97caa2f19720cb33a72f # v7.0.0
         continue-on-error: true
         with:
           token: ${{ secrets.CODECOV_TOKEN }} # not required for public repos

.github/workflows/long_running.yml

@@ -25,7 +25,7 @@ jobs:
       contents: read
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
     - name: Install stable
       uses: actions-rs/toolchain@16499b5e05bf2e26879000db0c1d13f7e13fa3af # v1.0.7
       with:

.github/workflows/scorecard.yml

@@ -22,7 +22,7 @@ jobs:
 
     steps:
       - name: 'Checkout code'
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
         with:
           persist-credentials: false
 
@@ -36,7 +36,7 @@ jobs:
 
       # Upload the results as artifacts.
       - name: 'Upload artifact'
-        uses: actions/upload-artifact@bbbca2ddaa5d8feaa63e36b76fdaad77386f024f # v7.0.0
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: SARIF file
           path: results.sarif
@@ -44,6 +44,6 @@ jobs:
 
       # Upload the results to GitHub's code scanning dashboard.
       - name: 'Upload to code-scanning'
-        uses: github/codeql-action/upload-sarif@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v4.35.2
+        uses: github/codeql-action/upload-sarif@87557b9c84dde89fdd9b10e88954ac2f4248e463 # v4.36.1
         with:
           sarif_file: results.sarif

.github/workflows/test.yml

@@ -27,7 +27,7 @@ jobs:
       checks: write
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
 
     - name: Install nightly
       uses: actions-rs/toolchain@16499b5e05bf2e26879000db0c1d13f7e13fa3af # v1.0.7
@@ -77,7 +77,7 @@ jobs:
     name: test-${{ matrix.features.label}}
 
     steps:
-    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+    - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
 
     - name: Install stable
       uses: actions-rs/toolchain@16499b5e05bf2e26879000db0c1d13f7e13fa3af # v1.0.7

Cargo.toml

@@ -65,7 +65,7 @@ tantivy-bitpacker = { version = "0.10", path = "./bitpacker" }
 common = { version = "0.11", path = "./common/", package = "tantivy-common" }
 tokenizer-api = { version = "0.7", path = "./tokenizer-api", package = "tantivy-tokenizer-api" }
 sketches-ddsketch = { version = "0.4", features = ["use_serde"] }
-datasketches = { git = "https://github.com/fulmicoton-dd/datasketches-rust", rev = "7635fb8" }
+datasketches = { version = "0.3.0", features = ["hll"] }
 futures-util = { version = "0.3.28", optional = true }
 futures-channel = { version = "0.3.28", optional = true }
 fnv = "1.0.7"
@@ -75,7 +75,7 @@ typetag = "0.2.21"
 winapi = "0.3.9"
 
 [dev-dependencies]
-binggan = "0.16.1"
+binggan = "0.17.0"
 rand = "0.9"
 maplit = "1.0.2"
 matches = "0.1.9"

benches/agg_bench.rs

@@ -66,6 +66,9 @@ fn bench_agg(mut group: InputGroup<Index>) {
     register!(group, terms_status_with_terms_zipf_1000_sub_agg);
     register!(group, terms_zipf_1000_with_terms_status_sub_agg);
     register!(group, terms_status_with_histogram);
+    register!(group, terms_status_with_date_histogram);
+    register!(group, terms_status_with_date_histogram_hard_bounds);
+    register!(group, terms_status_with_date_histogram_and_sibling_terms);
     register!(group, terms_zipf_1000);
     register!(group, terms_zipf_1000_with_histogram);
     register!(group, terms_zipf_1000_with_avg_sub_agg);
@@ -390,6 +393,57 @@ fn terms_status_with_histogram(index: &Index) {
     execute_agg(index, agg_req);
 }
 
+fn terms_status_with_date_histogram(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_few_terms_status" },
+            "aggs": {
+                "over_time": { "date_histogram": { "field": "timestamp", "fixed_interval": "1h" } }
+            }
+        }
+    });
+    execute_agg(index, agg_req);
+}
+
+/// Same fused terms × date_histogram, but with `hard_bounds`. The timestamps span 0..120h; the
+/// bounds drop only the first and last hour (ms: 1h=3_600_000, 119h=428_400_000), so almost every
+/// doc is in-bounds. This exercises the collector's hard-bounds path: `bounds.contains` runs per
+/// doc (the `all_docs_in_bounds` short-circuit is off) and the rare out-of-bounds doc takes the
+/// `term_counts` branch.
+fn terms_status_with_date_histogram_hard_bounds(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_few_terms_status" },
+            "aggs": {
+                "over_time": {
+                    "date_histogram": {
+                        "field": "timestamp",
+                        "fixed_interval": "1h",
+                        "hard_bounds": { "min": 3_600_000, "max": 428_400_000 }
+                    }
+                }
+            }
+        }
+    });
+    execute_agg(index, agg_req);
+}
+
+/// Same fused terms × date_histogram, but with a sibling terms aggregation next to it. The fused
+/// fast path should still trigger for `my_texts` (sibling aggregations are independent top-level
+/// aggregations, so they don't change its eligibility).
+fn terms_status_with_date_histogram_and_sibling_terms(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_few_terms_status" },
+            "aggs": {
+                "over_time": { "date_histogram": { "field": "timestamp", "fixed_interval": "1h" } }
+            }
+        },
+        "other_texts": { "terms": { "field": "text_few_terms" } }
+    });
+    execute_agg(index, agg_req);
+}
+
 fn terms_zipf_1000_with_histogram(index: &Index) {
     let agg_req = json!({
         "my_texts": {
@@ -783,7 +837,9 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
             doc_with_value /= 20;
         }
         let _val_max = 1_000_000.0;
-        for _ in 0..doc_with_value {
+        const SPAN_MS: i64 = 120 * 3600 * 1000; // 120 hours in ms
+        const NOISE_MS: i64 = 2 * 3600 * 1000; // ±2h noise
+        for i in 0..doc_with_value {
             let val: f64 = rng.random_range(0.0..1_000_000.0);
             let json = if rng.random_bool(0.1) {
                 // 10% are numeric values
@@ -791,6 +847,9 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
             } else {
                 json!({"mixed_type": many_terms_data.choose(&mut rng).unwrap().to_string()})
             };
+            let base_ms = (i as i64 * SPAN_MS) / doc_with_value as i64;
+            let noise_ms = rng.random_range(-NOISE_MS..NOISE_MS);
+            let ts_ms = (base_ms + noise_ms).clamp(0, SPAN_MS);
             index_writer.add_document(doc!(
                 single_term => "single_term",
                 text_field => "cool",
@@ -803,7 +862,7 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
                 score_field => val as u64,
                 score_field_f64 => lg_norm.sample(&mut rng),
                 score_field_i64 => val as i64,
-                date_field => DateTime::from_timestamp_millis((val * 1_000_000.) as i64),
+                date_field => DateTime::from_timestamp_millis(ts_ms),
             ))?;
             if cardinality == Cardinality::OptionalSparse {
                 for _ in 0..20 {

benches/bool_queries_with_range.rs

@@ -110,43 +110,31 @@ fn main() {
     // Prepare corpora with varying scenarios
     let scenarios = vec![
         (
-            "dense and 99% a".to_string(),
-            10_000_000,
-            0.99,
+            "dense and 0.1% a".to_string(),
+            5_000_000,
+            0.001,
             "dense",
             0,
             9,
         ),
+        ("dense and 1% a".to_string(), 5_000_000, 0.01, "dense", 0, 9),
+        ("dense and 10% a".to_string(), 5_000_000, 0.1, "dense", 0, 9),
         (
-            "dense and 99% a".to_string(),
-            10_000_000,
-            0.99,
-            "dense",
-            990,
-            999,
-        ),
-        (
-            "sparse and 99% a".to_string(),
-            10_000_000,
+            "sparse and 50% a".to_string(),
+            5_000_000,
             0.99,
             "sparse",
             0,
             9,
         ),
-        (
-            "sparse and 99% a".to_string(),
-            10_000_000,
-            0.99,
-            "sparse",
-            9_999_990,
-            9_999_999,
-        ),
     ];
 
     let mut runner = BenchRunner::new();
-    for (scenario_id, n, p_title_a, num_rand_distribution, range_low, range_high) in scenarios {
+    for (scenario_id, num_docs, p_title_a, num_rand_distribution, range_low, range_high) in
+        scenarios
+    {
         // Build index for this scenario
-        let bench_index = build_shared_indices(n, p_title_a, num_rand_distribution);
+        let bench_index = build_shared_indices(num_docs, p_title_a, num_rand_distribution);
 
         // Create benchmark group
         let mut group = runner.new_group();
@@ -158,7 +146,7 @@ fn main() {
         let field_names = ["num_rand", "num_asc", "num_rand_fast", "num_asc_fast"];
 
         // Define the three terms we want to test with
-        let terms = ["a", "b", "z"];
+        let terms = ["a"];
 
         // Generate all combinations of terms and field names
         let mut queries = Vec::new();
@@ -203,7 +191,7 @@ fn run_benchmark_tasks(
         bench_index,
         query_str,
         DocSetCollector,
-        "all results",
+        "all_results",
     );
 
     // Test top 100 by the field (if it's a FAST field)

bitpacker/Cargo.toml

@@ -18,5 +18,10 @@ homepage = "https://github.com/quickwit-oss/tantivy"
 bitpacking = { version = "0.9.2", default-features = false, features = ["bitpacker1x"] }
 
 [dev-dependencies]
+binggan = "0.17.0"
 rand = "0.9"
 proptest = "1"
+
+[[bench]]
+name = "bench"
+harness = false

bitpacker/benches/bench.rs

@@ -1,65 +1,110 @@
-#![feature(test)]
+use std::cell::RefCell;
 
-extern crate test;
+use binggan::{BenchRunner, black_box};
+use rand::rng;
+use rand::seq::IteratorRandom;
+use tantivy_bitpacker::{BitPacker, BitUnpacker, BlockedBitpacker};
 
-#[cfg(test)]
-mod tests {
-    use rand::rng;
-    use rand::seq::IteratorRandom;
-    use tantivy_bitpacker::{BitPacker, BitUnpacker, BlockedBitpacker};
-    use test::Bencher;
+fn create_bitpacked_data(bit_width: u8, num_els: u32) -> Vec<u8> {
+    let mut bitpacker = BitPacker::new();
+    let mut buffer = Vec::new();
+    for _ in 0..num_els {
+        bitpacker.write(0u64, bit_width, &mut buffer).unwrap();
+        bitpacker.flush(&mut buffer).unwrap();
+    }
+    buffer
+}
 
-    #[inline(never)]
-    fn create_bitpacked_data(bit_width: u8, num_els: u32) -> Vec<u8> {
-        let mut bitpacker = BitPacker::new();
-        let mut buffer = Vec::new();
-        for _ in 0..num_els {
-            // the values do not matter.
-            bitpacker.write(0u64, bit_width, &mut buffer).unwrap();
-            bitpacker.flush(&mut buffer).unwrap();
+const N: usize = 100_000;
+const MAX_VAL: u64 = 1_000;
+const BIT_WIDTH: u8 = 10; // 2^10 = 1024 > MAX_VAL
+
+fn create_packed_data() -> (BitUnpacker, Vec<u8>) {
+    let mut bitpacker = BitPacker::new();
+    let mut data = Vec::new();
+    for i in 0..N as u64 {
+        let val = i * MAX_VAL / N as u64;
+        bitpacker.write(val, BIT_WIDTH, &mut data).unwrap();
+    }
+    bitpacker.close(&mut data).unwrap();
+    (BitUnpacker::new(BIT_WIDTH), data)
+}
+
+fn bench_bitpacking() {
+    let mut runner = BenchRunner::new();
+    let bit_width = 3;
+    let num_els = 1_000_000u32;
+    let bit_unpacker = BitUnpacker::new(bit_width);
+    let data = create_bitpacked_data(bit_width, num_els);
+    let idxs: Vec<u32> = (0..num_els).choose_multiple(&mut rng(), 100_000);
+    runner.bench_function("bitpacking_read", move |_| {
+        let mut out = 0u64;
+        for &idx in &idxs {
+            out = out.wrapping_add(bit_unpacker.get(idx, &data[..]));
         }
-        buffer
-    }
+        black_box(out);
+    });
+}
 
-    #[bench]
-    fn bench_bitpacking_read(b: &mut Bencher) {
-        let bit_width = 3;
-        let num_els = 1_000_000u32;
-        let bit_unpacker = BitUnpacker::new(bit_width);
-        let data = create_bitpacked_data(bit_width, num_els);
-        let idxs: Vec<u32> = (0..num_els).choose_multiple(&mut rng(), 100_000);
-        b.iter(|| {
-            let mut out = 0u64;
-            for &idx in &idxs {
-                out = out.wrapping_add(bit_unpacker.get(idx, &data[..]));
-            }
-            out
-        });
+fn bench_blocked_bitpacker() {
+    let mut runner = BenchRunner::new();
+    let mut blocked_bitpacker = BlockedBitpacker::new();
+    for val in 0..=21500 {
+        blocked_bitpacker.add(val * val);
     }
-
-    #[bench]
-    fn bench_blockedbitp_read(b: &mut Bencher) {
+    runner.bench_function("blockedbitp_read", move |_| {
+        let mut out = 0u64;
+        for val in 0..=21500 {
+            out = out.wrapping_add(blocked_bitpacker.get(val));
+        }
+        black_box(out);
+    });
+    runner.bench_function("blockedbitp_create", |_| {
         let mut blocked_bitpacker = BlockedBitpacker::new();
         for val in 0..=21500 {
             blocked_bitpacker.add(val * val);
         }
-        b.iter(|| {
-            let mut out = 0u64;
-            for val in 0..=21500 {
-                out = out.wrapping_add(blocked_bitpacker.get(val));
-            }
-            out
-        });
-    }
-
-    #[bench]
-    fn bench_blockedbitp_create(b: &mut Bencher) {
-        b.iter(|| {
-            let mut blocked_bitpacker = BlockedBitpacker::new();
-            for val in 0..=21500 {
-                blocked_bitpacker.add(val * val);
-            }
-            blocked_bitpacker
-        });
-    }
+        black_box(blocked_bitpacker);
+    });
+}
+
+fn bench_filter_vec() {
+    let mut runner = BenchRunner::new();
+
+    let (unpacker, data) = create_packed_data();
+    let positions = RefCell::new(Vec::with_capacity(N));
+    runner.bench_function("filter_vec_dense", move |_| {
+        unpacker.get_ids_for_value_range(
+            250..=750,
+            0..N as u32,
+            &data,
+            &mut positions.borrow_mut(),
+        );
+        black_box(positions.borrow().len());
+    });
+
+    let (unpacker, data) = create_packed_data();
+    let positions = RefCell::new(Vec::with_capacity(N));
+    runner.bench_function("filter_vec_sparse", move |_| {
+        unpacker.get_ids_for_value_range(0..=50, 0..N as u32, &data, &mut positions.borrow_mut());
+        black_box(positions.borrow().len());
+    });
+
+    let (unpacker, data) = create_packed_data();
+    let positions = RefCell::new(Vec::with_capacity(N));
+    runner.bench_function("filter_vec_full", move |_| {
+        unpacker.get_ids_for_value_range(
+            0..=MAX_VAL,
+            0..N as u32,
+            &data,
+            &mut positions.borrow_mut(),
+        );
+        black_box(positions.borrow().len());
+    });
+}
+
+fn main() {
+    bench_bitpacking();
+    bench_blocked_bitpacker();
+    bench_filter_vec();
 }

bitpacker/src/filter_vec/mod.rs

@@ -1,8 +1,17 @@
+#[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+use std::arch::is_aarch64_feature_detected;
 use std::ops::RangeInclusive;
 
 #[cfg(target_arch = "x86_64")]
 mod avx2;
 
+#[cfg(target_arch = "aarch64")]
+mod neon;
+
+// SVE intrinsics are not exposed on aarch64-apple-darwin.
+#[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+mod sve;
+
 mod scalar;
 
 #[derive(Clone, Copy, Eq, PartialEq, Debug)]
@@ -10,6 +19,10 @@ mod scalar;
 enum FilterImplPerInstructionSet {
     #[cfg(target_arch = "x86_64")]
     AVX2 = 0u8,
+    #[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+    SVE = 3u8,
+    #[cfg(target_arch = "aarch64")]
+    Neon = 2u8,
     Scalar = 1u8,
 }
 
@@ -19,29 +32,57 @@ impl FilterImplPerInstructionSet {
         match *self {
             #[cfg(target_arch = "x86_64")]
             FilterImplPerInstructionSet::AVX2 => is_x86_feature_detected!("avx2"),
+            #[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+            FilterImplPerInstructionSet::SVE => is_aarch64_feature_detected!("sve"),
+            // TIL Neon is required on aarch 64.
+            #[cfg(target_arch = "aarch64")]
+            FilterImplPerInstructionSet::Neon => true,
             FilterImplPerInstructionSet::Scalar => true,
         }
     }
 }
 
-// List of available implementation in preferred order.
+// List of available implementations in preferred order.
 #[cfg(target_arch = "x86_64")]
 const IMPLS: [FilterImplPerInstructionSet; 2] = [
     FilterImplPerInstructionSet::AVX2,
     FilterImplPerInstructionSet::Scalar,
 ];
 
-#[cfg(not(target_arch = "x86_64"))]
+// Non-Apple aarch64: try SVE, NEON, Scalar.
+#[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+const IMPLS: [FilterImplPerInstructionSet; 3] = [
+    FilterImplPerInstructionSet::SVE,
+    FilterImplPerInstructionSet::Neon,
+    FilterImplPerInstructionSet::Scalar,
+];
+
+// Apple aarch64 (M-series): SVE not available; use NEON or Scalar.
+#[cfg(all(target_arch = "aarch64", target_vendor = "apple"))]
+const IMPLS: [FilterImplPerInstructionSet; 2] = [
+    FilterImplPerInstructionSet::Neon,
+    FilterImplPerInstructionSet::Scalar,
+];
+
+#[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
 const IMPLS: [FilterImplPerInstructionSet; 1] = [FilterImplPerInstructionSet::Scalar];
 
 impl FilterImplPerInstructionSet {
     #[inline]
-    #[allow(unused_variables)] // on non-x86_64, code is unused.
+    #[allow(unused_variables)]
     fn from(code: u8) -> FilterImplPerInstructionSet {
         #[cfg(target_arch = "x86_64")]
         if code == FilterImplPerInstructionSet::AVX2 as u8 {
             return FilterImplPerInstructionSet::AVX2;
         }
+        #[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+        if code == FilterImplPerInstructionSet::SVE as u8 {
+            return FilterImplPerInstructionSet::SVE;
+        }
+        #[cfg(target_arch = "aarch64")]
+        if code == FilterImplPerInstructionSet::Neon as u8 {
+            return FilterImplPerInstructionSet::Neon;
+        }
         FilterImplPerInstructionSet::Scalar
     }
 
@@ -50,6 +91,13 @@ impl FilterImplPerInstructionSet {
         match self {
             #[cfg(target_arch = "x86_64")]
             FilterImplPerInstructionSet::AVX2 => avx2::filter_vec_in_place(range, offset, output),
+            #[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+            // SAFETY: SVE availability was verified by is_available() before selecting this impl.
+            FilterImplPerInstructionSet::SVE => unsafe {
+                sve::filter_vec_in_place(range, offset, output)
+            },
+            #[cfg(target_arch = "aarch64")]
+            FilterImplPerInstructionSet::Neon => neon::filter_vec_in_place(range, offset, output),
             FilterImplPerInstructionSet::Scalar => {
                 scalar::filter_vec_in_place(range, offset, output)
             }
@@ -57,6 +105,12 @@ impl FilterImplPerInstructionSet {
     }
 }
 
+fn available_impls() -> impl Iterator<Item = FilterImplPerInstructionSet> {
+    IMPLS
+        .into_iter()
+        .filter(FilterImplPerInstructionSet::is_available)
+}
+
 #[inline]
 fn get_best_available_instruction_set() -> FilterImplPerInstructionSet {
     use std::sync::atomic::{AtomicU8, Ordering};
@@ -64,10 +118,7 @@ fn get_best_available_instruction_set() -> FilterImplPerInstructionSet {
     let instruction_set_byte: u8 = INSTRUCTION_SET_BYTE.load(Ordering::Relaxed);
     if instruction_set_byte == u8::MAX {
         // Let's initialize the instruction set and cache it.
-        let instruction_set = IMPLS
-            .into_iter()
-            .find(FilterImplPerInstructionSet::is_available)
-            .unwrap();
+        let instruction_set = available_impls().next().unwrap();
         INSTRUCTION_SET_BYTE.store(instruction_set as u8, Ordering::Relaxed);
         return instruction_set;
     }
@@ -80,12 +131,12 @@ pub fn filter_vec_in_place(range: RangeInclusive<u32>, offset: u32, output: &mut
 
 #[cfg(test)]
 mod tests {
+    use proptest::strategy::Strategy;
+
     use super::*;
 
     #[test]
     fn test_get_best_available_instruction_set() {
-        // This does not test much unfortunately.
-        // We just make sure the function returns without crashing and returns the same result.
         let instruction_set = get_best_available_instruction_set();
         assert_eq!(get_best_available_instruction_set(), instruction_set);
     }
@@ -102,6 +153,31 @@ mod tests {
         }
     }
 
+    #[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+    #[test]
+    fn test_instruction_set_to_code_from_code() {
+        for instruction_set in [
+            FilterImplPerInstructionSet::SVE,
+            FilterImplPerInstructionSet::Neon,
+            FilterImplPerInstructionSet::Scalar,
+        ] {
+            let code = instruction_set as u8;
+            assert_eq!(instruction_set, FilterImplPerInstructionSet::from(code));
+        }
+    }
+
+    #[cfg(all(target_arch = "aarch64", target_vendor = "apple"))]
+    #[test]
+    fn test_instruction_set_to_code_from_code() {
+        for instruction_set in [
+            FilterImplPerInstructionSet::Neon,
+            FilterImplPerInstructionSet::Scalar,
+        ] {
+            let code = instruction_set as u8;
+            assert_eq!(instruction_set, FilterImplPerInstructionSet::from(code));
+        }
+    }
+
     fn test_filter_impl_empty_aux(filter_impl: FilterImplPerInstructionSet) {
         let mut output = vec![];
         filter_impl.filter_vec_in_place(0..=u32::MAX, 0, &mut output);
@@ -126,11 +202,20 @@ mod tests {
         assert_eq!(&output, &[1, 3, 4, 5, 6, 7, 8]);
     }
 
+    fn test_filter_impl_empty_range_aux(filter_impl: FilterImplPerInstructionSet) {
+        // start > end: RangeInclusive::contains always returns false; output must be empty.
+        // The SVE path's wrapping_sub would otherwise produce a huge range_width.
+        let mut output = vec![3, 2, 1, 5, 11, 2, 5, 10, 2];
+        filter_impl.filter_vec_in_place(10..=5, 0, &mut output);
+        assert_eq!(&output, &[]);
+    }
+
     fn test_filter_impl_test_suite(filter_impl: FilterImplPerInstructionSet) {
         test_filter_impl_empty_aux(filter_impl);
         test_filter_impl_simple_aux(filter_impl);
         test_filter_impl_simple_aux_shifted(filter_impl);
         test_filter_impl_simple_outside_i32_range(filter_impl);
+        test_filter_impl_empty_range_aux(filter_impl);
     }
 
     #[test]
@@ -141,25 +226,60 @@ mod tests {
         }
     }
 
+    #[test]
+    #[cfg(all(target_arch = "aarch64", not(target_vendor = "apple")))]
+    fn test_filter_implementation_sve() {
+        if FilterImplPerInstructionSet::SVE.is_available() {
+            test_filter_impl_test_suite(FilterImplPerInstructionSet::SVE);
+        }
+    }
+
+    #[test]
+    #[cfg(target_arch = "aarch64")]
+    fn test_filter_implementation_neon() {
+        test_filter_impl_test_suite(FilterImplPerInstructionSet::Neon);
+    }
+
     #[test]
     fn test_filter_implementation_scalar() {
         test_filter_impl_test_suite(FilterImplPerInstructionSet::Scalar);
     }
 
-    #[cfg(target_arch = "x86_64")]
+    fn max_val_strategy() -> impl proptest::strategy::Strategy<Value = u32> {
+        proptest::prop_oneof![
+            0u32..10u32,
+            255u32..258u32,
+            proptest::prelude::Just(1u32 << 25),
+            proptest::prelude::Just(u32::MAX - 1),
+            proptest::prelude::Just(u32::MAX),
+        ]
+    }
+
+    fn vals_strategy() -> impl proptest::strategy::Strategy<Value = Vec<u32>> {
+        proptest::prop_oneof![
+            proptest::collection::vec(proptest::prelude::any::<u32>(), 0..300),
+            max_val_strategy()
+                .prop_flat_map(|max_val| { proptest::collection::vec(0..=max_val, 0..300) })
+        ]
+    }
+
     proptest::proptest! {
         #[test]
-        fn test_filter_compare_scalar_and_avx2_impl_proptest(
-            start in proptest::prelude::any::<u32>(),
-            end in proptest::prelude::any::<u32>(),
+        fn test_filter_compare_scalar_and_impls_impl_proptest(
+            start in 0u32..400u32,
+            end in 0u32..400u32,
             offset in 0u32..2u32,
-            mut vals in proptest::collection::vec(0..u32::MAX, 0..30)) {
-            if FilterImplPerInstructionSet::AVX2.is_available() {
-                let mut vals_clone = vals.clone();
-                FilterImplPerInstructionSet::AVX2.filter_vec_in_place(start..=end, offset, &mut vals);
-                FilterImplPerInstructionSet::Scalar.filter_vec_in_place(start..=end, offset, &mut vals_clone);
-                assert_eq!(&vals, &vals_clone);
-            }
+            vals in vals_strategy()) {
+                for implementation in available_impls() {
+                    if implementation == FilterImplPerInstructionSet::Scalar {
+                        continue;
+                    }
+                    let mut impl_output = vals.clone();
+                    let mut scalar_output = vals.clone();
+                    implementation.filter_vec_in_place(start..=end, offset, &mut impl_output);
+                    FilterImplPerInstructionSet::Scalar.filter_vec_in_place(start..=end, offset, &mut scalar_output);
+                    assert_eq!(&impl_output, &scalar_output);
+                }
        }
     }
 }

bitpacker/src/filter_vec/neon.rs

@@ -0,0 +1,118 @@
+use std::arch::aarch64::*;
+use std::ops::RangeInclusive;
+
+const NUM_LANES: usize = 4;
+
+// Compacts matching lanes to the front using a byte-level shuffle.
+// `mask` is a 4-bit value: bit k=1 means lane k should appear in the output.
+#[inline]
+#[target_feature(enable = "neon")]
+unsafe fn compact(data: uint32x4_t, mask: u8) -> uint32x4_t {
+    unsafe {
+        // SAFETY: mask is always in [0, 15] by construction (max sum of [1,2,4,8]).
+        // BYTE_SHUFFLE_TABLE has 16 entries, so this is always in bounds.
+        let shuffle = BYTE_SHUFFLE_TABLE.get_unchecked(mask as usize);
+        let shuffle_vec = vld1q_u8(shuffle.as_ptr());
+        vreinterpretq_u32_u8(vqtbl1q_u8(vreinterpretq_u8_u32(data), shuffle_vec))
+    }
+}
+
+// Safe (not unsafe) because NEON is mandatory on aarch64: no runtime feature check needed.
+#[inline(never)]
+pub fn filter_vec_in_place(range: RangeInclusive<u32>, offset: u32, output: &mut Vec<u32>) {
+    let num_words = output.len() / NUM_LANES;
+    let mut output_len = unsafe {
+        filter_vec_neon_aux(
+            output.as_ptr(),
+            range.clone(),
+            output.as_mut_ptr(),
+            offset,
+            num_words,
+        )
+    };
+    let remainder_start = num_words * NUM_LANES;
+    for i in remainder_start..output.len() {
+        let val = output[i];
+        output[output_len] = offset + i as u32;
+        output_len += if range.contains(&val) { 1 } else { 0 };
+    }
+    output.truncate(output_len);
+}
+
+#[target_feature(enable = "neon")]
+unsafe fn filter_vec_neon_aux(
+    input: *const u32,
+    range: RangeInclusive<u32>,
+    output: *mut u32,
+    offset: u32,
+    num_words: usize,
+) -> usize {
+    unsafe {
+        let mut input = input;
+        let mut output_tail = output;
+        let range_start_simd = vdupq_n_u32(*range.start());
+        let range_end_simd = vdupq_n_u32(*range.end());
+        let mut ids = vld1q_u32([offset, offset + 1, offset + 2, offset + 3].as_ptr());
+        let shift = vdupq_n_u32(NUM_LANES as u32);
+        let bit_weights = vld1q_u32([1u32, 2, 4, 8].as_ptr());
+
+        for _ in 0..num_words {
+            let word = vld1q_u32(input);
+
+            // Unsigned compares: CMHS (compare higher or same) tests `word >= start`
+            // and `end >= word`. ANDing both gives the inside-range mask directly,
+            // which is cheaper than computing `outside` and then negating.
+            let ge_start = vcgeq_u32(word, range_start_simd);
+            let le_end = vcleq_u32(word, range_end_simd);
+            // inside[k] = 0xFFFFFFFF if val[k] is in range, 0 otherwise.
+            let inside = vandq_u32(ge_start, le_end);
+
+            // Build the 4-bit mask: AND bit_weights with the inside lane mask, so each
+            // inside lane contributes its bit_weight (1, 2, 4, or 8). Summing yields the
+            // 4-bit mask in one addv.
+            let inside_bits = vandq_u32(bit_weights, inside);
+            let mask = vaddvq_u32(inside_bits) as u8;
+            // mask is mathematically bounded: max value is 1+2+4+8=15 (all lanes match)
+            debug_assert!(mask <= 15, "mask must fit in 4 bits: {}", mask);
+
+            // Count of matching lanes = popcount(mask). Derives the count directly from
+            // the mask instead of running a parallel SIMD reduction over `outside`.
+            let added_len = mask.count_ones() as usize;
+
+            // Safe because mask is guaranteed to be in [0, 15]
+            let filtered_ids = compact(ids, mask);
+            vst1q_u32(output_tail, filtered_ids);
+            output_tail = output_tail.add(added_len);
+            ids = vaddq_u32(ids, shift);
+            input = input.add(NUM_LANES);
+        }
+
+        output_tail.offset_from(output) as usize
+    }
+}
+
+// Byte shuffle patterns to compact matching lanes to the front of the vector.
+// Index is a 4-bit mask: bit k=1 means lane k (bytes 4k..4k+3) is in-range.
+// The j-th set bit determines which input lane goes to output position j.
+const BYTE_SHUFFLE_TABLE: [[u8; 16]; 16] = [
+    [
+        16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
+    ], // 0b0000: none
+    [0, 1, 2, 3, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16], // 0b0001: lane 0
+    [4, 5, 6, 7, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16], // 0b0010: lane 1
+    [0, 1, 2, 3, 4, 5, 6, 7, 16, 16, 16, 16, 16, 16, 16, 16],     // 0b0011: lanes 0,1
+    [8, 9, 10, 11, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16], // 0b0100: lane 2
+    [0, 1, 2, 3, 8, 9, 10, 11, 16, 16, 16, 16, 16, 16, 16, 16],   // 0b0101: lanes 0,2
+    [4, 5, 6, 7, 8, 9, 10, 11, 16, 16, 16, 16, 16, 16, 16, 16],   // 0b0110: lanes 1,2
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 16, 16, 16, 16],       // 0b0111: lanes 0,1,2
+    [
+        12, 13, 14, 15, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
+    ], // 0b1000: lane 3
+    [0, 1, 2, 3, 12, 13, 14, 15, 16, 16, 16, 16, 16, 16, 16, 16], // 0b1001: lanes 0,3
+    [4, 5, 6, 7, 12, 13, 14, 15, 16, 16, 16, 16, 16, 16, 16, 16], // 0b1010: lanes 1,3
+    [0, 1, 2, 3, 4, 5, 6, 7, 12, 13, 14, 15, 16, 16, 16, 16],     // 0b1011: lanes 0,1,3
+    [8, 9, 10, 11, 12, 13, 14, 15, 16, 16, 16, 16, 16, 16, 16, 16], // 0b1100: lanes 2,3
+    [0, 1, 2, 3, 8, 9, 10, 11, 12, 13, 14, 15, 16, 16, 16, 16],   // 0b1101: lanes 0,2,3
+    [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 16, 16, 16],   // 0b1110: lanes 1,2,3
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15],       // 0b1111: all lanes
+];

bitpacker/src/filter_vec/sve.rs

@@ -0,0 +1,260 @@
+use std::ops::RangeInclusive;
+
+// SVE vector length (in u32 lanes) is not a compile-time constant; query at runtime.
+// Safe to call only when SVE is confirmed available via is_aarch64_feature_detected!("sve").
+#[target_feature(enable = "sve")]
+unsafe fn num_lanes() -> usize {
+    let vl: usize;
+    unsafe {
+        core::arch::asm!(
+            "cntw {vl}",
+            vl = out(reg) vl,
+            options(nostack, nomem, preserves_flags),
+        );
+    }
+    vl
+}
+
+// SAFETY: caller must ensure SVE is available (checked via is_aarch64_feature_detected!("sve")).
+// Unlike NEON, SVE is optional on aarch64 and not guaranteed by the target architecture.
+pub unsafe fn filter_vec_in_place(range: RangeInclusive<u32>, offset: u32, output: &mut Vec<u32>) {
+    if range.start() > range.end() {
+        output.clear();
+        return;
+    }
+    let vl = unsafe { num_lanes() };
+    let num_words = output.len() / vl;
+    let range_start = *range.start();
+    // Unsigned subtraction trick: val ∈ [lo, hi] ↔ (val - lo) ≤ᵤ (hi - lo).
+    // Values below lo wrap around to large u32, so the single unsigned ≤ excludes them.
+    let range_width = range.end().wrapping_sub(range_start);
+    let mut output_len = unsafe {
+        filter_vec_sve_aux(
+            output.as_ptr(),
+            range_start,
+            range_width,
+            output.as_mut_ptr(),
+            offset,
+            num_words,
+            vl,
+        )
+    };
+    let remainder_start = num_words * vl;
+    for i in remainder_start..output.len() {
+        let val = output[i];
+        output[output_len] = offset + i as u32;
+        output_len += if range.contains(&val) { 1 } else { 0 };
+    }
+    output.truncate(output_len);
+}
+
+// Register allocation for the asm! blocks:
+//   z0        ids_a (index vector for first half of each pair, advances by step2 each iter)
+//   z1        range_width broadcast
+//   z2        range_start broadcast
+//   z3        step2 broadcast (2 * vl)
+//   z4        ids_b (index vector for second half, = ids_a + step, advances by step2)
+//   z5        scratch: loaded word_a, then compacted_a
+//   z6        scratch: loaded word_b, then compacted_b
+//   p0        all-true predicate (ptrue p0.s)
+//   p1        in-range mask for word_a
+//   p2        in-range mask for word_b
+#[target_feature(enable = "sve")]
+unsafe fn filter_vec_sve_aux(
+    input: *const u32,
+    range_start: u32,
+    range_width: u32,
+    output: *mut u32,
+    offset: u32,
+    num_words: usize,
+    vl: usize,
+) -> usize {
+    let num_pairs = num_words / 2;
+    let mut input_ptr = input;
+    let mut output_tail = output;
+
+    if num_pairs > 0 {
+        unsafe {
+            // We rely on asm! because the SVE intrinsics are not available in stable Rust.
+            // The code that follows was generated by Rustc nightly based on the intrinsics version
+            // at the bottom of this file.
+            core::arch::asm!(
+                // --- Setup ---
+                // All-true predicate for 32-bit lanes.
+                "ptrue p0.s",
+                // ids_a = [offset, offset+1, offset+2, ...]
+                "index z0.s, {offset:w}, #1",
+                // Broadcast scalars into SVE vectors.
+                "mov z1.s, {range_width:w}",
+                "mov z2.s, {range_start:w}",
+                // vl_gpr = number of 32-bit lanes (cntw).
+                "cntw {vl_gpr}",
+                // step2_bytes will first hold 2*vl (for the step2 vector), then 2*VL in bytes.
+                "lsl {step2_bytes}, {vl_gpr}, #1",
+                // z4 = step = [vl, vl, ...]; will become ids_b after the add below.
+                "mov z4.s, {vl_gpr:w}",
+                // z3 = step2 = [2*vl, 2*vl, ...], used to advance both id vectors each iter.
+                "mov z3.s, {step2_bytes:w}",
+                // Repurpose step2_bytes to hold the byte stride for advancing the input pointer
+                // by two full SVE vectors per iteration.
+                "rdvl {step2_bytes}, #2",
+                // ids_b = ids_a + step = [offset+vl, offset+vl+1, ...]
+                "add z4.s, z0.s, z4.s",
+
+                // --- Main loop: process two SVE vectors (ids_a and ids_b) per iteration ---
+                "0:",
+                // Load two consecutive SVE vectors from input.
+                "ld1w {{z5.s}}, p0/z, [{input}]",
+                "ld1w {{z6.s}}, p0/z, [{input}, #1, mul vl]",
+                // Advance input pointer by 2 * VL bytes.
+                "add {input}, {input}, {step2_bytes}",
+                // Unsigned shift: subtract range_start so in-range check becomes a single cmpu ≤.
+                "sub z5.s, z5.s, z2.s",
+                "sub z6.s, z6.s, z2.s",
+                // in_range: shifted value ≤ range_width  (unsigned, so values below lo also fail).
+                "cmphs p1.s, p0/z, z1.s, z5.s",
+                "cmphs p2.s, p0/z, z1.s, z6.s",
+                // Count matching lanes; both cntp calls have independent inputs for OOO parallelism.
+                "cntp {cnt_a}, p0, p1.s",
+                "compact z5.s, p1, z0.s",
+                "compact z6.s, p2, z4.s",
+                "cntp {cnt_b}, p0, p2.s",
+                // Advance id vectors for the next iteration.
+                "add z0.s, z0.s, z3.s",
+                "add z4.s, z4.s, z3.s",
+                // Store compacted ids. Only the first cnt_a / cnt_b slots are valid; the rest
+                // will be overwritten by subsequent iterations before the final truncate.
+                "str z5, [{out}]",
+                "st1w {{z6.s}}, p0, [{out}, {cnt_a}, lsl #2]",
+                "add {out}, {out}, {cnt_a}, lsl #2",
+                "add {out}, {out}, {cnt_b}, lsl #2",
+                "subs {pairs}, {pairs}, #1",
+                "b.ne 0b",
+
+                // --- Operands ---
+                input       = inout(reg) input_ptr,
+                out         = inout(reg) output_tail,
+                pairs       = inout(reg) num_pairs => _,
+                offset      = in(reg) offset,
+                range_start = in(reg) range_start,
+                range_width = in(reg) range_width,
+                vl_gpr      = out(reg) _,
+                step2_bytes = out(reg) _,
+                cnt_a       = out(reg) _,
+                cnt_b       = out(reg) _,
+                out("p0") _, out("p1") _, out("p2") _,
+                out("v0") _, out("v1") _, out("v2") _, out("v3") _,
+                out("v4") _, out("v5") _, out("v6") _,
+                options(nostack),
+            );
+        }
+    }
+
+    // Handle an odd trailing vector.
+    if num_words % 2 == 1 {
+        // ids_a for the odd word starts at offset + num_pairs * 2 * vl.
+        // input_ptr was advanced by the main loop and now points at the odd word.
+        let odd_offset =
+            offset.wrapping_add((num_pairs as u32).wrapping_mul(2).wrapping_mul(vl as u32));
+        unsafe {
+            core::arch::asm!(
+                "ptrue p0.s",
+                "index z0.s, {odd_offset:w}, #1",
+                "mov z1.s, {range_width:w}",
+                "mov z2.s, {range_start:w}",
+                "ld1w {{z3.s}}, p0/z, [{input}]",
+                "sub z3.s, z3.s, z2.s",
+                "cmphs p1.s, p0/z, z1.s, z3.s",
+                "cntp {cnt}, p0, p1.s",
+                "compact z0.s, p1, z0.s",
+                "str z0, [{out}]",
+                "add {out}, {out}, {cnt}, lsl #2",
+                odd_offset  = in(reg) odd_offset,
+                range_width = in(reg) range_width,
+                range_start = in(reg) range_start,
+                input       = in(reg) input_ptr,
+                out         = inout(reg) output_tail,
+                cnt         = out(reg) _,
+                out("p0") _, out("p1") _,
+                out("v0") _, out("v1") _, out("v2") _, out("v3") _,
+                options(nostack),
+            );
+        }
+    }
+
+    unsafe { output_tail.offset_from(output) as usize }
+}
+
+// SVE implements with intrinsics.
+//
+// #[target_feature(enable = "sve")]
+// unsafe fn filter_vec_sve_aux(
+//     input: *const u32,
+//     range_start: u32,
+//     range_width: u32,
+//     output: *mut u32,
+//     offset: u32,
+//     num_words: usize,
+//     vl: usize,
+// ) -> usize {
+//     unsafe {
+//         let all_true = svptrue_b32();
+//         let range_start_simd = svdup_n_u32(range_start);
+//         let range_width_simd = svdup_n_u32(range_width);
+//         // ids_a covers [offset .. offset+vl), ids_b covers the next vl ids.
+//         // Keeping them separate breaks the loop-carried dependency through ids so
+//         // both compact/cntp chains are fully independent within each unrolled body.
+//         let mut ids_a = svindex_u32(offset, 1);
+//         let step = svdup_n_u32(vl as u32);
+//         let step2 = svdup_n_u32(2 * vl as u32);
+//         let mut ids_b = svadd_u32_x(all_true, ids_a, step);
+
+//         let mut input = input;
+//         let mut output_tail = output;
+
+//         // Unrolled ×2: both cntp calls have independent inputs and execute in parallel.
+//         // The two output_tail updates are sequential but together cost 4+1+1=6 cy per
+//         // pair vs 5+5=10 cy for two scalar iterations, breaking the cntp latency chain.
+//         let num_pairs = num_words / 2;
+//         for _ in 0..num_pairs {
+//             let word_a = svld1_u32(all_true, input);
+//             let word_b = svld1_u32(all_true, input.add(vl));
+
+//             let shifted_a = svsub_u32_x(all_true, word_a, range_start_simd);
+//             let shifted_b = svsub_u32_x(all_true, word_b, range_start_simd);
+
+//             let in_range_a = svcmple_u32(all_true, shifted_a, range_width_simd);
+//             let in_range_b = svcmple_u32(all_true, shifted_b, range_width_simd);
+
+//             let compacted_a = svcompact_u32(in_range_a, ids_a);
+//             let compacted_b = svcompact_u32(in_range_b, ids_b);
+//             // cntp_a and cntp_b have independent inputs: OOO engine issues them in parallel.
+//             let added_len_a = svcntp_b32(all_true, in_range_a) as usize;
+//             let added_len_b = svcntp_b32(all_true, in_range_b) as usize;
+
+//             // Write the full vector — only the first added_len slots are valid.
+//             // Subsequent iterations overwrite the trailing zeros before truncate.
+//             svst1_u32(all_true, output_tail, compacted_a);
+//             output_tail = output_tail.add(added_len_a);
+//             svst1_u32(all_true, output_tail, compacted_b);
+//             output_tail = output_tail.add(added_len_b);
+
+//             ids_a = svadd_u32_x(all_true, ids_a, step2);
+//             ids_b = svadd_u32_x(all_true, ids_b, step2);
+//             input = input.add(2 * vl);
+//         }
+
+//         // Handle an odd trailing word.
+//         if num_words % 2 == 1 {
+//             let word = svld1_u32(all_true, input);
+//             let shifted = svsub_u32_x(all_true, word, range_start_simd);
+//             let in_range = svcmple_u32(all_true, shifted, range_width_simd);
+//             let added_len = svcntp_b32(all_true, in_range) as usize;
+//             let compacted_ids = svcompact_u32(in_range, ids_a);
+//             svst1_u32(all_true, output_tail, compacted_ids);
+//             output_tail = output_tail.add(added_len);
+//         }
+
+//         output_tail.offset_from(output) as usize
+//     }
+// }

columnar/Cargo.toml

@@ -23,7 +23,7 @@ downcast-rs = "2.0.1"
 proptest = "1"
 more-asserts = "0.3.1"
 rand = "0.9"
-binggan = "0.16.1"
+binggan = "0.17.0"
 
 [[bench]]
 name = "bench_merge"

columnar/src/block_accessor.rs

@@ -15,9 +15,37 @@ impl<T: PartialOrd + Copy + std::fmt::Debug + Send + Sync + 'static + Default>
 {
     #[inline]
     pub fn fetch_block<'a>(&'a mut self, docs: &'a [u32], accessor: &Column<T>) {
-        if accessor.index.get_cardinality().is_full() {
-            self.val_cache.resize(docs.len(), T::default());
-            accessor.values.get_vals(docs, &mut self.val_cache);
+        self.fetch_block_with_is_full(docs, accessor, accessor.index.get_cardinality().is_full());
+    }
+
+    /// Like [`Self::fetch_block`] but takes the column's fullness instead of querying
+    /// `accessor.index.get_cardinality()` each call — for callers that know it up front (e.g.
+    /// checked once at construction). `is_full` must equal
+    /// `accessor.index.get_cardinality().is_full()`.
+    #[inline]
+    pub fn fetch_block_with_is_full<'a>(
+        &'a mut self,
+        docs: &'a [u32],
+        accessor: &Column<T>,
+        is_full: bool,
+    ) {
+        if is_full {
+            // Skip the resize when already the right length (common case: fixed-size blocks).
+            if self.val_cache.len() != docs.len() {
+                self.val_cache.resize(docs.len(), T::default());
+            }
+            // When the docs form a contiguous ascending run we can fetch the values
+            // as a single range. This lets codecs (e.g. bitpacked) bulk-decode the
+            // slice instead of gathering value-by-value, and avoids per-value dynamic
+            // dispatch. `docs` is always sorted ascending and free of duplicates here,
+            // so comparing the endpoints is enough to detect contiguity.
+            if is_contiguous(docs) {
+                accessor
+                    .values
+                    .get_range(docs[0] as u64, &mut self.val_cache);
+            } else {
+                accessor.values.get_vals(docs, &mut self.val_cache);
+            }
         } else {
             self.docid_cache.clear();
             self.row_id_cache.clear();
@@ -158,6 +186,22 @@ impl<T: PartialOrd + Copy + std::fmt::Debug + Send + Sync + 'static + Default>
     }
 }
 
+/// Returns true if `docs` is a contiguous ascending run `[d, d + 1, ..., d + n - 1]`.
+///
+/// Assumes `docs` is sorted ascending and free of duplicates (the invariant for the
+/// doc blocks passed to `fetch_block`), so comparing the endpoints is sufficient.
+#[inline]
+fn is_contiguous(docs: &[u32]) -> bool {
+    let (Some(&first), Some(&last)) = (docs.first(), docs.last()) else {
+        return false;
+    };
+    debug_assert!(
+        docs.windows(2).all(|w| w[0] < w[1]),
+        "fetch_block requires docs sorted ascending without duplicates"
+    );
+    (last - first) as usize + 1 == docs.len()
+}
+
 /// Given two sorted lists of docids `docs` and `hits`, hits is a subset of `docs`.
 /// Return all docs that are not in `hits`.
 fn find_missing_docs<F>(docs: &[u32], hits: &[u32], mut callback: F)
@@ -288,4 +332,46 @@ mod tests {
         assert_eq!(accessor.docid_cache, vec![0]);
         assert_eq!(accessor.val_cache, vec![1]);
     }
+
+    #[test]
+    fn test_is_contiguous() {
+        assert!(!is_contiguous(&[]));
+        assert!(is_contiguous(&[5]));
+        assert!(is_contiguous(&[5, 6, 7, 8]));
+        assert!(is_contiguous(&[0, 1, 2]));
+        assert!(!is_contiguous(&[5, 7, 8]));
+        assert!(!is_contiguous(&[0, 1, 3]));
+    }
+
+    #[test]
+    fn test_fetch_block_contiguous_and_gather_match() {
+        use crate::column_index::ColumnIndex;
+        use crate::column_values::{
+            ALL_U64_CODEC_TYPES, serialize_and_load_u64_based_column_values,
+        };
+
+        let vals: Vec<u64> = (0..200u64).map(|i| i * 7 + 3).collect();
+        let values =
+            serialize_and_load_u64_based_column_values::<u64>(&&vals[..], &ALL_U64_CODEC_TYPES);
+        let column = Column {
+            index: ColumnIndex::Full,
+            values,
+        };
+
+        let check = |accessor: &mut ColumnBlockAccessor<u64>, docs: &[u32]| {
+            accessor.fetch_block(docs, &column);
+            let got: Vec<(u32, u64)> = accessor.iter_docid_vals(docs, &column).collect();
+            let expected: Vec<(u32, u64)> = docs.iter().map(|&d| (d, vals[d as usize])).collect();
+            assert_eq!(got, expected);
+        };
+
+        let mut accessor = ColumnBlockAccessor::<u64>::default();
+        // Contiguous block -> get_range fast path.
+        check(&mut accessor, &(10..74).collect::<Vec<u32>>());
+        // Non-contiguous block -> get_vals gather path.
+        check(&mut accessor, &[0, 5, 9, 100, 199]);
+        // Single doc and full span.
+        check(&mut accessor, &[42]);
+        check(&mut accessor, &(0..200).collect::<Vec<u32>>());
+    }
 }

columnar/src/column_values/mod.rs

@@ -119,8 +119,18 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync + DowncastSync {
     /// the segment's `maxdoc`.
     #[inline(always)]
     fn get_range(&self, start: u64, output: &mut [T]) {
-        for (out, idx) in output.iter_mut().zip(start..) {
+        let mut out_chunks = output.chunks_exact_mut(4);
+        let mut idx = start;
+        for out_x4 in out_chunks.by_ref() {
+            out_x4[0] = self.get_val(idx as u32);
+            out_x4[1] = self.get_val((idx + 1) as u32);
+            out_x4[2] = self.get_val((idx + 2) as u32);
+            out_x4[3] = self.get_val((idx + 3) as u32);
+            idx += 4;
+        }
+        for out in out_chunks.into_remainder() {
             *out = self.get_val(idx as u32);
+            idx += 1;
         }
     }
 

columnar/src/column_values/u64_based/tests.rs

@@ -121,6 +121,22 @@ pub(crate) fn create_and_validate<TColumnCodec: ColumnCodec>(
     reader.get_vals(&all_docs, &mut buffer);
     assert_eq!(vals, buffer);
 
+    // Validate `get_range` over the full column and a sub-range. The sub-range starts
+    // at a non-zero offset to exercise the entrance-ramp alignment of the batch decode.
+    buffer.resize(all_docs.len(), 0);
+    reader.get_range(0, &mut buffer);
+    assert_eq!(vals, buffer, "get_range (full) mismatch in data set {name}");
+    if vals.len() >= 2 {
+        let start = 1usize;
+        buffer.resize(vals.len() - start, 0);
+        reader.get_range(start as u64, &mut buffer);
+        assert_eq!(
+            &vals[start..],
+            &buffer[..],
+            "get_range (sub-range) mismatch in data set {name}"
+        );
+    }
+
     if !vals.is_empty() {
         let test_rand_idx = rand::rng().random_range(0..=vals.len() - 1);
         let expected_positions: Vec<u32> = vals

common/Cargo.toml

@@ -19,6 +19,6 @@ time = { version = "0.3.47", features = ["serde-well-known"] }
 serde = { version = "1.0.136", features = ["derive"] }
 
 [dev-dependencies]
-binggan = "0.16.1"
+binggan = "0.17.0"
 proptest = "1.0.0"
 rand = "0.9"

common/src/file_slice.rs

@@ -121,7 +121,7 @@ pub struct FileSlice {
 
 impl fmt::Debug for FileSlice {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        write!(f, "FileSlice({:?}, {:?})", &self.data, self.range)
+        write!(f, "FileSlice({:?}, {:?})", self.data, self.range)
     }
 }
 

query-grammar/src/query_grammar.rs

@@ -327,7 +327,9 @@ fn exists(inp: &str) -> IResult<&str, UserInputLeaf> {
             peek(alt((
                 value(
                     "",
-                    satisfy(|c: char| c.is_whitespace() || ESCAPE_IN_WORD.contains(&c)),
+                    satisfy(|c: char| {
+                        c.is_whitespace() || (ESCAPE_IN_WORD.contains(&c) && c != '\\')
+                    }),
                 ),
                 eof,
             ))),
@@ -345,7 +347,9 @@ fn exists_precond(inp: &str) -> IResult<&str, (), ()> {
             peek(alt((
                 value(
                     "",
-                    satisfy(|c: char| c.is_whitespace() || ESCAPE_IN_WORD.contains(&c)),
+                    satisfy(|c: char| {
+                        c.is_whitespace() || (ESCAPE_IN_WORD.contains(&c) && c != '\\')
+                    }),
                 ),
                 eof,
             ))), // we need to check this isn't a wildcard query
@@ -707,6 +711,7 @@ fn regex(inp: &str) -> IResult<&str, UserInputLeaf> {
             peek(alt((
                 value((), multispace1),
                 value((), char(')')),
+                value((), char('^')),
                 value((), eof),
             ))),
         ),
@@ -728,9 +733,10 @@ fn regex_infallible(inp: &str) -> JResult<&str, UserInputLeaf> {
             peek(alt((
                 value((), multispace1),
                 value((), char(')')),
+                value((), char('^')),
                 value((), eof),
             ))),
-            "expected whitespace, closing parenthesis, or end of input",
+            "expected whitespace, closing parenthesis, boost, or end of input",
         ),
     )(inp)
     {
@@ -773,6 +779,10 @@ fn leaf(inp: &str) -> IResult<&str, UserInputAst> {
                     value((), multispace1),
                     value((), char(')')),
                     value((), eof),
+                    value(
+                        (),
+                        satisfy(|c: char| ESCAPE_IN_WORD.contains(&c) && c != '\\'),
+                    ),
                 ))),
             ),
             |_| UserInputAst::from(UserInputLeaf::All),
@@ -805,6 +815,10 @@ fn leaf_infallible(inp: &str) -> JResult<&str, Option<UserInputAst>> {
                             value((), multispace1),
                             value((), char(')')),
                             value((), eof),
+                            value(
+                                (),
+                                satisfy(|c: char| ESCAPE_IN_WORD.contains(&c) && c != '\\'),
+                            ),
                         ))),
                     ),
                 ),
@@ -1751,6 +1765,8 @@ mod test {
         test_parse_query_to_ast_helper("*", "*");
         test_parse_query_to_ast_helper("(*)", "*");
         test_parse_query_to_ast_helper("(* )", "*");
+        // All query with boost
+        test_parse_query_to_ast_helper("*^2", "(*)^2");
     }
 
     #[test]
@@ -1813,6 +1829,7 @@ mod test {
         test_parse_query_to_ast_helper("a:b*", "\"a\":b*");
         test_parse_query_to_ast_helper("a:*b", "\"a\":*b");
         test_parse_query_to_ast_helper(r#"a:*def*"#, "\"a\":*def*");
+        test_parse_query_to_ast_helper("a:*\\:foo", "\"a\":*:foo");
     }
 
     #[test]
@@ -1877,6 +1894,8 @@ mod test {
             },
             _ => panic!("Expected a leaf"),
         }
+        // Regex followed by `^boost`
+        test_parse_query_to_ast_helper(r#"foo:/bar/^2"#, r#"("foo":/bar/)^2"#);
     }
 
     #[test]

src/aggregation/agg_data.rs

@@ -10,11 +10,11 @@ use crate::aggregation::accessor_helpers::{
 };
 use crate::aggregation::agg_req::{Aggregation, AggregationVariants, Aggregations};
 use crate::aggregation::bucket::{
-    build_segment_filter_collector, build_segment_range_collector, CompositeAggReqData,
-    CompositeAggregation, CompositeSourceAccessors, FilterAggReqData, HistogramAggReqData,
-    HistogramBounds, IncludeExcludeParam, MissingTermAggReqData, RangeAggReqData,
-    SegmentHistogramCollector, TermMissingAgg, TermsAggReqData, TermsAggregation,
-    TermsAggregationInternal,
+    build_segment_filter_collector, build_segment_histogram_collector,
+    build_segment_range_collector, CompositeAggReqData, CompositeAggregation,
+    CompositeSourceAccessors, FilterAggReqData, HistogramAggReqData, HistogramBounds,
+    IncludeExcludeParam, MissingTermAggReqData, RangeAggReqData, TermMissingAgg, TermsAggReqData,
+    TermsAggregation, TermsAggregationInternal,
 };
 use crate::aggregation::metric::{
     build_segment_stats_collector, AverageAggregation, CardinalityAggReqData,
@@ -41,7 +41,7 @@ pub struct AggregationsSegmentCtx {
 
 impl AggregationsSegmentCtx {
     pub(crate) fn push_term_req_data(&mut self, data: TermsAggReqData) -> usize {
-        self.per_request.term_req_data.push(Some(Box::new(data)));
+        self.per_request.term_req_data.push(data);
         self.per_request.term_req_data.len() - 1
     }
     pub(crate) fn push_cardinality_req_data(&mut self, data: CardinalityAggReqData) -> usize {
@@ -61,31 +61,25 @@ impl AggregationsSegmentCtx {
         self.per_request.missing_term_req_data.len() - 1
     }
     pub(crate) fn push_histogram_req_data(&mut self, data: HistogramAggReqData) -> usize {
-        self.per_request
-            .histogram_req_data
-            .push(Some(Box::new(data)));
+        self.per_request.histogram_req_data.push(data);
         self.per_request.histogram_req_data.len() - 1
     }
     pub(crate) fn push_range_req_data(&mut self, data: RangeAggReqData) -> usize {
-        self.per_request.range_req_data.push(Some(Box::new(data)));
+        self.per_request.range_req_data.push(data);
         self.per_request.range_req_data.len() - 1
     }
     pub(crate) fn push_filter_req_data(&mut self, data: FilterAggReqData) -> usize {
-        self.per_request.filter_req_data.push(Some(Box::new(data)));
+        self.per_request.filter_req_data.push(data);
         self.per_request.filter_req_data.len() - 1
     }
     pub(crate) fn push_composite_req_data(&mut self, data: CompositeAggReqData) -> usize {
-        self.per_request
-            .composite_req_data
-            .push(Some(Box::new(data)));
+        self.per_request.composite_req_data.push(data);
         self.per_request.composite_req_data.len() - 1
     }
 
     #[inline]
     pub(crate) fn get_term_req_data(&self, idx: usize) -> &TermsAggReqData {
-        self.per_request.term_req_data[idx]
-            .as_deref()
-            .expect("term_req_data slot is empty (taken)")
+        &self.per_request.term_req_data[idx]
     }
     #[inline]
     pub(crate) fn get_cardinality_req_data(&self, idx: usize) -> &CardinalityAggReqData {
@@ -103,116 +97,6 @@ impl AggregationsSegmentCtx {
     pub(crate) fn get_missing_term_req_data(&self, idx: usize) -> &MissingTermAggReqData {
         &self.per_request.missing_term_req_data[idx]
     }
-    #[inline]
-    pub(crate) fn get_histogram_req_data(&self, idx: usize) -> &HistogramAggReqData {
-        self.per_request.histogram_req_data[idx]
-            .as_deref()
-            .expect("histogram_req_data slot is empty (taken)")
-    }
-    #[inline]
-    pub(crate) fn get_range_req_data(&self, idx: usize) -> &RangeAggReqData {
-        self.per_request.range_req_data[idx]
-            .as_deref()
-            .expect("range_req_data slot is empty (taken)")
-    }
-    #[inline]
-    pub(crate) fn get_composite_req_data(&self, idx: usize) -> &CompositeAggReqData {
-        self.per_request.composite_req_data[idx]
-            .as_deref()
-            .expect("composite_req_data slot is empty (taken)")
-    }
-
-    // ---------- mutable getters ----------
-
-    #[inline]
-    pub(crate) fn get_metric_req_data_mut(&mut self, idx: usize) -> &mut MetricAggReqData {
-        &mut self.per_request.stats_metric_req_data[idx]
-    }
-
-    #[inline]
-    pub(crate) fn get_cardinality_req_data_mut(
-        &mut self,
-        idx: usize,
-    ) -> &mut CardinalityAggReqData {
-        &mut self.per_request.cardinality_req_data[idx]
-    }
-
-    #[inline]
-    pub(crate) fn get_histogram_req_data_mut(&mut self, idx: usize) -> &mut HistogramAggReqData {
-        self.per_request.histogram_req_data[idx]
-            .as_deref_mut()
-            .expect("histogram_req_data slot is empty (taken)")
-    }
-
-    // ---------- take / put (terms, histogram, range) ----------
-
-    /// Move out the boxed Histogram request at `idx`, leaving `None`.
-    #[inline]
-    pub(crate) fn take_histogram_req_data(&mut self, idx: usize) -> Box<HistogramAggReqData> {
-        self.per_request.histogram_req_data[idx]
-            .take()
-            .expect("histogram_req_data slot is empty (taken)")
-    }
-
-    /// Put back a Histogram request into an empty slot at `idx`.
-    #[inline]
-    pub(crate) fn put_back_histogram_req_data(
-        &mut self,
-        idx: usize,
-        value: Box<HistogramAggReqData>,
-    ) {
-        debug_assert!(self.per_request.histogram_req_data[idx].is_none());
-        self.per_request.histogram_req_data[idx] = Some(value);
-    }
-
-    /// Move out the boxed Range request at `idx`, leaving `None`.
-    #[inline]
-    pub(crate) fn take_range_req_data(&mut self, idx: usize) -> Box<RangeAggReqData> {
-        self.per_request.range_req_data[idx]
-            .take()
-            .expect("range_req_data slot is empty (taken)")
-    }
-
-    /// Put back a Range request into an empty slot at `idx`.
-    #[inline]
-    pub(crate) fn put_back_range_req_data(&mut self, idx: usize, value: Box<RangeAggReqData>) {
-        debug_assert!(self.per_request.range_req_data[idx].is_none());
-        self.per_request.range_req_data[idx] = Some(value);
-    }
-
-    /// Move out the boxed Filter request at `idx`, leaving `None`.
-    #[inline]
-    pub(crate) fn take_filter_req_data(&mut self, idx: usize) -> Box<FilterAggReqData> {
-        self.per_request.filter_req_data[idx]
-            .take()
-            .expect("filter_req_data slot is empty (taken)")
-    }
-
-    /// Put back a Filter request into an empty slot at `idx`.
-    #[inline]
-    pub(crate) fn put_back_filter_req_data(&mut self, idx: usize, value: Box<FilterAggReqData>) {
-        debug_assert!(self.per_request.filter_req_data[idx].is_none());
-        self.per_request.filter_req_data[idx] = Some(value);
-    }
-
-    /// Move out the Composite request at `idx`.
-    #[inline]
-    pub(crate) fn take_composite_req_data(&mut self, idx: usize) -> Box<CompositeAggReqData> {
-        self.per_request.composite_req_data[idx]
-            .take()
-            .expect("composite_req_data slot is empty (taken)")
-    }
-
-    /// Put back a Composite request into an empty slot at `idx`.
-    #[inline]
-    pub(crate) fn put_back_composite_req_data(
-        &mut self,
-        idx: usize,
-        value: Box<CompositeAggReqData>,
-    ) {
-        debug_assert!(self.per_request.composite_req_data[idx].is_none());
-        self.per_request.composite_req_data[idx] = Some(value);
-    }
 }
 
 /// Each type of aggregation has its own request data struct. This struct holds
@@ -223,15 +107,14 @@ impl AggregationsSegmentCtx {
 /// for a node with [AggKind::Terms]).
 #[derive(Default)]
 pub struct PerRequestAggSegCtx {
-    // Box for cheap take/put - Only necessary for bucket aggs that have sub-aggregations
     /// TermsAggReqData contains the request data for a terms aggregation.
-    pub term_req_data: Vec<Option<Box<TermsAggReqData>>>,
+    pub term_req_data: Vec<TermsAggReqData>,
     /// HistogramAggReqData contains the request data for a histogram aggregation.
-    pub histogram_req_data: Vec<Option<Box<HistogramAggReqData>>>,
+    pub histogram_req_data: Vec<HistogramAggReqData>,
     /// RangeAggReqData contains the request data for a range aggregation.
-    pub range_req_data: Vec<Option<Box<RangeAggReqData>>>,
+    pub range_req_data: Vec<RangeAggReqData>,
     /// FilterAggReqData contains the request data for a filter aggregation.
-    pub filter_req_data: Vec<Option<Box<FilterAggReqData>>>,
+    pub filter_req_data: Vec<FilterAggReqData>,
     /// Shared by avg, min, max, sum, stats, extended_stats, count
     pub stats_metric_req_data: Vec<MetricAggReqData>,
     /// CardinalityAggReqData contains the request data for a cardinality aggregation.
@@ -241,7 +124,7 @@ pub struct PerRequestAggSegCtx {
     /// MissingTermAggReqData contains the request data for a missing term aggregation.
     pub missing_term_req_data: Vec<MissingTermAggReqData>,
     /// CompositeAggReqData contains the request data for a composite aggregation.
-    pub composite_req_data: Vec<Option<Box<CompositeAggReqData>>>,
+    pub composite_req_data: Vec<CompositeAggReqData>,
 
     /// Request tree used to build collectors.
     pub agg_tree: Vec<AggRefNode>,
@@ -252,22 +135,22 @@ impl PerRequestAggSegCtx {
     fn get_memory_consumption(&self) -> usize {
         self.term_req_data
             .iter()
-            .map(|b| b.as_ref().unwrap().get_memory_consumption())
+            .map(|t| t.get_memory_consumption())
             .sum::<usize>()
             + self
                 .histogram_req_data
                 .iter()
-                .map(|b| b.as_ref().unwrap().get_memory_consumption())
+                .map(|t| t.get_memory_consumption())
                 .sum::<usize>()
             + self
                 .range_req_data
                 .iter()
-                .map(|b| b.as_ref().unwrap().get_memory_consumption())
+                .map(|t| t.get_memory_consumption())
                 .sum::<usize>()
             + self
                 .filter_req_data
                 .iter()
-                .map(|b| b.as_ref().unwrap().get_memory_consumption())
+                .map(|t| t.get_memory_consumption())
                 .sum::<usize>()
             + self
                 .stats_metric_req_data
@@ -292,7 +175,7 @@ impl PerRequestAggSegCtx {
             + self
                 .composite_req_data
                 .iter()
-                .map(|b| b.as_ref().map(|d| d.get_memory_consumption()).unwrap_or(0))
+                .map(|t| t.get_memory_consumption())
                 .sum::<usize>()
             + self.agg_tree.len() * std::mem::size_of::<AggRefNode>()
     }
@@ -301,40 +184,16 @@ impl PerRequestAggSegCtx {
         let idx = node.idx_in_req_data;
         let kind = node.kind;
         match kind {
-            AggKind::Terms => self.term_req_data[idx]
-                .as_deref()
-                .expect("term_req_data slot is empty (taken)")
-                .name
-                .as_str(),
+            AggKind::Terms => self.term_req_data[idx].name.as_str(),
             AggKind::Cardinality => &self.cardinality_req_data[idx].name,
             AggKind::StatsKind(_) => &self.stats_metric_req_data[idx].name,
             AggKind::TopHits => &self.top_hits_req_data[idx].name,
             AggKind::MissingTerm => &self.missing_term_req_data[idx].name,
-            AggKind::Histogram => self.histogram_req_data[idx]
-                .as_deref()
-                .expect("histogram_req_data slot is empty (taken)")
-                .name
-                .as_str(),
-            AggKind::DateHistogram => self.histogram_req_data[idx]
-                .as_deref()
-                .expect("histogram_req_data slot is empty (taken)")
-                .name
-                .as_str(),
-            AggKind::Range => self.range_req_data[idx]
-                .as_deref()
-                .expect("range_req_data slot is empty (taken)")
-                .name
-                .as_str(),
-            AggKind::Filter => self.filter_req_data[idx]
-                .as_deref()
-                .expect("filter_req_data slot is empty (taken)")
-                .name
-                .as_str(),
-            AggKind::Composite => self.composite_req_data[idx]
-                .as_deref()
-                .expect("composite_req_data slot is empty (taken)")
-                .name
-                .as_str(),
+            AggKind::Histogram => self.histogram_req_data[idx].name.as_str(),
+            AggKind::DateHistogram => self.histogram_req_data[idx].name.as_str(),
+            AggKind::Range => self.range_req_data[idx].name.as_str(),
+            AggKind::Filter => self.filter_req_data[idx].name.as_str(),
+            AggKind::Composite => self.composite_req_data[idx].name.as_str(),
         }
     }
 
@@ -412,7 +271,7 @@ pub(crate) fn build_segment_agg_collector(
             Ok(Box::new(TermMissingAgg::new(req, node)?))
         }
         AggKind::Cardinality => {
-            let req_data = &mut req.get_cardinality_req_data_mut(node.idx_in_req_data);
+            let req_data = req.get_cardinality_req_data(node.idx_in_req_data);
             // For str columns, choose the per-bucket entries representation
             // based on the segment's column.max_value():
             //   * small (< BITSET_MAX_TERM_ORD): `BitSet`, pre-allocated, no promotion machinery.
@@ -459,7 +318,7 @@ pub(crate) fn build_segment_agg_collector(
                     SegmentExtendedStatsCollector::from_req(req_data, sigma),
                 )),
                 StatsType::Percentiles => {
-                    let req_data = req.get_metric_req_data_mut(node.idx_in_req_data);
+                    let req_data = req.get_metric_req_data(node.idx_in_req_data);
                     Ok(Box::new(
                         SegmentPercentilesCollector::from_req_and_validate(
                             req_data.field_type,
@@ -479,12 +338,8 @@ pub(crate) fn build_segment_agg_collector(
                 req_data.segment_ordinal,
             )))
         }
-        AggKind::Histogram => Ok(Box::new(SegmentHistogramCollector::from_req_and_validate(
-            req, node,
-        )?)),
-        AggKind::DateHistogram => Ok(Box::new(SegmentHistogramCollector::from_req_and_validate(
-            req, node,
-        )?)),
+        AggKind::Histogram => build_segment_histogram_collector(req, node),
+        AggKind::DateHistogram => build_segment_histogram_collector(req, node),
         AggKind::Range => Ok(build_segment_range_collector(req, node)?),
         AggKind::Filter => build_segment_filter_collector(req, node),
         AggKind::Composite => Ok(Box::new(
@@ -799,23 +654,18 @@ fn build_nodes(
             let schema = reader.schema();
             let tokenizers = &data.context.tokenizers;
             let query = filter_req.parse_query(schema, tokenizers)?;
-            let evaluator = crate::aggregation::bucket::DocumentQueryEvaluator::new(
-                query,
-                schema.clone(),
-                reader,
-            )?;
-
-            // Pre-allocate buffer for batch filtering
-            let max_doc = reader.max_doc();
-            let buffer_capacity = crate::docset::COLLECT_BLOCK_BUFFER_LEN.min(max_doc as usize);
-            let matching_docs_buffer = Vec::with_capacity(buffer_capacity);
+            let evaluator =
+                std::rc::Rc::new(crate::aggregation::bucket::DocumentQueryEvaluator::new(
+                    query,
+                    schema.clone(),
+                    reader,
+                )?);
 
             let idx_in_req_data = data.push_filter_req_data(FilterAggReqData {
                 name: agg_name.to_string(),
                 req: filter_req.clone(),
                 segment_reader: reader.clone(),
                 evaluator,
-                matching_docs_buffer,
                 is_top_level,
             });
             let children = build_children(&req.sub_aggregation, reader, segment_ordinal, data)?;

src/aggregation/agg_req.rs

@@ -115,6 +115,71 @@ pub fn get_fast_field_names(aggs: &Aggregations) -> HashSet<String> {
     fast_field_names
 }
 
+/// Validates that all fields referenced in the aggregation request exist in the schema
+/// and are configured as fast fields.
+///
+/// This is a convenience function for upfront validation before executing aggregations.
+/// Returns an error if any field doesn't exist or is not a fast field.
+///
+/// Validation is intentionally opt-in rather than baked into aggregation execution: the
+/// default lenient behavior (returning empty results for missing fields) supports
+/// schema evolution and federated queries where the same request runs against segments
+/// or indices with different schemas.
+///
+/// # Example
+/// ```
+/// use tantivy::aggregation::agg_req::{Aggregations, validate_aggregation_fields_exist};
+/// use tantivy::schema::{Schema, FAST};
+/// use tantivy::Index;
+///
+/// # fn main() -> tantivy::Result<()> {
+/// // Create a simple index
+/// let mut schema_builder = Schema::builder();
+/// schema_builder.add_f64_field("price", FAST);
+/// let schema = schema_builder.build();
+/// let index = Index::create_in_ram(schema);
+///
+/// // Parse aggregation request
+/// let agg_req: Aggregations = serde_json::from_str(r#"{
+///     "avg_price": { "avg": { "field": "price" } }
+/// }"#)?;
+///
+/// let reader = index.reader()?;
+/// let searcher = reader.searcher();
+///
+/// // Validate fields before executing
+/// for segment_reader in searcher.segment_readers() {
+///     validate_aggregation_fields_exist(&agg_req, segment_reader)?;
+/// }
+/// # Ok(())
+/// # }
+/// ```
+pub fn validate_aggregation_fields_exist(
+    aggs: &Aggregations,
+    reader: &crate::SegmentReader,
+) -> crate::Result<()> {
+    let field_names = get_fast_field_names(aggs);
+    let schema = reader.schema();
+
+    for field_name in field_names {
+        // Check if the field is either directly in the schema or could be part of a json field
+        // present in the schema, and verify it's a fast field.
+        if let Some((field, _path)) = schema.find_field(&field_name) {
+            let field_type = schema.get_field_entry(field).field_type();
+            if !field_type.is_fast() {
+                return Err(crate::TantivyError::SchemaError(format!(
+                    "Field '{}' is not a fast field. Aggregations require fast fields.",
+                    field_name
+                )));
+            }
+        } else {
+            return Err(crate::TantivyError::FieldNotFound(field_name));
+        }
+    }
+
+    Ok(())
+}
+
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 /// All aggregation types.
 pub enum AggregationVariants {
@@ -234,6 +299,12 @@ impl AggregationVariants {
             _ => None,
         }
     }
+    pub(crate) fn as_sum(&self) -> Option<&SumAggregation> {
+        match &self {
+            AggregationVariants::Sum(sum) => Some(sum),
+            _ => None,
+        }
+    }
 }
 
 #[cfg(test)]

src/aggregation/agg_tests.rs

@@ -1436,3 +1436,46 @@ fn test_aggregation_on_json_object_mixed_numerical_segments() {
         )
     );
 }
+
+#[test]
+fn test_aggregation_field_validation_helper() {
+    // Test the standalone validation helper function for field validation
+    let index = get_test_index_2_segments(false).unwrap();
+    let reader = index.reader().unwrap();
+    let searcher = reader.searcher();
+    let segment_reader = searcher.segment_reader(0);
+
+    // Test with invalid field
+    let agg_req: Aggregations = serde_json::from_str(
+        r#"{
+        "avg_test": {
+            "avg": { "field": "nonexistent_field" }
+        }
+    }"#,
+    )
+    .unwrap();
+
+    let result =
+        crate::aggregation::agg_req::validate_aggregation_fields_exist(&agg_req, segment_reader);
+    assert!(result.is_err());
+    match result {
+        Err(crate::TantivyError::FieldNotFound(field_name)) => {
+            assert_eq!(field_name, "nonexistent_field");
+        }
+        _ => panic!("Expected FieldNotFound error, got: {:?}", result),
+    }
+
+    // Test with valid field
+    let agg_req: Aggregations = serde_json::from_str(
+        r#"{
+        "avg_test": {
+            "avg": { "field": "score" }
+        }
+    }"#,
+    )
+    .unwrap();
+
+    let result =
+        crate::aggregation::agg_req::validate_aggregation_fields_exist(&agg_req, segment_reader);
+    assert!(result.is_ok());
+}

src/aggregation/bucket/composite/accessors.rs

@@ -16,6 +16,7 @@ use crate::{SegmentReader, TantivyError};
 
 /// Contains all information required by the SegmentCompositeCollector to perform the
 /// composite aggregation on a segment.
+#[derive(Debug, Clone)]
 pub struct CompositeAggReqData {
     /// The name of the aggregation.
     pub name: String,
@@ -34,6 +35,7 @@ impl CompositeAggReqData {
 }
 
 /// Accessors for a single column in a composite source.
+#[derive(Debug, Clone)]
 pub struct CompositeAccessor {
     /// The fast field column
     pub column: Column<u64>,
@@ -48,6 +50,7 @@ pub struct CompositeAccessor {
 }
 
 /// Accessors to all the columns that belong to the field of a composite source.
+#[derive(Debug, Clone)]
 pub struct CompositeSourceAccessors {
     /// The accessors for this source
     pub accessors: Vec<CompositeAccessor>,
@@ -358,7 +361,7 @@ impl PrecomputedDateInterval {
 ///
 /// Some column types (term, IP) might not have an exact representation of the
 /// specified after key
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub enum PrecomputedAfterKey {
     /// The after key could be exactly represented in the column space.
     Exact(u64),

src/aggregation/bucket/composite/collector.rs

@@ -118,7 +118,7 @@ impl InternalValueRepr {
 pub struct SegmentCompositeCollector {
     /// One DynArrayHeapMap per parent bucket.
     parent_buckets: Vec<DynArrayHeapMap<InternalValueRepr, CompositeBucketCollector>>,
-    accessor_idx: usize,
+    req_data: CompositeAggReqData,
     sub_agg: Option<BufferedSubAggs<HighCardSubAggBuffer>>,
     bucket_id_provider: BucketIdProvider,
     /// Number of sources, needed when creating new DynArrayHeapMaps.
@@ -132,10 +132,7 @@ impl SegmentAggregationCollector for SegmentCompositeCollector {
         results: &mut IntermediateAggregationResults,
         parent_bucket_id: BucketId,
     ) -> crate::Result<()> {
-        let name = agg_data
-            .get_composite_req_data(self.accessor_idx)
-            .name
-            .clone();
+        let name = self.req_data.name.clone();
 
         let buckets = self.add_intermediate_bucket_result(agg_data, parent_bucket_id)?;
         results.push(
@@ -153,12 +150,11 @@ impl SegmentAggregationCollector for SegmentCompositeCollector {
         agg_data: &mut AggregationsSegmentCtx,
     ) -> crate::Result<()> {
         let mem_pre = self.get_memory_consumption(parent_bucket_id);
-        let composite_agg_data = agg_data.take_composite_req_data(self.accessor_idx);
 
         for doc in docs {
             let mut visitor = CompositeKeyVisitor {
                 doc_id: *doc,
-                composite_agg_data: &composite_agg_data,
+                composite_agg_data: &self.req_data,
                 buckets: &mut self.parent_buckets[parent_bucket_id as usize],
                 sub_agg: &mut self.sub_agg,
                 bucket_id_provider: &mut self.bucket_id_provider,
@@ -166,7 +162,6 @@ impl SegmentAggregationCollector for SegmentCompositeCollector {
             };
             visitor.visit(0, true)?;
         }
-        agg_data.put_back_composite_req_data(self.accessor_idx, composite_agg_data);
 
         if let Some(sub_agg) = &mut self.sub_agg {
             sub_agg.check_flush_local(agg_data)?;
@@ -221,7 +216,13 @@ impl SegmentCompositeCollector {
         req_data: &mut AggregationsSegmentCtx,
         node: &AggRefNode,
     ) -> crate::Result<Self> {
-        validate_req(req_data, node.idx_in_req_data)?;
+        let composite_req_data =
+            req_data.per_request.composite_req_data[node.idx_in_req_data].clone();
+        validate_req(&composite_req_data)?;
+        req_data
+            .context
+            .limits
+            .add_memory_consumed(composite_req_data.get_memory_consumption() as u64)?;
 
         let has_sub_aggregations = !node.children.is_empty();
         let sub_agg = if has_sub_aggregations {
@@ -231,12 +232,11 @@ impl SegmentCompositeCollector {
             None
         };
 
-        let composite_req_data = req_data.get_composite_req_data(node.idx_in_req_data);
         let num_sources = composite_req_data.req.sources.len();
 
         Ok(SegmentCompositeCollector {
             parent_buckets: vec![DynArrayHeapMap::try_new(num_sources)?],
-            accessor_idx: node.idx_in_req_data,
+            req_data: composite_req_data,
             sub_agg,
             bucket_id_provider: BucketIdProvider::default(),
             num_sources,
@@ -258,7 +258,7 @@ impl SegmentCompositeCollector {
         let mut dict: FxHashMap<Vec<CompositeIntermediateKey>, IntermediateCompositeBucketEntry> =
             Default::default();
         dict.reserve(heap_map.size());
-        let composite_data = agg_data.get_composite_req_data(self.accessor_idx);
+        let composite_data = &self.req_data;
         for (key_internal_repr, agg) in heap_map.into_iter() {
             let key = resolve_key(&key_internal_repr, composite_data)?;
             let mut sub_aggregation_res = IntermediateAggregationResults::default();
@@ -298,8 +298,7 @@ impl SegmentCompositeCollector {
     }
 }
 
-fn validate_req(req_data: &mut AggregationsSegmentCtx, accessor_idx: usize) -> crate::Result<()> {
-    let composite_data = req_data.get_composite_req_data(accessor_idx);
+fn validate_req(composite_data: &CompositeAggReqData) -> crate::Result<()> {
     let req = &composite_data.req;
     if req.sources.is_empty() {
         return Err(TantivyError::InvalidArgument(

src/aggregation/bucket/filter.rs

@@ -1,4 +1,5 @@
 use std::fmt::Debug;
+use std::rc::Rc;
 
 use common::BitSet;
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
@@ -396,6 +397,7 @@ impl PartialEq for FilterAggregation {
 
 /// Request data for filter aggregation
 /// This struct holds the per-segment data needed to execute a filter aggregation
+#[derive(Clone)]
 pub struct FilterAggReqData {
     /// The name of the filter aggregation
     pub name: String,
@@ -403,22 +405,20 @@ pub struct FilterAggReqData {
     pub req: FilterAggregation,
     /// The segment reader
     pub segment_reader: SegmentReader,
-    /// Document evaluator for the filter query (precomputed BitSet)
-    /// This is built once when the request data is created
-    pub evaluator: DocumentQueryEvaluator,
-    /// Reusable buffer for matching documents to minimize allocations during collection
-    pub matching_docs_buffer: Vec<DocId>,
+    /// Document evaluator for the filter query (precomputed BitSet).
+    /// Wrapped in `Rc` so cloning the request data does not duplicate the (potentially large)
+    /// underlying BitSet.
+    pub evaluator: Rc<DocumentQueryEvaluator>,
     /// True if this filter aggregation is at the top level of the aggregation tree (not nested).
     pub is_top_level: bool,
 }
 
 impl FilterAggReqData {
     pub(crate) fn get_memory_consumption(&self) -> usize {
-        // Estimate: name + segment reader reference + bitset + buffer capacity
+        // Estimate: name + segment reader reference + bitset
         self.name.len()
         + std::mem::size_of::<SegmentReader>()
         + self.evaluator.bitset.len() / 8 // BitSet memory (bits to bytes)
-        + self.matching_docs_buffer.capacity() * std::mem::size_of::<DocId>()
         + std::mem::size_of::<bool>()
     }
 }
@@ -509,8 +509,10 @@ pub struct SegmentFilterCollector<B: SubAggBuffer> {
     /// Sub-aggregation collectors
     sub_aggregations: Option<BufferedSubAggs<B>>,
     bucket_id_provider: BucketIdProvider,
-    /// Accessor index for this filter aggregation (to access FilterAggReqData)
-    accessor_idx: usize,
+    /// Per-segment filter request data, owned by this collector.
+    req_data: FilterAggReqData,
+    /// Reusable buffer for matching documents to minimize allocations during collection.
+    matching_docs_buffer: Vec<DocId>,
 }
 
 impl<B: SubAggBuffer> SegmentFilterCollector<B> {
@@ -518,6 +520,7 @@ impl<B: SubAggBuffer> SegmentFilterCollector<B> {
     pub(crate) fn from_req_and_validate(
         req: &mut AggregationsSegmentCtx,
         node: &AggRefNode,
+        req_data: FilterAggReqData,
     ) -> crate::Result<Self> {
         // Build sub-aggregation collectors if any
         let sub_agg_collector = if !node.children.is_empty() {
@@ -527,11 +530,15 @@ impl<B: SubAggBuffer> SegmentFilterCollector<B> {
         };
         let sub_agg_collector = sub_agg_collector.map(BufferedSubAggs::new);
 
+        let max_doc = req_data.segment_reader.max_doc();
+        let buffer_capacity = crate::docset::COLLECT_BLOCK_BUFFER_LEN.min(max_doc as usize);
+
         Ok(SegmentFilterCollector {
             parent_buckets: Vec::new(),
             sub_aggregations: sub_agg_collector,
-            accessor_idx: node.idx_in_req_data,
+            req_data,
             bucket_id_provider: BucketIdProvider::default(),
+            matching_docs_buffer: Vec::with_capacity(buffer_capacity),
         })
     }
 }
@@ -540,18 +547,23 @@ pub(crate) fn build_segment_filter_collector(
     req: &mut AggregationsSegmentCtx,
     node: &AggRefNode,
 ) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
-    let is_top_level = req.per_request.filter_req_data[node.idx_in_req_data]
-        .as_ref()
-        .expect("filter_req_data slot is empty")
-        .is_top_level;
+    let req_data = req.per_request.filter_req_data[node.idx_in_req_data].clone();
+    req.context
+        .limits
+        .add_memory_consumed(req_data.get_memory_consumption() as u64)?;
+    let is_top_level = req_data.is_top_level;
 
     if is_top_level {
         Ok(Box::new(
-            SegmentFilterCollector::<LowCardSubAggBuffer>::from_req_and_validate(req, node)?,
+            SegmentFilterCollector::<LowCardSubAggBuffer>::from_req_and_validate(
+                req, node, req_data,
+            )?,
         ))
     } else {
         Ok(Box::new(
-            SegmentFilterCollector::<HighCardSubAggBuffer>::from_req_and_validate(req, node)?,
+            SegmentFilterCollector::<HighCardSubAggBuffer>::from_req_and_validate(
+                req, node, req_data,
+            )?,
         ))
     }
 }
@@ -561,7 +573,7 @@ impl<B: SubAggBuffer> Debug for SegmentFilterCollector<B> {
         f.debug_struct("SegmentFilterCollector")
             .field("buckets", &self.parent_buckets)
             .field("has_sub_aggs", &self.sub_aggregations.is_some())
-            .field("accessor_idx", &self.accessor_idx)
+            .field("name", &self.req_data.name)
             .finish()
     }
 }
@@ -598,11 +610,7 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentFilterCollector<B>
         };
 
         // Get the name of this filter aggregation
-        let name = agg_data.per_request.filter_req_data[self.accessor_idx]
-            .as_ref()
-            .expect("filter_req_data slot is empty")
-            .name
-            .clone();
+        let name = self.req_data.name.clone();
 
         results.push(
             name,
@@ -623,27 +631,24 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentFilterCollector<B>
         }
 
         let mut bucket = self.parent_buckets[parent_bucket_id as usize];
-        // Take the request data to avoid borrow checker issues with sub-aggregations
-        let mut req = agg_data.take_filter_req_data(self.accessor_idx);
 
         // Use batch filtering with O(1) BitSet lookups
-        req.matching_docs_buffer.clear();
-        req.evaluator
-            .filter_batch(docs, &mut req.matching_docs_buffer);
+        self.matching_docs_buffer.clear();
+        self.req_data
+            .evaluator
+            .filter_batch(docs, &mut self.matching_docs_buffer);
 
-        bucket.doc_count += req.matching_docs_buffer.len() as u64;
+        bucket.doc_count += self.matching_docs_buffer.len() as u64;
 
         // Batch process sub-aggregations if we have matches
-        if !req.matching_docs_buffer.is_empty() {
+        if !self.matching_docs_buffer.is_empty() {
             if let Some(sub_aggs) = &mut self.sub_aggregations {
-                for &doc_id in &req.matching_docs_buffer {
+                for &doc_id in &self.matching_docs_buffer {
                     sub_aggs.push(bucket.bucket_id, doc_id);
                 }
             }
         }
 
-        // Put the request data back
-        agg_data.put_back_filter_req_data(self.accessor_idx, req);
         if let Some(sub_aggs) = &mut self.sub_aggregations {
             sub_aggs.check_flush_local(agg_data)?;
         }

src/aggregation/bucket/histogram/histogram.rs

@@ -21,6 +21,7 @@ use crate::TantivyError;
 
 /// Contains all information required by the SegmentHistogramCollector to perform the
 /// histogram or date_histogram aggregation on a segment.
+#[derive(Debug, Clone)]
 pub struct HistogramAggReqData {
     /// The column accessor to access the fast field values.
     pub accessor: Column<u64>,
@@ -243,19 +244,52 @@ impl Display for HistogramBounds {
 }
 
 impl HistogramBounds {
-    fn contains(&self, val: f64) -> bool {
+    pub(crate) fn contains(&self, val: f64) -> bool {
         val >= self.min && val <= self.max
     }
 }
 
-#[derive(Default, Clone, Debug, PartialEq)]
-pub(crate) struct SegmentHistogramBucketEntry {
-    pub key: f64,
-    pub doc_count: u64,
-    pub bucket_id: BucketId,
+/// The per-bucket identifier stored in a [`SegmentHistogramBucketEntry`].
+///
+/// It is [`BucketId`] when the histogram has sub aggregations (which key their state by it), and
+/// the zero-sized `()` when it does not. Without sub aggregations the id is never read, so storing
+/// `()` drops 8 bytes per bucket (24 -> 16) and turns id assignment into a no-op.
+pub trait BucketIdSlot: Copy + Default + std::fmt::Debug + PartialEq {
+    /// Assigns the next id from the provider, called once when a bucket is first filled.
+    fn assign(provider: &mut BucketIdProvider) -> Self;
+    /// Resolves to the `BucketId` for sub-aggregation bookkeeping.
+    ///
+    /// Only ever called for the [`BucketId`] slot: the `()` slot is used exactly when there are no
+    /// sub aggregations, so every call site is guarded by `sub_agg.is_some()` and is dead for `()`.
+    fn to_bucket_id(self) -> BucketId;
+}
+impl BucketIdSlot for BucketId {
+    #[inline(always)]
+    fn assign(provider: &mut BucketIdProvider) -> Self {
+        provider.next_bucket_id()
+    }
+    #[inline(always)]
+    fn to_bucket_id(self) -> BucketId {
+        self
+    }
+}
+impl BucketIdSlot for () {
+    #[inline(always)]
+    fn assign(_provider: &mut BucketIdProvider) -> Self {}
+    #[inline(always)]
+    fn to_bucket_id(self) -> BucketId {
+        unreachable!("bucket ids are only resolved when sub aggregations are present")
+    }
 }
 
-impl SegmentHistogramBucketEntry {
+#[derive(Default, Clone, Debug, PartialEq)]
+pub(crate) struct SegmentHistogramBucketEntry<B> {
+    pub key: f64,
+    pub doc_count: u64,
+    pub bucket_id: B,
+}
+
+impl<B: BucketIdSlot> SegmentHistogramBucketEntry<B> {
     pub(crate) fn into_intermediate_bucket_entry(
         self,
         sub_aggregation: &mut Option<HighCardBufferedSubAggs>,
@@ -268,7 +302,7 @@ impl SegmentHistogramBucketEntry {
                 .add_intermediate_aggregation_result(
                     agg_data,
                     &mut sub_aggregation_res,
-                    self.bucket_id,
+                    self.bucket_id.to_bucket_id(),
                 )?;
         }
         Ok(IntermediateHistogramBucketEntry {
@@ -279,39 +313,147 @@ impl SegmentHistogramBucketEntry {
     }
 }
 
-#[derive(Clone, Debug, Default)]
-struct HistogramBuckets {
-    pub buckets: FxHashMap<i64, SegmentHistogramBucketEntry>,
+/// The contiguous bucket range a histogram can span, derived from the column min/max (clamped to
+/// the histogram bounds). Buckets in `[base_pos, base_pos + len)` can be stored in a flat `Vec`
+/// indexed by `bucket_pos - base_pos`, avoiding the hash map on the hot path.
+#[derive(Clone, Copy, Debug)]
+pub(crate) struct DenseRange {
+    /// `bucket_pos` mapped to index 0 of the dense `Vec`.
+    pub(crate) base_pos: i64,
+    /// Number of bucket positions in the range.
+    pub(crate) len: usize,
 }
-impl HistogramBuckets {
+
+/// Storage for the histogram buckets of a single parent bucket.
+///
+/// Starts out sparse (a hash map keyed by `bucket_pos`). Once enough distinct buckets have been
+/// filled that we are clearly going to cover most of the column's theoretical range, it switches
+/// to a dense `Vec` indexed by `bucket_pos - base_pos`, which removes hashing from the hot loop.
+#[derive(Clone, Debug)]
+enum HistogramBuckets<B> {
+    Sparse(FxHashMap<i64, SegmentHistogramBucketEntry<B>>),
+    Dense {
+        base_pos: i64,
+        /// One slot per bucket position; a slot with `doc_count == 0` has not been hit yet.
+        buckets: Vec<SegmentHistogramBucketEntry<B>>,
+    },
+}
+impl<B> Default for HistogramBuckets<B> {
+    fn default() -> Self {
+        HistogramBuckets::Sparse(FxHashMap::default())
+    }
+}
+impl<B: BucketIdSlot> HistogramBuckets<B> {
     fn memory_consumption(&self) -> u64 {
-        self.buckets.capacity() as u64 * std::mem::size_of::<SegmentHistogramBucketEntry>() as u64
+        let num_slots = match self {
+            HistogramBuckets::Sparse(map) => map.capacity(),
+            HistogramBuckets::Dense { buckets, .. } => buckets.capacity(),
+        };
+        num_slots as u64 * std::mem::size_of::<SegmentHistogramBucketEntry<B>>() as u64
+    }
+
+    /// Switches from sparse to dense storage once the dense `Vec` would use no more memory than the
+    /// hash map does now, so the switch never increases memory. Called at block boundaries.
+    ///
+    /// The `Vec` holds one `Entry` per bucket position in the range. The map additionally stores
+    /// the key and a control byte per slot, at a load factor of 7/16..7/8, so for a dense histogram
+    /// its footprint grows past the `Vec` well before full coverage. And since the `Vec` never
+    /// grows afterwards while the map would keep growing, dense only gets relatively cheaper — so
+    /// no upper bound on the range is needed: a large but sparse range simply never crosses over.
+    #[inline]
+    fn maybe_densify(&mut self, dense_range: Option<DenseRange>) {
+        let Some(range) = dense_range else { return };
+        let HistogramBuckets::Sparse(map) = self else {
+            return;
+        };
+        let dense_bytes = range
+            .len
+            .saturating_mul(std::mem::size_of::<SegmentHistogramBucketEntry<B>>());
+        let sparse_bytes = map
+            .capacity()
+            .saturating_mul(std::mem::size_of::<(i64, SegmentHistogramBucketEntry<B>)>() + 1);
+        if dense_bytes > sparse_bytes {
+            return;
+        }
+        let map = std::mem::take(map);
+        let mut buckets = vec![SegmentHistogramBucketEntry::<B>::default(); range.len];
+        for (bucket_pos, entry) in map {
+            buckets[(bucket_pos - range.base_pos) as usize] = entry;
+        }
+        *self = HistogramBuckets::Dense {
+            base_pos: range.base_pos,
+            buckets,
+        };
+    }
+
+    /// Returns the bucket entry for `bucket_pos`, setting its key (and `bucket_id`, when `B` is
+    /// [`BucketId`]) on first use.
+    ///
+    /// For the dense variant `bucket_pos` is guaranteed to be inside the range, since it is
+    /// derived from the column min/max that bounds every value (see [`compute_dense_range`]).
+    #[inline]
+    fn get_or_create(
+        &mut self,
+        bucket_pos: i64,
+        bucket_id_provider: &mut BucketIdProvider,
+        key_from_pos: impl FnOnce(i64) -> f64,
+    ) -> &mut SegmentHistogramBucketEntry<B> {
+        match self {
+            HistogramBuckets::Sparse(map) => {
+                map.entry(bucket_pos)
+                    .or_insert_with(|| SegmentHistogramBucketEntry {
+                        key: key_from_pos(bucket_pos),
+                        doc_count: 0,
+                        bucket_id: B::assign(bucket_id_provider),
+                    })
+            }
+            HistogramBuckets::Dense { base_pos, buckets } => {
+                let idx = (bucket_pos - *base_pos) as usize;
+                debug_assert!(idx < buckets.len(), "bucket_pos outside the dense range");
+                let entry = &mut buckets[idx];
+                if entry.doc_count == 0 {
+                    entry.key = key_from_pos(bucket_pos);
+                    entry.bucket_id = B::assign(bucket_id_provider);
+                }
+                entry
+            }
+        }
+    }
+
+    /// Consumes the storage, yielding all non-empty bucket entries.
+    fn into_filled_entries(self) -> Vec<SegmentHistogramBucketEntry<B>> {
+        match self {
+            HistogramBuckets::Sparse(map) => map.into_values().collect(),
+            HistogramBuckets::Dense { buckets, .. } => {
+                buckets.into_iter().filter(|b| b.doc_count > 0).collect()
+            }
+        }
     }
 }
 
 /// The collector puts values from the fast field into the correct buckets and does a conversion to
 /// the correct datatype.
 #[derive(Debug)]
-pub struct SegmentHistogramCollector {
+pub struct SegmentHistogramCollector<B> {
     /// The buckets containing the aggregation data.
     /// One Histogram bucket per parent bucket id.
-    parent_buckets: Vec<HistogramBuckets>,
+    parent_buckets: Vec<HistogramBuckets<B>>,
     sub_agg: Option<HighCardBufferedSubAggs>,
-    accessor_idx: usize,
+    req_data: HistogramAggReqData,
     bucket_id_provider: BucketIdProvider,
+    /// Theoretical bucket range derived from the column min/max, if dense `Vec` storage is
+    /// viable. `None` keeps every parent bucket in the sparse hash map.
+    dense_range: Option<DenseRange>,
 }
 
-impl SegmentAggregationCollector for SegmentHistogramCollector {
+impl<B: BucketIdSlot> SegmentAggregationCollector for SegmentHistogramCollector<B> {
     fn add_intermediate_aggregation_result(
         &mut self,
         agg_data: &AggregationsSegmentCtx,
         results: &mut IntermediateAggregationResults,
         parent_bucket_id: BucketId,
     ) -> crate::Result<()> {
-        let name = agg_data
-            .get_histogram_req_data(self.accessor_idx)
-            .name
-            .clone();
+        let name = self.req_data.name.clone();
         // TODO: avoid prepare_max_bucket here and handle empty buckets.
         self.prepare_max_bucket(parent_bucket_id, agg_data)?;
         let histogram = std::mem::take(&mut self.parent_buckets[parent_bucket_id as usize]);
@@ -328,10 +470,13 @@ impl SegmentAggregationCollector for SegmentHistogramCollector {
         docs: &[crate::DocId],
         agg_data: &mut AggregationsSegmentCtx,
     ) -> crate::Result<()> {
-        let req = agg_data.take_histogram_req_data(self.accessor_idx);
         let mem_pre = self.get_memory_consumption(parent_bucket_id);
-        let buckets = &mut self.parent_buckets[parent_bucket_id as usize].buckets;
+        let dense_range = self.dense_range;
+        let store = &mut self.parent_buckets[parent_bucket_id as usize];
+        // Upgrade to dense storage before processing the block if the buckets are dense enough.
+        store.maybe_densify(dense_range);
 
+        let req = &self.req_data;
         let bounds = req.bounds;
         let interval = req.req.interval;
         let offset = req.offset;
@@ -340,31 +485,42 @@ impl SegmentAggregationCollector for SegmentHistogramCollector {
         agg_data
             .column_block_accessor
             .fetch_block(docs, &req.accessor);
-        for (doc, val) in agg_data
-            .column_block_accessor
-            .iter_docid_vals(docs, &req.accessor)
-        {
-            let val = f64_from_fastfield_u64(val, req.field_type);
-            let bucket_pos = get_bucket_pos(val);
-            if bounds.contains(val) {
-                let bucket = buckets.entry(bucket_pos).or_insert_with(|| {
-                    let key = get_bucket_key_from_pos(bucket_pos as f64, interval, offset);
-                    SegmentHistogramBucketEntry {
-                        key,
-                        doc_count: 0,
-                        bucket_id: self.bucket_id_provider.next_bucket_id(),
-                    }
-                });
-                bucket.doc_count += 1;
-                if let Some(sub_agg) = &mut self.sub_agg {
-                    sub_agg.push(bucket.bucket_id, doc);
+        // special path for nested buckets
+        if let Some(sub_agg) = &mut self.sub_agg {
+            for (doc, val) in agg_data
+                .column_block_accessor
+                .iter_docid_vals(docs, &req.accessor)
+            {
+                let val = f64_from_fastfield_u64(val, req.field_type);
+                if bounds.contains(val) {
+                    let bucket = store.get_or_create(
+                        get_bucket_pos(val),
+                        &mut self.bucket_id_provider,
+                        |pos| get_bucket_key_from_pos(pos as f64, interval, offset),
+                    );
+                    bucket.doc_count += 1;
+                    sub_agg.push(bucket.bucket_id.to_bucket_id(), doc);
+                }
+            }
+        } else {
+            for val in agg_data.column_block_accessor.iter_vals() {
+                let val = f64_from_fastfield_u64(val, req.field_type);
+                if bounds.contains(val) {
+                    let bucket = store.get_or_create(
+                        get_bucket_pos(val),
+                        &mut self.bucket_id_provider,
+                        |pos| get_bucket_key_from_pos(pos as f64, interval, offset),
+                    );
+                    bucket.doc_count += 1;
                 }
             }
         }
-        agg_data.put_back_histogram_req_data(self.accessor_idx, req);
 
-        let mem_delta = self.get_memory_consumption(parent_bucket_id) - mem_pre;
-        if mem_delta > 0 {
+        // `checked_sub` is `None` when densifying shrank the accounted memory; only account growth.
+        if let Some(mem_delta) = self
+            .get_memory_consumption(parent_bucket_id)
+            .checked_sub(mem_pre)
+        {
             agg_data.context.limits.add_memory_consumed(mem_delta)?;
         }
 
@@ -388,9 +544,7 @@ impl SegmentAggregationCollector for SegmentHistogramCollector {
         _agg_data: &AggregationsSegmentCtx,
     ) -> crate::Result<()> {
         while self.parent_buckets.len() <= max_bucket as usize {
-            self.parent_buckets.push(HistogramBuckets {
-                buckets: FxHashMap::default(),
-            });
+            self.parent_buckets.push(HistogramBuckets::default());
         }
         Ok(())
     }
@@ -407,7 +561,7 @@ impl SegmentAggregationCollector for SegmentHistogramCollector {
     }
 }
 
-impl SegmentHistogramCollector {
+impl<B: BucketIdSlot> SegmentHistogramCollector<B> {
     fn get_memory_consumption(&self, parent_bucket_id: BucketId) -> u64 {
         self.parent_buckets[parent_bucket_id as usize].memory_consumption()
     }
@@ -416,21 +570,19 @@ impl SegmentHistogramCollector {
     fn add_intermediate_bucket_result(
         &mut self,
         agg_data: &AggregationsSegmentCtx,
-        histogram: HistogramBuckets,
+        histogram: HistogramBuckets<B>,
     ) -> crate::Result<IntermediateBucketResult> {
-        let mut buckets = Vec::with_capacity(histogram.buckets.len());
+        let filled = histogram.into_filled_entries();
+        let mut buckets = Vec::with_capacity(filled.len());
 
-        for bucket in histogram.buckets.into_values() {
+        for bucket in filled {
             let bucket_res = bucket.into_intermediate_bucket_entry(&mut self.sub_agg, agg_data);
 
             buckets.push(bucket_res?);
         }
         buckets.sort_unstable_by(|b1, b2| b1.key.total_cmp(&b2.key));
 
-        let is_date_agg = agg_data
-            .get_histogram_req_data(self.accessor_idx)
-            .field_type
-            == ColumnType::DateTime;
+        let is_date_agg = self.req_data.field_type == ColumnType::DateTime;
         Ok(IntermediateBucketResult::Histogram {
             buckets,
             is_date_agg,
@@ -446,32 +598,175 @@ impl SegmentHistogramCollector {
         } else {
             None
         };
-        let req_data = agg_data.get_histogram_req_data_mut(node.idx_in_req_data);
-        req_data.req.validate()?;
-        if req_data.field_type == ColumnType::DateTime && !req_data.is_date_histogram {
-            req_data.req.normalize_date_time();
-        }
-        req_data.bounds = req_data.req.hard_bounds.unwrap_or(HistogramBounds {
-            min: f64::MIN,
-            max: f64::MAX,
-        });
-        req_data.offset = req_data.req.offset.unwrap_or(0.0);
+        let mut req_data = agg_data.per_request.histogram_req_data[node.idx_in_req_data].clone();
+        normalize_histogram_req(&mut req_data)?;
+        agg_data
+            .context
+            .limits
+            .add_memory_consumed(req_data.get_memory_consumption() as u64)?;
+        let dense_range = compute_dense_range(
+            &req_data.accessor,
+            req_data.field_type,
+            req_data.req.interval,
+            req_data.offset,
+            req_data.bounds,
+        );
         let sub_agg = sub_agg.map(BufferedSubAggs::new);
 
         Ok(Self {
             parent_buckets: Default::default(),
             sub_agg,
-            accessor_idx: node.idx_in_req_data,
+            req_data,
             bucket_id_provider: BucketIdProvider::default(),
+            dense_range,
         })
     }
 }
 
+impl SegmentHistogramCollector<()> {
+    /// Builds a histogram collector whose parent `t` is a dense histogram filled from
+    /// `counts[t * num_time_buckets .. (t + 1) * num_time_buckets]` (row-major). Used by the fused
+    /// terms×histogram collector to turn its flat 2D counters into the regular intermediate result,
+    /// so cross-segment merging is shared with the general path.
+    pub(crate) fn from_dense_rows(
+        req_data: HistogramAggReqData,
+        base_pos: i64,
+        num_time_buckets: usize,
+        counts: &[u32],
+    ) -> Self {
+        let interval = req_data.req.interval;
+        let offset = req_data.offset;
+        let num_parents = counts.len().checked_div(num_time_buckets).unwrap_or(0);
+        let parent_buckets = (0..num_parents)
+            .map(|t| {
+                let row = &counts[t * num_time_buckets..(t + 1) * num_time_buckets];
+                let buckets = row
+                    .iter()
+                    .enumerate()
+                    .map(|(b, &doc_count)| SegmentHistogramBucketEntry {
+                        key: get_bucket_key_from_pos(
+                            (base_pos + b as i64) as f64,
+                            interval,
+                            offset,
+                        ),
+                        doc_count: doc_count as u64,
+                        bucket_id: (),
+                    })
+                    .collect();
+                HistogramBuckets::Dense { base_pos, buckets }
+            })
+            .collect();
+        Self {
+            parent_buckets,
+            sub_agg: None,
+            req_data,
+            bucket_id_provider: BucketIdProvider::default(),
+            dense_range: None,
+        }
+    }
+}
+
+/// Validates and normalizes a histogram request in place: applies date ns-normalization (for a
+/// `histogram` on a date column) and resolves `bounds`/`offset` from the request.
+fn normalize_histogram_req(req_data: &mut HistogramAggReqData) -> crate::Result<()> {
+    req_data.req.validate()?;
+    if req_data.field_type == ColumnType::DateTime && !req_data.is_date_histogram {
+        req_data.req.normalize_date_time();
+    }
+    req_data.bounds = req_data.req.hard_bounds.unwrap_or(HistogramBounds {
+        min: f64::MIN,
+        max: f64::MAX,
+    });
+    req_data.offset = req_data.req.offset.unwrap_or(0.0);
+    // Drop `hard_bounds` that can't exclude any value (the column's range already sits inside
+    // them): the per-doc `bounds.contains` check is then a no-op, so collapsing to the unbounded
+    // sentinel lets the histogram hot loop skip it and the fused term×histogram path derive
+    // per-term counts from the grid. Only this collect-time filter is touched — empty-bucket
+    // emission reads `req.hard_bounds` directly (see `get_req_min_max`), and `hard_bounds` only
+    // ever clips that range, so a wider-than-data bound leaves the result unchanged.
+    if req_data.req.hard_bounds.is_some() {
+        let col_min = f64_from_fastfield_u64(req_data.accessor.min_value(), req_data.field_type);
+        let col_max = f64_from_fastfield_u64(req_data.accessor.max_value(), req_data.field_type);
+        if col_min >= req_data.bounds.min && col_max <= req_data.bounds.max {
+            req_data.bounds = HistogramBounds {
+                min: f64::MIN,
+                max: f64::MAX,
+            };
+        }
+    }
+    Ok(())
+}
+
+/// Clones and normalizes (resolving interval/offset/bounds) the histogram request at `node`, and
+/// returns it together with its dense bucket range — or `None` if the column has no usable range.
+/// Used by the fused terms×histogram collector, which then owns the normalized request.
+pub(crate) fn prepare_histogram_dense_range(
+    agg_data: &AggregationsSegmentCtx,
+    node: &AggRefNode,
+) -> crate::Result<Option<(HistogramAggReqData, DenseRange)>> {
+    let mut req_data = agg_data.per_request.histogram_req_data[node.idx_in_req_data].clone();
+    normalize_histogram_req(&mut req_data)?;
+    let dense_range = compute_dense_range(
+        &req_data.accessor,
+        req_data.field_type,
+        req_data.req.interval,
+        req_data.offset,
+        req_data.bounds,
+    );
+    Ok(dense_range.map(|range| (req_data, range)))
+}
+
+/// Builds a boxed histogram (or date histogram) segment collector, picking the bucket-id storage
+/// based on whether there are sub aggregations: `()` (no id stored) when there are none, otherwise
+/// [`BucketId`].
+pub(crate) fn build_segment_histogram_collector(
+    agg_data: &mut AggregationsSegmentCtx,
+    node: &AggRefNode,
+) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
+    if node.children.is_empty() {
+        Ok(Box::new(
+            SegmentHistogramCollector::<()>::from_req_and_validate(agg_data, node)?,
+        ))
+    } else {
+        Ok(Box::new(
+            SegmentHistogramCollector::<BucketId>::from_req_and_validate(agg_data, node)?,
+        ))
+    }
+}
+
 #[inline]
-fn get_bucket_pos_f64(val: f64, interval: f64, offset: f64) -> f64 {
+pub(crate) fn get_bucket_pos_f64(val: f64, interval: f64, offset: f64) -> f64 {
     ((val - offset) / interval).floor()
 }
 
+/// Computes the dense bucket range for a column from its min/max value (clamped to the histogram
+/// bounds), or `None` if there are no values within bounds (or the range overflows `usize`).
+///
+/// There is no upper bound on the range: whether dense storage is actually used is decided later,
+/// per parent bucket, by [`HistogramBuckets::maybe_densify`] based on the memory it would save.
+///
+/// The column min/max bound every value the collector can see, so a `Vec` sized to this range can
+/// be indexed by `bucket_pos - base_pos` without any out-of-bounds check on the hot path.
+fn compute_dense_range(
+    accessor: &Column<u64>,
+    field_type: ColumnType,
+    interval: f64,
+    offset: f64,
+    bounds: HistogramBounds,
+) -> Option<DenseRange> {
+    let col_min = f64_from_fastfield_u64(accessor.min_value(), field_type);
+    let col_max = f64_from_fastfield_u64(accessor.max_value(), field_type);
+    let lo = col_min.max(bounds.min);
+    let hi = col_max.min(bounds.max);
+    if lo > hi {
+        return None;
+    }
+    let base_pos = get_bucket_pos_f64(lo, interval, offset) as i64;
+    let top_pos = get_bucket_pos_f64(hi, interval, offset) as i64;
+    let len = usize::try_from(top_pos.checked_sub(base_pos)?.checked_add(1)?).ok()?;
+    (len > 0).then_some(DenseRange { base_pos, len })
+}
+
 #[inline]
 fn get_bucket_key_from_pos(bucket_pos: f64, interval: f64, offset: f64) -> f64 {
     bucket_pos * interval + offset
@@ -776,6 +1071,62 @@ mod tests {
         Ok(())
     }
 
+    #[test]
+    fn histogram_dense_storage_test() -> crate::Result<()> {
+        histogram_dense_storage_test_with_opt(false)?;
+        histogram_dense_storage_test_with_opt(true)?;
+        Ok(())
+    }
+
+    /// Exercises the switch from sparse hash map to dense `Vec` storage. The switch happens at a
+    /// block boundary (a block is `COLLECT_BLOCK_BUFFER_LEN` = 64 docs), so we need many docs in a
+    /// single segment, densely covering the bucket range. `with_sub_agg` toggles the `iter_vals`
+    /// fast path vs. the `iter_docid_vals` path used when there is a sub aggregation.
+    fn histogram_dense_storage_test_with_opt(with_sub_agg: bool) -> crate::Result<()> {
+        let num_buckets = 50usize;
+        let docs_per_bucket = 10usize;
+        // Value `k` repeated `docs_per_bucket` times for each bucket `k`, so every value in bucket
+        // `k` equals `k` and the per-bucket average is exactly `k`.
+        let values: Vec<f64> = (0..num_buckets * docs_per_bucket)
+            .map(|i| (i % num_buckets) as f64)
+            .collect();
+        // `merge_segments = true` collapses the per-value segments into a single segment with all
+        // the docs, which is collected in 64-doc blocks and therefore switches to dense storage.
+        let index = get_test_index_from_values(true, &values)?;
+
+        let agg_req: Aggregations = serde_json::from_value(if with_sub_agg {
+            json!({
+                "histogram": {
+                    "histogram": { "field": "score_f64", "interval": 1.0 },
+                    "aggs": { "avg": { "avg": { "field": "score_f64" } } }
+                }
+            })
+        } else {
+            json!({
+                "histogram": {
+                    "histogram": { "field": "score_f64", "interval": 1.0 }
+                }
+            })
+        })
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+
+        for k in 0..num_buckets {
+            assert_eq!(res["histogram"]["buckets"][k]["key"], k as f64);
+            assert_eq!(
+                res["histogram"]["buckets"][k]["doc_count"],
+                docs_per_bucket as u64
+            );
+            if with_sub_agg {
+                assert_eq!(res["histogram"]["buckets"][k]["avg"]["value"], k as f64);
+            }
+        }
+        assert_eq!(res["histogram"]["buckets"][num_buckets], Value::Null);
+
+        Ok(())
+    }
+
     #[test]
     fn histogram_memory_limit() -> crate::Result<()> {
         let index = get_test_index_with_num_docs(true, 100)?;
@@ -1070,6 +1421,55 @@ mod tests {
         Ok(())
     }
 
+    #[test]
+    fn histogram_non_binding_hard_bounds_test_multi_segment() -> crate::Result<()> {
+        histogram_non_binding_hard_bounds_test_with_opt(false)
+    }
+    #[test]
+    fn histogram_non_binding_hard_bounds_test_single_segment() -> crate::Result<()> {
+        histogram_non_binding_hard_bounds_test_with_opt(true)
+    }
+    /// `hard_bounds` wider than the data (here with mid-interval edges, to cover the "bound cuts a
+    /// bucket" case) can't exclude any value, so the result must be identical to the same request
+    /// without bounds. Guards the normalization that collapses such bounds to the unbounded
+    /// sentinel so the hot loop / fused path can skip the per-doc bounds check.
+    fn histogram_non_binding_hard_bounds_test_with_opt(merge_segments: bool) -> crate::Result<()> {
+        let values = vec![10.0, 12.0, 14.0, 16.0, 10.0, 13.0, 10.0, 12.0];
+        let index = get_test_index_from_values(merge_segments, &values)?;
+
+        // Mid-interval edges, but wider than the data range [10, 16] -> they exclude nothing.
+        let with_bounds: Aggregations = serde_json::from_value(json!({
+            "histogram": {
+                "histogram": {
+                    "field": "score_f64",
+                    "interval": 1.0,
+                    "hard_bounds": { "min": 9.5, "max": 16.5 }
+                }
+            }
+        }))
+        .unwrap();
+        let no_bounds: Aggregations = serde_json::from_value(json!({
+            "histogram": {
+                "histogram": { "field": "score_f64", "interval": 1.0 }
+            }
+        }))
+        .unwrap();
+
+        let res_bounds = exec_request(with_bounds, &index)?;
+        let res_plain = exec_request(no_bounds, &index)?;
+        // Dropping a non-binding bound must not change anything.
+        assert_eq!(res_bounds, res_plain);
+
+        // Sanity: buckets span the data range with gaps filled (min_doc_count defaults to 0).
+        assert_eq!(res_bounds["histogram"]["buckets"][0]["key"], 10.0);
+        assert_eq!(res_bounds["histogram"]["buckets"][0]["doc_count"], 3);
+        assert_eq!(res_bounds["histogram"]["buckets"][6]["key"], 16.0);
+        assert_eq!(res_bounds["histogram"]["buckets"][6]["doc_count"], 1);
+        assert_eq!(res_bounds["histogram"]["buckets"][7], Value::Null);
+
+        Ok(())
+    }
+
     #[test]
     fn histogram_empty_result_behaviour_test_single_segment() -> crate::Result<()> {
         histogram_empty_result_behaviour_test_with_opt(true)

src/aggregation/bucket/range.rs

@@ -23,6 +23,7 @@ use crate::TantivyError;
 
 /// Contains all information required by the SegmentRangeCollector to perform the
 /// range aggregation on a segment.
+#[derive(Debug, Clone)]
 pub struct RangeAggReqData {
     /// The column accessor to access the fast field values.
     pub accessor: Column<u64>,
@@ -161,7 +162,7 @@ pub struct SegmentRangeCollector<B: SubAggBuffer> {
     /// One for each ParentBucketId
     parent_buckets: Vec<Vec<SegmentRangeAndBucketEntry>>,
     column_type: ColumnType,
-    pub(crate) accessor_idx: usize,
+    pub(crate) req_data: RangeAggReqData,
     sub_agg: Option<BufferedSubAggs<B>>,
     /// Here things get a bit weird. We need to assign unique bucket ids across all
     /// parent buckets. So we keep track of the next available bucket id here.
@@ -184,7 +185,7 @@ impl<B: SubAggBuffer> Debug for SegmentRangeCollector<B> {
         f.debug_struct("SegmentRangeCollector")
             .field("parent_buckets_len", &self.parent_buckets.len())
             .field("column_type", &self.column_type)
-            .field("accessor_idx", &self.accessor_idx)
+            .field("name", &self.req_data.name)
             .field("has_sub_agg", &self.sub_agg.is_some())
             .finish()
     }
@@ -239,10 +240,7 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
     ) -> crate::Result<()> {
         self.prepare_max_bucket(parent_bucket_id, agg_data)?;
         let field_type = self.column_type;
-        let name = agg_data
-            .get_range_req_data(self.accessor_idx)
-            .name
-            .to_string();
+        let name = self.req_data.name.to_string();
 
         let buckets = std::mem::take(&mut self.parent_buckets[parent_bucket_id as usize]);
 
@@ -281,17 +279,15 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
         docs: &[crate::DocId],
         agg_data: &mut AggregationsSegmentCtx,
     ) -> crate::Result<()> {
-        let req = agg_data.take_range_req_data(self.accessor_idx);
-
         agg_data
             .column_block_accessor
-            .fetch_block(docs, &req.accessor);
+            .fetch_block(docs, &self.req_data.accessor);
 
         let buckets = &mut self.parent_buckets[parent_bucket_id as usize];
 
         for (doc, val) in agg_data
             .column_block_accessor
-            .iter_docid_vals(docs, &req.accessor)
+            .iter_docid_vals(docs, &self.req_data.accessor)
         {
             let bucket_pos = get_bucket_pos(val, buckets);
             let bucket = &mut buckets[bucket_pos];
@@ -301,7 +297,6 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
             }
         }
 
-        agg_data.put_back_range_req_data(self.accessor_idx, req);
         if let Some(sub_agg) = self.sub_agg.as_mut() {
             sub_agg.check_flush_local(agg_data)?;
         }
@@ -319,10 +314,10 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
     fn prepare_max_bucket(
         &mut self,
         max_bucket: BucketId,
-        agg_data: &AggregationsSegmentCtx,
+        _agg_data: &AggregationsSegmentCtx,
     ) -> crate::Result<()> {
         while self.parent_buckets.len() <= max_bucket as usize {
-            let new_buckets = self.create_new_buckets(agg_data)?;
+            let new_buckets = self.create_new_buckets()?;
             self.parent_buckets.push(new_buckets);
         }
 
@@ -346,8 +341,11 @@ pub(crate) fn build_segment_range_collector(
     agg_data: &mut AggregationsSegmentCtx,
     node: &AggRefNode,
 ) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
-    let accessor_idx = node.idx_in_req_data;
-    let req_data = agg_data.get_range_req_data(node.idx_in_req_data);
+    let req_data = agg_data.per_request.range_req_data[node.idx_in_req_data].clone();
+    agg_data
+        .context
+        .limits
+        .add_memory_consumed(req_data.get_memory_consumption() as u64)?;
     let field_type = req_data.field_type;
 
     // TODO: A better metric instead of is_top_level would be the number of buckets expected.
@@ -365,7 +363,7 @@ pub(crate) fn build_segment_range_collector(
         Ok(Box::new(SegmentRangeCollector::<LowCardSubAggBuffer> {
             sub_agg: sub_agg.map(LowCardBufferedSubAggs::new),
             column_type: field_type,
-            accessor_idx,
+            req_data,
             parent_buckets: Vec::new(),
             bucket_id_provider: BucketIdProvider::default(),
             limits: agg_data.context.limits.clone(),
@@ -374,7 +372,7 @@ pub(crate) fn build_segment_range_collector(
         Ok(Box::new(SegmentRangeCollector::<HighCardSubAggBuffer> {
             sub_agg: sub_agg.map(BufferedSubAggs::new),
             column_type: field_type,
-            accessor_idx,
+            req_data,
             parent_buckets: Vec::new(),
             bucket_id_provider: BucketIdProvider::default(),
             limits: agg_data.context.limits.clone(),
@@ -383,12 +381,9 @@ pub(crate) fn build_segment_range_collector(
 }
 
 impl<B: SubAggBuffer> SegmentRangeCollector<B> {
-    pub(crate) fn create_new_buckets(
-        &mut self,
-        agg_data: &AggregationsSegmentCtx,
-    ) -> crate::Result<Vec<SegmentRangeAndBucketEntry>> {
+    pub(crate) fn create_new_buckets(&mut self) -> crate::Result<Vec<SegmentRangeAndBucketEntry>> {
         let field_type = self.column_type;
-        let req_data = agg_data.get_range_req_data(self.accessor_idx);
+        let req_data = &self.req_data;
         // The range input on the request is f64.
         // We need to convert to u64 ranges, because we read the values as u64.
         // The mapping from the conversion is monotonic so ordering is preserved.
@@ -563,17 +558,16 @@ mod tests {
         get_test_index_with_num_docs,
     };
 
-    pub fn get_collector_from_ranges(
-        ranges: Vec<RangeAggregationRange>,
+    pub fn build_test_buckets(
+        ranges: &[RangeAggregationRange],
         field_type: ColumnType,
-    ) -> SegmentRangeCollector<HighCardSubAggBuffer> {
+    ) -> Vec<SegmentRangeAndBucketEntry> {
         let req = RangeAggregation {
             field: "dummy".to_string(),
-            ranges,
+            ranges: ranges.to_vec(),
             ..Default::default()
         };
-        // Build buckets directly as in from_req_and_validate without AggregationsData
-        let buckets: Vec<_> = extend_validate_ranges(&req.ranges, &field_type)
+        extend_validate_ranges(&req.ranges, &field_type)
             .expect("unexpected error in extend_validate_ranges")
             .iter()
             .map(|range| {
@@ -604,16 +598,7 @@ mod tests {
                     },
                 }
             })
-            .collect();
-
-        SegmentRangeCollector {
-            parent_buckets: vec![buckets],
-            column_type: field_type,
-            accessor_idx: 0,
-            sub_agg: None,
-            bucket_id_provider: Default::default(),
-            limits: AggregationLimitsGuard::default(),
-        }
+            .collect()
     }
 
     #[test]
@@ -856,10 +841,10 @@ mod tests {
 
     #[test]
     fn bucket_test_extend_range_hole() {
-        let buckets = vec![(10f64..20f64).into(), (30f64..40f64).into()];
-        let collector = get_collector_from_ranges(buckets, ColumnType::F64);
+        let buckets = [(10f64..20f64).into(), (30f64..40f64).into()];
+        let parent_buckets = [build_test_buckets(&buckets, ColumnType::F64)];
 
-        let buckets = collector.parent_buckets[0].clone();
+        let buckets = parent_buckets[0].clone();
         assert_eq!(buckets[0].range.start, u64::MIN);
         assert_eq!(buckets[0].range.end, 10f64.to_u64());
         assert_eq!(buckets[1].range.start, 10f64.to_u64());
@@ -875,14 +860,14 @@ mod tests {
     fn bucket_test_range_conversion_special_case() {
         // the monotonic conversion between f64 and u64, does not map f64::MIN.to_u64() ==
         // u64::MIN, but the into trait converts f64::MIN/MAX to None
-        let buckets = vec![
+        let buckets = [
             (f64::MIN..10f64).into(),
             (10f64..20f64).into(),
             (20f64..f64::MAX).into(),
         ];
-        let collector = get_collector_from_ranges(buckets, ColumnType::F64);
+        let parent_buckets = [build_test_buckets(&buckets, ColumnType::F64)];
 
-        let buckets = collector.parent_buckets[0].clone();
+        let buckets = parent_buckets[0].clone();
         assert_eq!(buckets[0].range.start, u64::MIN);
         assert_eq!(buckets[0].range.end, 10f64.to_u64());
         assert_eq!(buckets[1].range.start, 10f64.to_u64());
@@ -894,28 +879,28 @@ mod tests {
 
     #[test]
     fn bucket_range_test_negative_vals() {
-        let buckets = vec![(-10f64..-1f64).into()];
-        let collector = get_collector_from_ranges(buckets, ColumnType::F64);
+        let buckets = [(-10f64..-1f64).into()];
+        let parent_buckets = [build_test_buckets(&buckets, ColumnType::F64)];
 
-        let buckets = collector.parent_buckets[0].clone();
+        let buckets = parent_buckets[0].clone();
         assert_eq!(&buckets[0].bucket.key.to_string(), "*--10");
         assert_eq!(&buckets[buckets.len() - 1].bucket.key.to_string(), "-1-*");
     }
     #[test]
     fn bucket_range_test_positive_vals() {
-        let buckets = vec![(0f64..10f64).into()];
-        let collector = get_collector_from_ranges(buckets, ColumnType::F64);
+        let buckets = [(0f64..10f64).into()];
+        let parent_buckets = [build_test_buckets(&buckets, ColumnType::F64)];
 
-        let buckets = collector.parent_buckets[0].clone();
+        let buckets = parent_buckets[0].clone();
         assert_eq!(&buckets[0].bucket.key.to_string(), "*-0");
         assert_eq!(&buckets[buckets.len() - 1].bucket.key.to_string(), "10-*");
     }
 
     #[test]
     fn range_binary_search_test_u64() {
-        let check_ranges = |ranges: Vec<RangeAggregationRange>| {
-            let collector = get_collector_from_ranges(ranges, ColumnType::U64);
-            let search = |val: u64| get_bucket_pos(val, &collector.parent_buckets[0]);
+        let check_ranges = |ranges: &[RangeAggregationRange]| {
+            let parent_buckets = [build_test_buckets(ranges, ColumnType::U64)];
+            let search = |val: u64| get_bucket_pos(val, &parent_buckets[0]);
 
             assert_eq!(search(u64::MIN), 0);
             assert_eq!(search(9), 0);
@@ -928,7 +913,7 @@ mod tests {
         };
 
         let ranges = vec![(10.0..100.0).into()];
-        check_ranges(ranges);
+        check_ranges(&ranges);
 
         let ranges = vec![
             RangeAggregationRange {
@@ -938,7 +923,7 @@ mod tests {
             },
             (10.0..100.0).into(),
         ];
-        check_ranges(ranges);
+        check_ranges(&ranges);
 
         let ranges = vec![
             RangeAggregationRange {
@@ -953,15 +938,15 @@ mod tests {
                 from: Some(100.0),
             },
         ];
-        check_ranges(ranges);
+        check_ranges(&ranges);
     }
 
     #[test]
     fn range_binary_search_test_f64() {
-        let ranges = vec![(10.0..100.0).into()];
+        let ranges = [(10.0..100.0).into()];
 
-        let collector = get_collector_from_ranges(ranges, ColumnType::F64);
-        let search = |val: u64| get_bucket_pos(val, &collector.parent_buckets[0]);
+        let parent_buckets = [build_test_buckets(&ranges, ColumnType::F64)];
+        let search = |val: u64| get_bucket_pos(val, &parent_buckets[0]);
 
         assert_eq!(search(u64::MIN), 0);
         assert_eq!(search(9f64.to_u64()), 0);

src/aggregation/bucket/term_agg/mod.rs

@@ -29,6 +29,8 @@ use crate::aggregation::{format_date, BucketId, Key};
 use crate::error::DataCorruption;
 use crate::TantivyError;
 
+mod term_histogram;
+
 /// Contains all information required by the SegmentTermCollector to perform the
 /// terms aggregation on a segment.
 #[derive(Debug, Clone)]
@@ -374,9 +376,21 @@ pub(crate) fn build_segment_term_collector(
     // Let's see if we can use a vec to aggregate our data
     // instead of a hashmap.
     let col_max_value = terms_req_data.accessor.max_value();
-    let max_term_id: u64 =
+    let max_column_val: u64 =
         col_max_value.max(terms_req_data.missing_value_for_accessor.unwrap_or(0u64));
 
+    // Fused fast path: low-cardinality terms × a single `histogram`/`date_histogram` leaf over full
+    // columns with a small enough bucket grid. Anything else falls through to the general path.
+    if let Some(collector) = term_histogram::maybe_build_collector(
+        req_data,
+        node,
+        &terms_req_data,
+        max_column_val,
+        is_top_level,
+    )? {
+        return Ok(collector);
+    }
+
     let sub_agg_collector = if has_sub_aggregations {
         Some(build_segment_agg_collectors(req_data, &node.children)?)
     } else {
@@ -385,30 +399,30 @@ pub(crate) fn build_segment_term_collector(
 
     let mut bucket_id_provider = BucketIdProvider::default();
     // Decide which bucket storage is best suited for this aggregation.
-    if is_top_level && max_term_id < MAX_NUM_TERMS_FOR_VEC && !has_sub_aggregations {
-        let term_buckets = VecTermBucketsNoAgg::new(max_term_id + 1, &mut bucket_id_provider);
+    if is_top_level && max_column_val < MAX_NUM_TERMS_FOR_VEC && !has_sub_aggregations {
+        let term_buckets = VecTermBucketsNoAgg::new(max_column_val + 1, &mut bucket_id_provider);
         let collector: SegmentTermCollector<_, HighCardSubAggBuffer> = SegmentTermCollector {
             parent_buckets: vec![term_buckets],
             sub_agg: None,
             bucket_id_provider,
-            max_term_id,
+            max_term_id: max_column_val,
             terms_req_data,
         };
         Ok(Box::new(collector))
-    } else if is_top_level && max_term_id < MAX_NUM_TERMS_FOR_VEC {
-        let term_buckets = VecTermBuckets::new(max_term_id + 1, &mut bucket_id_provider);
+    } else if is_top_level && max_column_val < MAX_NUM_TERMS_FOR_VEC {
+        let term_buckets = VecTermBuckets::new(max_column_val + 1, &mut bucket_id_provider);
         let sub_agg = sub_agg_collector.map(LowCardBufferedSubAggs::new);
         let collector: SegmentTermCollector<_, LowCardSubAggBuffer> = SegmentTermCollector {
             parent_buckets: vec![term_buckets],
             sub_agg,
             bucket_id_provider,
-            max_term_id,
+            max_term_id: max_column_val,
             terms_req_data,
         };
         Ok(Box::new(collector))
-    } else if max_term_id < 8_000_000 && is_top_level {
+    } else if max_column_val < 8_000_000 && is_top_level {
         let term_buckets: PagedTermMap =
-            PagedTermMap::new(max_term_id + 1, &mut bucket_id_provider);
+            PagedTermMap::new(max_column_val + 1, &mut bucket_id_provider);
         // Build sub-aggregation blueprint (flat pairs)
         let sub_agg = sub_agg_collector.map(BufferedSubAggs::new);
         let collector: SegmentTermCollector<PagedTermMap, HighCardSubAggBuffer> =
@@ -416,7 +430,7 @@ pub(crate) fn build_segment_term_collector(
                 parent_buckets: vec![term_buckets],
                 sub_agg,
                 bucket_id_provider,
-                max_term_id,
+                max_term_id: max_column_val,
                 terms_req_data,
             };
         Ok(Box::new(collector))
@@ -429,7 +443,7 @@ pub(crate) fn build_segment_term_collector(
                 parent_buckets: vec![term_buckets],
                 sub_agg,
                 bucket_id_provider,
-                max_term_id,
+                max_term_id: max_column_val,
                 terms_req_data,
             };
         Ok(Box::new(collector))

src/aggregation/bucket/term_agg/term_histogram.rs

@@ -0,0 +1,585 @@
+//! Fused collector for the very common shape `terms` (low cardinality) × a single
+//! `histogram`/`date_histogram` sub-aggregation with nothing nested below it.
+//!
+//! See [`SegmentTermHistogramCollector`] for the approach and [`maybe_build_collector`] for the
+//! conditions under which it is used.
+
+use columnar::ColumnBlockAccessor;
+
+use super::{Bucket, SegmentTermCollector, TermsAggReqData, VecTermBuckets};
+use crate::aggregation::agg_data::{AggKind, AggRefNode, AggregationsSegmentCtx};
+use crate::aggregation::bucket::{
+    get_bucket_pos_f64, prepare_histogram_dense_range, HistogramAggReqData,
+    SegmentHistogramCollector,
+};
+use crate::aggregation::buffered_sub_aggs::LowCardSubAggBuffer;
+use crate::aggregation::intermediate_agg_result::{
+    IntermediateAggregationResult, IntermediateAggregationResults,
+};
+use crate::aggregation::segment_agg_result::{BucketIdProvider, SegmentAggregationCollector};
+use crate::aggregation::{f64_from_fastfield_u64, BucketId};
+
+/// Maximum number of cells (`num_terms × num_time_buckets`) in the fused flat 2D grid. Above this
+/// the grid would be too large/cache-unfriendly, so we fall back to the general buffered path.
+/// `1 << 14` cells = 128 KB of `u64` counters, comfortably L2-resident.
+///
+/// Since we are only at the top-level, this won't be multiplied by any parent buckets.
+const MAX_FUSED_GRID_BUCKETS: usize = 16384;
+
+/// Fused collector for `terms` (low cardinality) × a single `histogram`/`date_histogram` leaf with
+/// nothing nested below it, when the resulting `num_terms × num_time_buckets` grid is small (see
+/// [`MAX_FUSED_GRID_BUCKETS`]).
+///
+/// It keeps a flat, fully dense 2D counter grid (`counts[term * num_time_buckets + bucket]`) and a
+/// per-term total. A single pass reads both the term and histogram columns in document order and
+/// bumps the counters directly — no doc-id buffering, no per-term scattered re-fetch, no dynamic
+/// dispatch on flush, no per-bucket key/id storage during collection (keys are derived from the
+/// index at the end).
+///
+/// At result time the flat grid is expanded back into the regular term map + histogram storage and
+/// handed to the shared intermediate-result builders, so cross-segment merging is identical to the
+/// general path.
+#[derive(Debug)]
+pub(crate) struct SegmentTermHistogramCollector {
+    /// Per-term count of docs *outside* `hard_bounds` (still in `doc_count`, but in no bucket).
+    /// Per-term total = this + the term's `counts` row-sum; left empty when there are no hard
+    /// bounds (every doc is in-bounds, so there's no remainder to track).
+    term_counts: Vec<u32>,
+    /// Flattened `[num_terms * num_time_buckets]` histogram counters (`u32`, see
+    /// `term_counts`).
+    ///
+    /// Each term id get its own contiguous slice of `num_time_buckets` histogram counter.
+    /// When we count all docs (#nofilter), we can derive the per-term total as the sum over that
+    /// term's slice.
+    counts: Vec<u32>,
+    /// Histogram buckets per term (the dense time-range length).
+    num_time_buckets: usize,
+    /// `bucket_pos` mapped to time-bucket index 0.
+    base_pos: i64,
+    terms_req_data: TermsAggReqData,
+    /// The (cloned, normalized) histogram request: its column + interval/offset/bounds.
+    hist_req_data: HistogramAggReqData,
+    /// Private block accessors for both columns. We read them together, so each needs its own
+    /// (the shared `agg_data` scratch accessor only holds one block at a time). Owning them keeps
+    /// `collect` independent of `agg_data`.
+    term_block: ColumnBlockAccessor<u64>,
+    hist_block: ColumnBlockAccessor<u64>,
+    /// No hard bounds, so every doc is in-bounds.
+    all_docs_in_bounds: bool,
+    /// Both columns are full (fused-path precondition); cached so `collect` skips the per-block
+    /// cardinality lookup in `fetch_block`.
+    is_full: bool,
+}
+
+impl SegmentAggregationCollector for SegmentTermHistogramCollector {
+    fn add_intermediate_aggregation_result(
+        &mut self,
+        agg_data: &AggregationsSegmentCtx,
+        results: &mut IntermediateAggregationResults,
+        parent_bucket_id: BucketId,
+    ) -> crate::Result<()> {
+        debug_assert_eq!(
+            parent_bucket_id, 0,
+            "fused term-histogram collector is top-level only"
+        );
+        // Expand the flat grid back into the regular structures and reuse the shared builders, so
+        // ordering/cut-off/dict handling and cross-segment merging match the general path exactly.
+        let mut bucket_id_provider = BucketIdProvider::default();
+        // Per-term total = histogram row-sum (in-bounds) + `term_counts` (out-of-bounds remainder,
+        // empty when there are no hard bounds).
+        let term_buckets = VecTermBuckets {
+            buckets: self
+                .counts
+                .chunks_exact(self.num_time_buckets)
+                .enumerate()
+                .map(|(term_id, row)| {
+                    let in_bounds: u32 = row.iter().sum();
+                    let out_of_bounds = self.term_counts.get(term_id).copied().unwrap_or(0);
+                    Bucket {
+                        count: in_bounds + out_of_bounds,
+                        bucket_id: bucket_id_provider.next_bucket_id(),
+                    }
+                })
+                .collect(),
+        };
+        let mut histogram = SegmentHistogramCollector::<()>::from_dense_rows(
+            self.hist_req_data.clone(),
+            self.base_pos,
+            self.num_time_buckets,
+            &self.counts,
+        );
+        let name = self.terms_req_data.name.clone();
+        let bucket = SegmentTermCollector::<VecTermBuckets, LowCardSubAggBuffer>::into_intermediate_bucket_result(
+            &self.terms_req_data,
+            Some(&mut histogram as &mut dyn SegmentAggregationCollector),
+            term_buckets,
+            agg_data,
+        )?;
+        results.push(name, IntermediateAggregationResult::Bucket(bucket))?;
+        Ok(())
+    }
+
+    #[inline]
+    fn collect(
+        &mut self,
+        parent_bucket_id: BucketId,
+        docs: &[crate::DocId],
+        _agg_data: &mut AggregationsSegmentCtx,
+    ) -> crate::Result<()> {
+        debug_assert_eq!(
+            parent_bucket_id, 0,
+            "fused term-histogram collector is top-level only"
+        );
+
+        // Fetch both columns into our own accessors (we read them together, so they can't share the
+        // single `agg_data` scratch accessor). The collector owns all its inputs, so `collect`
+        // doesn't touch `agg_data`.
+        self.term_block
+            .fetch_block_with_is_full(docs, &self.terms_req_data.accessor, self.is_full);
+        self.hist_block
+            .fetch_block_with_is_full(docs, &self.hist_req_data.accessor, self.is_full);
+
+        // Hoist the loop-invariant fields into locals: the optimizer can't prove the
+        // `self.counts`/`self.term_counts` writes don't alias these `self` fields, so it can't keep
+        // them in registers and re-reads them from memory every iteration — ~15% slower on
+        // `terms_status_with_date_histogram` when read straight from `self`.
+        // Note: check which are actually relevant.
+        let field_type = self.hist_req_data.field_type;
+        let bounds = self.hist_req_data.bounds;
+        let interval = self.hist_req_data.req.interval;
+        let offset = self.hist_req_data.offset;
+        let base_pos = self.base_pos;
+        let num_time_buckets = self.num_time_buckets;
+        let all_docs_in_bounds = self.all_docs_in_bounds;
+        let term_counts = &mut self.term_counts;
+        let counts = &mut self.counts;
+
+        // Both columns are full (checked at construction), so values align with `docs` positionally
+        // and are read together in one pass.
+        // In-bounds docs bump the `counts` grid, out-of-bounds bump `term_counts`; deriving the
+        // total at flush avoids a per-doc `term_counts` RMW that serializes on
+        // store-to-load forwarding.
+        for (term_id, hist_raw) in self.term_block.iter_vals().zip(self.hist_block.iter_vals()) {
+            let term_id = term_id as usize;
+            let val = f64_from_fastfield_u64(hist_raw, field_type);
+            if all_docs_in_bounds || bounds.contains(val) {
+                let bucket = (get_bucket_pos_f64(val, interval, offset) as i64 - base_pos) as usize;
+                debug_assert!(
+                    bucket < num_time_buckets,
+                    "histogram bucket outside dense range"
+                );
+                counts[term_id * num_time_buckets + bucket] += 1;
+            } else {
+                term_counts[term_id] += 1;
+            }
+        }
+        Ok(())
+    }
+
+    fn flush(&mut self, _agg_data: &mut AggregationsSegmentCtx) -> crate::Result<()> {
+        // Nothing is buffered: `collect` writes the flat grid directly.
+        Ok(())
+    }
+
+    fn prepare_max_bucket(
+        &mut self,
+        _max_bucket: BucketId,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> crate::Result<()> {
+        // Top-level: the flat grid is allocated up front.
+        Ok(())
+    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        None
+    }
+}
+
+/// Builds the fused terms×histogram collector for a single top-level parent, when the shape is
+/// eligible. Returns `Ok(None)` to fall back to the general buffered terms path.
+///
+/// Eligibility: top-level, low-cardinality terms over a full column with no missing/include-exclude
+/// handling; a single `histogram`/`date_histogram` leaf (no nesting below it) over a full column;
+/// and a `num_terms × num_time_buckets` grid no larger than [`MAX_FUSED_GRID_BUCKETS`].
+pub(super) fn maybe_build_collector(
+    agg_data: &mut AggregationsSegmentCtx,
+    node: &AggRefNode,
+    terms_req_data: &TermsAggReqData,
+    col_max_val: u64,
+    is_top_level: bool,
+) -> crate::Result<Option<Box<dyn SegmentAggregationCollector>>> {
+    // Both columns must be full (one value per doc) so their values align positionally with `docs`
+    // and we can zip them. Requiring full columns also makes the terms agg's `missing` config a
+    // no-op (`fetch_block_with_missing` early-returns on full columns), so we needn't check for it.
+    //
+    // We don't cap the term cardinality here: the flat grid is bounded by the total cell count
+    // (`num_terms * num_time_buckets <= MAX_FUSED_GRID_BUCKETS`) checked below, which subsumes it.
+    //
+    // We only allow this at the top-level, since we don't know how many buckets are created. We
+    // are less likely to get enough docs for the preallocation to be worth and there's a risk of
+    // using too much memory. We could check the maximum theoretical buckets up-front and pass
+    // them down.
+    let fuseable = is_top_level
+        // TODO: We can easily support this
+        && terms_req_data.allowed_term_ids.is_none()
+        && terms_req_data.accessor.get_cardinality().is_full()
+        // The flat counters are `u32`, bumped once per value, so no count can exceed the column's
+        // value count. (Essentially always true here: the column is full, so its value count
+        // equals the doc count, and `DocId` is `u32`.)
+        && terms_req_data.accessor.values.num_vals() < u32::MAX
+        && node.children.len() == 1
+        && matches!(
+            node.children[0].kind,
+            AggKind::Histogram | AggKind::DateHistogram
+        )
+        && node.children[0].children.is_empty()
+        && agg_data.per_request.histogram_req_data[node.children[0].idx_in_req_data]
+            .accessor
+            .get_cardinality()
+            .is_full();
+    if !fuseable {
+        return Ok(None);
+    }
+
+    // Clone + normalize the histogram request and get its dense bucket range; only take the fused
+    // path when the flat `num_terms × num_time_buckets` grid is small enough.
+    let Some((hist_req_data, range)) = prepare_histogram_dense_range(agg_data, &node.children[0])?
+    else {
+        return Ok(None);
+    };
+    let num_terms = col_max_val.saturating_add(1) as usize;
+    if num_terms.saturating_mul(range.len) > MAX_FUSED_GRID_BUCKETS {
+        return Ok(None);
+    }
+
+    // No hard bounds means every doc is in-bounds, letting `collect` short-circuit the bounds
+    // check — and leaving `term_counts` (the out-of-bounds remainder) unused, so we skip allocating
+    // it.
+    let all_docs_in_bounds =
+        hist_req_data.bounds.min == f64::MIN && hist_req_data.bounds.max == f64::MAX;
+    let counts = vec![0u32; num_terms * range.len];
+    let term_counts = if all_docs_in_bounds {
+        Vec::new()
+    } else {
+        vec![0u32; num_terms]
+    };
+    // Charge both grids to the aggregation memory limit.
+    agg_data.context.limits.add_memory_consumed(
+        ((counts.len() + term_counts.len()) * std::mem::size_of::<u32>()) as u64,
+    )?;
+    Ok(Some(Box::new(SegmentTermHistogramCollector {
+        term_counts,
+        counts,
+        num_time_buckets: range.len,
+        base_pos: range.base_pos,
+        terms_req_data: terms_req_data.clone(),
+        hist_req_data,
+        term_block: ColumnBlockAccessor::default(),
+        hist_block: ColumnBlockAccessor::default(),
+        all_docs_in_bounds,
+        is_full: terms_req_data.accessor.get_cardinality().is_full(),
+    })))
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::aggregation::agg_req::Aggregations;
+    use crate::aggregation::tests::{
+        exec_request, exec_request_with_query_and_memory_limit,
+        get_test_index_from_values_and_terms,
+    };
+    use crate::aggregation::AggregationLimitsGuard;
+
+    /// Hand-computed correctness check for the fused terms×histogram fast path
+    /// ([`super::SegmentTermHistogramCollector`]): low-cardinality terms × a histogram leaf over
+    /// full columns, exercised single- and multi-segment.
+    #[test]
+    fn fused_term_histogram_test() -> crate::Result<()> {
+        fused_term_histogram_with_opt(false)?;
+        fused_term_histogram_with_opt(true)?;
+        Ok(())
+    }
+
+    fn fused_term_histogram_with_opt(merge_segments: bool) -> crate::Result<()> {
+        // 300 docs: term = {a, b, c} by i % 3, histogram value = i % 20 (interval 1 => buckets
+        // 0..19). gcd(3, 20) = 1, so every (term, bucket) pair occurs exactly 300 / 60 = 5 times.
+        let docs: Vec<(f64, String)> = (0..300u64)
+            .map(|i| {
+                (
+                    (i % 20) as f64,
+                    ["a", "b", "c"][(i % 3) as usize].to_string(),
+                )
+            })
+            .collect();
+        // Two segments, to also exercise cross-segment merging of the fused per-term histograms.
+        let segments = vec![docs[..150].to_vec(), docs[150..].to_vec()];
+        let index = get_test_index_from_values_and_terms(merge_segments, &segments)?;
+
+        let agg_req: Aggregations = serde_json::from_value(serde_json::json!({
+            "by_term": {
+                "terms": { "field": "string_id", "order": { "_key": "asc" } },
+                "aggs": {
+                    "histo": { "histogram": { "field": "score_f64", "interval": 1.0 } }
+                }
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+
+        for (term_idx, term) in ["a", "b", "c"].iter().enumerate() {
+            assert_eq!(res["by_term"]["buckets"][term_idx]["key"], *term);
+            assert_eq!(res["by_term"]["buckets"][term_idx]["doc_count"], 100);
+            let histo = &res["by_term"]["buckets"][term_idx]["histo"]["buckets"];
+            for b in 0..20usize {
+                assert_eq!(histo[b]["key"], b as f64, "term {term} bucket {b}");
+                assert_eq!(histo[b]["doc_count"], 5, "term {term} bucket {b}");
+            }
+            assert_eq!(histo[20], serde_json::Value::Null);
+        }
+        assert_eq!(res["by_term"]["buckets"][3], serde_json::Value::Null);
+
+        Ok(())
+    }
+
+    /// A `missing` config on a *full* term column still takes the fused path (the string sentinel
+    /// is just `col_max + 1`, so the column stays low-cardinality). Since no doc is missing, the
+    /// real term buckets must be exactly as without `missing`.
+    #[test]
+    fn fused_term_histogram_with_missing_on_full_column() -> crate::Result<()> {
+        let docs: Vec<(f64, String)> = (0..300u64)
+            .map(|i| {
+                (
+                    (i % 20) as f64,
+                    ["a", "b", "c"][(i % 3) as usize].to_string(),
+                )
+            })
+            .collect();
+        let index = get_test_index_from_values_and_terms(true, &[docs])?;
+
+        let agg_req: Aggregations = serde_json::from_value(serde_json::json!({
+            "by_term": {
+                "terms": { "field": "string_id", "missing": "MISSING", "order": { "_key": "asc" } },
+                "aggs": {
+                    "histo": { "histogram": { "field": "score_f64", "interval": 1.0 } }
+                }
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+
+        // Column is full, so "MISSING" never applies: a, b, c are unchanged (100 docs, 5 per
+        // bucket).
+        for (term_idx, term) in ["a", "b", "c"].iter().enumerate() {
+            assert_eq!(res["by_term"]["buckets"][term_idx]["key"], *term);
+            assert_eq!(res["by_term"]["buckets"][term_idx]["doc_count"], 100);
+            let histo = &res["by_term"]["buckets"][term_idx]["histo"]["buckets"];
+            for b in 0..20usize {
+                assert_eq!(histo[b]["doc_count"], 5, "term {term} bucket {b}");
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Term cardinality above the general path's `MAX_NUM_TERMS_FOR_VEC` (100) still fuses: the
+    /// flat grid is bounded by the total cell count (`num_terms * num_time_buckets`), not the
+    /// term count.
+    #[test]
+    fn fused_term_histogram_many_terms() -> crate::Result<()> {
+        let num_terms = 150usize;
+        let docs_per_term = 2usize;
+        // All docs share histogram value 0 (a single bucket), so the grid is 150 x 1 = 150 cells.
+        let docs: Vec<(f64, String)> = (0..num_terms * docs_per_term)
+            .map(|i| (0.0, format!("t{:03}", i % num_terms)))
+            .collect();
+        let index = get_test_index_from_values_and_terms(true, &[docs])?;
+
+        let agg_req: Aggregations = serde_json::from_value(serde_json::json!({
+            "by_term": {
+                "terms": { "field": "string_id", "size": 1000, "order": { "_key": "asc" } },
+                "aggs": {
+                    "histo": { "histogram": { "field": "score_f64", "interval": 1.0 } }
+                }
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+
+        let buckets = res["by_term"]["buckets"].as_array().unwrap();
+        assert_eq!(buckets.len(), num_terms);
+        for (i, bucket) in buckets.iter().enumerate() {
+            assert_eq!(bucket["key"], format!("t{i:03}"));
+            assert_eq!(bucket["doc_count"], docs_per_term as u64);
+            assert_eq!(bucket["histo"]["buckets"][0]["key"], 0.0);
+            assert_eq!(
+                bucket["histo"]["buckets"][0]["doc_count"],
+                docs_per_term as u64
+            );
+        }
+
+        Ok(())
+    }
+
+    /// `hard_bounds` exercises the non-derived `term_counts` branch: a term's `doc_count` must
+    /// count *every* doc with that term, including docs whose histogram value is outside the
+    /// bounds (those are excluded from the histogram buckets but still counted for the term). This
+    /// is the case where the per-doc `term_counts` increment cannot be replaced by the grid
+    /// row-sum.
+    #[test]
+    fn fused_term_histogram_with_hard_bounds() -> crate::Result<()> {
+        // 300 docs: term = {a, b, c} by i % 3, value = i % 20. Per term: 100 docs, each value in
+        // 0..=19 occurring 5 times.
+        let docs: Vec<(f64, String)> = (0..300u64)
+            .map(|i| {
+                (
+                    (i % 20) as f64,
+                    ["a", "b", "c"][(i % 3) as usize].to_string(),
+                )
+            })
+            .collect();
+        let index = get_test_index_from_values_and_terms(true, &[docs])?;
+
+        // hard_bounds [5, 14] (inclusive) keeps only values 5..=14 in the histogram (10 buckets);
+        // values 0..=4 and 15..=19 are out of bounds.
+        let agg_req: Aggregations = serde_json::from_value(serde_json::json!({
+            "by_term": {
+                "terms": { "field": "string_id", "order": { "_key": "asc" } },
+                "aggs": {
+                    "histo": {
+                        "histogram": {
+                            "field": "score_f64",
+                            "interval": 1.0,
+                            "hard_bounds": { "min": 5.0, "max": 14.0 }
+                        }
+                    }
+                }
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+
+        for (term_idx, term) in ["a", "b", "c"].iter().enumerate() {
+            assert_eq!(res["by_term"]["buckets"][term_idx]["key"], *term);
+            // doc_count includes the 50 per-term docs whose value is outside [5, 14].
+            assert_eq!(res["by_term"]["buckets"][term_idx]["doc_count"], 100);
+            let histo = &res["by_term"]["buckets"][term_idx]["histo"]["buckets"];
+            for b in 0..10usize {
+                let key = 5 + b;
+                assert_eq!(histo[b]["key"], key as f64, "term {term} bucket key {key}");
+                assert_eq!(histo[b]["doc_count"], 5, "term {term} bucket {key}");
+            }
+            // Only the 10 in-bounds buckets exist.
+            assert_eq!(histo[10], serde_json::Value::Null);
+        }
+
+        Ok(())
+    }
+
+    /// Non-binding `hard_bounds` (wider than the data, with mid-interval edges) must still produce
+    /// exact results via the derive-from-grid path: since no doc is out of bounds, normalization
+    /// drops the bound, every doc lands in the dense range, and each term's total equals its
+    /// histogram row-sum. This is the case that previously fell back to the per-doc counter only
+    /// because `bounds != [MIN, MAX]`.
+    #[test]
+    fn fused_term_histogram_with_non_binding_hard_bounds() -> crate::Result<()> {
+        // 300 docs: term = {a, b, c} by i % 3, value = i % 20. Data values span [0, 19].
+        let docs: Vec<(f64, String)> = (0..300u64)
+            .map(|i| {
+                (
+                    (i % 20) as f64,
+                    ["a", "b", "c"][(i % 3) as usize].to_string(),
+                )
+            })
+            .collect();
+        let index = get_test_index_from_values_and_terms(true, &[docs])?;
+
+        // Bounds wider than [0, 19], with mid-interval edges -> they exclude nothing.
+        let agg_req: Aggregations = serde_json::from_value(serde_json::json!({
+            "by_term": {
+                "terms": { "field": "string_id", "order": { "_key": "asc" } },
+                "aggs": {
+                    "histo": {
+                        "histogram": {
+                            "field": "score_f64",
+                            "interval": 1.0,
+                            "hard_bounds": { "min": -0.5, "max": 19.5 }
+                        }
+                    }
+                }
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+
+        for (term_idx, term) in ["a", "b", "c"].iter().enumerate() {
+            assert_eq!(res["by_term"]["buckets"][term_idx]["key"], *term);
+            // Every doc is in-bounds, so the per-term total is the full 100 (as without bounds).
+            assert_eq!(res["by_term"]["buckets"][term_idx]["doc_count"], 100);
+            let histo = &res["by_term"]["buckets"][term_idx]["histo"]["buckets"];
+            for b in 0..20usize {
+                assert_eq!(histo[b]["key"], b as f64, "term {term} bucket {b}");
+                assert_eq!(histo[b]["doc_count"], 5, "term {term} bucket {b}");
+            }
+            assert_eq!(histo[20], serde_json::Value::Null);
+        }
+
+        Ok(())
+    }
+
+    /// Regression: with hard bounds the fused path allocates `term_counts` (one `u32`/term) on top
+    /// of the grid, and that allocation must be charged to the memory limit. With many terms and a
+    /// single time bucket the two are equal in size, so a limit admitting the grid alone but not
+    /// grid + `term_counts` must fail.
+    #[test]
+    fn fused_term_histogram_hard_bounds_charges_term_counts() -> crate::Result<()> {
+        // 16k distinct terms, one doc each; values alternate in/out of the single-bucket bounds
+        // [5, 5] so the bounds bind and `term_counts` is allocated. num_terms=16000,
+        // num_time_buckets=1 => `counts` and `term_counts` are ~64 KB each.
+        let docs: Vec<(f64, String)> = (0..16_000u64)
+            .map(|i| (if i % 2 == 0 { 5.0 } else { 10.0 }, format!("t{i:05}")))
+            .collect();
+        let index = get_test_index_from_values_and_terms(true, &[docs])?;
+
+        let agg_req: Aggregations = serde_json::from_value(serde_json::json!({
+            "by_term": {
+                "terms": { "field": "string_id" },
+                "aggs": {
+                    "histo": {
+                        "histogram": {
+                            "field": "score_f64",
+                            "interval": 1.0,
+                            "hard_bounds": { "min": 5.0, "max": 5.0 }
+                        }
+                    }
+                }
+            }
+        }))
+        .unwrap();
+
+        // ~96 KB admits the grid (~64 KB) but not grid + `term_counts` (~128 KB).
+        let err = exec_request_with_query_and_memory_limit(
+            agg_req,
+            &index,
+            None,
+            AggregationLimitsGuard::new(Some(96_000), None),
+        )
+        .unwrap_err();
+        assert!(
+            err.to_string().contains("memory limit was exceeded"),
+            "expected a memory-limit error, got: {err}"
+        );
+
+        Ok(())
+    }
+}

src/aggregation/buffered_sub_aggs.rs

@@ -138,6 +138,7 @@ impl SubAggBuffer for HighCardSubAggBuffer {
         }
     }
 
+    #[inline]
     fn push(&mut self, bucket_id: BucketId, doc_id: DocId) {
         let idx = bucket_id % NUM_PARTITIONS as u32;
         let slot = &mut self.partitions[idx as usize];
@@ -196,6 +197,7 @@ impl SubAggBuffer for LowCardSubAggBuffer {
         }
     }
 
+    #[inline]
     fn push(&mut self, bucket_id: BucketId, doc_id: DocId) {
         let idx = bucket_id as usize;
         if self.per_bucket_docs.len() <= idx {

src/aggregation/intermediate_agg_result.rs

@@ -377,7 +377,22 @@ impl IntermediateMetricResult {
                 MetricResult::ExtendedStats(intermediate_stats.finalize())
             }
             IntermediateMetricResult::Sum(intermediate_sum) => {
-                MetricResult::Sum(intermediate_sum.finalize().into())
+                // By default match Elasticsearch: empty / all-missing sum
+                // buckets serialize as `"value": 0`, not `"value": null`.
+                // The non-ES `none_if_no_match` flag on `SumAggregation`
+                // opts into SQL-style `null` for downstream consumers.
+                let none_if_no_match = req
+                    .agg
+                    .as_sum()
+                    .and_then(|sum| sum.none_if_no_match)
+                    .unwrap_or(false);
+                let value = intermediate_sum.finalize();
+                if none_if_no_match {
+                    MetricResult::Sum(value.into())
+                } else {
+                    let value = Some(value.unwrap_or(0.0));
+                    MetricResult::Sum(value.into())
+                }
             }
             IntermediateMetricResult::Percentiles(percentiles) => MetricResult::Percentiles(
                 percentiles

src/aggregation/metric/cardinality.rs

@@ -166,8 +166,13 @@ impl CouponCache {
         let should_use_dense =
             highest_term_ord < 1_000_000u64 || highest_term_ord < num_terms as u64 * 3u64;
         if should_use_dense {
-            let mut coupon_map: Vec<Coupon> = vec![Coupon::EMPTY; highest_term_ord as usize + 1];
-            for (term_ord, coupon) in term_ords.into_iter().zip(coupons.into_iter()) {
+            // We don't really care about the value here. We will populate all the values we will
+            // read anyway.
+            let uninitialized_coupon = Coupon::from_hash(0);
+            let mut coupon_map: Vec<Coupon> =
+                vec![uninitialized_coupon; highest_term_ord as usize + 1];
+
+            for (term_ord, coupon) in term_ords.into_iter().zip(coupons) {
                 coupon_map[term_ord as usize] = coupon;
             }
             CouponCache::Dense {
@@ -821,7 +826,7 @@ impl<'de> Deserialize<'de> for CardinalityCollector {
 impl CardinalityCollector {
     fn new(salt: u8) -> Self {
         Self {
-            sketch: HllSketch::new(LG_K, HllType::Hll4),
+            sketch: HllSketch::new(LG_K, HllType::Hll8),
             salt,
         }
     }
@@ -852,7 +857,7 @@ impl CardinalityCollector {
         let mut union = HllUnion::new(LG_K);
         union.update(&self.sketch);
         union.update(&right.sketch);
-        self.sketch = union.to_sketch(HllType::Hll4);
+        self.sketch = union.to_sketch(HllType::Hll8);
         Ok(())
     }
 }

src/aggregation/metric/sum.rs

@@ -27,6 +27,16 @@ pub struct SumAggregation {
     /// { "field": "my_numbers", "missing": "10.0" }
     #[serde(default, deserialize_with = "deserialize_option_f64")]
     pub missing: Option<f64>,
+    /// Non-Elasticsearch extension. When `Some(true)`, the serialized result
+    /// returns `"value": null` if no values were collected (all documents had
+    /// missing/NULL values for the field), matching the behavior of `min`,
+    /// `max`, and `avg`. When `None` or `Some(false)` (the default) the
+    /// result returns `"value": 0`, matching Elasticsearch.
+    ///
+    /// Intended for SQL-style consumers where `SUM` of zero rows is `NULL`
+    /// and must be distinguishable from a bucket that genuinely sums to `0`.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub none_if_no_match: Option<bool>,
 }
 
 impl SumAggregation {
@@ -35,6 +45,7 @@ impl SumAggregation {
         Self {
             field: field_name,
             missing: None,
+            none_if_no_match: None,
         }
     }
     /// Returns the field name the aggregation is computed on.
@@ -59,8 +70,104 @@ impl IntermediateSum {
     pub fn merge_fruits(&mut self, other: IntermediateSum) {
         self.stats.merge_fruits(other.stats);
     }
-    /// Computes the final minimum value.
+    /// Computes the final sum value.
+    ///
+    /// Returns `None` when no values were collected, matching the Rust-side
+    /// behavior of `IntermediateMin`, `IntermediateMax`, and
+    /// `IntermediateAvg`. The Elasticsearch-vs-SQL choice for the
+    /// user-visible result is made at the boundary in
+    /// [`IntermediateMetricResult::into_final_metric_result`]: by default
+    /// `None` is coerced to `Some(0.0)` to match Elasticsearch
+    /// (`"value": 0`), and the [`SumAggregation::none_if_no_match`] flag
+    /// opts out of that coercion for SQL-style consumers.
     pub fn finalize(&self) -> Option<f64> {
-        Some(self.stats.finalize().sum)
+        let stats = self.stats.finalize();
+        if stats.count == 0 {
+            None
+        } else {
+            Some(stats.sum)
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_sum_finalize_returns_none_when_no_values() {
+        // Default IntermediateSum has count=0 — finalize should return None,
+        // matching MIN/MAX/AVG behavior for all-NULL groups.
+        let sum = IntermediateSum::default();
+        assert_eq!(sum.finalize(), None);
+    }
+
+    #[test]
+    fn test_sum_finalize_returns_value_when_has_values() {
+        let mut sum = IntermediateSum::default();
+        // Merge in a result that has actual values
+        let stats = IntermediateStats {
+            count: 3,
+            sum: 42.0,
+            min: 10.0,
+            max: 20.0,
+            ..Default::default()
+        };
+        let other = IntermediateSum::from_stats(stats);
+        sum.merge_fruits(other);
+        assert_eq!(sum.finalize(), Some(42.0));
+    }
+
+    #[test]
+    fn test_sum_merge_two_empty_still_none() {
+        let mut a = IntermediateSum::default();
+        let b = IntermediateSum::default();
+        a.merge_fruits(b);
+        assert_eq!(a.finalize(), None);
+    }
+
+    #[test]
+    fn test_sum_aggregation_empty_index_default_matches_es() -> crate::Result<()> {
+        use serde_json::json;
+
+        use crate::aggregation::agg_req::Aggregations;
+        use crate::aggregation::tests::{exec_request, get_test_index_from_terms};
+
+        // Empty index — sum has no values to collect.
+        let values: Vec<Vec<&str>> = vec![];
+        let index = get_test_index_from_terms(false, &values)?;
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "score_sum": { "sum": { "field": "score" } }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        // Default: match Elasticsearch — empty sum serializes as 0, not null.
+        assert_eq!(res["score_sum"]["value"], 0.0);
+        Ok(())
+    }
+
+    #[test]
+    fn test_sum_aggregation_empty_index_none_if_no_match_opt_in() -> crate::Result<()> {
+        use serde_json::json;
+
+        use crate::aggregation::agg_req::Aggregations;
+        use crate::aggregation::tests::{exec_request, get_test_index_from_terms};
+
+        let values: Vec<Vec<&str>> = vec![];
+        let index = get_test_index_from_terms(false, &values)?;
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "score_sum": { "sum": { "field": "score", "none_if_no_match": true } }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        // Opt-in non-ES extension — empty sum serializes as null.
+        assert!(
+            res["score_sum"]["value"].is_null(),
+            "expected null, got {:?}",
+            res["score_sum"]["value"]
+        );
+        Ok(())
     }
 }

src/collector/sort_key/sort_by_static_fast_value.rs

@@ -52,7 +52,7 @@ impl<T: FastValue> SortKeyComputer for SortByStaticFastValue<T> {
         if schema_type != T::to_type() {
             return Err(crate::TantivyError::SchemaError(format!(
                 "Field `{}` is of type {schema_type:?}, not of the type {:?}.",
-                &self.field,
+                self.field,
                 T::to_type()
             )));
         }

src/index/segment_reader.rs

@@ -6,6 +6,7 @@ use common::{ByteCount, HasLen};
 use fnv::FnvHashMap;
 use itertools::Itertools;
 
+use crate::directory::error::OpenReadError;
 use crate::directory::{CompositeFile, FileSlice};
 use crate::error::DataCorruption;
 use crate::fastfield::{intersect_alive_bitsets, AliveBitSet, FacetReader, FastFieldReaders};
@@ -159,12 +160,10 @@ impl SegmentReader {
         let postings_file = segment.open_read(SegmentComponent::Postings)?;
         let postings_composite = CompositeFile::open(&postings_file)?;
 
-        let positions_composite = {
-            if let Ok(positions_file) = segment.open_read(SegmentComponent::Positions) {
-                CompositeFile::open(&positions_file)?
-            } else {
-                CompositeFile::empty()
-            }
+        let positions_composite = match segment.open_read(SegmentComponent::Positions) {
+            Ok(positions_file) => CompositeFile::open(&positions_file)?,
+            Err(OpenReadError::FileDoesNotExist(_)) => CompositeFile::empty(),
+            Err(open_read_error) => return Err(open_read_error.into()),
         };
 
         let schema = segment.schema();
@@ -323,7 +322,7 @@ impl SegmentReader {
                             // Without expand dots enabled dots need to be escaped.
                             let escaped_json_path = json_path.replace('.', "\\.");
                             let full_path = format!("{field_name}.{escaped_json_path}");
-                            let full_path_unescaped = format!("{}.{}", field_name, &json_path);
+                            let full_path_unescaped = format!("{}.{}", field_name, json_path);
                             map_to_canonical.insert(full_path_unescaped, full_path.to_string());
                             full_path
                         } else {

src/postings/block_segment_postings.rs

@@ -287,6 +287,33 @@ impl BlockSegmentPostings {
         doc
     }
 
+    /// Returns the number of documents with a doc id strictly smaller than `target`
+    /// (i.e. the *rank* of `target` in this posting list).
+    ///
+    /// This jumps to the block that may contain `target` through the skip list, so no
+    /// skipped block is decoded; a single block is then decoded to locate `target`
+    /// within it. The cost is therefore `O(number_of_skip_list_entries)` plus one block
+    /// decode, rather than `O(doc_freq)`.
+    ///
+    /// Like [`Self::seek`], the underlying cursor only ever moves forward. This method
+    /// must be called with **non-decreasing** `target` values (galloping); calling it
+    /// with a `target` smaller than a previous one yields an incorrect result. `target`
+    /// must be a valid doc id (i.e. `target <= TERMINATED`), exactly as for `seek`.
+    ///
+    /// Edge cases: returns `0` when `target` is smaller than every doc id, and
+    /// `doc_freq()` when `target` is larger than every doc id.
+    pub fn rank(&mut self, target: DocId) -> u32 {
+        if self.doc_freq == 0 {
+            return 0;
+        }
+        // `within` = number of docs in the landed block with a doc id < target.
+        let within = self.seek(target);
+        // `remaining_docs` counts the landed block and everything after it, so the
+        // difference is the number of docs in all blocks strictly before it.
+        let docs_before_block = self.doc_freq - self.skip_reader.remaining_docs();
+        docs_before_block + within as u32
+    }
+
     pub(crate) fn position_offset(&self) -> u64 {
         self.skip_reader.position_offset()
     }
@@ -568,4 +595,38 @@ mod tests {
         assert_eq!(block_segments.docs(), &[1, 3, 5]);
         Ok(())
     }
+
+    #[test]
+    fn test_block_segment_postings_rank() -> crate::Result<()> {
+        // ~8 blocks worth of docs so the skip list is actually exercised.
+        let docs: Vec<DocId> = (0..1000u32).map(|i| i * 3).collect();
+        let mut block_postings = build_block_postings(&docs[..])?;
+        let doc_freq = block_postings.doc_freq();
+
+        // rank(target) must equal the number of docs strictly below target.
+        // Targets are queried in non-decreasing order, as the API requires.
+        // `target` values must be a valid doc id (<= TERMINATED) and non-decreasing.
+        let targets = [
+            0u32, 1, 2, 3, 4, 299, 300, 301, 1500, 2996, 2997, 3000, 10_000,
+        ];
+        for &target in &targets {
+            let expected = docs.iter().filter(|&&d| d < target).count() as u32;
+            assert_eq!(
+                block_postings.rank(target),
+                expected,
+                "rank({target}) mismatch"
+            );
+        }
+
+        // Edge cases: below the first doc -> 0, above the last doc -> doc_freq.
+        let mut fresh = build_block_postings(&docs[..])?;
+        assert_eq!(fresh.rank(0), 0);
+        let mut fresh = build_block_postings(&docs[..])?;
+        assert_eq!(fresh.rank(1_000_000), doc_freq);
+
+        // Empty postings: rank is always 0.
+        let mut empty = BlockSegmentPostings::empty();
+        assert_eq!(empty.rank(42), 0);
+        Ok(())
+    }
 }

src/postings/recorder.rs

@@ -275,8 +275,9 @@ impl Recorder for TfAndPositionRecorder {
 mod tests {
 
     use common::write_u32_vint;
+    use stacker::MemoryArena;
 
-    use super::{BufferLender, VInt32Reader};
+    use super::{BufferLender, Recorder, TermFrequencyRecorder, VInt32Reader};
 
     #[test]
     fn test_buffer_lender() {
@@ -314,4 +315,98 @@ mod tests {
         let res: Vec<u32> = VInt32Reader::new(&buffer[..]).collect();
         assert_eq!(&res[..], &vals[..]);
     }
+
+    // ── TermFrequencyRecorder ─────────────────────────────────────────────────
+
+    #[test]
+    fn term_frequency_recorder_has_term_freq() {
+        let rec = TermFrequencyRecorder::default();
+        assert!(
+            rec.has_term_freq(),
+            "TermFrequencyRecorder must advertise term-frequency support"
+        );
+    }
+
+    #[test]
+    fn term_frequency_recorder_term_doc_freq_single_doc() {
+        let mut arena = MemoryArena::default();
+        let mut rec = TermFrequencyRecorder::default();
+
+        // Record one document with two term occurrences.
+        rec.new_doc(0, &mut arena);
+        rec.record_position(0, &mut arena);
+        rec.record_position(1, &mut arena);
+        rec.close_doc(&mut arena);
+
+        assert_eq!(
+            rec.term_doc_freq(),
+            Some(1),
+            "term_doc_freq should be 1 after recording one document"
+        );
+    }
+
+    #[test]
+    fn term_frequency_recorder_term_doc_freq_multiple_docs() {
+        let mut arena = MemoryArena::default();
+        let mut rec = TermFrequencyRecorder::default();
+
+        // Three documents with 1, 3, and 2 occurrences respectively.
+        for (doc, tf) in [(0u32, 1u32), (5, 3), (10, 2)] {
+            rec.new_doc(doc, &mut arena);
+            for pos in 0..tf {
+                rec.record_position(pos, &mut arena);
+            }
+            rec.close_doc(&mut arena);
+        }
+
+        assert_eq!(
+            rec.term_doc_freq(),
+            Some(3),
+            "term_doc_freq should equal the number of documents recorded"
+        );
+    }
+
+    #[test]
+    fn term_frequency_recorder_zero_docs() {
+        let rec = TermFrequencyRecorder::default();
+        assert_eq!(
+            rec.term_doc_freq(),
+            Some(0),
+            "term_doc_freq should be 0 before any document is recorded"
+        );
+    }
+
+    #[test]
+    fn term_frequency_recorder_single_occurrence_per_doc() {
+        let mut arena = MemoryArena::default();
+        let mut rec = TermFrequencyRecorder::default();
+
+        // Each document has exactly one occurrence — the minimum non-trivial case.
+        for doc in [1u32, 2, 100] {
+            rec.new_doc(doc, &mut arena);
+            rec.record_position(0, &mut arena);
+            rec.close_doc(&mut arena);
+        }
+
+        assert_eq!(rec.term_doc_freq(), Some(3));
+    }
+
+    #[test]
+    fn term_frequency_recorder_high_frequency_doc() {
+        let mut arena = MemoryArena::default();
+        let mut rec = TermFrequencyRecorder::default();
+
+        // A document where the term appears many times.
+        rec.new_doc(42, &mut arena);
+        for pos in 0..1000 {
+            rec.record_position(pos, &mut arena);
+        }
+        rec.close_doc(&mut arena);
+
+        assert_eq!(
+            rec.term_doc_freq(),
+            Some(1),
+            "term_doc_freq counts documents, not occurrences"
+        );
+    }
 }

src/postings/skip.rs

@@ -187,6 +187,12 @@ impl SkipReader {
         self.last_doc_in_block
     }
 
+    /// Number of docs from the start of the current block to the end of the postings
+    /// (i.e. the current block plus every block after it).
+    pub(crate) fn remaining_docs(&self) -> u32 {
+        self.remaining_docs
+    }
+
     pub fn position_offset(&self) -> u64 {
         self.position_offset
     }

src/query/const_score_query.rs

@@ -1,6 +1,6 @@
 use std::fmt;
 
-use crate::docset::COLLECT_BLOCK_BUFFER_LEN;
+use crate::docset::{SeekDangerResult, COLLECT_BLOCK_BUFFER_LEN};
 use crate::query::{EnableScoring, Explanation, Query, Scorer, Weight};
 use crate::{DocId, DocSet, Score, SegmentReader, TantivyError, Term};
 
@@ -119,6 +119,10 @@ impl<TDocSet: DocSet> DocSet for ConstScorer<TDocSet> {
         self.docset.seek(target)
     }
 
+    fn seek_danger(&mut self, target: DocId) -> SeekDangerResult {
+        self.docset.seek_danger(target)
+    }
+
     fn fill_buffer(&mut self, buffer: &mut [DocId; COLLECT_BLOCK_BUFFER_LEN]) -> usize {
         self.docset.fill_buffer(buffer)
     }

src/query/range_query/fast_field_range_doc_set.rs

@@ -3,6 +3,7 @@ use std::ops::RangeInclusive;
 
 use columnar::Column;
 
+use crate::docset::SeekDangerResult;
 use crate::{DocId, DocSet, TERMINATED};
 
 /// Helper to have a cursor over a vec of docids
@@ -184,6 +185,37 @@ impl<T: Send + Sync + PartialOrd + Copy + Debug + 'static> DocSet for RangeDocSe
         doc
     }
 
+    /// `seek_danger` only needs to answer whether `target` itself matches, so it does a cheap
+    /// point lookup on the column instead of scanning forward to materialize the next match (the
+    /// expensive part of a regular `seek`).
+    fn seek_danger(&mut self, target: DocId) -> SeekDangerResult {
+        // Covers `target == TERMINATED` and any target past the last doc: no match is possible.
+        if target >= self.column.num_docs() {
+            return SeekDangerResult::SeekLowerBound(TERMINATED);
+        }
+
+        if self.is_last_seek_distance_large(target) {
+            self.reset_fetch_range();
+        }
+        self.last_seek_pos_opt = Some(target);
+
+        let is_match = self
+            .column
+            .values_for_doc(target)
+            .any(|value| self.value_range.contains(&value));
+        if is_match {
+            // Leave the docset in a valid state positioned on `target`, so `doc()` returns it and a
+            // following `advance()` resumes the scan right after it.
+            self.loaded_docs.get_cleared_data().push(target);
+            self.next_fetch_start = target + 1;
+            SeekDangerResult::Found
+        } else {
+            // `target` is not in the docset. The next match is strictly greater than `target`, so
+            // `target + 1` is a valid lower bound. We may leave the docset in an invalid state.
+            SeekDangerResult::SeekLowerBound(target + 1)
+        }
+    }
+
     fn size_hint(&self) -> u32 {
         // TODO: Implement a better size hint
         self.column.num_docs() / 10
@@ -209,12 +241,148 @@ impl<T: Send + Sync + PartialOrd + Copy + Debug + 'static> DocSet for RangeDocSe
 
 #[cfg(test)]
 mod tests {
-    use std::ops::Bound;
+    use std::ops::{Bound, RangeInclusive};
 
+    use columnar::Column;
+
+    use super::RangeDocSet;
     use crate::collector::Count;
     use crate::directory::RamDirectory;
+    use crate::docset::{SeekDangerResult, TERMINATED};
     use crate::query::RangeQuery;
-    use crate::{schema, IndexBuilder, TantivyDocument, Term};
+    use crate::{schema, DocSet, Index, IndexBuilder, TantivyDocument, Term};
+
+    /// Builds a single-segment index where doc `i` carries `values_for_doc(i)` in a u64 fast
+    /// field, then returns its column so we can drive a `RangeDocSet` directly.
+    fn build_u64_column(
+        num_docs: usize,
+        values_for_doc: impl Fn(usize) -> Vec<u64>,
+    ) -> Column<u64> {
+        let mut schema_builder = schema::SchemaBuilder::new();
+        let value_field = schema_builder.add_u64_field("value", schema::FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        {
+            let mut writer = index.writer_for_tests().unwrap();
+            for i in 0..num_docs {
+                let mut doc = TantivyDocument::new();
+                for v in values_for_doc(i) {
+                    doc.add_u64(value_field, v);
+                }
+                writer.add_document(doc).unwrap();
+            }
+            writer.commit().unwrap();
+        }
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(searcher.segment_readers().len(), 1);
+        searcher
+            .segment_reader(0)
+            .fast_fields()
+            .u64("value")
+            .unwrap()
+    }
+
+    fn range_docset(
+        value_range: RangeInclusive<u64>,
+        num_docs: usize,
+        values_for_doc: impl Fn(usize) -> Vec<u64>,
+    ) -> RangeDocSet<u64> {
+        RangeDocSet::new(value_range, build_u64_column(num_docs, values_for_doc))
+    }
+
+    #[test]
+    fn seek_danger_found_leaves_valid_state() {
+        // Even docs match the range, odd docs do not.
+        let mut docset = range_docset(0..=0, 100, |i| vec![(i % 2) as u64]);
+
+        // Matching target: `Found`, and the docset is positioned exactly on it.
+        assert_eq!(docset.seek_danger(10), SeekDangerResult::Found);
+        assert_eq!(docset.doc(), 10);
+        // A following advance resumes the scan right after the found doc.
+        assert_eq!(docset.advance(), 12);
+        assert_eq!(docset.doc(), 12);
+    }
+
+    #[test]
+    fn seek_danger_miss_returns_lower_bound() {
+        let mut docset = range_docset(0..=0, 100, |i| vec![(i % 2) as u64]);
+
+        // Odd target does not match: lower bound is strictly greater than the target and never
+        // skips past the next real match (here doc 12, the first even doc after 11).
+        match docset.seek_danger(11) {
+            SeekDangerResult::SeekLowerBound(lower_bound) => {
+                assert!(lower_bound > 11);
+                assert!(lower_bound <= 12);
+            }
+            SeekDangerResult::Found => panic!("11 should not match"),
+        }
+        // After a miss we may be in an invalid state; another seek_danger recovers it.
+        assert_eq!(docset.seek_danger(12), SeekDangerResult::Found);
+        assert_eq!(docset.doc(), 12);
+    }
+
+    #[test]
+    fn seek_danger_terminated_and_out_of_bounds() {
+        let mut docset = range_docset(0..=0, 10, |i| vec![(i % 2) as u64]);
+        assert_eq!(
+            docset.seek_danger(TERMINATED),
+            SeekDangerResult::SeekLowerBound(TERMINATED)
+        );
+        // A target past the last doc has no possible match either.
+        assert_eq!(
+            docset.seek_danger(10),
+            SeekDangerResult::SeekLowerBound(TERMINATED)
+        );
+    }
+
+    #[test]
+    fn seek_danger_multivalued() {
+        // Doc `i` holds values [i, i+1]; the range {5} matches docs 4 and 5.
+        let mut docset = range_docset(5..=5, 20, |i| vec![i as u64, i as u64 + 1]);
+
+        assert_eq!(docset.seek_danger(4), SeekDangerResult::Found);
+        assert_eq!(docset.doc(), 4);
+        assert_eq!(docset.advance(), 5);
+        // No further match after doc 5.
+        assert_eq!(docset.advance(), TERMINATED);
+    }
+
+    #[test]
+    fn seek_danger_matches_seek() {
+        // Cross-check seek_danger against the true next match for every target, on a column with a
+        // few sparse matches.
+        let matches = [3u32, 7, 50, 51, 99];
+        let num_docs = 100;
+        let values_for_doc = |i: usize| {
+            vec![if matches.contains(&(i as u32)) {
+                1u64
+            } else {
+                0u64
+            }]
+        };
+
+        for target in 0..num_docs as u32 {
+            // The first matching doc greater than or equal to `target`, i.e. what `seek` returns.
+            let expected = matches
+                .iter()
+                .copied()
+                .find(|&m| m >= target)
+                .unwrap_or(TERMINATED);
+
+            let mut danger = range_docset(1..=1, num_docs, values_for_doc);
+            match danger.seek_danger(target) {
+                SeekDangerResult::Found => {
+                    assert_eq!(expected, target, "target {target} reported Found");
+                    assert_eq!(danger.doc(), target);
+                }
+                SeekDangerResult::SeekLowerBound(lower_bound) => {
+                    assert_ne!(expected, target, "target {target} should have been Found");
+                    assert!(lower_bound > target);
+                    // The lower bound must never skip past the true next match.
+                    assert!(lower_bound <= expected);
+                }
+            }
+        }
+    }
 
     #[test]
     fn range_query_fast_optional_field_minimum() {

src/store/index/skip_index.rs

@@ -94,13 +94,7 @@ impl SkipIndex {
             byte_range: 0..first_layer_len,
         };
         for layer in &self.layers {
-            if let Some(checkpoint) =
-                layer.seek_start_at_offset(target, cur_checkpoint.byte_range.start)
-            {
-                cur_checkpoint = checkpoint;
-            } else {
-                return None;
-            }
+            cur_checkpoint = layer.seek_start_at_offset(target, cur_checkpoint.byte_range.start)?;
         }
         Some(cur_checkpoint)
     }

sstable/src/dictionary.rs

@@ -14,11 +14,8 @@ use itertools::Itertools;
 use tantivy_fst::Automaton;
 use tantivy_fst::automaton::AlwaysMatch;
 
-use crate::sstable_index_v3::SSTableIndexV3Empty;
 use crate::streamer::{Streamer, StreamerBuilder};
-use crate::{
-    BlockAddr, DeltaReader, Reader, SSTable, SSTableIndex, SSTableIndexV3, TermOrdinal, VoidSSTable,
-};
+use crate::{BlockAddr, DeltaReader, Reader, SSTable, SSTableIndex, TermOrdinal, VoidSSTable};
 
 /// An SSTable is a sorted map that associates sorted `&[u8]` keys
 /// to any kind of typed values.
@@ -288,33 +285,7 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
         let (sstable_slice, index_slice) = main_slice.split(index_offset as usize);
         let sstable_index_bytes = index_slice.read_bytes()?;
 
-        let sstable_index = match version {
-            2 => SSTableIndex::V2(
-                crate::sstable_index_v2::SSTableIndex::load(sstable_index_bytes).map_err(|_| {
-                    io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption")
-                })?,
-            ),
-            3 => {
-                let (sstable_index_bytes, mut footerv3_len_bytes) = sstable_index_bytes.rsplit(8);
-                let store_offset = u64::deserialize(&mut footerv3_len_bytes)?;
-                if store_offset != 0 {
-                    SSTableIndex::V3(
-                        SSTableIndexV3::load(sstable_index_bytes, store_offset).map_err(|_| {
-                            io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption")
-                        })?,
-                    )
-                } else {
-                    // if store_offset is zero, there is no index, so we build a pseudo-index
-                    // assuming a single block of sstable covering everything.
-                    SSTableIndex::V3Empty(SSTableIndexV3Empty::load(index_offset as usize))
-                }
-            }
-            _ => {
-                return Err(io::Error::other(format!(
-                    "Unsupported sstable version, expected one of [2, 3], found {version}"
-                )));
-            }
-        };
+        let sstable_index = SSTableIndex::open(version, index_offset, sstable_index_bytes)?;
 
         Ok(Dictionary {
             sstable_slice,
@@ -525,10 +496,15 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
 
         // Open the block for the first ordinal.
         let mut bytes = Vec::new();
-        let mut current_block_addr = self.sstable_index.get_block_with_ord(ord);
+        let (mut current_block_addr, block_id) = self.sstable_index.get_and_locate_with_ord(ord);
         let mut current_sstable_delta_reader =
             self.sstable_delta_reader_block(current_block_addr.clone())?;
         let mut current_block_ordinal = current_block_addr.first_ordinal;
+        let mut current_block_end_bound = self
+            .sstable_index
+            .get_block(block_id + 1)
+            .map(|block_addr| block_addr.first_ordinal)
+            .unwrap_or(u64::MAX);
 
         loop {
             // move to the ord inside the current block
@@ -557,17 +533,19 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
                 }
             };
 
-            // TODO optimization: it is silly to do a binary search to get the block every single
-            // time.
-            //
-            // Check if block changed for new term_ord
-            let new_block_addr = self.sstable_index.get_block_with_ord(next_ord);
-            if new_block_addr != current_block_addr {
+            if next_ord >= current_block_end_bound {
+                let (new_block_addr, block_id) =
+                    self.sstable_index.get_and_locate_with_ord(next_ord);
                 current_block_addr = new_block_addr;
                 current_block_ordinal = current_block_addr.first_ordinal;
                 current_sstable_delta_reader =
                     self.sstable_delta_reader_block(current_block_addr.clone())?;
                 bytes.clear();
+                current_block_end_bound = self
+                    .sstable_index
+                    .get_block(block_id + 1)
+                    .map(|block_addr| block_addr.first_ordinal)
+                    .unwrap_or(u64::MAX)
             }
             ord = next_ord;
         }

sstable/src/index/mod.rs

@@ -0,0 +1,319 @@
+pub(crate) mod v2;
+pub(crate) mod v3;
+
+use std::io::{self, Read, Write};
+use std::ops::Range;
+
+use common::{BinarySerializable, FixedSize, OwnedBytes};
+use tantivy_fst::{Automaton, MapBuilder};
+
+use crate::{TermOrdinal, common_prefix_len};
+
+#[derive(Debug, Clone)]
+pub enum SSTableIndex {
+    V2(v2::SSTableIndex),
+    V3(v3::SSTableIndexV3),
+    V3Empty(v3::SSTableIndexV3Empty),
+}
+
+impl SSTableIndex {
+    pub(crate) fn open(
+        version: u32,
+        index_offset: u64,
+        index_bytes: OwnedBytes,
+    ) -> io::Result<Self> {
+        let index = match version {
+            2 => {
+                SSTableIndex::V2(v2::SSTableIndex::load(index_bytes).map_err(|_| {
+                    io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption")
+                })?)
+            }
+            3 => {
+                let (index_bytes, mut footerv3_len_bytes) = index_bytes.rsplit(8);
+                let store_offset = u64::deserialize(&mut footerv3_len_bytes)?;
+                if store_offset != 0 {
+                    SSTableIndex::V3(v3::SSTableIndexV3::load(index_bytes, store_offset).map_err(
+                        |_| io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption"),
+                    )?)
+                } else {
+                    // if store_offset is zero, there is no index, so we build a pseudo-index
+                    // assuming a single block of sstable covering everything.
+                    SSTableIndex::V3Empty(v3::SSTableIndexV3Empty::load(index_offset as usize))
+                }
+            }
+            _ => {
+                return Err(io::Error::other(format!(
+                    "Unsupported sstable version, expected one of [2, 3], found {version}"
+                )));
+            }
+        };
+        Ok(index)
+    }
+
+    /// Get the [`BlockAddr`] of the requested block.
+    pub(crate) fn get_block(&self, block_id: u64) -> Option<BlockAddr> {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.get_block(block_id as usize),
+            SSTableIndex::V3(v3_index) => v3_index.get_block(block_id),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block(block_id),
+        }
+    }
+
+    /// Get the block id of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub(crate) fn locate_with_key(&self, key: &[u8]) -> Option<u64> {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.locate_with_key(key).map(|i| i as u64),
+            SSTableIndex::V3(v3_index) => v3_index.locate_with_key(key),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.locate_with_key(key),
+        }
+    }
+
+    /// Get the [`BlockAddr`] of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub fn get_block_with_key(&self, key: &[u8]) -> Option<BlockAddr> {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.get_block_with_key(key),
+            SSTableIndex::V3(v3_index) => v3_index.get_block_with_key(key),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block_with_key(key),
+        }
+    }
+
+    pub(crate) fn locate_with_ord(&self, ord: TermOrdinal) -> u64 {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.locate_with_ord(ord) as u64,
+            SSTableIndex::V3(v3_index) => v3_index.locate_with_ord(ord),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.locate_with_ord(ord),
+        }
+    }
+
+    /// Get the [`BlockAddr`] of the block containing the `ord`-th term.
+    pub(crate) fn get_block_with_ord(&self, ord: TermOrdinal) -> BlockAddr {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.get_block_with_ord(ord),
+            SSTableIndex::V3(v3_index) => v3_index.get_block_with_ord(ord),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block_with_ord(ord),
+        }
+    }
+
+    pub(crate) fn get_and_locate_with_ord(&self, ord: TermOrdinal) -> (BlockAddr, u64) {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.get_and_locate_with_ord(ord),
+            SSTableIndex::V3(v3_index) => v3_index.get_and_locate_with_ord(ord),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_and_locate_with_ord(ord),
+        }
+    }
+
+    pub fn get_block_for_automaton<'a>(
+        &'a self,
+        automaton: &'a impl Automaton,
+    ) -> impl Iterator<Item = (u64, BlockAddr)> + 'a {
+        match self {
+            SSTableIndex::V2(v2_index) => {
+                BlockIter::V2(v2_index.get_block_for_automaton(automaton))
+            }
+            SSTableIndex::V3(v3_index) => {
+                BlockIter::V3(v3_index.get_block_for_automaton(automaton))
+            }
+            SSTableIndex::V3Empty(v3_empty) => {
+                BlockIter::V3Empty(std::iter::once((0, v3_empty.block_addr.clone())))
+            }
+        }
+    }
+}
+
+enum BlockIter<V2, V3, T> {
+    V2(V2),
+    V3(V3),
+    V3Empty(std::iter::Once<T>),
+}
+
+impl<V2: Iterator<Item = T>, V3: Iterator<Item = T>, T> Iterator for BlockIter<V2, V3, T> {
+    type Item = T;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        match self {
+            BlockIter::V2(v2) => v2.next(),
+            BlockIter::V3(v3) => v3.next(),
+            BlockIter::V3Empty(once) => once.next(),
+        }
+    }
+}
+
+#[derive(Clone, Eq, PartialEq, Debug)]
+pub struct BlockAddr {
+    pub first_ordinal: u64,
+    pub byte_range: Range<usize>,
+}
+
+impl BlockAddr {
+    fn to_block_start(&self) -> BlockStartAddr {
+        BlockStartAddr {
+            first_ordinal: self.first_ordinal,
+            byte_range_start: self.byte_range.start,
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct BlockStartAddr {
+    first_ordinal: u64,
+    byte_range_start: usize,
+}
+
+impl BlockStartAddr {
+    fn to_block_addr(&self, byte_range_end: usize) -> BlockAddr {
+        BlockAddr {
+            first_ordinal: self.first_ordinal,
+            byte_range: self.byte_range_start..byte_range_end,
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub(crate) struct BlockMeta {
+    /// Any byte string that is lexicographically greater or equal to
+    /// the last key in the block,
+    /// and yet strictly smaller than the first key in the next block.
+    pub last_key_or_greater: Vec<u8>,
+    pub block_addr: BlockAddr,
+}
+
+impl BinarySerializable for BlockStartAddr {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        let start = self.byte_range_start as u64;
+        start.serialize(writer)?;
+        self.first_ordinal.serialize(writer)
+    }
+
+    fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
+        let byte_range_start = u64::deserialize(reader)? as usize;
+        let first_ordinal = u64::deserialize(reader)?;
+        Ok(BlockStartAddr {
+            first_ordinal,
+            byte_range_start,
+        })
+    }
+
+    // Provided method
+    fn num_bytes(&self) -> u64 {
+        BlockStartAddr::SIZE_IN_BYTES as u64
+    }
+}
+
+impl FixedSize for BlockStartAddr {
+    const SIZE_IN_BYTES: usize = 2 * u64::SIZE_IN_BYTES;
+}
+
+/// Given that left < right,
+/// mutates `left into a shorter byte string left'` that
+/// matches `left <= left' < right`.
+fn find_shorter_str_in_between(left: &mut Vec<u8>, right: &[u8]) {
+    assert!(&left[..] < right);
+    let common_len = common_prefix_len(left, right);
+    if left.len() == common_len {
+        return;
+    }
+    // It is possible to do one character shorter in some case,
+    // but it is not worth the extra complexity
+    for pos in (common_len + 1)..left.len() {
+        if left[pos] != u8::MAX {
+            left[pos] += 1;
+            left.truncate(pos + 1);
+            return;
+        }
+    }
+}
+
+#[derive(Default)]
+pub struct SSTableIndexBuilder {
+    blocks: Vec<BlockMeta>,
+}
+
+impl SSTableIndexBuilder {
+    /// In order to make the index as light as possible, we
+    /// try to find a shorter alternative to the last key of the last block
+    /// that is still smaller than the next key.
+    pub(crate) fn shorten_last_block_key_given_next_key(&mut self, next_key: &[u8]) {
+        if let Some(last_block) = self.blocks.last_mut() {
+            find_shorter_str_in_between(&mut last_block.last_key_or_greater, next_key);
+        }
+    }
+
+    pub fn add_block(&mut self, last_key: &[u8], byte_range: Range<usize>, first_ordinal: u64) {
+        self.blocks.push(BlockMeta {
+            last_key_or_greater: last_key.to_vec(),
+            block_addr: BlockAddr {
+                byte_range,
+                first_ordinal,
+            },
+        })
+    }
+
+    pub fn serialize<W: std::io::Write>(&self, wrt: W) -> io::Result<u64> {
+        if self.blocks.len() <= 1 {
+            return Ok(0);
+        }
+        let counting_writer = common::CountingWriter::wrap(wrt);
+        let mut map_builder = MapBuilder::new(counting_writer).map_err(fst_error_to_io_error)?;
+        for (i, block) in self.blocks.iter().enumerate() {
+            map_builder
+                .insert(&block.last_key_or_greater, i as u64)
+                .map_err(fst_error_to_io_error)?;
+        }
+        let counting_writer = map_builder.into_inner().map_err(fst_error_to_io_error)?;
+        let written_bytes = counting_writer.written_bytes();
+        let mut wrt = counting_writer.finish();
+
+        let mut block_store_writer = v3::BlockAddrStoreWriter::new();
+        for block in &self.blocks {
+            block_store_writer.write_block_meta(block.block_addr.clone())?;
+        }
+        block_store_writer.serialize(&mut wrt)?;
+
+        Ok(written_bytes)
+    }
+}
+
+fn fst_error_to_io_error(error: tantivy_fst::Error) -> io::Error {
+    match error {
+        tantivy_fst::Error::Fst(fst_error) => io::Error::other(fst_error),
+        tantivy_fst::Error::Io(ioerror) => ioerror,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #[track_caller]
+    fn test_find_shorter_str_in_between_aux(left: &[u8], right: &[u8]) {
+        let mut left_buf = left.to_vec();
+        super::find_shorter_str_in_between(&mut left_buf, right);
+        assert!(left_buf.len() <= left.len());
+        assert!(left <= &left_buf);
+        assert!(&left_buf[..] < right);
+    }
+
+    #[test]
+    fn test_find_shorter_str_in_between() {
+        test_find_shorter_str_in_between_aux(b"", b"hello");
+        test_find_shorter_str_in_between_aux(b"abc", b"abcd");
+        test_find_shorter_str_in_between_aux(b"abcd", b"abd");
+        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[1]);
+        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[0, 0, 1]);
+        test_find_shorter_str_in_between_aux(&[0, 0, 255, 255, 255, 0u8], &[0, 1]);
+    }
+
+    use proptest::prelude::*;
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(100))]
+        #[test]
+        fn test_proptest_find_shorter_str(left in any::<Vec<u8>>(), right in any::<Vec<u8>>()) {
+            if left < right {
+                test_find_shorter_str_in_between_aux(&left, &right);
+            }
+        }
+    }
+}

sstable/src/index/v2.rs

@@ -77,6 +77,13 @@ impl SSTableIndex {
         self.get_block(self.locate_with_ord(ord)).unwrap()
     }
 
+    pub(crate) fn get_and_locate_with_ord(&self, ord: TermOrdinal) -> (BlockAddr, u64) {
+        let location = self.locate_with_ord(ord);
+        // locate_with_ord always returns an index within range
+        let block_addr = self.get_block(location).unwrap();
+        (block_addr, location as u64)
+    }
+
     pub(crate) fn get_block_for_automaton<'a>(
         &'a self,
         automaton: &'a impl Automaton,

sstable/src/index/v3.rs

@@ -1,106 +1,14 @@
 use std::io::{self, Read, Write};
-use std::ops::Range;
 use std::sync::Arc;
 
 use common::{BinarySerializable, FixedSize, OwnedBytes};
 use tantivy_bitpacker::{BitPacker, compute_num_bits};
 use tantivy_fst::raw::Fst;
-use tantivy_fst::{Automaton, IntoStreamer, Map, MapBuilder, Streamer};
+use tantivy_fst::{Automaton, IntoStreamer, Map, Streamer};
 
+use super::{BlockAddr, BlockStartAddr};
 use crate::block_match_automaton::can_block_match_automaton;
-use crate::{SSTableDataCorruption, TermOrdinal, common_prefix_len};
-
-#[derive(Debug, Clone)]
-pub enum SSTableIndex {
-    V2(crate::sstable_index_v2::SSTableIndex),
-    V3(SSTableIndexV3),
-    V3Empty(SSTableIndexV3Empty),
-}
-
-impl SSTableIndex {
-    /// Get the [`BlockAddr`] of the requested block.
-    pub(crate) fn get_block(&self, block_id: u64) -> Option<BlockAddr> {
-        match self {
-            SSTableIndex::V2(v2_index) => v2_index.get_block(block_id as usize),
-            SSTableIndex::V3(v3_index) => v3_index.get_block(block_id),
-            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block(block_id),
-        }
-    }
-
-    /// Get the block id of the block that would contain `key`.
-    ///
-    /// Returns None if `key` is lexicographically after the last key recorded.
-    pub(crate) fn locate_with_key(&self, key: &[u8]) -> Option<u64> {
-        match self {
-            SSTableIndex::V2(v2_index) => v2_index.locate_with_key(key).map(|i| i as u64),
-            SSTableIndex::V3(v3_index) => v3_index.locate_with_key(key),
-            SSTableIndex::V3Empty(v3_empty) => v3_empty.locate_with_key(key),
-        }
-    }
-
-    /// Get the [`BlockAddr`] of the block that would contain `key`.
-    ///
-    /// Returns None if `key` is lexicographically after the last key recorded.
-    pub fn get_block_with_key(&self, key: &[u8]) -> Option<BlockAddr> {
-        match self {
-            SSTableIndex::V2(v2_index) => v2_index.get_block_with_key(key),
-            SSTableIndex::V3(v3_index) => v3_index.get_block_with_key(key),
-            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block_with_key(key),
-        }
-    }
-
-    pub(crate) fn locate_with_ord(&self, ord: TermOrdinal) -> u64 {
-        match self {
-            SSTableIndex::V2(v2_index) => v2_index.locate_with_ord(ord) as u64,
-            SSTableIndex::V3(v3_index) => v3_index.locate_with_ord(ord),
-            SSTableIndex::V3Empty(v3_empty) => v3_empty.locate_with_ord(ord),
-        }
-    }
-
-    /// Get the [`BlockAddr`] of the block containing the `ord`-th term.
-    pub(crate) fn get_block_with_ord(&self, ord: TermOrdinal) -> BlockAddr {
-        match self {
-            SSTableIndex::V2(v2_index) => v2_index.get_block_with_ord(ord),
-            SSTableIndex::V3(v3_index) => v3_index.get_block_with_ord(ord),
-            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block_with_ord(ord),
-        }
-    }
-
-    pub fn get_block_for_automaton<'a>(
-        &'a self,
-        automaton: &'a impl Automaton,
-    ) -> impl Iterator<Item = (u64, BlockAddr)> + 'a {
-        match self {
-            SSTableIndex::V2(v2_index) => {
-                BlockIter::V2(v2_index.get_block_for_automaton(automaton))
-            }
-            SSTableIndex::V3(v3_index) => {
-                BlockIter::V3(v3_index.get_block_for_automaton(automaton))
-            }
-            SSTableIndex::V3Empty(v3_empty) => {
-                BlockIter::V3Empty(std::iter::once((0, v3_empty.block_addr.clone())))
-            }
-        }
-    }
-}
-
-enum BlockIter<V2, V3, T> {
-    V2(V2),
-    V3(V3),
-    V3Empty(std::iter::Once<T>),
-}
-
-impl<V2: Iterator<Item = T>, V3: Iterator<Item = T>, T> Iterator for BlockIter<V2, V3, T> {
-    type Item = T;
-
-    fn next(&mut self) -> Option<Self::Item> {
-        match self {
-            BlockIter::V2(v2) => v2.next(),
-            BlockIter::V3(v3) => v3.next(),
-            BlockIter::V3Empty(once) => once.next(),
-        }
-    }
-}
+use crate::{SSTableDataCorruption, TermOrdinal};
 
 #[derive(Debug, Clone)]
 pub struct SSTableIndexV3 {
@@ -160,6 +68,11 @@ impl SSTableIndexV3 {
         self.block_addr_store.binary_search_ord(ord).1
     }
 
+    pub(crate) fn get_and_locate_with_ord(&self, ord: TermOrdinal) -> (BlockAddr, u64) {
+        let (location, block_addr) = self.block_addr_store.binary_search_ord(ord);
+        (block_addr, location)
+    }
+
     pub(crate) fn get_block_for_automaton<'a>(
         &'a self,
         automaton: &'a impl Automaton,
@@ -216,7 +129,7 @@ impl<A: Automaton> Iterator for GetBlockForAutomaton<'_, A> {
 
 #[derive(Debug, Clone)]
 pub struct SSTableIndexV3Empty {
-    block_addr: BlockAddr,
+    pub block_addr: BlockAddr,
 }
 
 impl SSTableIndexV3Empty {
@@ -230,8 +143,8 @@ impl SSTableIndexV3Empty {
     }
 
     /// Get the [`BlockAddr`] of the requested block.
-    pub(crate) fn get_block(&self, _block_id: u64) -> Option<BlockAddr> {
-        Some(self.block_addr.clone())
+    pub(crate) fn get_block(&self, block_id: u64) -> Option<BlockAddr> {
+        (block_id == 0).then(|| self.block_addr.clone())
     }
 
     /// Get the block id of the block that would contain `key`.
@@ -256,146 +169,9 @@ impl SSTableIndexV3Empty {
     pub(crate) fn get_block_with_ord(&self, _ord: TermOrdinal) -> BlockAddr {
         self.block_addr.clone()
     }
-}
-#[derive(Clone, Eq, PartialEq, Debug)]
-pub struct BlockAddr {
-    pub first_ordinal: u64,
-    pub byte_range: Range<usize>,
-}
 
-impl BlockAddr {
-    fn to_block_start(&self) -> BlockStartAddr {
-        BlockStartAddr {
-            first_ordinal: self.first_ordinal,
-            byte_range_start: self.byte_range.start,
-        }
-    }
-}
-
-#[derive(Debug, Clone, PartialEq, Eq)]
-struct BlockStartAddr {
-    first_ordinal: u64,
-    byte_range_start: usize,
-}
-
-impl BlockStartAddr {
-    fn to_block_addr(&self, byte_range_end: usize) -> BlockAddr {
-        BlockAddr {
-            first_ordinal: self.first_ordinal,
-            byte_range: self.byte_range_start..byte_range_end,
-        }
-    }
-}
-
-#[derive(Debug, Clone)]
-pub(crate) struct BlockMeta {
-    /// Any byte string that is lexicographically greater or equal to
-    /// the last key in the block,
-    /// and yet strictly smaller than the first key in the next block.
-    pub last_key_or_greater: Vec<u8>,
-    pub block_addr: BlockAddr,
-}
-
-impl BinarySerializable for BlockStartAddr {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
-        let start = self.byte_range_start as u64;
-        start.serialize(writer)?;
-        self.first_ordinal.serialize(writer)
-    }
-
-    fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
-        let byte_range_start = u64::deserialize(reader)? as usize;
-        let first_ordinal = u64::deserialize(reader)?;
-        Ok(BlockStartAddr {
-            first_ordinal,
-            byte_range_start,
-        })
-    }
-
-    // Provided method
-    fn num_bytes(&self) -> u64 {
-        BlockStartAddr::SIZE_IN_BYTES as u64
-    }
-}
-
-impl FixedSize for BlockStartAddr {
-    const SIZE_IN_BYTES: usize = 2 * u64::SIZE_IN_BYTES;
-}
-
-/// Given that left < right,
-/// mutates `left into a shorter byte string left'` that
-/// matches `left <= left' < right`.
-fn find_shorter_str_in_between(left: &mut Vec<u8>, right: &[u8]) {
-    assert!(&left[..] < right);
-    let common_len = common_prefix_len(left, right);
-    if left.len() == common_len {
-        return;
-    }
-    // It is possible to do one character shorter in some case,
-    // but it is not worth the extra complexity
-    for pos in (common_len + 1)..left.len() {
-        if left[pos] != u8::MAX {
-            left[pos] += 1;
-            left.truncate(pos + 1);
-            return;
-        }
-    }
-}
-
-#[derive(Default)]
-pub struct SSTableIndexBuilder {
-    blocks: Vec<BlockMeta>,
-}
-
-impl SSTableIndexBuilder {
-    /// In order to make the index as light as possible, we
-    /// try to find a shorter alternative to the last key of the last block
-    /// that is still smaller than the next key.
-    pub(crate) fn shorten_last_block_key_given_next_key(&mut self, next_key: &[u8]) {
-        if let Some(last_block) = self.blocks.last_mut() {
-            find_shorter_str_in_between(&mut last_block.last_key_or_greater, next_key);
-        }
-    }
-
-    pub fn add_block(&mut self, last_key: &[u8], byte_range: Range<usize>, first_ordinal: u64) {
-        self.blocks.push(BlockMeta {
-            last_key_or_greater: last_key.to_vec(),
-            block_addr: BlockAddr {
-                byte_range,
-                first_ordinal,
-            },
-        })
-    }
-
-    pub fn serialize<W: std::io::Write>(&self, wrt: W) -> io::Result<u64> {
-        if self.blocks.len() <= 1 {
-            return Ok(0);
-        }
-        let counting_writer = common::CountingWriter::wrap(wrt);
-        let mut map_builder = MapBuilder::new(counting_writer).map_err(fst_error_to_io_error)?;
-        for (i, block) in self.blocks.iter().enumerate() {
-            map_builder
-                .insert(&block.last_key_or_greater, i as u64)
-                .map_err(fst_error_to_io_error)?;
-        }
-        let counting_writer = map_builder.into_inner().map_err(fst_error_to_io_error)?;
-        let written_bytes = counting_writer.written_bytes();
-        let mut wrt = counting_writer.finish();
-
-        let mut block_store_writer = BlockAddrStoreWriter::new();
-        for block in &self.blocks {
-            block_store_writer.write_block_meta(block.block_addr.clone())?;
-        }
-        block_store_writer.serialize(&mut wrt)?;
-
-        Ok(written_bytes)
-    }
-}
-
-fn fst_error_to_io_error(error: tantivy_fst::Error) -> io::Error {
-    match error {
-        tantivy_fst::Error::Fst(fst_error) => io::Error::other(fst_error),
-        tantivy_fst::Error::Io(ioerror) => ioerror,
+    pub(crate) fn get_and_locate_with_ord(&self, _ord: TermOrdinal) -> (BlockAddr, u64) {
+        (self.block_addr.clone(), 0)
     }
 }
 
@@ -647,14 +423,14 @@ fn binary_search(max: u64, cmp_fn: impl Fn(u64) -> std::cmp::Ordering) -> Result
     Err(left)
 }
 
-struct BlockAddrStoreWriter {
+pub(crate) struct BlockAddrStoreWriter {
     buffer_block_metas: Vec<u8>,
     buffer_addrs: Vec<u8>,
     block_addrs: Vec<BlockAddr>,
 }
 
 impl BlockAddrStoreWriter {
-    fn new() -> Self {
+    pub(crate) fn new() -> Self {
         BlockAddrStoreWriter {
             buffer_block_metas: Vec::new(),
             buffer_addrs: Vec::new(),
@@ -662,7 +438,7 @@ impl BlockAddrStoreWriter {
         }
     }
 
-    fn flush_block(&mut self) -> io::Result<()> {
+    pub(crate) fn flush_block(&mut self) -> io::Result<()> {
         if self.block_addrs.is_empty() {
             return Ok(());
         }
@@ -741,7 +517,7 @@ impl BlockAddrStoreWriter {
         Ok(())
     }
 
-    fn write_block_meta(&mut self, block_addr: BlockAddr) -> io::Result<()> {
+    pub(crate) fn write_block_meta(&mut self, block_addr: BlockAddr) -> io::Result<()> {
         self.block_addrs.push(block_addr);
         if self.block_addrs.len() >= STORE_BLOCK_LEN {
             self.flush_block()?;
@@ -749,7 +525,7 @@ impl BlockAddrStoreWriter {
         Ok(())
     }
 
-    fn serialize<W: std::io::Write>(&mut self, wrt: &mut W) -> io::Result<()> {
+    pub(crate) fn serialize<W: std::io::Write>(&mut self, wrt: &mut W) -> io::Result<()> {
         self.flush_block()?;
         let len = self.buffer_block_metas.len() as u64;
         len.serialize(wrt)?;
@@ -824,8 +600,9 @@ mod tests {
     use common::OwnedBytes;
 
     use super::*;
-    use crate::SSTableDataCorruption;
     use crate::block_match_automaton::tests::EqBuffer;
+    use crate::index::BlockMeta;
+    use crate::{SSTableDataCorruption, SSTableIndexBuilder};
 
     #[test]
     fn test_sstable_index() {
@@ -874,36 +651,7 @@ mod tests {
         assert!(matches!(data_corruption_err, SSTableDataCorruption));
     }
 
-    #[track_caller]
-    fn test_find_shorter_str_in_between_aux(left: &[u8], right: &[u8]) {
-        let mut left_buf = left.to_vec();
-        super::find_shorter_str_in_between(&mut left_buf, right);
-        assert!(left_buf.len() <= left.len());
-        assert!(left <= &left_buf);
-        assert!(&left_buf[..] < right);
-    }
-
-    #[test]
-    fn test_find_shorter_str_in_between() {
-        test_find_shorter_str_in_between_aux(b"", b"hello");
-        test_find_shorter_str_in_between_aux(b"abc", b"abcd");
-        test_find_shorter_str_in_between_aux(b"abcd", b"abd");
-        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[1]);
-        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[0, 0, 1]);
-        test_find_shorter_str_in_between_aux(&[0, 0, 255, 255, 255, 0u8], &[0, 1]);
-    }
-
-    use proptest::prelude::*;
-
-    proptest! {
-        #![proptest_config(ProptestConfig::with_cases(100))]
-        #[test]
-        fn test_proptest_find_shorter_str(left in any::<Vec<u8>>(), right in any::<Vec<u8>>()) {
-            if left < right {
-                test_find_shorter_str_in_between_aux(&left, &right);
-            }
-        }
-    }
+    //    use proptest::prelude::*;
 
     #[test]
     fn test_find_best_slop() {

sstable/src/lib.rs

@@ -47,9 +47,8 @@ pub mod merge;
 mod streamer;
 pub mod value;
 
-mod sstable_index_v3;
-pub use sstable_index_v3::{BlockAddr, SSTableIndex, SSTableIndexBuilder, SSTableIndexV3};
-mod sstable_index_v2;
+mod index;
+pub use index::{BlockAddr, SSTableIndex, SSTableIndexBuilder};
 pub(crate) mod vint;
 pub use dictionary::{Dictionary, TermOrdHit};
 pub use streamer::{Streamer, StreamerBuilder};

stacker/Cargo.toml

@@ -8,7 +8,7 @@ repository = "https://github.com/quickwit-oss/tantivy"
 description = "term hashmap used for indexing"
 
 [dependencies]
-murmurhash32 = "0.3"
+murmurhash32 = "0.4"
 common = { version = "0.11", path = "../common/", package = "tantivy-common" }
 ahash = { version = "0.8.11", default-features = false, optional = true }
 
@@ -27,7 +27,7 @@ rand = "0.9"
 zipf = "7.0.0"
 rustc-hash = "2.1.0"
 proptest = "1.2.0"
-binggan = { version = "0.16.1" }
+binggan = { version = "0.17.0" }
 rand_distr = "0.5"
 
 [features]