Update fs4 requirement from 0.13.1 to 1.1.0

Updates the requirements on [fs4](https://github.com/al8n/fs4) to permit the latest version. - [Release notes](https://github.com/al8n/fs4/releases) - [Changelog](https://github.com/al8n/fs4/blob/main/CHANGELOG.md) - [Commits](https://github.com/al8n/fs4/commits/1.1.0) --- updated-dependencies: - dependency-name: fs4 dependency-version: 1.1.0 dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com>
Enable BMW for single-scorer boolean queries by removing early return in scorer_union (#2915 )
2026-05-26 13:10:41 +00:00 · 2026-04-30 20:04:14 +00:00 · 2026-04-28 14:49:53 -07:00 · 2026-04-28 16:59:59 +02:00 · 2026-04-28 16:59:59 +02:00 · 2026-04-28 11:57:36 +02:00
75 changed files with 7762 additions and 490 deletions
--- a/.claude/skills/update-changelog/SKILL.md
+++ b/.claude/skills/update-changelog/SKILL.md
@@ -0,0 +1,87 @@
+---
+name: update-changelog
+description: Update CHANGELOG.md with merged PRs since the last changelog update, categorized by type
+---
+
+# Update Changelog
+
+This skill updates CHANGELOG.md with merged PRs that aren't already listed.
+
+## Step 1: Determine the changelog scope
+
+Read `CHANGELOG.md` to identify the current unreleased version section at the top (e.g., `Tantivy 0.26 (Unreleased)`).
+
+Collect all PR numbers already mentioned in the unreleased section by extracting `#NNNN` references.
+
+## Step 2: Find merged PRs not yet in the changelog
+
+Use `gh` to list recently merged PRs from the upstream repo:
+
+```bash
+gh pr list --repo quickwit-oss/tantivy --state merged --limit 100 --json number,title,author,labels,mergedAt
+```
+
+Filter out any PRs whose number already appears in the unreleased section of the changelog.
+
+## Step 3: Consolidate related PRs
+
+Before categorizing, group PRs that belong to the same logical change. This is critical for producing a clean changelog. Use PR descriptions, titles, cross-references, and the files touched to identify relationships.
+
+**Merge follow-up PRs into the original:**
+- If a PR is a bugfix, refinement, or follow-up to another PR in the same unreleased cycle, combine them into a single changelog entry with multiple `[#N](url)` links.
+- Also consolidate PRs that touch the same feature area even if not explicitly linked — e.g., a PR fixing an edge case in a new API should be folded into the entry for the PR that introduced that API.
+
+**Filter out bugfixes on unreleased features:**
+- If a bugfix PR fixes something introduced by another PR in the **same unreleased version**, it must NOT appear as a separate Bugfixes entry. Instead, silently fold it into the original feature/improvement entry. The changelog should describe the final shipped state, not the development history.
+- To detect this: check if the bugfix PR references or reverts changes from another PR in the same release cycle, or if it touches code that was newly added (not present in the previous release).
+
+## Step 4: Review the actual code diff
+
+**Do not rely on PR titles or descriptions alone.** For every candidate PR, run `gh pr diff <number> --repo quickwit-oss/tantivy` and read the actual changes. PR titles are often misleading — the diff is the source of truth.
+
+**What to look for in the diff:**
+- Does it change observable behavior, public API surface, or performance characteristics?
+- Is the change something a user of the library would notice or need to know about?
+- Could the change break existing code (API changes, removed features)?
+
+**Skip PRs where the diff reveals the change is not meaningful enough for the changelog** — e.g., cosmetic renames, trivial visibility tweaks, test-only changes, etc.
+
+## Step 5: Categorize each PR group
+
+For each PR (or consolidated group) that survived the diff review, determine its category:
+
+- **Bugfixes** — fixes to behavior that existed in the **previous release**. NOT fixes to features introduced in this release cycle.
+- **Features/Improvements** — new features, API additions, new options, improvements that change user-facing behavior or add new capabilities.
+- **Performance** — optimizations, speed improvements, memory reductions. **If a PR adds new API whose primary purpose is enabling a performance optimization, categorize it as Performance, not Features.** The deciding question is: does a user benefit from this because of new functionality, or because things got faster/leaner? For example, a new trait method that exists solely to enable cheaper intersection ordering is Performance, not a Feature.
+
+If a PR doesn't clearly fit any category (e.g., CI-only changes, internal refactors with no user-facing impact, dependency bumps with no behavior change), skip it — not everything belongs in the changelog.
+
+When unclear, use your best judgment or ask the user.
+
+## Step 6: Format entries
+
+Each entry must follow this exact format:
+
+```
+- Description [#NUMBER](https://github.com/quickwit-oss/tantivy/pull/NUMBER)(@author)
+```
+
+Rules:
+- The description should be concise and describe the user-facing change (not the implementation). Describe the final shipped state, not the incremental development steps.
+- Use sub-categories with bold headers when multiple entries relate to the same area (e.g., `- **Aggregation**` with indented entries beneath). Follow the existing grouping style in the changelog.
+- Author is the GitHub username from the PR, prefixed with `@`. For consolidated entries, include all contributing authors.
+- For consolidated PRs, list all PR links in a single entry: `[#100](url) [#110](url)` (see existing entries for examples).
+
+## Step 7: Present changes to the user
+
+Show the user the proposed changelog entries grouped by category **before** editing the file. Ask for confirmation or adjustments.
+
+## Step 8: Update CHANGELOG.md
+
+Insert the new entries into the appropriate sections of the unreleased version block. If a section doesn't exist yet, create it following the order: Bugfixes, Features/Improvements, Performance.
+
+Append new entries at the end of each section (before the next section header or version header).
+
+## Step 9: Verify
+
+Read back the updated unreleased section and display it to the user for final review.
--- a/.github/dependabot.yml
+++ b/.github/dependabot.yml
@@ -6,6 +6,8 @@ updates:
    interval: daily
    time: "20:00"
  open-pull-requests-limit: 10
+  cooldown:
+    default-days: 2

 - package-ecosystem: "github-actions"
  directory: "/"
@@ -13,3 +15,5 @@ updates:
    interval: daily
    time: "20:00"
  open-pull-requests-limit: 10
+  cooldown:
+    default-days: 2
--- a/.github/workflows/coverage.yml
+++ b/.github/workflows/coverage.yml
@@ -4,6 +4,9 @@ on:
  push:
    branches: [main]

+permissions:
+  contents: read
+
 # Ensures that we cancel running jobs for the same PR / same workflow.
 concurrency:
  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
@@ -12,16 +15,20 @@ concurrency:
 jobs:
  coverage:
    runs-on: ubuntu-latest
+
+    permissions:
+      contents: read
+
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      - name: Install Rust
        run: rustup toolchain install nightly-2025-12-01 --profile minimal --component llvm-tools-preview
-      - uses: Swatinem/rust-cache@v2
-      - uses: taiki-e/install-action@cargo-llvm-cov
+      - uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
+      - uses: taiki-e/install-action@e4b3a0453201addddc06d3a72db90326aad87084 # cargo-llvm-cov
      - 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@v3
+        uses: codecov/codecov-action@57e3a136b779b570ffcdbf80b3bdc90e7fab3de2 # v6.0.0
        continue-on-error: true
        with:
          token: ${{ secrets.CODECOV_TOKEN }} # not required for public repos
--- a/.github/workflows/long_running.yml
+++ b/.github/workflows/long_running.yml
@@ -8,6 +8,9 @@ env:
  CARGO_TERM_COLOR: always
  NUM_FUNCTIONAL_TEST_ITERATIONS: 20000

+permissions:
+  contents: read
+
 # Ensures that we cancel running jobs for the same PR / same workflow.
 concurrency:
  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
@@ -18,10 +21,13 @@ jobs:

    runs-on: ubuntu-latest

+    permissions:
+      contents: read
+
    steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
    - name: Install stable
-      uses: actions-rs/toolchain@v1
+      uses: actions-rs/toolchain@16499b5e05bf2e26879000db0c1d13f7e13fa3af # v1.0.7
      with:
          toolchain: stable
          profile: minimal
--- a/.github/workflows/scorecard.yml
+++ b/.github/workflows/scorecard.yml
@@ -0,0 +1,49 @@
+name: OpenSSF Scorecard
+
+on:
+  schedule:
+    - cron: '0 0 * * 0'
+  push:
+    branches:
+      - main
+
+permissions:
+  contents: read
+
+jobs:
+  analysis:
+    name: Scorecards analysis
+    runs-on: ubuntu-latest
+    permissions:
+      # Needed to upload the results to code-scanning dashboard.
+      security-events: write
+      # Needed to publish results
+      id-token: write
+
+    steps:
+      - name: 'Checkout code'
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: 'Run analysis'
+        uses: ossf/scorecard-action@4eaacf0543bb3f2c246792bd56e8cdeffafb205a # v2.4.3
+        with:
+          results_file: results.sarif
+          results_format: sarif
+          repo_token: ${{ secrets.GITHUB_TOKEN }}
+          publish_results: true
+
+      # Upload the results as artifacts.
+      - name: 'Upload artifact'
+        uses: actions/upload-artifact@bbbca2ddaa5d8feaa63e36b76fdaad77386f024f # v7.0.0
+        with:
+          name: SARIF file
+          path: results.sarif
+          retention-days: 5
+
+      # Upload the results to GitHub's code scanning dashboard.
+      - name: 'Upload to code-scanning'
+        uses: github/codeql-action/upload-sarif@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v4.35.2
+        with:
+          sarif_file: results.sarif
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -9,6 +9,9 @@ on:
 env:
  CARGO_TERM_COLOR: always

+permissions:
+  contents: read
+
 # Ensures that we cancel running jobs for the same PR / same workflow.
 concurrency:
  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
@@ -19,23 +22,27 @@ jobs:

    runs-on: ubuntu-latest

+    permissions:
+      contents: read
+      checks: write
+
    steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

    - name: Install nightly
-      uses: actions-rs/toolchain@v1
+      uses: actions-rs/toolchain@16499b5e05bf2e26879000db0c1d13f7e13fa3af # v1.0.7
      with:
            toolchain: nightly
            profile: minimal
            components: rustfmt
    - name: Install stable
-      uses: actions-rs/toolchain@v1
+      uses: actions-rs/toolchain@16499b5e05bf2e26879000db0c1d13f7e13fa3af # v1.0.7
      with:
            toolchain: stable
            profile: minimal
            components: clippy

-    - uses: Swatinem/rust-cache@v2
+    - uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1

    - name: Check Formatting
      run: cargo +nightly fmt --all -- --check
@@ -47,7 +54,7 @@ jobs:
    - name: Check Bench Compilation
      run: cargo +nightly bench --no-run --profile=dev --all-features

-    - uses: actions-rs/clippy-check@v1
+    - uses: actions-rs/clippy-check@b5b5f21f4797c02da247df37026fcd0a5024aa4d # v1.0.7
      with:
        toolchain: stable
        token: ${{ secrets.GITHUB_TOKEN }}
@@ -57,6 +64,9 @@ jobs:

    runs-on: ubuntu-latest

+    permissions:
+      contents: read
+
    strategy:
      matrix:
        features:
@@ -67,17 +77,17 @@ jobs:
    name: test-${{ matrix.features.label}}

    steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

    - name: Install stable
-      uses: actions-rs/toolchain@v1
+      uses: actions-rs/toolchain@16499b5e05bf2e26879000db0c1d13f7e13fa3af # v1.0.7
      with:
            toolchain: stable
            profile: minimal
            override: true

-    - uses: taiki-e/install-action@nextest
-    - uses: Swatinem/rust-cache@v2
+    - uses: taiki-e/install-action@56cc9adf3a3e2c23eafb56e8acaf9d0373cb845a # nextest
+    - uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1

    - name: Run tests
      run: |
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,58 @@
+Tantivy 0.26.1
+================================
+
+## Performance
+- Fix quadratic runtime in nested term and composite aggregations: memory accounting scanned all parent buckets on every collect instead of just the current parent (@PSeitz @fulmicoton)
+
+Tantivy 0.26 (Unreleased)
+================================
+
+## Bugfixes
+- Align float query coercion during search with the columnar coercion rules [#2692](https://github.com/quickwit-oss/tantivy/pull/2692)(@fulmicoton)
+- Fix lenient elastic range queries with trailing closing parentheses [#2816](https://github.com/quickwit-oss/tantivy/pull/2816)(@evance-br)
+- Fix intersection `seek()` advancing below current doc id [#2812](https://github.com/quickwit-oss/tantivy/pull/2812)(@fulmicoton)
+- Fix phrase query prefixed with `*` [#2751](https://github.com/quickwit-oss/tantivy/pull/2751)(@Darkheir)
+- Fix `vint` buffer overflow during index creation [#2778](https://github.com/quickwit-oss/tantivy/pull/2778)(@rebasedming)
+- Fix integer overflow in `ExpUnrolledLinkedList` for large datasets [#2735](https://github.com/quickwit-oss/tantivy/pull/2735)(@mdashti)
+- Fix integer overflow in segment sorting and merge policy truncation [#2846](https://github.com/quickwit-oss/tantivy/pull/2846)(@anaslimem)
+- Fix merging of intermediate aggregation results [#2719](https://github.com/quickwit-oss/tantivy/pull/2719)(@PSeitz)
+- Fix deduplicate doc counts in term aggregation for multi-valued fields [#2854](https://github.com/quickwit-oss/tantivy/pull/2854)(@nuri-yoo)
+
+## Features/Improvements
+- **Aggregation**
+    - Add filter aggregation [#2711](https://github.com/quickwit-oss/tantivy/pull/2711)(@mdashti)
+    - Add include/exclude filtering for term aggregations [#2717](https://github.com/quickwit-oss/tantivy/pull/2717)(@PSeitz)
+    - Add public accessors for intermediate aggregation results [#2829](https://github.com/quickwit-oss/tantivy/pull/2829)(@congx4)
+    - Replace HyperLogLog++ with Apache DataSketches HLL for cardinality aggregation [#2837](https://github.com/quickwit-oss/tantivy/pull/2837) [#2842](https://github.com/quickwit-oss/tantivy/pull/2842)(@congx4)
+    - Add composite aggregation [#2856](https://github.com/quickwit-oss/tantivy/pull/2856)(@fulmicoton)
+- **Fast Fields**
+    - Add fast field fallback for `TermQuery` when the field is not indexed [#2693](https://github.com/quickwit-oss/tantivy/pull/2693)(@PSeitz-dd)
+    - Add fast field support for `Bytes` values [#2830](https://github.com/quickwit-oss/tantivy/pull/2830)(@mdashti)
+- **Query Parser**
+    - Add support for regexes in the query grammar [#2677](https://github.com/quickwit-oss/tantivy/pull/2677) [#2818](https://github.com/quickwit-oss/tantivy/pull/2818)(@Darkheir)
+    - Deduplicate queries in query parser [#2698](https://github.com/quickwit-oss/tantivy/pull/2698)(@PSeitz-dd)
+- Add erased `SortKeyComputer` for sorting on column types unknown until runtime [#2770](https://github.com/quickwit-oss/tantivy/pull/2770) [#2790](https://github.com/quickwit-oss/tantivy/pull/2790)(@stuhood @PSeitz)
+- Add natural-order-with-none-highest support in `TopDocs::order_by` [#2780](https://github.com/quickwit-oss/tantivy/pull/2780)(@stuhood)
+- Move stemming behing `stemmer` feature flag [#2791](https://github.com/quickwit-oss/tantivy/pull/2791)(@fulmicoton)
+- Make `DeleteMeta`, `AddOperation`, `advance_deletes`, `with_max_doc`, `serializer` module, and `delete_queue` public [#2762](https://github.com/quickwit-oss/tantivy/pull/2762) [#2765](https://github.com/quickwit-oss/tantivy/pull/2765) [#2766](https://github.com/quickwit-oss/tantivy/pull/2766) [#2835](https://github.com/quickwit-oss/tantivy/pull/2835)(@philippemnoel @PSeitz)
+- Make `Language` hashable [#2763](https://github.com/quickwit-oss/tantivy/pull/2763)(@philippemnoel)
+- Improve `space_usage` reporting for JSON fields and columnar data [#2761](https://github.com/quickwit-oss/tantivy/pull/2761)(@PSeitz-dd)
+- Split `Term` into `Term` and `IndexingTerm` [#2744](https://github.com/quickwit-oss/tantivy/pull/2744) [#2750](https://github.com/quickwit-oss/tantivy/pull/2750)(@PSeitz-dd @PSeitz)
+
+## Performance
+- **Aggregation**
+    - Large speed up and memory reduction for nested high cardinality aggregations by using one collector per request instead of one per bucket, and adding `PagedTermMap` for faster medium cardinality term aggregations [#2715](https://github.com/quickwit-oss/tantivy/pull/2715) [#2759](https://github.com/quickwit-oss/tantivy/pull/2759)(@PSeitz @PSeitz-dd)
+    - Optimize low-cardinality term aggregations by using a `Vec` instead of a `HashMap` [#2740](https://github.com/quickwit-oss/tantivy/pull/2740)(@fulmicoton-dd)
+- Optimize `ExistsQuery` for a high number of dynamic columns [#2694](https://github.com/quickwit-oss/tantivy/pull/2694)(@PSeitz-dd)
+- Add lazy scorers to stop score evaluation early when a doc won't reach the top-K threshold [#2726](https://github.com/quickwit-oss/tantivy/pull/2726) [#2777](https://github.com/quickwit-oss/tantivy/pull/2777)(@fulmicoton @stuhood)
+- Add `DocSet::cost()` and use it to order scorers in intersections [#2707](https://github.com/quickwit-oss/tantivy/pull/2707)(@PSeitz)
+- Add `collect_block` support for collector wrappers [#2727](https://github.com/quickwit-oss/tantivy/pull/2727)(@stuhood)
+- Optimize saturated posting lists by replacing them with `AllScorer` in boolean queries [#2745](https://github.com/quickwit-oss/tantivy/pull/2745) [#2760](https://github.com/quickwit-oss/tantivy/pull/2760) [#2774](https://github.com/quickwit-oss/tantivy/pull/2774)(@fulmicoton @mdashti @trinity-1686a)
+- Add `seek_danger` on `DocSet` for more efficient intersections [#2538](https://github.com/quickwit-oss/tantivy/pull/2538) [#2810](https://github.com/quickwit-oss/tantivy/pull/2810)(@PSeitz @stuhood @fulmicoton)
+- Skip column traversal in `RangeDocSet` when query range does not overlap with column bounds [#2783](https://github.com/quickwit-oss/tantivy/pull/2783)(@ChangRui-Ryan)
+- Speed up exclude queries by supporting multiple excluded `DocSet`s without intermediate union [#2825](https://github.com/quickwit-oss/tantivy/pull/2825)(@PSeitz)
+- Improve union performance for non-score unions with `fill_buffer` and optimized `TinySet` [#2863](https://github.com/quickwit-oss/tantivy/pull/2863)(@PSeitz)
+
 Tantivy 0.25
 ================================

--- a/Cargo.toml
+++ b/Cargo.toml
@@ -11,7 +11,7 @@ repository = "https://github.com/quickwit-oss/tantivy"
 readme = "README.md"
 keywords = ["search", "information", "retrieval"]
 edition = "2021"
-rust-version = "1.85"
+rust-version = "1.86"
 exclude = ["benches/*.json", "benches/*.txt"]

 [dependencies]
@@ -27,13 +27,13 @@ regex = { version = "1.5.5", default-features = false, features = [
 aho-corasick = "1.0"
 tantivy-fst = "0.5"
 memmap2 = { version = "0.9.0", optional = true }
-lz4_flex = { version = "0.12", default-features = false, optional = true }
+lz4_flex = { version = "0.13", default-features = false, optional = true }
 zstd = { version = "0.13", optional = true, default-features = false }
 tempfile = { version = "3.12.0", optional = true }
 log = "0.4.16"
 serde = { version = "1.0.219", features = ["derive"] }
 serde_json = "1.0.140"
-fs4 = { version = "0.13.1", optional = true }
+fs4 = { version = "1.1.0", optional = true }
 levenshtein_automata = "0.2.1"
 uuid = { version = "1.0.0", features = ["v4", "serde"] }
 crossbeam-channel = "0.5.4"
@@ -57,15 +57,15 @@ measure_time = "0.9.0"
 arc-swap = "1.5.0"
 bon = "3.3.1"

-columnar = { version = "0.6", path = "./columnar", package = "tantivy-columnar" }
-sstable = { version = "0.6", path = "./sstable", package = "tantivy-sstable", optional = true }
-stacker = { version = "0.6", path = "./stacker", package = "tantivy-stacker" }
-query-grammar = { version = "0.25.0", path = "./query-grammar", package = "tantivy-query-grammar" }
-tantivy-bitpacker = { version = "0.9", path = "./bitpacker" }
-common = { version = "0.10", path = "./common/", package = "tantivy-common" }
-tokenizer-api = { version = "0.6", path = "./tokenizer-api", package = "tantivy-tokenizer-api" }
-sketches-ddsketch = { git = "https://github.com/quickwit-oss/rust-sketches-ddsketch.git", rev = "555caf1", features = ["use_serde"] }
-datasketches = "0.2.0"
+columnar = { version = "0.7", path = "./columnar", package = "tantivy-columnar" }
+sstable = { version = "0.7", path = "./sstable", package = "tantivy-sstable", optional = true }
+stacker = { version = "0.7", path = "./stacker", package = "tantivy-stacker" }
+query-grammar = { version = "0.26.0", path = "./query-grammar", package = "tantivy-query-grammar" }
+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" }
 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 = "0.14.2"
+binggan = "0.16.1"
 rand = "0.9"
 maplit = "1.0.2"
 matches = "0.1.9"
@@ -92,7 +92,7 @@ postcard = { version = "1.0.4", features = [
 ], default-features = false }

 [target.'cfg(not(windows))'.dev-dependencies]
-criterion = { version = "0.5", default-features = false }
+criterion = { version = "0.8", default-features = false }

 [dev-dependencies.fail]
 version = "0.5.0"
@@ -201,3 +201,11 @@ harness = false
 [[bench]]
 name = "regex_all_terms"
 harness = false
+
+[[bench]]
+name = "query_parser_nested"
+harness = false
+
+[[bench]]
+name = "intersection_bench"
+harness = false
--- a/README.md
+++ b/README.md
@@ -1,6 +1,7 @@
 [![Docs](https://docs.rs/tantivy/badge.svg)](https://docs.rs/crate/tantivy/)
 [![Build Status](https://github.com/quickwit-oss/tantivy/actions/workflows/test.yml/badge.svg)](https://github.com/quickwit-oss/tantivy/actions/workflows/test.yml)
 [![codecov](https://codecov.io/gh/quickwit-oss/tantivy/branch/main/graph/badge.svg)](https://codecov.io/gh/quickwit-oss/tantivy)
+[![OpenSSF Scorecard](https://api.scorecard.dev/projects/github.com/quickwit-oss/tantivy/badge)](https://scorecard.dev/viewer/?uri=github.com/quickwit-oss/tantivy)
 [![Join the chat at https://discord.gg/MT27AG5EVE](https://shields.io/discord/908281611840282624?label=chat%20on%20discord)](https://discord.gg/MT27AG5EVE)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 [![Crates.io](https://img.shields.io/crates/v/tantivy.svg)](https://crates.io/crates/tantivy)
--- a/benches/agg_bench.rs
+++ b/benches/agg_bench.rs
@@ -10,7 +10,7 @@ use tantivy::aggregation::agg_req::Aggregations;
 use tantivy::aggregation::AggregationCollector;
 use tantivy::query::{AllQuery, TermQuery};
 use tantivy::schema::{IndexRecordOption, Schema, TextFieldIndexing, FAST, STRING};
-use tantivy::{doc, Index, Term};
+use tantivy::{doc, DateTime, Index, Term};

 #[global_allocator]
 pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
@@ -63,6 +63,8 @@ fn bench_agg(mut group: InputGroup<Index>) {
    register!(group, terms_all_unique_with_avg_sub_agg);
    register!(group, terms_many_with_avg_sub_agg);
    register!(group, terms_status_with_avg_sub_agg);
+    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_zipf_1000);
    register!(group, terms_zipf_1000_with_histogram);
@@ -70,8 +72,20 @@ fn bench_agg(mut group: InputGroup<Index>) {

    register!(group, terms_many_json_mixed_type_with_avg_sub_agg);

+    register!(group, composite_term_many_page_1000);
+    register!(group, composite_term_many_page_1000_with_avg_sub_agg);
+    register!(group, composite_term_few);
+    register!(group, composite_histogram);
+    register!(group, composite_histogram_calendar);
+
    register!(group, cardinality_agg);
    register!(group, terms_status_with_cardinality_agg);
+    register!(group, terms_100_buckets_with_cardinality_agg);
+    register!(group, terms_many_with_single_term_order_by_cardinality_agg);
+    register!(
+        group,
+        terms_many_with_nested_terms_double_order_by_cardinality_agg
+    );

    register!(group, range_agg);
    register!(group, range_agg_with_avg_sub_agg);
@@ -163,6 +177,22 @@ fn terms_status_with_cardinality_agg(index: &Index) {
    let agg_req = json!({
        "my_texts": {
            "terms": { "field": "text_few_terms_status" },
+            "aggs": {
+                "cardinality": {
+                    "cardinality": {
+                        "field": "text_few_terms_status"
+                    },
+                }
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+
+fn terms_100_buckets_with_cardinality_agg(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_1000_terms_zipf", "size": 100 },
            "aggs": {
                "cardinality": {
                    "cardinality": {
@@ -175,6 +205,60 @@ fn terms_status_with_cardinality_agg(index: &Index) {
    execute_agg(index, agg_req);
 }

+fn terms_many_with_single_term_order_by_cardinality_agg(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_many_terms" },
+            "aggs": {
+                "nested_terms": {
+                    "terms": {
+                        "field": "single_term",
+                        "order": { "cardinality": "desc" }
+                    },
+                    "aggs": {
+                        "cardinality": {
+                            "cardinality": { "field": "text_many_terms" }
+                        }
+                    }
+                }
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+
+// Two-level terms ordered by cardinality at each level: a high-card outer terms
+// (text_many_terms) ordered by a cardinality sub-agg, with a nested low-card terms
+// (text_few_terms_status) also ordered by a cardinality sub-agg, plus an avg.
+fn terms_many_with_nested_terms_double_order_by_cardinality_agg(index: &Index) {
+    let agg_req = json!({
+        "by_ip": {
+            "terms": {
+                "field": "text_many_terms",
+                "size": 50,
+                "order": { "distinct_path": "desc" }
+            },
+            "aggs": {
+                "distinct_path": {
+                    "cardinality": { "field": "text_few_terms" }
+                },
+                "by_asn": {
+                    "terms": {
+                        "field": " single_term",
+                        "size": 10,
+                        "order": { "distinct_path2": "desc" }
+                    },
+                    "aggs": {
+                        "avg_botscore": { "avg": { "field": "score" } },
+                        "distinct_path2": { "cardinality": { "field": "text_few_terms" } }
+                    }
+                }
+            }
+        }
+    });
+    execute_agg(index, agg_req);
+}
+
 fn terms_7(index: &Index) {
    let agg_req = json!({
        "my_texts": { "terms": { "field": "text_few_terms_status" } },
@@ -247,6 +331,30 @@ fn terms_all_unique_with_avg_sub_agg(index: &Index) {
    });
    execute_agg(index, agg_req);
 }
+fn terms_status_with_terms_zipf_1000_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_few_terms_status" },
+            "aggs": {
+                "nested_terms": { "terms": { "field": "text_1000_terms_zipf" } }
+            }
+        }
+    });
+    execute_agg(index, agg_req);
+}
+
+fn terms_zipf_1000_with_terms_status_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_1000_terms_zipf" },
+            "aggs": {
+                "nested_terms": { "terms": { "field": "text_few_terms_status" } }
+            }
+        }
+    });
+    execute_agg(index, agg_req);
+}
+
 fn terms_status_with_histogram(index: &Index) {
    let agg_req = json!({
        "my_texts": {
@@ -314,6 +422,75 @@ fn terms_many_json_mixed_type_with_avg_sub_agg(index: &Index) {
    execute_agg(index, agg_req);
 }

+fn composite_term_few(index: &Index) {
+    let agg_req = json!({
+        "my_ctf": {
+            "composite": {
+                "sources": [
+                    { "text_few_terms": { "terms": { "field": "text_few_terms" } } }
+                ],
+                "size": 1000
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+fn composite_term_many_page_1000(index: &Index) {
+    let agg_req = json!({
+        "my_ctmp1000": {
+            "composite": {
+                "sources": [
+                    { "text_many_terms": { "terms": { "field": "text_many_terms" } } }
+                ],
+                "size": 1000
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+fn composite_term_many_page_1000_with_avg_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "my_ctmp1000wasa": {
+            "composite": {
+                "sources": [
+                    { "text_many_terms": { "terms": { "field": "text_many_terms" } } }
+                ],
+                "size": 1000,
+            },
+            "aggs": {
+                "average_f64": { "avg": { "field": "score_f64" } }
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+fn composite_histogram(index: &Index) {
+    let agg_req = json!({
+        "my_ch": {
+            "composite": {
+                "sources": [
+                    { "f64_histogram": { "histogram": { "field": "score_f64", "interval": 1 } } }
+                ],
+                "size": 1000
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+fn composite_histogram_calendar(index: &Index) {
+    let agg_req = json!({
+        "my_chc": {
+            "composite": {
+                "sources": [
+                    { "time_histogram": { "date_histogram": { "field": "timestamp", "calendar_interval": "month" } } }
+                ],
+                "size": 1000
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+
 fn execute_agg(index: &Index, agg_req: serde_json::Value) {
    let agg_req: Aggregations = serde_json::from_value(agg_req).unwrap();
    let collector = get_collector(agg_req);
@@ -491,11 +668,13 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
            TextFieldIndexing::default().set_index_option(IndexRecordOption::WithFreqs),
        )
        .set_stored();
-    let text_field = schema_builder.add_text_field("text", text_fieldtype);
+    let text_field = schema_builder.add_text_field("text", text_fieldtype.clone());
+    let single_term = schema_builder.add_text_field("single_term", FAST);
    let json_field = schema_builder.add_json_field("json", FAST);
    let text_field_all_unique_terms =
        schema_builder.add_text_field("text_all_unique_terms", STRING | FAST);
    let text_field_many_terms = schema_builder.add_text_field("text_many_terms", STRING | FAST);
+    let text_field_few_terms = schema_builder.add_text_field("text_few_terms", STRING | FAST);
    let text_field_few_terms_status =
        schema_builder.add_text_field("text_few_terms_status", STRING | FAST);
    let text_field_1000_terms_zipf =
@@ -504,6 +683,7 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
    let score_field = schema_builder.add_u64_field("score", score_fieldtype.clone());
    let score_field_f64 = schema_builder.add_f64_field("score_f64", score_fieldtype.clone());
    let score_field_i64 = schema_builder.add_i64_field("score_i64", score_fieldtype);
+    let date_field = schema_builder.add_date_field("timestamp", FAST);
    // use tmp dir
    let index = if reuse_index {
        Index::create_in_dir("agg_bench", schema_builder.build())?
@@ -523,6 +703,7 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
    let log_level_distribution =
        WeightedIndex::new(status_field_data.iter().map(|item| item.1)).unwrap();

+    let few_terms_data = ["INFO", "ERROR", "WARN", "DEBUG"];
    let lg_norm = rand_distr::LogNormal::new(2.996f64, 0.979f64).unwrap();

    let many_terms_data = (0..150_000)
@@ -552,12 +733,16 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
            index_writer.add_document(doc!(
                json_field => json!({"mixed_type": 10.0}),
                json_field => json!({"mixed_type": 10.0}),
+                single_term => "single_term",
+                single_term => "single_term",
                text_field => "cool",
                text_field => "cool",
                text_field_all_unique_terms => "cool",
                text_field_all_unique_terms => "coolo",
                text_field_many_terms => "cool",
                text_field_many_terms => "cool",
+                text_field_few_terms => "cool",
+                text_field_few_terms => "cool",
                text_field_few_terms_status => log_level_sample_a,
                text_field_few_terms_status => log_level_sample_b,
                text_field_1000_terms_zipf => term_1000_a.as_str(),
@@ -584,15 +769,18 @@ fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
                json!({"mixed_type": many_terms_data.choose(&mut rng).unwrap().to_string()})
            };
            index_writer.add_document(doc!(
+                single_term => "single_term",
                text_field => "cool",
                json_field => json,
                text_field_all_unique_terms => format!("unique_term_{}", rng.random::<u64>()),
                text_field_many_terms => many_terms_data.choose(&mut rng).unwrap().to_string(),
+                text_field_few_terms => few_terms_data.choose(&mut rng).unwrap().to_string(),
                text_field_few_terms_status => status_field_data[log_level_distribution.sample(&mut rng)].0,
                text_field_1000_terms_zipf => terms_1000[zipf_1000.sample(&mut rng) as usize - 1].as_str(),
                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),
            ))?;
            if cardinality == Cardinality::OptionalSparse {
                for _ in 0..20 {
--- a/benches/and_or_queries.rs
+++ b/benches/and_or_queries.rs
@@ -22,7 +22,7 @@ use rand::rngs::StdRng;
 use rand::SeedableRng;
 use tantivy::collector::sort_key::SortByStaticFastValue;
 use tantivy::collector::{Collector, Count, TopDocs};
-use tantivy::query::{Query, QueryParser};
+use tantivy::query::QueryParser;
 use tantivy::schema::{Schema, FAST, TEXT};
 use tantivy::{doc, Index, Order, ReloadPolicy, Searcher};

@@ -38,7 +38,7 @@ struct BenchIndex {
 /// return two BenchIndex views:
 /// - single_field: QueryParser defaults to only "body"
 /// - multi_field:  QueryParser defaults to ["title", "body"]
-fn build_shared_indices(num_docs: usize, p_a: f32, p_b: f32, p_c: f32) -> (BenchIndex, BenchIndex) {
+fn build_index(num_docs: usize, terms: &[(&str, f32)]) -> (BenchIndex, BenchIndex) {
    // Unified schema (two text fields)
    let mut schema_builder = Schema::builder();
    let f_title = schema_builder.add_text_field("title", TEXT);
@@ -55,32 +55,17 @@ fn build_shared_indices(num_docs: usize, p_a: f32, p_b: f32, p_c: f32) -> (Bench
    {
        let mut writer = index.writer_with_num_threads(1, 500_000_000).unwrap();
        for _ in 0..num_docs {
-            let has_a = rng.random_bool(p_a as f64);
-            let has_b = rng.random_bool(p_b as f64);
-            let has_c = rng.random_bool(p_c as f64);
            let score = rng.random_range(0u64..100u64);
            let score2 = rng.random_range(0u64..100_000u64);
            let mut title_tokens: Vec<&str> = Vec::new();
            let mut body_tokens: Vec<&str> = Vec::new();
-            if has_a {
-                if rng.random_bool(0.1) {
-                    title_tokens.push("a");
-                } else {
-                    body_tokens.push("a");
-                }
-            }
-            if has_b {
-                if rng.random_bool(0.1) {
-                    title_tokens.push("b");
-                } else {
-                    body_tokens.push("b");
-                }
-            }
-            if has_c {
-                if rng.random_bool(0.1) {
-                    title_tokens.push("c");
-                } else {
-                    body_tokens.push("c");
+            for &(tok, prob) in terms {
+                if rng.random_bool(prob as f64) {
+                    if rng.random_bool(0.1) {
+                        title_tokens.push(tok);
+                    } else {
+                        body_tokens.push(tok);
+                    }
                }
            }
            if title_tokens.is_empty() && body_tokens.is_empty() {
@@ -110,59 +95,97 @@ fn build_shared_indices(num_docs: usize, p_a: f32, p_b: f32, p_c: f32) -> (Bench
    let qp_single = QueryParser::for_index(&index, vec![f_body]);
    let qp_multi = QueryParser::for_index(&index, vec![f_title, f_body]);

-    let single_view = BenchIndex {
+    let only_title = BenchIndex {
        index: index.clone(),
        searcher: searcher.clone(),
        query_parser: qp_single,
    };
-    let multi_view = BenchIndex {
+    let title_and_body = BenchIndex {
        index,
        searcher,
        query_parser: qp_multi,
    };
-    (single_view, multi_view)
+    (only_title, title_and_body)
+}
+
+fn format_pct(p: f32) -> String {
+    let pct = (p as f64) * 100.0;
+    let rounded = (pct * 1_000_000.0).round() / 1_000_000.0;
+    if rounded.fract() <= 0.001 {
+        format!("{}%", rounded as u64)
+    } else {
+        format!("{}%", rounded)
+    }
+}
+
+fn query_label(query_str: &str, term_pcts: &[(&str, String)]) -> String {
+    let mut label = query_str.to_string();
+    for (term, pct) in term_pcts {
+        label = label.replace(term, pct);
+    }
+    label.replace(' ', "_")
 }

 fn main() {
-    // Prepare corpora with varying selectivity. Build one index per corpus
-    // and derive two views (single-field vs multi-field) from it.
-    let scenarios = vec![
+    // terms with varying selectivity, ordered from rarest to most common.
+    // With 1M docs, we expect:
+    // a: 0.01% (100), b: 1% (10k), c: 5% (50k), d: 15% (150k), e: 30% (300k)
+    let num_docs = 1_000_000;
+    let terms: &[(&str, f32)] = &[
+        ("a", 0.0001),
+        ("b", 0.01),
+        ("c", 0.05),
+        ("d", 0.15),
+        ("e", 0.30),
+    ];
+
+    let queries: &[(&str, &[&str])] = &[
        (
-            "N=1M, p(a)=5%, p(b)=1%, p(c)=15%".to_string(),
-            1_000_000,
-            0.05,
-            0.01,
-            0.15,
+            "only_union",
+            &["c OR b", "c OR b OR d", "c OR e", "e OR a"] as &[&str],
        ),
        (
-            "N=1M, p(a)=1%, p(b)=1%, p(c)=15%".to_string(),
-            1_000_000,
-            0.01,
-            0.01,
-            0.15,
+            "only_intersection",
+            &["+c +b", "+c +b +d", "+c +e", "+e +a"] as &[&str],
+        ),
+        (
+            "union_intersection",
+            &["+c +(b OR d)", "+e +(c OR a)", "+(c OR b) +(d OR e)"] as &[&str],
        ),
    ];

-    let queries = &["a", "+a +b", "+a +b +c", "a OR b", "a OR b OR c"];
-
    let mut runner = BenchRunner::new();
-    for (label, n, pa, pb, pc) in scenarios {
-        let (single_view, multi_view) = build_shared_indices(n, pa, pb, pc);
+    let (only_title, title_and_body) = build_index(num_docs, terms);
+    let term_pcts: Vec<(&str, String)> = terms
+        .iter()
+        .map(|&(term, p)| (term, format_pct(p)))
+        .collect();

-        for (view_name, bench_index) in [("single_field", single_view), ("multi_field", multi_view)]
-        {
-            // Single-field group: default field is body only
-            let mut group = runner.new_group();
-            group.set_name(format!("{} — {}", view_name, label));
-            for query_str in queries {
+    for (view_name, bench_index) in [
+        ("single_field", only_title),
+        ("multi_field", title_and_body),
+    ] {
+        for (category_name, category_queries) in queries {
+            for query_str in *category_queries {
+                let mut group = runner.new_group();
+                let query_label = query_label(query_str, &term_pcts);
+                group.set_name(format!("{}_{}_{}", view_name, category_name, query_label));
                add_bench_task(&mut group, &bench_index, query_str, Count, "count");
                add_bench_task(
                    &mut group,
                    &bench_index,
                    query_str,
                    TopDocs::with_limit(10).order_by_score(),
-                    "top10",
+                    "top10_inv_idx",
                );
+                add_bench_task(
+                    &mut group,
+                    &bench_index,
+                    query_str,
+                    (Count, TopDocs::with_limit(10).order_by_score()),
+                    "count+top10",
+                );
+
                add_bench_task(
                    &mut group,
                    &bench_index,
@@ -180,39 +203,47 @@ fn main() {
                    )),
                    "top10_by_2ff",
                );
+
+                group.run();
            }
-            group.run();
        }
    }
 }

+trait FruitCount {
+    fn count(&self) -> usize;
+}
+
+impl FruitCount for usize {
+    fn count(&self) -> usize {
+        *self
+    }
+}
+
+impl<T> FruitCount for Vec<T> {
+    fn count(&self) -> usize {
+        self.len()
+    }
+}
+
+impl<A: FruitCount, B> FruitCount for (A, B) {
+    fn count(&self) -> usize {
+        self.0.count()
+    }
+}
+
 fn add_bench_task<C: Collector + 'static>(
    bench_group: &mut BenchGroup,
    bench_index: &BenchIndex,
    query_str: &str,
    collector: C,
    collector_name: &str,
-) {
-    let task_name = format!("{}_{}", query_str.replace(" ", "_"), collector_name);
+) where
+    C::Fruit: FruitCount,
+{
    let query = bench_index.query_parser.parse_query(query_str).unwrap();
-    let search_task = SearchTask {
-        searcher: bench_index.searcher.clone(),
-        collector,
-        query,
-    };
-    bench_group.register(task_name, move |_| black_box(search_task.run()));
-}
-
-struct SearchTask<C: Collector> {
-    searcher: Searcher,
-    collector: C,
-    query: Box<dyn Query>,
-}
-
-impl<C: Collector> SearchTask<C> {
-    #[inline(never)]
-    pub fn run(&self) -> usize {
-        self.searcher.search(&self.query, &self.collector).unwrap();
-        1
-    }
+    let searcher = bench_index.searcher.clone();
+    bench_group.register(collector_name.to_string(), move |_| {
+        black_box(searcher.search(&query, &collector).unwrap().count())
+    });
 }
--- a/benches/intersection_bench.rs
+++ b/benches/intersection_bench.rs
@@ -0,0 +1,149 @@
+// Benchmarks top-K intersection of term scorers (block_wand_intersection).
+//
+// What's measured:
+// - Conjunctive queries (+a +b, +a +b +c) with top-10 by score
+// - Varying doc-frequency balance between terms (balanced, skewed, very skewed)
+// - Realistic term frequencies (geometric distribution, mostly low)
+// - 1M-doc single segment
+//
+// Run with: cargo bench --bench intersection_bench
+
+use binggan::{black_box, BenchRunner};
+use rand::prelude::*;
+use rand::rngs::StdRng;
+use rand::SeedableRng;
+use tantivy::collector::TopDocs;
+use tantivy::query::QueryParser;
+use tantivy::schema::{Schema, TEXT};
+use tantivy::{doc, Index, ReloadPolicy, Searcher};
+
+const NUM_DOCS: usize = 1_000_000;
+
+struct BenchIndex {
+    searcher: Searcher,
+    query_parser: QueryParser,
+}
+
+/// Generate term frequency from a geometric-like distribution.
+/// Most values are 1, a few are 2-3, rarely higher.
+/// p controls the decay: higher p → more weight on tf=1.
+fn random_term_freq(rng: &mut StdRng, p: f64) -> u32 {
+    let mut tf = 1u32;
+    while tf < 10 && rng.random_bool(1.0 - p) {
+        tf += 1;
+    }
+    tf
+}
+
+/// Build an index with three terms (a, b, c) with given doc-frequency probabilities.
+/// Each term occurrence has a realistic term frequency (geometric distribution).
+/// Field length is padded with filler tokens to create varied fieldnorms.
+fn build_index(p_a: f64, p_b: f64, p_c: f64) -> BenchIndex {
+    let mut schema_builder = Schema::builder();
+    let body = schema_builder.add_text_field("body", TEXT);
+    let schema = schema_builder.build();
+    let index = Index::create_in_ram(schema);
+
+    let mut rng = StdRng::from_seed([42u8; 32]);
+
+    {
+        let mut writer = index.writer_with_num_threads(1, 500_000_000).unwrap();
+        for _ in 0..NUM_DOCS {
+            let mut tokens: Vec<String> = Vec::new();
+
+            if rng.random_bool(p_a) {
+                let tf = random_term_freq(&mut rng, 0.7);
+                for _ in 0..tf {
+                    tokens.push("aaa".to_string());
+                }
+            }
+            if rng.random_bool(p_b) {
+                let tf = random_term_freq(&mut rng, 0.7);
+                for _ in 0..tf {
+                    tokens.push("bbb".to_string());
+                }
+            }
+            if rng.random_bool(p_c) {
+                let tf = random_term_freq(&mut rng, 0.7);
+                for _ in 0..tf {
+                    tokens.push("ccc".to_string());
+                }
+            }
+
+            // Pad with filler to create varied field lengths (5-30 tokens).
+            let filler_count = rng.random_range(5u32..30u32);
+            for _ in 0..filler_count {
+                tokens.push("filler".to_string());
+            }
+
+            let text = tokens.join(" ");
+            writer.add_document(doc!(body => text)).unwrap();
+        }
+        writer.commit().unwrap();
+    }
+
+    let reader = index
+        .reader_builder()
+        .reload_policy(ReloadPolicy::Manual)
+        .try_into()
+        .unwrap();
+    let searcher = reader.searcher();
+    let query_parser = QueryParser::for_index(&index, vec![body]);
+
+    BenchIndex {
+        searcher,
+        query_parser,
+    }
+}
+
+fn main() {
+    // Scenarios: (label, p_a, p_b, p_c)
+    //
+    // "balanced":    all terms ~10% → intersection ~1% of docs
+    // "skewed":      one common (50%), one rare (2%) → intersection ~1%
+    // "very_skewed": one very common (80%), one very rare (0.5%) → intersection ~0.4%
+    // "three_balanced": three terms ~20% each → intersection ~0.8%
+    // "three_skewed":   50% / 10% / 2% → intersection ~0.1%
+    let scenarios: Vec<(&str, f64, f64, f64)> = vec![
+        ("balanced_10%_10%", 0.10, 0.10, 0.0),
+        ("skewed_50%_2%", 0.50, 0.02, 0.0),
+        ("very_skewed_80%_0.5%", 0.80, 0.005, 0.0),
+        ("three_balanced_20%_20%_20%", 0.20, 0.20, 0.20),
+        ("three_skewed_50%_10%_2%", 0.50, 0.10, 0.02),
+    ];
+
+    let mut runner = BenchRunner::new();
+
+    for (label, p_a, p_b, p_c) in &scenarios {
+        let bench_index = build_index(*p_a, *p_b, *p_c);
+
+        let mut group = runner.new_group();
+        group.set_name(format!("intersection — {label}"));
+
+        // Two-term intersection
+        if *p_a > 0.0 && *p_b > 0.0 {
+            let query_str = "+aaa +bbb";
+            let query = bench_index.query_parser.parse_query(query_str).unwrap();
+            let searcher = bench_index.searcher.clone();
+            group.register(format!("{query_str} top10"), move |_| {
+                let collector = TopDocs::with_limit(10).order_by_score();
+                black_box(searcher.search(&query, &collector).unwrap());
+                1usize
+            });
+        }
+
+        // Three-term intersection
+        if *p_c > 0.0 {
+            let query_str = "+aaa +bbb +ccc";
+            let query = bench_index.query_parser.parse_query(query_str).unwrap();
+            let searcher = bench_index.searcher.clone();
+            group.register(format!("{query_str} top10"), move |_| {
+                let collector = TopDocs::with_limit(10).order_by_score();
+                black_box(searcher.search(&query, &collector).unwrap());
+                1usize
+            });
+        }
+
+        group.run();
+    }
+}
--- a/benches/query_parser_nested.rs
+++ b/benches/query_parser_nested.rs
@@ -0,0 +1,35 @@
+// Benchmark for the query grammar parsing deeply nested queries.
+//
+// Regression guard for https://github.com/quickwit-oss/tantivy/issues/2498:
+// at depth 20/21 the old parser took 0.87 s / 1.72 s respectively because
+// `ast()` retried `occur_leaf` on backtrack, giving O(2^n) time. With the
+// fix parsing is linear and completes in microseconds.
+//
+// Run with: `cargo bench --bench query_parser_nested`.
+
+use binggan::{black_box, BenchRunner};
+use tantivy::query_grammar::parse_query;
+
+fn nested_query(depth: usize, leading_plus: bool) -> String {
+    let leading = "(".repeat(depth);
+    let trailing = ")".repeat(depth);
+    let prefix = if leading_plus { "+" } else { "" };
+    format!("{prefix}{leading}title:test{trailing}")
+}
+
+fn main() {
+    let mut runner = BenchRunner::new();
+
+    for depth in [20, 21] {
+        for leading_plus in [false, true] {
+            let query = nested_query(depth, leading_plus);
+            let label = format!(
+                "parse_nested_depth_{depth}_{}",
+                if leading_plus { "plus" } else { "plain" },
+            );
+            runner.bench_function(&label, move |_| {
+                black_box(parse_query(black_box(&query)).unwrap());
+            });
+        }
+    }
+}
--- a/benches/str_search_and_get.rs
+++ b/benches/str_search_and_get.rs
@@ -45,7 +45,7 @@ fn build_shared_indices(num_docs: usize, distribution: &str) -> BenchIndex {
        match distribution {
            "dense_random" => {
                for _doc_id in 0..num_docs {
-                    let suffix = rng.gen_range(0u64..1000u64);
+                    let suffix = rng.random_range(0u64..1000u64);
                    let str_val = format!("str_{:03}", suffix);

                    writer
@@ -71,7 +71,7 @@ fn build_shared_indices(num_docs: usize, distribution: &str) -> BenchIndex {
            }
            "sparse_random" => {
                for _doc_id in 0..num_docs {
-                    let suffix = rng.gen_range(0u64..1000000u64);
+                    let suffix = rng.random_range(0u64..1000000u64);
                    let str_val = format!("str_{:07}", suffix);

                    writer
--- a/bitpacker/Cargo.toml
+++ b/bitpacker/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-bitpacker"
-version = "0.9.0"
+version = "0.10.0"
 edition = "2024"
 authors = ["Paul Masurel <paul.masurel@gmail.com>"]
 license = "MIT"
--- a/columnar/Cargo.toml
+++ b/columnar/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-columnar"
-version = "0.6.0"
+version = "0.7.0"
 edition = "2024"
 license = "MIT"
 homepage = "https://github.com/quickwit-oss/tantivy"
@@ -12,10 +12,10 @@ categories = ["database-implementations", "data-structures", "compression"]
 itertools = "0.14.0"
 fastdivide = "0.4.0"

-stacker = { version= "0.6", path = "../stacker", package="tantivy-stacker"}
-sstable = { version= "0.6", path = "../sstable", package = "tantivy-sstable" }
-common = { version= "0.10", path = "../common", package = "tantivy-common" }
-tantivy-bitpacker = { version= "0.9", path = "../bitpacker/" }
+stacker = { version= "0.7", path = "../stacker", package="tantivy-stacker"}
+sstable = { version= "0.7", path = "../sstable", package = "tantivy-sstable" }
+common = { version= "0.11", path = "../common", package = "tantivy-common" }
+tantivy-bitpacker = { version= "0.10", path = "../bitpacker/" }
 serde = "1.0.152"
 downcast-rs = "2.0.1"

@@ -23,7 +23,7 @@ downcast-rs = "2.0.1"
 proptest = "1"
 more-asserts = "0.3.1"
 rand = "0.9"
-binggan = "0.14.0"
+binggan = "0.16.1"

 [[bench]]
 name = "bench_merge"
--- a/columnar/src/block_accessor.rs
+++ b/columnar/src/block_accessor.rs
@@ -33,14 +33,14 @@ impl<T: PartialOrd + Copy + std::fmt::Debug + Send + Sync + 'static + Default>
        &mut self,
        docs: &[u32],
        accessor: &Column<T>,
-        missing: Option<T>,
+        missing_opt: Option<T>,
    ) {
        self.fetch_block(docs, accessor);
        // no missing values
        if accessor.index.get_cardinality().is_full() {
            return;
        }
-        let Some(missing) = missing else {
+        let Some(missing) = missing_opt else {
            return;
        };

@@ -58,6 +58,78 @@ impl<T: PartialOrd + Copy + std::fmt::Debug + Send + Sync + 'static + Default>
        }
    }

+    /// Like `fetch_block_with_missing`, but deduplicates (doc_id, value) pairs
+    /// so that each unique value per document is returned only once.
+    ///
+    /// This is necessary for correct document counting in aggregations,
+    /// where multi-valued fields can produce duplicate entries that inflate counts.
+    #[inline]
+    pub fn fetch_block_with_missing_unique_per_doc(
+        &mut self,
+        docs: &[u32],
+        accessor: &Column<T>,
+        missing: Option<T>,
+    ) where
+        T: Ord,
+    {
+        self.fetch_block_with_missing(docs, accessor, missing);
+        if accessor.index.get_cardinality().is_multivalue() {
+            self.dedup_docid_val_pairs();
+        }
+    }
+
+    /// Removes duplicate (doc_id, value) pairs from the caches.
+    ///
+    /// After `fetch_block`, entries are sorted by doc_id, but values within
+    /// the same doc may not be sorted (e.g. `(0,1), (0,2), (0,1)`).
+    /// We group consecutive entries by doc_id, sort values within each group
+    /// if it has more than 2 elements, then deduplicate adjacent pairs.
+    ///
+    /// Skips entirely if no doc_id appears more than once in the block.
+    fn dedup_docid_val_pairs(&mut self)
+    where T: Ord {
+        if self.docid_cache.len() <= 1 {
+            return;
+        }
+
+        // Quick check: if no consecutive doc_ids are equal, no dedup needed.
+        let has_multivalue = self.docid_cache.windows(2).any(|w| w[0] == w[1]);
+        if !has_multivalue {
+            return;
+        }
+
+        // Sort values within each doc_id group so duplicates become adjacent.
+        let mut start = 0;
+        while start < self.docid_cache.len() {
+            let doc = self.docid_cache[start];
+            let mut end = start + 1;
+            while end < self.docid_cache.len() && self.docid_cache[end] == doc {
+                end += 1;
+            }
+            if end - start > 2 {
+                self.val_cache[start..end].sort();
+            }
+            start = end;
+        }
+
+        // Now duplicates are adjacent — deduplicate in place.
+        let mut write = 0;
+        for read in 1..self.docid_cache.len() {
+            if self.docid_cache[read] != self.docid_cache[write]
+                || self.val_cache[read] != self.val_cache[write]
+            {
+                write += 1;
+                if write != read {
+                    self.docid_cache[write] = self.docid_cache[read];
+                    self.val_cache[write] = self.val_cache[read];
+                }
+            }
+        }
+        let new_len = write + 1;
+        self.docid_cache.truncate(new_len);
+        self.val_cache.truncate(new_len);
+    }
+
    #[inline]
    pub fn iter_vals(&self) -> impl Iterator<Item = T> + '_ {
        self.val_cache.iter().cloned()
@@ -119,6 +191,7 @@ where F: FnMut(u32) {
 }

 #[cfg(test)]
+#[allow(clippy::field_reassign_with_default)]
 mod tests {
    use super::*;

@@ -163,4 +236,56 @@ mod tests {

        assert_eq!(missing_docs, vec![1, 2, 3, 4, 5]);
    }
+
+    #[test]
+    fn test_dedup_docid_val_pairs_consecutive() {
+        let mut accessor = ColumnBlockAccessor::<u64>::default();
+        accessor.docid_cache = vec![0, 0, 2, 3];
+        accessor.val_cache = vec![10, 10, 10, 10];
+        accessor.dedup_docid_val_pairs();
+        assert_eq!(accessor.docid_cache, vec![0, 2, 3]);
+        assert_eq!(accessor.val_cache, vec![10, 10, 10]);
+    }
+
+    #[test]
+    fn test_dedup_docid_val_pairs_non_consecutive() {
+        // (0,1), (0,2), (0,1) — duplicate value not adjacent
+        let mut accessor = ColumnBlockAccessor::<u64>::default();
+        accessor.docid_cache = vec![0, 0, 0];
+        accessor.val_cache = vec![1, 2, 1];
+        accessor.dedup_docid_val_pairs();
+        assert_eq!(accessor.docid_cache, vec![0, 0]);
+        assert_eq!(accessor.val_cache, vec![1, 2]);
+    }
+
+    #[test]
+    fn test_dedup_docid_val_pairs_multi_doc() {
+        // doc 0: values [3, 1, 3], doc 1: values [5, 5]
+        let mut accessor = ColumnBlockAccessor::<u64>::default();
+        accessor.docid_cache = vec![0, 0, 0, 1, 1];
+        accessor.val_cache = vec![3, 1, 3, 5, 5];
+        accessor.dedup_docid_val_pairs();
+        assert_eq!(accessor.docid_cache, vec![0, 0, 1]);
+        assert_eq!(accessor.val_cache, vec![1, 3, 5]);
+    }
+
+    #[test]
+    fn test_dedup_docid_val_pairs_no_duplicates() {
+        let mut accessor = ColumnBlockAccessor::<u64>::default();
+        accessor.docid_cache = vec![0, 0, 1];
+        accessor.val_cache = vec![1, 2, 3];
+        accessor.dedup_docid_val_pairs();
+        assert_eq!(accessor.docid_cache, vec![0, 0, 1]);
+        assert_eq!(accessor.val_cache, vec![1, 2, 3]);
+    }
+
+    #[test]
+    fn test_dedup_docid_val_pairs_single_element() {
+        let mut accessor = ColumnBlockAccessor::<u64>::default();
+        accessor.docid_cache = vec![0];
+        accessor.val_cache = vec![1];
+        accessor.dedup_docid_val_pairs();
+        assert_eq!(accessor.docid_cache, vec![0]);
+        assert_eq!(accessor.val_cache, vec![1]);
+    }
 }
--- a/columnar/src/column_values/mod.rs
+++ b/columnar/src/column_values/mod.rs
@@ -31,7 +31,7 @@ pub use u64_based::{
    serialize_and_load_u64_based_column_values, serialize_u64_based_column_values,
 };
 pub use u128_based::{
-    CompactSpaceU64Accessor, open_u128_as_compact_u64, open_u128_mapped,
+    CompactHit, CompactSpaceU64Accessor, open_u128_as_compact_u64, open_u128_mapped,
    serialize_column_values_u128,
 };
 pub use vec_column::VecColumn;
--- a/columnar/src/column_values/u128_based/compact_space/mod.rs
+++ b/columnar/src/column_values/u128_based/compact_space/mod.rs
@@ -292,6 +292,19 @@ impl BinarySerializable for IPCodecParams {
    }
 }

+/// Represents the result of looking up a u128 value in the compact space.
+///
+/// If a value is outside the compact space, the next compact value is returned.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum CompactHit {
+    /// The value exists in the compact space
+    Exact(u32),
+    /// The value does not exist in the compact space, but the next higher value does
+    Next(u32),
+    /// The value is greater than the maximum compact value
+    AfterLast,
+}
+
 /// Exposes the compact space compressed values as u64.
 ///
 /// This allows faster access to the values, as u64 is faster to work with than u128.
@@ -309,6 +322,11 @@ impl CompactSpaceU64Accessor {
    pub fn compact_to_u128(&self, compact: u32) -> u128 {
        self.0.compact_to_u128(compact)
    }
+
+    /// Finds the next compact space value for a given u128 value.
+    pub fn u128_to_next_compact(&self, value: u128) -> CompactHit {
+        self.0.u128_to_next_compact(value)
+    }
 }

 impl ColumnValues<u64> for CompactSpaceU64Accessor {
@@ -441,6 +459,21 @@ impl CompactSpaceDecompressor {
        self.params.compact_space.u128_to_compact(value)
    }

+    /// Finds the next compact space value for a given u128 value.
+    pub fn u128_to_next_compact(&self, value: u128) -> CompactHit {
+        match self.u128_to_compact(value) {
+            Ok(compact) => CompactHit::Exact(compact),
+            Err(pos) => {
+                if pos >= self.params.compact_space.ranges_mapping.len() {
+                    CompactHit::AfterLast
+                } else {
+                    let next_range = &self.params.compact_space.ranges_mapping[pos];
+                    CompactHit::Next(next_range.compact_start)
+                }
+            }
+        }
+    }
+
    fn compact_to_u128(&self, compact: u32) -> u128 {
        self.params.compact_space.compact_to_u128(compact)
    }
@@ -823,6 +856,41 @@ mod tests {
        let _data = test_aux_vals(vals);
    }

+    #[test]
+    fn test_u128_to_next_compact() {
+        let vals = &[100u128, 200u128, 1_000_000_000u128, 1_000_000_100u128];
+        let mut data = test_aux_vals(vals);
+
+        let _header = U128Header::deserialize(&mut data);
+        let decomp = CompactSpaceDecompressor::open(data).unwrap();
+
+        // Test value that's already in a range
+        let compact_100 = decomp.u128_to_compact(100).unwrap();
+        assert_eq!(
+            decomp.u128_to_next_compact(100),
+            CompactHit::Exact(compact_100)
+        );
+
+        // Test value between two ranges
+        let compact_million = decomp.u128_to_compact(1_000_000_000).unwrap();
+        assert_eq!(
+            decomp.u128_to_next_compact(250),
+            CompactHit::Next(compact_million)
+        );
+
+        // Test value before the first range
+        assert_eq!(
+            decomp.u128_to_next_compact(50),
+            CompactHit::Next(compact_100)
+        );
+
+        // Test value after the last range
+        assert_eq!(
+            decomp.u128_to_next_compact(10_000_000_000),
+            CompactHit::AfterLast
+        );
+    }
+
    use proptest::prelude::*;

    fn num_strategy() -> impl Strategy<Value = u128> {
--- a/columnar/src/column_values/u128_based/mod.rs
+++ b/columnar/src/column_values/u128_based/mod.rs
@@ -7,7 +7,7 @@ mod compact_space;

 use common::{BinarySerializable, OwnedBytes, VInt};
 pub use compact_space::{
-    CompactSpaceCompressor, CompactSpaceDecompressor, CompactSpaceU64Accessor,
+    CompactHit, CompactSpaceCompressor, CompactSpaceDecompressor, CompactSpaceU64Accessor,
 };

 use crate::column_values::monotonic_map_column;
--- a/columnar/src/lib.rs
+++ b/columnar/src/lib.rs
@@ -59,7 +59,7 @@ pub struct RowAddr {
    pub row_id: RowId,
 }

-pub use sstable::Dictionary;
+pub use sstable::{Dictionary, TermOrdHit};
 pub type Streamer<'a> = sstable::Streamer<'a, VoidSSTable>;

 pub use common::DateTime;
--- a/common/Cargo.toml
+++ b/common/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-common"
-version = "0.10.0"
+version = "0.11.0"
 authors = ["Paul Masurel <paul@quickwit.io>", "Pascal Seitz <pascal@quickwit.io>"]
 license = "MIT"
 edition = "2024"
@@ -19,6 +19,6 @@ time = { version = "0.3.47", features = ["serde-well-known"] }
 serde = { version = "1.0.136", features = ["derive"] }

 [dev-dependencies]
-binggan = "0.14.0"
+binggan = "0.16.1"
 proptest = "1.0.0"
 rand = "0.9"
--- a/common/src/bitset.rs
+++ b/common/src/bitset.rs
@@ -47,6 +47,9 @@ impl TinySet {
        TinySet(val)
    }

+    /// An empty `TinySet` constant.
+    pub const EMPTY: TinySet = TinySet(0u64);
+
    /// Returns an empty `TinySet`.
    #[inline]
    pub fn empty() -> TinySet {
@@ -153,7 +156,22 @@ impl TinySet {
            None
        } else {
            let lowest = self.0.trailing_zeros();
-            self.0 ^= TinySet::singleton(lowest).0;
+            // Kernighan's trick: `n &= n - 1` clears the lowest set bit
+            // without depending on `lowest`. This lets the CPU execute
+            // `trailing_zeros` and the bit-clear in parallel instead of
+            // serializing them.
+            //
+            // The previous form `self.0 ^= 1 << lowest` needs the result of
+            // `trailing_zeros` before it can shift, creating a dependency chain:
+            //   ARM64: rbit → clz → lsl → eor
+            //   x86:   tzcnt → btc
+            //
+            // With Kernighan's trick the clear path is independent of the count:
+            //   ARM64: sub → and  (trailing_zeros runs in parallel)
+            //   x86:   blsr       (tzcnt runs in parallel)
+            //
+            // https://godbolt.org/z/fnfrP1T5f
+            self.0 &= self.0 - 1;
            Some(lowest)
        }
    }
--- a/query-grammar/Cargo.toml
+++ b/query-grammar/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-query-grammar"
-version = "0.25.0"
+version = "0.26.0"
 authors = ["Paul Masurel <paul.masurel@gmail.com>"]
 license = "MIT"
 categories = ["database-implementations", "data-structures"]
--- a/query-grammar/src/query_grammar.rs
+++ b/query-grammar/src/query_grammar.rs
@@ -1045,18 +1045,43 @@ fn operand_leaf(inp: &str) -> IResult<&str, (Option<BinaryOperand>, Option<Occur
 }

 fn ast(inp: &str) -> IResult<&str, UserInputAst> {
-    let boolean_expr = map_res(
-        separated_pair(occur_leaf, multispace1, many1(operand_leaf)),
-        |(left, right)| aggregate_binary_expressions(left, right),
-    );
-    let single_leaf = map(occur_leaf, |(occur, ast)| {
-        if occur == Some(Occur::MustNot) {
-            ast.unary(Occur::MustNot)
-        } else {
-            ast
-        }
-    });
-    delimited(multispace0, alt((boolean_expr, single_leaf)), multispace0)(inp)
+    // Parse `occur_leaf` once, then conditionally extend into a boolean
+    // expression. The previous implementation used `alt((boolean_expr,
+    // single_leaf))` which, when the input was a single leaf with no
+    // following operand, would parse `occur_leaf` once for `boolean_expr`,
+    // fail at `multispace1`, backtrack, then re-parse `occur_leaf` for
+    // `single_leaf`. With recursively-nested groups like `(+(+(+a)))`, that
+    // doubling at every level produced O(2^n) parse time. Parsing once and
+    // peeking ahead for the operand keeps it O(n).
+    delimited(
+        multispace0,
+        |inp| {
+            let (rest, first) = occur_leaf(inp)?;
+            // Only fall back on `Err::Error` (recoverable), mirroring
+            // `alt`'s behaviour. `Err::Failure` and `Err::Incomplete`
+            // must propagate so cut points and streaming needs are not
+            // accidentally swallowed if they are ever introduced in the
+            // operand parsers.
+            match preceded(multispace1, many1(operand_leaf))(rest) {
+                Ok((rest, more)) => {
+                    let combined = aggregate_binary_expressions(first, more)
+                        .map_err(|_| nom::Err::Error(Error::new(inp, ErrorKind::MapRes)))?;
+                    Ok((rest, combined))
+                }
+                Err(nom::Err::Error(_)) => {
+                    let (occur, ast) = first;
+                    let single = if occur == Some(Occur::MustNot) {
+                        ast.unary(Occur::MustNot)
+                    } else {
+                        ast
+                    };
+                    Ok((rest, single))
+                }
+                Err(e) => Err(e),
+            }
+        },
+        multispace0,
+    )(inp)
 }

 fn ast_infallible(inp: &str) -> JResult<&str, UserInputAst> {
@@ -1891,4 +1916,23 @@ mod test {
            r#"(+"field":'happy tax payer' +"other_field":1)"#,
        );
    }
+
+    // Regression test for https://github.com/quickwit-oss/tantivy/issues/2498:
+    // deeply nested parenthesized queries used to take O(2^n) time because the
+    // top-level `ast()` parser tried `boolean_expr` first and re-parsed the
+    // inner `occur_leaf` when it backtracked to `single_leaf`. Depth 60 would
+    // take ~10^18 operations under the regression; with the fix it parses
+    // instantly. We use `test_parse_query_to_ast_helper` so this test would
+    // never finish if the regression returned.
+    #[test]
+    fn test_parse_deeply_nested_query() {
+        let depth = 60;
+        let leading: String = "(".repeat(depth);
+        let trailing: String = ")".repeat(depth);
+        let query = format!("{leading}title:test{trailing}");
+        test_parse_query_to_ast_helper(&query, r#""title":test"#);
+
+        let query_with_plus = format!("+{leading}title:test{trailing}");
+        test_parse_query_to_ast_helper(&query_with_plus, r#""title":test"#);
+    }
 }
--- a/src/aggregation/agg_data.rs
+++ b/src/aggregation/agg_data.rs
@@ -10,9 +10,10 @@ 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, FilterAggReqData,
-    HistogramAggReqData, HistogramBounds, IncludeExcludeParam, MissingTermAggReqData,
-    RangeAggReqData, SegmentHistogramCollector, TermMissingAgg, TermsAggReqData, TermsAggregation,
+    build_segment_filter_collector, build_segment_range_collector, CompositeAggReqData,
+    CompositeAggregation, CompositeSourceAccessors, FilterAggReqData, HistogramAggReqData,
+    HistogramBounds, IncludeExcludeParam, MissingTermAggReqData, RangeAggReqData,
+    SegmentHistogramCollector, TermMissingAgg, TermsAggReqData, TermsAggregation,
    TermsAggregationInternal,
 };
 use crate::aggregation::metric::{
@@ -73,6 +74,12 @@ impl AggregationsSegmentCtx {
        self.per_request.filter_req_data.push(Some(Box::new(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.len() - 1
+    }

    #[inline]
    pub(crate) fn get_term_req_data(&self, idx: usize) -> &TermsAggReqData {
@@ -108,6 +115,12 @@ impl AggregationsSegmentCtx {
            .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 ----------

@@ -181,6 +194,25 @@ impl AggregationsSegmentCtx {
        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
@@ -208,6 +240,8 @@ pub struct PerRequestAggSegCtx {
    pub top_hits_req_data: Vec<TopHitsAggReqData>,
    /// 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>>>,

    /// Request tree used to build collectors.
    pub agg_tree: Vec<AggRefNode>,
@@ -255,6 +289,11 @@ impl PerRequestAggSegCtx {
                .iter()
                .map(|t| t.get_memory_consumption())
                .sum::<usize>()
+            + self
+                .composite_req_data
+                .iter()
+                .map(|b| b.as_ref().map(|d| d.get_memory_consumption()).unwrap_or(0))
+                .sum::<usize>()
            + self.agg_tree.len() * std::mem::size_of::<AggRefNode>()
    }

@@ -291,6 +330,11 @@ impl PerRequestAggSegCtx {
                .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(),
        }
    }

@@ -417,6 +461,11 @@ pub(crate) fn build_segment_agg_collector(
        )?)),
        AggKind::Range => Ok(build_segment_range_collector(req, node)?),
        AggKind::Filter => build_segment_filter_collector(req, node),
+        AggKind::Composite => Ok(Box::new(
+            crate::aggregation::bucket::SegmentCompositeCollector::from_req_and_validate(
+                req, node,
+            )?,
+        )),
    }
 }

@@ -447,6 +496,7 @@ pub enum AggKind {
    DateHistogram,
    Range,
    Filter,
+    Composite,
 }

 impl AggKind {
@@ -462,6 +512,7 @@ impl AggKind {
            AggKind::DateHistogram => "DateHistogram",
            AggKind::Range => "Range",
            AggKind::Filter => "Filter",
+            AggKind::Composite => "Composite",
        }
    }
 }
@@ -709,6 +760,14 @@ fn build_nodes(
                children,
            }])
        }
+        AggregationVariants::Composite(composite_req) => Ok(vec![build_composite_node(
+            agg_name,
+            reader,
+            segment_ordinal,
+            data,
+            &req.sub_aggregation,
+            composite_req,
+        )?]),
        AggregationVariants::Filter(filter_req) => {
            // Build the query and evaluator upfront
            let schema = reader.schema();
@@ -743,6 +802,35 @@ fn build_nodes(
    }
 }

+fn build_composite_node(
+    agg_name: &str,
+    reader: &SegmentReader,
+    _segment_ordinal: SegmentOrdinal,
+    data: &mut AggregationsSegmentCtx,
+    sub_aggs: &Aggregations,
+    req: &CompositeAggregation,
+) -> crate::Result<AggRefNode> {
+    let mut composite_accessors = Vec::with_capacity(req.sources.len());
+    for source in &req.sources {
+        let source_after_key_opt = req.after.get(source.name()).map(|k| &k.0);
+        let source_accessor =
+            CompositeSourceAccessors::build_for_source(reader, source, source_after_key_opt)?;
+        composite_accessors.push(source_accessor);
+    }
+    let agg = CompositeAggReqData {
+        name: agg_name.to_string(),
+        req: req.clone(),
+        composite_accessors,
+    };
+    let idx = data.push_composite_req_data(agg);
+    let children = build_children(sub_aggs, reader, _segment_ordinal, data)?;
+    Ok(AggRefNode {
+        kind: AggKind::Composite,
+        idx_in_req_data: idx,
+        children,
+    })
+}
+
 fn build_children(
    aggs: &Aggregations,
    reader: &SegmentReader,
@@ -897,8 +985,12 @@ fn build_terms_or_cardinality_nodes(
                    let str_col = str_dict_column
                        .as_ref()
                        .expect("str_dict_column must exist for string column");
-                    allowed_term_ids =
-                        build_allowed_term_ids_for_str(str_col, &req.include, &req.exclude)?;
+                    allowed_term_ids = build_allowed_term_ids_for_str(
+                        str_col,
+                        &req.include,
+                        &req.exclude,
+                        missing.is_some(),
+                    )?;
                };
                let idx_in_req_data = data.push_term_req_data(TermsAggReqData {
                    accessor,
@@ -937,16 +1029,21 @@ fn build_terms_or_cardinality_nodes(

 /// Builds a single BitSet of allowed term ordinals for a string dictionary column according to
 /// include/exclude parameters.
+///
+/// When `reserve_missing_sentinel` is true, the bitset will have 1 additional slot for the missing
+/// term ordinal
 fn build_allowed_term_ids_for_str(
    str_col: &StrColumn,
    include: &Option<IncludeExcludeParam>,
    exclude: &Option<IncludeExcludeParam>,
+    reserve_missing_sentinel: bool,
 ) -> crate::Result<Option<BitSet>> {
    let mut allowed: Option<BitSet> = None;
-    let num_terms = str_col.dictionary().num_terms() as u32;
+    let missing_sentinel_adjustment = if reserve_missing_sentinel { 1 } else { 0 };
+    let allowed_capacity = str_col.dictionary().num_terms() as u32 + missing_sentinel_adjustment;
    if let Some(include) = include {
        // add matches
-        allowed = Some(BitSet::with_max_value(num_terms));
+        allowed = Some(BitSet::with_max_value(allowed_capacity));
        let allowed = allowed.as_mut().unwrap();
        for_each_matching_term_ord(str_col, include, |ord| allowed.insert(ord))?;
    };
@@ -954,7 +1051,7 @@ fn build_allowed_term_ids_for_str(
    if let Some(exclude) = exclude {
        if allowed.is_none() {
            // Start with all terms allowed
-            allowed = Some(BitSet::with_max_value_and_full(num_terms));
+            allowed = Some(BitSet::with_max_value_and_full(allowed_capacity));
        }
        let allowed = allowed.as_mut().unwrap();
        for_each_matching_term_ord(str_col, exclude, |ord| allowed.remove(ord))?;
--- a/src/aggregation/agg_req.rs
+++ b/src/aggregation/agg_req.rs
@@ -32,8 +32,8 @@ use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};

 use super::bucket::{
-    DateHistogramAggregationReq, FilterAggregation, HistogramAggregation, RangeAggregation,
-    TermsAggregation,
+    CompositeAggregation, DateHistogramAggregationReq, FilterAggregation, HistogramAggregation,
+    RangeAggregation, TermsAggregation,
 };
 use super::metric::{
    AverageAggregation, CardinalityAggregationReq, CountAggregation, ExtendedStatsAggregation,
@@ -134,6 +134,9 @@ pub enum AggregationVariants {
    /// Filter documents into a single bucket.
    #[serde(rename = "filter")]
    Filter(FilterAggregation),
+    /// Multi-dimensional, paginable bucket aggregation.
+    #[serde(rename = "composite")]
+    Composite(CompositeAggregation),

    // Metric aggregation types
    /// Computes the average of the extracted values.
@@ -180,6 +183,11 @@ impl AggregationVariants {
            AggregationVariants::Histogram(histogram) => vec![histogram.field.as_str()],
            AggregationVariants::DateHistogram(histogram) => vec![histogram.field.as_str()],
            AggregationVariants::Filter(filter) => filter.get_fast_field_names(),
+            AggregationVariants::Composite(composite) => composite
+                .sources
+                .iter()
+                .map(|source| source.field())
+                .collect(),
            AggregationVariants::Average(avg) => vec![avg.field_name()],
            AggregationVariants::Count(count) => vec![count.field_name()],
            AggregationVariants::Max(max) => vec![max.field_name()],
@@ -214,6 +222,12 @@ impl AggregationVariants {
            _ => None,
        }
    }
+    pub(crate) fn as_composite(&self) -> Option<&CompositeAggregation> {
+        match &self {
+            AggregationVariants::Composite(composite) => Some(composite),
+            _ => None,
+        }
+    }
    pub(crate) fn as_percentile(&self) -> Option<&PercentilesAggregationReq> {
        match &self {
            AggregationVariants::Percentiles(percentile_req) => Some(percentile_req),
--- a/src/aggregation/agg_result.rs
+++ b/src/aggregation/agg_result.rs
@@ -9,10 +9,12 @@ use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};

 use super::bucket::GetDocCount;
+use super::intermediate_agg_result::CompositeIntermediateKey;
 use super::metric::{
    ExtendedStats, PercentilesMetricResult, SingleMetricResult, Stats, TopHitsMetricResult,
 };
 use super::{AggregationError, Key};
+use crate::aggregation::bucket::AfterKey;
 use crate::TantivyError;

 #[derive(Clone, Default, Debug, PartialEq, Serialize, Deserialize)]
@@ -158,6 +160,14 @@ pub enum BucketResult {
    },
    /// This is the filter result - a single bucket with sub-aggregations
    Filter(FilterBucketResult),
+    /// This is the composite result
+    Composite {
+        /// The buckets
+        buckets: Vec<CompositeBucketEntry>,
+        /// The key to start after when paginating
+        #[serde(skip_serializing_if = "FxHashMap::is_empty")]
+        after_key: FxHashMap<String, AfterKey>,
+    },
 }

 impl BucketResult {
@@ -179,6 +189,9 @@ impl BucketResult {
                // Only count sub-aggregation buckets
                filter_result.sub_aggregations.get_bucket_count()
            }
+            BucketResult::Composite { buckets, .. } => {
+                buckets.iter().map(|bucket| bucket.get_bucket_count()).sum()
+            }
        }
    }
 }
@@ -195,7 +208,8 @@ pub enum BucketEntries<T> {
 }

 impl<T> BucketEntries<T> {
-    fn iter<'a>(&'a self) -> Box<dyn Iterator<Item = &'a T> + 'a> {
+    /// Iterate over all bucket entries.
+    pub fn iter<'a>(&'a self) -> Box<dyn Iterator<Item = &'a T> + 'a> {
        match self {
            BucketEntries::Vec(vec) => Box::new(vec.iter()),
            BucketEntries::HashMap(map) => Box::new(map.values()),
@@ -337,3 +351,87 @@ pub struct FilterBucketResult {
    #[serde(flatten)]
    pub sub_aggregations: AggregationResults,
 }
+
+/// Note the type information loss compared to `CompositeIntermediateKey`.
+/// Pagination is performed using `AfterKey`, which encodes type information.
+#[derive(Clone, Debug, Serialize, Deserialize)]
+#[serde(untagged)]
+pub enum CompositeKey {
+    /// Boolean key
+    Bool(bool),
+    /// String key
+    Str(String),
+    /// `i64` key
+    I64(i64),
+    /// `u64` key
+    U64(u64),
+    /// `f64` key
+    F64(f64),
+    /// Null key
+    Null,
+}
+impl Eq for CompositeKey {}
+impl std::hash::Hash for CompositeKey {
+    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
+        core::mem::discriminant(self).hash(state);
+        match self {
+            Self::Bool(val) => val.hash(state),
+            Self::Str(text) => text.hash(state),
+            Self::F64(val) => val.to_bits().hash(state),
+            Self::U64(val) => val.hash(state),
+            Self::I64(val) => val.hash(state),
+            Self::Null => {}
+        }
+    }
+}
+impl PartialEq for CompositeKey {
+    fn eq(&self, other: &Self) -> bool {
+        match (self, other) {
+            (Self::Bool(l), Self::Bool(r)) => l == r,
+            (Self::Str(l), Self::Str(r)) => l == r,
+            (Self::F64(l), Self::F64(r)) => l.to_bits() == r.to_bits(),
+            (Self::I64(l), Self::I64(r)) => l == r,
+            (Self::U64(l), Self::U64(r)) => l == r,
+            (Self::Null, Self::Null) => true,
+            _ => false,
+        }
+    }
+}
+impl From<CompositeIntermediateKey> for CompositeKey {
+    fn from(value: CompositeIntermediateKey) -> Self {
+        match value {
+            CompositeIntermediateKey::Str(s) => Self::Str(s),
+            CompositeIntermediateKey::IpAddr(s) => {
+                if let Some(ip) = s.to_ipv4_mapped() {
+                    Self::Str(ip.to_string())
+                } else {
+                    Self::Str(s.to_string())
+                }
+            }
+            CompositeIntermediateKey::F64(f) => Self::F64(f),
+            CompositeIntermediateKey::Bool(f) => Self::Bool(f),
+            CompositeIntermediateKey::U64(f) => Self::U64(f),
+            CompositeIntermediateKey::I64(f) => Self::I64(f),
+            CompositeIntermediateKey::DateTime(f) => Self::I64(f / 1_000_000), // ns to ms
+            CompositeIntermediateKey::Null => Self::Null,
+        }
+    }
+}
+
+/// Composite bucket entry with a multi-dimensional key.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct CompositeBucketEntry {
+    /// The identifier of the bucket.
+    pub key: FxHashMap<String, CompositeKey>,
+    /// Number of documents in the bucket.
+    pub doc_count: u64,
+    #[serde(flatten)]
+    /// Sub-aggregations in this bucket.
+    pub sub_aggregation: AggregationResults,
+}
+
+impl CompositeBucketEntry {
+    pub(crate) fn get_bucket_count(&self) -> u64 {
+        1 + self.sub_aggregation.get_bucket_count()
+    }
+}
--- a/src/aggregation/bucket/composite/accessors.rs
+++ b/src/aggregation/bucket/composite/accessors.rs
@@ -0,0 +1,518 @@
+use std::net::Ipv6Addr;
+
+use columnar::column_values::{CompactHit, CompactSpaceU64Accessor};
+use columnar::{Column, ColumnType, MonotonicallyMappableToU64, StrColumn, TermOrdHit};
+
+use crate::aggregation::accessor_helpers::get_numeric_or_date_column_types;
+use crate::aggregation::bucket::composite::numeric_types::num_proj;
+use crate::aggregation::bucket::composite::numeric_types::num_proj::ProjectedNumber;
+use crate::aggregation::bucket::composite::ToTypePaginationOrder;
+use crate::aggregation::bucket::{
+    parse_into_milliseconds, CalendarInterval, CompositeAggregation, CompositeAggregationSource,
+    MissingOrder, Order,
+};
+use crate::aggregation::intermediate_agg_result::CompositeIntermediateKey;
+use crate::{SegmentReader, TantivyError};
+
+/// Contains all information required by the SegmentCompositeCollector to perform the
+/// composite aggregation on a segment.
+pub struct CompositeAggReqData {
+    /// The name of the aggregation.
+    pub name: String,
+    /// The normalized term aggregation request.
+    pub req: CompositeAggregation,
+    /// Accessors for each source, each source can have multiple accessors (columns).
+    pub composite_accessors: Vec<CompositeSourceAccessors>,
+}
+
+impl CompositeAggReqData {
+    /// Estimate the memory consumption of this struct in bytes.
+    pub fn get_memory_consumption(&self) -> usize {
+        std::mem::size_of::<Self>()
+            + self.composite_accessors.len() * std::mem::size_of::<CompositeSourceAccessors>()
+    }
+}
+
+/// Accessors for a single column in a composite source.
+pub struct CompositeAccessor {
+    /// The fast field column
+    pub column: Column<u64>,
+    /// The column type
+    pub column_type: ColumnType,
+    /// Term dictionary if the column type is Str
+    ///
+    /// Only used by term sources
+    pub str_dict_column: Option<StrColumn>,
+    /// Parsed date interval for date histogram sources
+    pub date_histogram_interval: PrecomputedDateInterval,
+}
+
+/// Accessors to all the columns that belong to the field of a composite source.
+pub struct CompositeSourceAccessors {
+    /// The accessors for this source
+    pub accessors: Vec<CompositeAccessor>,
+    /// The key after which to start collecting results. Applies to the first
+    /// column of the source.
+    pub after_key: PrecomputedAfterKey,
+
+    /// The column index the after_key applies to. The after_key only applies to
+    /// one column. Columns before should be skipped. Columns after should be
+    /// kept without comparison to the after_key.
+    pub after_key_accessor_idx: usize,
+
+    /// Whether to skip missing values because of the after_key. Skipping only
+    /// applies if the value for previous columns were exactly equal to the
+    /// corresponding after keys (is_on_after_key).
+    pub skip_missing: bool,
+
+    /// The after key was set to null to indicate that the last collected key
+    /// was a missing value.
+    pub is_after_key_explicit_missing: bool,
+}
+
+impl CompositeSourceAccessors {
+    /// Creates a new set of accessors for the composite source.
+    ///
+    /// Precomputes some values to make collection faster.
+    pub fn build_for_source(
+        reader: &SegmentReader,
+        source: &CompositeAggregationSource,
+        // First option is None when no after key was set in the query, the
+        // second option is None when the after key was set but its value for
+        // this source was set to `null`
+        source_after_key_opt: Option<&CompositeIntermediateKey>,
+    ) -> crate::Result<Self> {
+        let is_after_key_explicit_missing = source_after_key_opt
+            .map(|after_key| matches!(after_key, CompositeIntermediateKey::Null))
+            .unwrap_or(false);
+        let mut skip_missing = false;
+        if let Some(CompositeIntermediateKey::Null) = source_after_key_opt {
+            if !source.missing_bucket() {
+                return Err(TantivyError::InvalidArgument(
+                    "the 'after' key for a source cannot be null when 'missing_bucket' is false"
+                        .to_string(),
+                ));
+            }
+        } else if source_after_key_opt.is_some() {
+            // if missing buckets come first and we have a non null after key, we skip missing
+            if MissingOrder::First == source.missing_order() {
+                skip_missing = true;
+            }
+            if MissingOrder::Default == source.missing_order() && Order::Asc == source.order() {
+                skip_missing = true;
+            }
+        };
+
+        match source {
+            CompositeAggregationSource::Terms(source) => {
+                let allowed_column_types = [
+                    ColumnType::I64,
+                    ColumnType::U64,
+                    ColumnType::F64,
+                    ColumnType::Str,
+                    ColumnType::DateTime,
+                    ColumnType::Bool,
+                    ColumnType::IpAddr,
+                    // ColumnType::Bytes Unsupported
+                ];
+                let mut columns_and_types = reader
+                    .fast_fields()
+                    .u64_lenient_for_type_all(Some(&allowed_column_types), &source.field)?;
+
+                // Sort columns by their pagination order and determine which to skip
+                columns_and_types.sort_by_key(|(_, col_type): &(Column, ColumnType)| {
+                    col_type.column_pagination_order()
+                });
+                if source.order == Order::Desc {
+                    columns_and_types.reverse();
+                }
+                let after_key_accessor_idx = find_first_column_to_collect(
+                    &columns_and_types,
+                    source_after_key_opt,
+                    source.missing_order,
+                    source.order,
+                )?;
+
+                let source_collectors: Vec<CompositeAccessor> = columns_and_types
+                    .into_iter()
+                    .map(|(column, column_type)| {
+                        Ok(CompositeAccessor {
+                            column,
+                            column_type,
+                            str_dict_column: reader.fast_fields().str(&source.field)?,
+                            date_histogram_interval: PrecomputedDateInterval::NotApplicable,
+                        })
+                    })
+                    .collect::<crate::Result<_>>()?;
+
+                let after_key = if let Some(first_col) =
+                    source_collectors.get(after_key_accessor_idx)
+                {
+                    match source_after_key_opt {
+                        Some(after_key) => PrecomputedAfterKey::precompute(
+                            first_col,
+                            after_key,
+                            &source.field,
+                            source.missing_order,
+                            source.order,
+                        )?,
+                        None => {
+                            precompute_missing_after_key(false, source.missing_order, source.order)
+                        }
+                    }
+                } else {
+                    // if no columns, we don't care about the after_key
+                    PrecomputedAfterKey::Next(0)
+                };
+
+                Ok(CompositeSourceAccessors {
+                    accessors: source_collectors,
+                    is_after_key_explicit_missing,
+                    skip_missing,
+                    after_key,
+                    after_key_accessor_idx,
+                })
+            }
+            CompositeAggregationSource::Histogram(source) => {
+                let column_and_types: Vec<(Column, ColumnType)> =
+                    reader.fast_fields().u64_lenient_for_type_all(
+                        Some(get_numeric_or_date_column_types()),
+                        &source.field,
+                    )?;
+                let source_collectors: Vec<CompositeAccessor> = column_and_types
+                    .into_iter()
+                    .map(|(column, column_type)| {
+                        Ok(CompositeAccessor {
+                            column,
+                            column_type,
+                            str_dict_column: None,
+                            date_histogram_interval: PrecomputedDateInterval::NotApplicable,
+                        })
+                    })
+                    .collect::<crate::Result<_>>()?;
+                let after_key = match source_after_key_opt {
+                    Some(CompositeIntermediateKey::F64(key)) => {
+                        let normalized_key = *key / source.interval;
+                        num_proj::f64_to_i64(normalized_key).into()
+                    }
+                    Some(CompositeIntermediateKey::Null) => {
+                        precompute_missing_after_key(true, source.missing_order, source.order)
+                    }
+                    None => precompute_missing_after_key(true, source.missing_order, source.order),
+                    _ => {
+                        return Err(crate::TantivyError::InvalidArgument(
+                            "After key type invalid for interval composite source".to_string(),
+                        ));
+                    }
+                };
+                Ok(CompositeSourceAccessors {
+                    accessors: source_collectors,
+                    is_after_key_explicit_missing,
+                    skip_missing,
+                    after_key,
+                    after_key_accessor_idx: 0,
+                })
+            }
+            CompositeAggregationSource::DateHistogram(source) => {
+                let column_and_types = reader
+                    .fast_fields()
+                    .u64_lenient_for_type_all(Some(&[ColumnType::DateTime]), &source.field)?;
+                let date_histogram_interval =
+                    PrecomputedDateInterval::from_date_histogram_source_intervals(
+                        &source.fixed_interval,
+                        source.calendar_interval,
+                    )?;
+                let source_collectors: Vec<CompositeAccessor> = column_and_types
+                    .into_iter()
+                    .map(|(column, column_type)| {
+                        Ok(CompositeAccessor {
+                            column,
+                            column_type,
+                            str_dict_column: None,
+                            date_histogram_interval,
+                        })
+                    })
+                    .collect::<crate::Result<_>>()?;
+                let after_key = match source_after_key_opt {
+                    Some(CompositeIntermediateKey::DateTime(key)) => {
+                        PrecomputedAfterKey::Exact(key.to_u64())
+                    }
+                    Some(CompositeIntermediateKey::Null) => {
+                        precompute_missing_after_key(true, source.missing_order, source.order)
+                    }
+                    None => precompute_missing_after_key(true, source.missing_order, source.order),
+                    _ => {
+                        return Err(crate::TantivyError::InvalidArgument(
+                            "After key type invalid for interval composite source".to_string(),
+                        ));
+                    }
+                };
+                Ok(CompositeSourceAccessors {
+                    accessors: source_collectors,
+                    is_after_key_explicit_missing,
+                    skip_missing,
+                    after_key,
+                    after_key_accessor_idx: 0,
+                })
+            }
+        }
+    }
+}
+
+/// Finds the index of the first column we should start collecting from to
+/// resume the pagination from the after_key.
+fn find_first_column_to_collect<T>(
+    sorted_columns: &[(T, ColumnType)],
+    after_key_opt: Option<&CompositeIntermediateKey>,
+    missing_order: MissingOrder,
+    order: Order,
+) -> crate::Result<usize> {
+    let after_key = match after_key_opt {
+        None => return Ok(0), // No pagination, start from beginning
+        Some(key) => key,
+    };
+    // Handle null after_key (we were on a missing value last time)
+    if matches!(after_key, CompositeIntermediateKey::Null) {
+        return match (missing_order, order) {
+            // Missing values come first, so all columns remain
+            (MissingOrder::First, _) | (MissingOrder::Default, Order::Asc) => Ok(0),
+            // Missing values come last, so all columns are done
+            (MissingOrder::Last, _) | (MissingOrder::Default, Order::Desc) => {
+                Ok(sorted_columns.len())
+            }
+        };
+    }
+    // Find the first column whose type order matches or follows the after_key's
+    // type in the pagination sequence
+    let after_key_column_order = after_key.column_pagination_order();
+    for (idx, (_, col_type)) in sorted_columns.iter().enumerate() {
+        let col_order = col_type.column_pagination_order();
+        let is_first_to_collect = match order {
+            Order::Asc => col_order >= after_key_column_order,
+            Order::Desc => col_order <= after_key_column_order,
+        };
+        if is_first_to_collect {
+            return Ok(idx);
+        }
+    }
+    // All columns are before the after_key, nothing left to collect
+    Ok(sorted_columns.len())
+}
+
+fn precompute_missing_after_key(
+    is_after_key_explicit_missing: bool,
+    missing_order: MissingOrder,
+    order: Order,
+) -> PrecomputedAfterKey {
+    let after_last = PrecomputedAfterKey::AfterLast;
+    let before_first = PrecomputedAfterKey::Next(0);
+    match (is_after_key_explicit_missing, missing_order, order) {
+        (true, MissingOrder::First, Order::Asc) => before_first,
+        (true, MissingOrder::First, Order::Desc) => after_last,
+        (true, MissingOrder::Last, Order::Asc) => after_last,
+        (true, MissingOrder::Last, Order::Desc) => before_first,
+        (true, MissingOrder::Default, Order::Asc) => before_first,
+        (true, MissingOrder::Default, Order::Desc) => after_last,
+        (false, _, Order::Asc) => before_first,
+        (false, _, Order::Desc) => after_last,
+    }
+}
+
+/// A parsed representation of the date interval for date histogram sources
+#[derive(Clone, Copy, Debug)]
+pub enum PrecomputedDateInterval {
+    /// This is not a date histogram source
+    NotApplicable,
+    /// Source was configured with a fixed interval
+    FixedNanoseconds(i64),
+    /// Source was configured with a calendar interval
+    Calendar(CalendarInterval),
+}
+
+impl PrecomputedDateInterval {
+    /// Validates the date histogram source interval fields and parses a date interval from them.
+    pub fn from_date_histogram_source_intervals(
+        fixed_interval: &Option<String>,
+        calendar_interval: Option<CalendarInterval>,
+    ) -> crate::Result<Self> {
+        match (fixed_interval, calendar_interval) {
+            (Some(_), Some(_)) | (None, None) => Err(TantivyError::InvalidArgument(
+                "date histogram source must one and only one of fixed_interval or \
+                 calendar_interval set"
+                    .to_string(),
+            )),
+            (Some(fixed_interval), None) => {
+                let fixed_interval_ms = parse_into_milliseconds(fixed_interval)?;
+                Ok(PrecomputedDateInterval::FixedNanoseconds(
+                    fixed_interval_ms * 1_000_000,
+                ))
+            }
+            (None, Some(calendar_interval)) => {
+                Ok(PrecomputedDateInterval::Calendar(calendar_interval))
+            }
+        }
+    }
+}
+
+/// The after key projected to the u64 column space
+///
+/// Some column types (term, IP) might not have an exact representation of the
+/// specified after key
+#[derive(Debug)]
+pub enum PrecomputedAfterKey {
+    /// The after key could be exactly represented in the column space.
+    Exact(u64),
+    /// The after key could not be exactly represented exactly represented, so
+    /// this is the next closest one.
+    Next(u64),
+    /// The after key could not be represented in the column space, it is
+    /// greater than all value
+    AfterLast,
+}
+
+impl From<CompactHit> for PrecomputedAfterKey {
+    fn from(hit: CompactHit) -> Self {
+        match hit {
+            CompactHit::Exact(ord) => PrecomputedAfterKey::Exact(ord as u64),
+            CompactHit::Next(ord) => PrecomputedAfterKey::Next(ord as u64),
+            CompactHit::AfterLast => PrecomputedAfterKey::AfterLast,
+        }
+    }
+}
+
+impl From<TermOrdHit> for PrecomputedAfterKey {
+    fn from(hit: TermOrdHit) -> Self {
+        match hit {
+            TermOrdHit::Exact(ord) => PrecomputedAfterKey::Exact(ord),
+            // TermOrdHit represents AfterLast as Next(u64::MAX), we keep it as is
+            TermOrdHit::Next(ord) => PrecomputedAfterKey::Next(ord),
+        }
+    }
+}
+
+impl<T: MonotonicallyMappableToU64> From<ProjectedNumber<T>> for PrecomputedAfterKey {
+    fn from(num: ProjectedNumber<T>) -> Self {
+        match num {
+            ProjectedNumber::Exact(number) => PrecomputedAfterKey::Exact(number.to_u64()),
+            ProjectedNumber::Next(number) => PrecomputedAfterKey::Next(number.to_u64()),
+            ProjectedNumber::AfterLast => PrecomputedAfterKey::AfterLast,
+        }
+    }
+}
+
+// /!\ These operators only makes sense if both values are in the same column space
+impl PrecomputedAfterKey {
+    pub fn equals(&self, column_value: u64) -> bool {
+        match self {
+            PrecomputedAfterKey::Exact(v) => *v == column_value,
+            PrecomputedAfterKey::Next(_) => false,
+            PrecomputedAfterKey::AfterLast => false,
+        }
+    }
+
+    pub fn gt(&self, column_value: u64) -> bool {
+        match self {
+            PrecomputedAfterKey::Exact(v) => *v > column_value,
+            PrecomputedAfterKey::Next(v) => *v > column_value,
+            PrecomputedAfterKey::AfterLast => true,
+        }
+    }
+
+    pub fn lt(&self, column_value: u64) -> bool {
+        match self {
+            PrecomputedAfterKey::Exact(v) => *v < column_value,
+            // a value equal to the next is greater than the after key
+            PrecomputedAfterKey::Next(v) => *v <= column_value,
+            PrecomputedAfterKey::AfterLast => false,
+        }
+    }
+
+    fn precompute_ip_addr(column: &Column<u64>, key: &Ipv6Addr) -> crate::Result<Self> {
+        let compact_space_accessor = column
+            .values
+            .clone()
+            .downcast_arc::<CompactSpaceU64Accessor>()
+            .map_err(|_| {
+                TantivyError::AggregationError(crate::aggregation::AggregationError::InternalError(
+                    "type mismatch: could not downcast to CompactSpaceU64Accessor".to_string(),
+                ))
+            })?;
+        let ip_u128 = key.to_bits();
+        let ip_next_compact = compact_space_accessor.u128_to_next_compact(ip_u128);
+        Ok(ip_next_compact.into())
+    }
+
+    fn precompute_term_ord(
+        str_dict_column: &Option<StrColumn>,
+        key: &str,
+        field: &str,
+    ) -> crate::Result<Self> {
+        let dict = str_dict_column
+            .as_ref()
+            .expect("dictionary missing for str accessor")
+            .dictionary();
+        let next_ord = dict.term_ord_or_next(key).map_err(|_| {
+            TantivyError::InvalidArgument(format!(
+                "failed to lookup after_key '{}' for field '{}'",
+                key, field
+            ))
+        })?;
+        Ok(next_ord.into())
+    }
+
+    /// Projects the after key into the column space of the given accessor.
+    ///
+    /// The computed after key will not take care of skipping entire columns
+    /// when the after key type is ordered after the accessor's type, that
+    /// should be performed earlier.
+    pub fn precompute(
+        composite_accessor: &CompositeAccessor,
+        source_after_key: &CompositeIntermediateKey,
+        field: &str,
+        missing_order: MissingOrder,
+        order: Order,
+    ) -> crate::Result<Self> {
+        use CompositeIntermediateKey as CIKey;
+        let precomputed_key = match (composite_accessor.column_type, source_after_key) {
+            (ColumnType::Bytes, _) => panic!("unsupported"),
+            // null after key
+            (_, CIKey::Null) => precompute_missing_after_key(false, missing_order, order),
+            // numerical
+            (ColumnType::I64, CIKey::I64(k)) => PrecomputedAfterKey::Exact(k.to_u64()),
+            (ColumnType::I64, CIKey::U64(k)) => num_proj::u64_to_i64(*k).into(),
+            (ColumnType::I64, CIKey::F64(k)) => num_proj::f64_to_i64(*k).into(),
+            (ColumnType::U64, CIKey::I64(k)) => num_proj::i64_to_u64(*k).into(),
+            (ColumnType::U64, CIKey::U64(k)) => PrecomputedAfterKey::Exact(*k),
+            (ColumnType::U64, CIKey::F64(k)) => num_proj::f64_to_u64(*k).into(),
+            (ColumnType::F64, CIKey::I64(k)) => num_proj::i64_to_f64(*k).into(),
+            (ColumnType::F64, CIKey::U64(k)) => num_proj::u64_to_f64(*k).into(),
+            (ColumnType::F64, CIKey::F64(k)) => PrecomputedAfterKey::Exact(k.to_u64()),
+            // boolean
+            (ColumnType::Bool, CIKey::Bool(key)) => PrecomputedAfterKey::Exact(key.to_u64()),
+            // string
+            (ColumnType::Str, CIKey::Str(key)) => PrecomputedAfterKey::precompute_term_ord(
+                &composite_accessor.str_dict_column,
+                key,
+                field,
+            )?,
+            // date time
+            (ColumnType::DateTime, CIKey::DateTime(key)) => {
+                PrecomputedAfterKey::Exact(key.to_u64())
+            }
+            // ip address
+            (ColumnType::IpAddr, CIKey::IpAddr(key)) => {
+                PrecomputedAfterKey::precompute_ip_addr(&composite_accessor.column, key)?
+            }
+            // assume the column's type is ordered after the after_key's type
+            _ => PrecomputedAfterKey::keep_all(order),
+        };
+        Ok(precomputed_key)
+    }
+
+    fn keep_all(order: Order) -> Self {
+        match order {
+            Order::Asc => PrecomputedAfterKey::Next(0),
+            Order::Desc => PrecomputedAfterKey::Next(u64::MAX),
+        }
+    }
+}
--- a/src/aggregation/bucket/composite/calendar_interval.rs
+++ b/src/aggregation/bucket/composite/calendar_interval.rs
@@ -0,0 +1,136 @@
+use time::convert::{Day, Nanosecond};
+use time::{Time, UtcDateTime};
+
+const NS_IN_DAY: i64 = Nanosecond::per_t::<i128>(Day) as i64;
+
+/// Computes the timestamp in nanoseconds corresponding to the beginning of the
+/// year (January 1st at midnight UTC).
+pub(super) fn try_year_bucket(timestamp_ns: i64) -> crate::Result<i64> {
+    year_bucket_using_time_crate(timestamp_ns).map_err(|e| {
+        crate::TantivyError::InvalidArgument(format!(
+            "Failed to compute year bucket for timestamp {}: {e}",
+            timestamp_ns
+        ))
+    })
+}
+
+/// Computes the timestamp in nanoseconds corresponding to the beginning of the
+/// month (1st at midnight UTC).
+pub(super) fn try_month_bucket(timestamp_ns: i64) -> crate::Result<i64> {
+    month_bucket_using_time_crate(timestamp_ns).map_err(|e| {
+        crate::TantivyError::InvalidArgument(format!(
+            "Failed to compute month bucket for timestamp {}: {e}",
+            timestamp_ns
+        ))
+    })
+}
+
+/// Computes the timestamp in nanoseconds corresponding to the beginning of the
+/// week (Monday at midnight UTC).
+pub(super) fn week_bucket(timestamp_ns: i64) -> i64 {
+    // 1970-01-01 was a Thursday (weekday = 4)
+    let days_since_epoch = timestamp_ns.div_euclid(NS_IN_DAY);
+    // Find the weekday: 0=Monday, ..., 6=Sunday
+    let weekday = (days_since_epoch + 3).rem_euclid(7);
+    let monday_days_since_epoch = days_since_epoch - weekday;
+    monday_days_since_epoch * NS_IN_DAY
+}
+
+fn year_bucket_using_time_crate(timestamp_ns: i64) -> Result<i64, time::Error> {
+    let timestamp_ns = UtcDateTime::from_unix_timestamp_nanos(timestamp_ns as i128)?
+        .replace_ordinal(1)?
+        .replace_time(Time::MIDNIGHT)
+        .unix_timestamp_nanos();
+    Ok(timestamp_ns as i64)
+}
+
+fn month_bucket_using_time_crate(timestamp_ns: i64) -> Result<i64, time::Error> {
+    let timestamp_ns = UtcDateTime::from_unix_timestamp_nanos(timestamp_ns as i128)?
+        .replace_day(1)?
+        .replace_time(Time::MIDNIGHT)
+        .unix_timestamp_nanos();
+    Ok(timestamp_ns as i64)
+}
+
+#[cfg(test)]
+mod tests {
+    use time::format_description::well_known::Iso8601;
+    use time::UtcDateTime;
+
+    use super::*;
+
+    fn ts_ns(iso: &str) -> i64 {
+        UtcDateTime::parse(iso, &Iso8601::DEFAULT)
+            .unwrap()
+            .unix_timestamp_nanos() as i64
+    }
+
+    #[test]
+    fn test_year_bucket() {
+        let ts = ts_ns("1970-01-01T00:00:00Z");
+        let res = try_year_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("1970-01-01T00:00:00Z"));
+
+        let ts = ts_ns("1970-06-01T10:00:01.010Z");
+        let res = try_year_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("1970-01-01T00:00:00Z"));
+
+        let ts = ts_ns("2008-12-31T23:59:59.999999999Z"); // leap year
+        let res = try_year_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("2008-01-01T00:00:00Z"));
+
+        let ts = ts_ns("2008-01-01T00:00:00Z"); // leap year
+        let res = try_year_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("2008-01-01T00:00:00Z"));
+
+        let ts = ts_ns("2010-12-31T23:59:59.999999999Z");
+        let res = try_year_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("2010-01-01T00:00:00Z"));
+
+        let ts = ts_ns("1972-06-01T00:10:00Z");
+        let res = try_year_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("1972-01-01T00:00:00Z"));
+    }
+
+    #[test]
+    fn test_month_bucket() {
+        let ts = ts_ns("1970-01-15T00:00:00Z");
+        let res = try_month_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("1970-01-01T00:00:00Z"));
+
+        let ts = ts_ns("1970-02-01T00:00:00Z");
+        let res = try_month_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("1970-02-01T00:00:00Z"));
+
+        let ts = ts_ns("2000-01-31T23:59:59.999999999Z");
+        let res = try_month_bucket(ts).unwrap();
+        assert_eq!(res, ts_ns("2000-01-01T00:00:00Z"));
+    }
+
+    #[test]
+    fn test_week_bucket() {
+        let ts = ts_ns("1970-01-05T00:00:00Z"); // Monday
+        let res = week_bucket(ts);
+        assert_eq!(res, ts_ns("1970-01-05T00:00:00Z"));
+
+        let ts = ts_ns("1970-01-05T23:59:59Z"); // Monday
+        let res = week_bucket(ts);
+        assert_eq!(res, ts_ns("1970-01-05T00:00:00Z"));
+
+        let ts = ts_ns("1970-01-07T01:13:00Z"); // Wednesday
+        let res = week_bucket(ts);
+        assert_eq!(res, ts_ns("1970-01-05T00:00:00Z"));
+
+        let ts = ts_ns("1970-01-11T23:59:59.999999999Z"); // Sunday
+        let res = week_bucket(ts);
+        assert_eq!(res, ts_ns("1970-01-05T00:00:00Z"));
+
+        let ts = ts_ns("2025-10-16T10:41:59.010Z"); // Thursday
+        let res = week_bucket(ts);
+        assert_eq!(res, ts_ns("2025-10-13T00:00:00Z"));
+
+        let ts = ts_ns("1970-01-01T00:00:00Z"); // Thursday
+        let res = week_bucket(ts);
+        assert_eq!(res, ts_ns("1969-12-29T00:00:00Z")); // Negative
+    }
+}
--- a/src/aggregation/bucket/composite/collector.rs
+++ b/src/aggregation/bucket/composite/collector.rs
@@ -0,0 +1,660 @@
+use std::fmt::Debug;
+use std::mem;
+use std::net::Ipv6Addr;
+
+use columnar::column_values::CompactSpaceU64Accessor;
+use columnar::{
+    Column, ColumnType, Dictionary, MonotonicallyMappableToU128, MonotonicallyMappableToU64,
+    NumericalValue, StrColumn,
+};
+use rustc_hash::FxHashMap;
+use smallvec::SmallVec;
+
+use crate::aggregation::agg_data::{
+    build_segment_agg_collectors, AggRefNode, AggregationsSegmentCtx,
+};
+use crate::aggregation::bucket::composite::accessors::{
+    CompositeAccessor, CompositeAggReqData, PrecomputedDateInterval,
+};
+use crate::aggregation::bucket::composite::calendar_interval;
+use crate::aggregation::bucket::composite::map::{DynArrayHeapMap, MAX_DYN_ARRAY_SIZE};
+use crate::aggregation::bucket::{
+    CalendarInterval, CompositeAggregationSource, MissingOrder, Order,
+};
+use crate::aggregation::buffered_sub_aggs::{BufferedSubAggs, HighCardSubAggBuffer};
+use crate::aggregation::intermediate_agg_result::{
+    CompositeIntermediateKey, IntermediateAggregationResult, IntermediateAggregationResults,
+    IntermediateBucketResult, IntermediateCompositeBucketEntry, IntermediateCompositeBucketResult,
+};
+use crate::aggregation::segment_agg_result::{BucketIdProvider, SegmentAggregationCollector};
+use crate::aggregation::BucketId;
+use crate::TantivyError;
+
+#[derive(Clone, Debug)]
+struct CompositeBucketCollector {
+    count: u32,
+    bucket_id: BucketId,
+}
+
+/// Compact sortable representation of a single source value within a composite key.
+///
+/// The struct encodes both the column identity and the fast field value in a way
+/// that preserves the desired sort order via the derived `Ord` implementation
+/// (fields are compared top-to-bottom: `sort_key` first, then `encoded_value`).
+///
+/// ## `sort_key` encoding
+/// - `0` — missing value, sorted first
+/// - `1..=254` — present value; the original accessor index is `sort_key - 1`
+/// - `u8::MAX` (255) — missing value, sorted last
+///
+/// ## `encoded_value` encoding
+/// - `0` when the field is missing
+/// - The raw u64 fast-field representation when order is ascending
+/// - Bitwise NOT of the raw u64 when order is descending
+#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Default, Hash)]
+struct InternalValueRepr {
+    /// Column index biased by +1 (so 0 and u8::MAX are reserved for missing sentinels).
+    sort_key: u8,
+    /// Fast field value, possibly bit-flipped for descending order.
+    encoded_value: u64,
+}
+
+impl InternalValueRepr {
+    #[inline]
+    fn new_term(raw: u64, accessor_idx: u8, order: Order) -> Self {
+        let encoded_value = match order {
+            Order::Asc => raw,
+            Order::Desc => !raw,
+        };
+        InternalValueRepr {
+            sort_key: accessor_idx + 1,
+            encoded_value,
+        }
+    }
+
+    /// For histogram sources the column index is irrelevant (always 1).
+    #[inline]
+    fn new_histogram(raw: u64, order: Order) -> Self {
+        let encoded_value = match order {
+            Order::Asc => raw,
+            Order::Desc => !raw,
+        };
+        InternalValueRepr {
+            sort_key: 1,
+            encoded_value,
+        }
+    }
+
+    #[inline]
+    fn new_missing(order: Order, missing_order: MissingOrder) -> Self {
+        let sort_key = match (missing_order, order) {
+            (MissingOrder::First, _) | (MissingOrder::Default, Order::Asc) => 0,
+            (MissingOrder::Last, _) | (MissingOrder::Default, Order::Desc) => u8::MAX,
+        };
+        InternalValueRepr {
+            sort_key,
+            encoded_value: 0,
+        }
+    }
+
+    /// Decode back to `(accessor_idx, raw_value)`.
+    /// Returns `None` when the value represents a missing field.
+    #[inline]
+    fn decode(self, order: Order) -> Option<(u8, u64)> {
+        if self.sort_key == 0 || self.sort_key == u8::MAX {
+            return None;
+        }
+        let raw = match order {
+            Order::Asc => self.encoded_value,
+            Order::Desc => !self.encoded_value,
+        };
+        Some((self.sort_key - 1, raw))
+    }
+}
+
+/// The collector puts values from the fast field into the correct buckets and
+/// does a conversion to the correct datatype.
+#[derive(Debug)]
+pub struct SegmentCompositeCollector {
+    /// One DynArrayHeapMap per parent bucket.
+    parent_buckets: Vec<DynArrayHeapMap<InternalValueRepr, CompositeBucketCollector>>,
+    accessor_idx: usize,
+    sub_agg: Option<BufferedSubAggs<HighCardSubAggBuffer>>,
+    bucket_id_provider: BucketIdProvider,
+    /// Number of sources, needed when creating new DynArrayHeapMaps.
+    num_sources: usize,
+}
+
+impl SegmentAggregationCollector for SegmentCompositeCollector {
+    fn add_intermediate_aggregation_result(
+        &mut self,
+        agg_data: &AggregationsSegmentCtx,
+        results: &mut IntermediateAggregationResults,
+        parent_bucket_id: BucketId,
+    ) -> crate::Result<()> {
+        let name = agg_data
+            .get_composite_req_data(self.accessor_idx)
+            .name
+            .clone();
+
+        let buckets = self.add_intermediate_bucket_result(agg_data, parent_bucket_id)?;
+        results.push(
+            name,
+            IntermediateAggregationResult::Bucket(IntermediateBucketResult::Composite { buckets }),
+        )?;
+
+        Ok(())
+    }
+
+    fn collect(
+        &mut self,
+        parent_bucket_id: BucketId,
+        docs: &[crate::DocId],
+        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,
+                buckets: &mut self.parent_buckets[parent_bucket_id as usize],
+                sub_agg: &mut self.sub_agg,
+                bucket_id_provider: &mut self.bucket_id_provider,
+                sub_level_values: SmallVec::new(),
+            };
+            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)?;
+        }
+
+        let mem_delta = self.get_memory_consumption(parent_bucket_id) - mem_pre;
+        if mem_delta > 0 {
+            agg_data.context.limits.add_memory_consumed(mem_delta)?;
+        }
+
+        Ok(())
+    }
+
+    fn flush(&mut self, agg_data: &mut AggregationsSegmentCtx) -> crate::Result<()> {
+        if let Some(sub_agg) = &mut self.sub_agg {
+            sub_agg.flush(agg_data)?;
+        }
+        Ok(())
+    }
+
+    fn prepare_max_bucket(
+        &mut self,
+        max_bucket: BucketId,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> crate::Result<()> {
+        let required_len = max_bucket as usize + 1;
+        while self.parent_buckets.len() < required_len {
+            let map = DynArrayHeapMap::try_new(self.num_sources)?;
+            self.parent_buckets.push(map);
+        }
+        Ok(())
+    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        // Composite is a multi-bucket agg with no single value to extract.
+        None
+    }
+}
+
+impl SegmentCompositeCollector {
+    fn get_memory_consumption(&self, parent_bucket_id: BucketId) -> u64 {
+        self.parent_buckets[parent_bucket_id as usize].memory_consumption()
+    }
+
+    pub(crate) fn from_req_and_validate(
+        req_data: &mut AggregationsSegmentCtx,
+        node: &AggRefNode,
+    ) -> crate::Result<Self> {
+        validate_req(req_data, node.idx_in_req_data)?;
+
+        let has_sub_aggregations = !node.children.is_empty();
+        let sub_agg = if has_sub_aggregations {
+            let sub_agg_collector = build_segment_agg_collectors(req_data, &node.children)?;
+            Some(BufferedSubAggs::new(sub_agg_collector))
+        } else {
+            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,
+            sub_agg,
+            bucket_id_provider: BucketIdProvider::default(),
+            num_sources,
+        })
+    }
+
+    #[inline]
+    fn add_intermediate_bucket_result(
+        &mut self,
+        agg_data: &AggregationsSegmentCtx,
+        parent_bucket_id: BucketId,
+    ) -> crate::Result<IntermediateCompositeBucketResult> {
+        let empty_map = DynArrayHeapMap::try_new(self.num_sources)?;
+        let heap_map = mem::replace(
+            &mut self.parent_buckets[parent_bucket_id as usize],
+            empty_map,
+        );
+
+        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);
+        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();
+            if let Some(sub_agg) = &mut self.sub_agg {
+                sub_agg
+                    .get_sub_agg_collector()
+                    .add_intermediate_aggregation_result(
+                        agg_data,
+                        &mut sub_aggregation_res,
+                        agg.bucket_id,
+                    )?;
+            }
+
+            dict.insert(
+                key,
+                IntermediateCompositeBucketEntry {
+                    doc_count: agg.count,
+                    sub_aggregation: sub_aggregation_res,
+                },
+            );
+        }
+
+        Ok(IntermediateCompositeBucketResult {
+            entries: dict,
+            target_size: composite_data.req.size,
+            orders: composite_data
+                .req
+                .sources
+                .iter()
+                .map(|source| match source {
+                    CompositeAggregationSource::Terms(t) => (t.order, t.missing_order),
+                    CompositeAggregationSource::Histogram(h) => (h.order, h.missing_order),
+                    CompositeAggregationSource::DateHistogram(d) => (d.order, d.missing_order),
+                })
+                .collect(),
+        })
+    }
+}
+
+fn validate_req(req_data: &mut AggregationsSegmentCtx, accessor_idx: usize) -> crate::Result<()> {
+    let composite_data = req_data.get_composite_req_data(accessor_idx);
+    let req = &composite_data.req;
+    if req.sources.is_empty() {
+        return Err(TantivyError::InvalidArgument(
+            "composite aggregation must have at least one source".to_string(),
+        ));
+    }
+    if req.size == 0 {
+        return Err(TantivyError::InvalidArgument(
+            "composite aggregation 'size' must be > 0".to_string(),
+        ));
+    }
+
+    if composite_data.composite_accessors.len() > MAX_DYN_ARRAY_SIZE {
+        return Err(TantivyError::InvalidArgument(format!(
+            "composite aggregation source supports maximum {MAX_DYN_ARRAY_SIZE} sources",
+        )));
+    }
+
+    let column_types_for_sources = composite_data.composite_accessors.iter().map(|item| {
+        item.accessors
+            .iter()
+            .map(|a| a.column_type)
+            .collect::<Vec<_>>()
+    });
+
+    for column_types in column_types_for_sources {
+        if column_types.contains(&ColumnType::Bytes) {
+            return Err(TantivyError::InvalidArgument(
+                "composite aggregation does not support 'bytes' field type".to_string(),
+            ));
+        }
+    }
+    Ok(())
+}
+
+fn collect_bucket_with_limit(
+    doc_id: crate::DocId,
+    limit_num_buckets: usize,
+    buckets: &mut DynArrayHeapMap<InternalValueRepr, CompositeBucketCollector>,
+    key: &[InternalValueRepr],
+    sub_agg: &mut Option<BufferedSubAggs<HighCardSubAggBuffer>>,
+    bucket_id_provider: &mut BucketIdProvider,
+) {
+    let mut record_in_bucket = |bucket: &mut CompositeBucketCollector| {
+        bucket.count += 1;
+        if let Some(sub_agg) = sub_agg {
+            sub_agg.push(bucket.bucket_id, doc_id);
+        }
+    };
+
+    // We still have room for buckets, just insert
+    if buckets.size() < limit_num_buckets {
+        let bucket = buckets.get_or_insert_with(key, || CompositeBucketCollector {
+            count: 0,
+            bucket_id: bucket_id_provider.next_bucket_id(),
+        });
+        record_in_bucket(bucket);
+        return;
+    }
+
+    // Map is full, but we can still update the bucket if it already exists
+    if let Some(bucket) = buckets.get_mut(key) {
+        record_in_bucket(bucket);
+        return;
+    }
+
+    // Check if the item qualifies to enter the top-k, and evict the highest if it does
+    if let Some(highest_key) = buckets.peek_highest() {
+        if key < highest_key {
+            buckets.evict_highest();
+            let bucket = buckets.get_or_insert_with(key, || CompositeBucketCollector {
+                count: 0,
+                bucket_id: bucket_id_provider.next_bucket_id(),
+            });
+            record_in_bucket(bucket);
+        }
+    }
+}
+
+/// Converts the composite key from its internal column space representation
+/// (segment specific) into its intermediate form.
+fn resolve_key(
+    internal_key: &[InternalValueRepr],
+    agg_data: &CompositeAggReqData,
+) -> crate::Result<Vec<CompositeIntermediateKey>> {
+    internal_key
+        .iter()
+        .enumerate()
+        .map(|(idx, val)| {
+            resolve_internal_value_repr(
+                *val,
+                &agg_data.req.sources[idx],
+                &agg_data.composite_accessors[idx].accessors,
+            )
+        })
+        .collect()
+}
+
+fn resolve_internal_value_repr(
+    internal_value_repr: InternalValueRepr,
+    source: &CompositeAggregationSource,
+    composite_accessors: &[CompositeAccessor],
+) -> crate::Result<CompositeIntermediateKey> {
+    let decoded_value_opt = match source {
+        CompositeAggregationSource::Terms(source) => internal_value_repr.decode(source.order),
+        CompositeAggregationSource::Histogram(source) => internal_value_repr.decode(source.order),
+        CompositeAggregationSource::DateHistogram(source) => {
+            internal_value_repr.decode(source.order)
+        }
+    };
+    let Some((decoded_accessor_idx, val)) = decoded_value_opt else {
+        return Ok(CompositeIntermediateKey::Null);
+    };
+    let key = match source {
+        CompositeAggregationSource::Terms(_) => {
+            let CompositeAccessor {
+                column_type,
+                str_dict_column,
+                column,
+                ..
+            } = &composite_accessors[decoded_accessor_idx as usize];
+            resolve_term(val, column_type, str_dict_column, column)?
+        }
+        CompositeAggregationSource::Histogram(source) => {
+            CompositeIntermediateKey::F64(i64::from_u64(val) as f64 * source.interval)
+        }
+        CompositeAggregationSource::DateHistogram(_) => {
+            CompositeIntermediateKey::DateTime(i64::from_u64(val))
+        }
+    };
+
+    Ok(key)
+}
+
+fn resolve_term(
+    val: u64,
+    column_type: &ColumnType,
+    str_dict_column: &Option<StrColumn>,
+    column: &Column,
+) -> crate::Result<CompositeIntermediateKey> {
+    let key = if *column_type == ColumnType::Str {
+        let fallback_dict = Dictionary::empty();
+        let term_dict = str_dict_column
+            .as_ref()
+            .map(|el| el.dictionary())
+            .unwrap_or_else(|| &fallback_dict);
+
+        let mut buffer = Vec::new();
+        term_dict.ord_to_term(val, &mut buffer)?;
+        CompositeIntermediateKey::Str(
+            String::from_utf8(buffer.to_vec()).expect("could not convert to String"),
+        )
+    } else if *column_type == ColumnType::DateTime {
+        let val = i64::from_u64(val);
+        CompositeIntermediateKey::DateTime(val)
+    } else if *column_type == ColumnType::Bool {
+        let val = bool::from_u64(val);
+        CompositeIntermediateKey::Bool(val)
+    } else if *column_type == ColumnType::IpAddr {
+        let compact_space_accessor = column
+            .values
+            .clone()
+            .downcast_arc::<CompactSpaceU64Accessor>()
+            .map_err(|_| {
+                TantivyError::AggregationError(crate::aggregation::AggregationError::InternalError(
+                    "Type mismatch: Could not downcast to CompactSpaceU64Accessor".to_string(),
+                ))
+            })?;
+        let val: u128 = compact_space_accessor.compact_to_u128(val as u32);
+        let val = Ipv6Addr::from_u128(val);
+        CompositeIntermediateKey::IpAddr(val)
+    } else if *column_type == ColumnType::U64 {
+        CompositeIntermediateKey::U64(val)
+    } else if *column_type == ColumnType::I64 {
+        CompositeIntermediateKey::I64(i64::from_u64(val))
+    } else {
+        let val = f64::from_u64(val);
+        let val: NumericalValue = val.into();
+
+        match val.normalize() {
+            NumericalValue::U64(val) => CompositeIntermediateKey::U64(val),
+            NumericalValue::I64(val) => CompositeIntermediateKey::I64(val),
+            NumericalValue::F64(val) => CompositeIntermediateKey::F64(val),
+        }
+    };
+    Ok(key)
+}
+
+/// Browse through the cardinal product obtained by the different values of the doc composite key
+/// sources.
+///
+/// For each of those tuple-key, that are after the limit key, we call collect_bucket_with_limit.
+struct CompositeKeyVisitor<'a> {
+    doc_id: crate::DocId,
+    composite_agg_data: &'a CompositeAggReqData,
+    buckets: &'a mut DynArrayHeapMap<InternalValueRepr, CompositeBucketCollector>,
+    sub_agg: &'a mut Option<BufferedSubAggs<HighCardSubAggBuffer>>,
+    bucket_id_provider: &'a mut BucketIdProvider,
+    sub_level_values: SmallVec<[InternalValueRepr; MAX_DYN_ARRAY_SIZE]>,
+}
+
+impl CompositeKeyVisitor<'_> {
+    /// Depth-first walk of the accessors to build the composite key combinations
+    /// and update the buckets.
+    ///
+    /// `source_idx` is the current source index in the recursion.
+    /// `is_on_after_key` tracks whether we still need to consider the after_key
+    /// for pruning at this level and below.
+    fn visit(&mut self, source_idx: usize, is_on_after_key: bool) -> crate::Result<()> {
+        if source_idx == self.composite_agg_data.req.sources.len() {
+            if !is_on_after_key {
+                collect_bucket_with_limit(
+                    self.doc_id,
+                    self.composite_agg_data.req.size as usize,
+                    self.buckets,
+                    &self.sub_level_values,
+                    self.sub_agg,
+                    self.bucket_id_provider,
+                );
+            }
+            return Ok(());
+        }
+
+        let current_level_accessors = &self.composite_agg_data.composite_accessors[source_idx];
+        let current_level_source = &self.composite_agg_data.req.sources[source_idx];
+        let mut missing = true;
+        for (accessor_idx, accessor) in current_level_accessors.accessors.iter().enumerate() {
+            let values = accessor.column.values_for_doc(self.doc_id);
+            for value in values {
+                missing = false;
+                match current_level_source {
+                    CompositeAggregationSource::Terms(_) => {
+                        let preceeds_after_key_type =
+                            accessor_idx < current_level_accessors.after_key_accessor_idx;
+                        if is_on_after_key && preceeds_after_key_type {
+                            break;
+                        }
+                        let matches_after_key_type =
+                            accessor_idx == current_level_accessors.after_key_accessor_idx;
+
+                        if matches_after_key_type && is_on_after_key {
+                            let should_skip = match current_level_source.order() {
+                                Order::Asc => current_level_accessors.after_key.gt(value),
+                                Order::Desc => current_level_accessors.after_key.lt(value),
+                            };
+                            if should_skip {
+                                continue;
+                            }
+                        }
+                        self.sub_level_values.push(InternalValueRepr::new_term(
+                            value,
+                            accessor_idx as u8,
+                            current_level_source.order(),
+                        ));
+                        let still_on_after_key = matches_after_key_type
+                            && current_level_accessors.after_key.equals(value);
+                        self.visit(source_idx + 1, is_on_after_key && still_on_after_key)?;
+                        self.sub_level_values.pop();
+                    }
+                    CompositeAggregationSource::Histogram(source) => {
+                        let float_value = match accessor.column_type {
+                            ColumnType::U64 => value as f64,
+                            ColumnType::I64 => i64::from_u64(value) as f64,
+                            ColumnType::DateTime => i64::from_u64(value) as f64 / 1_000_000.,
+                            ColumnType::F64 => f64::from_u64(value),
+                            _ => {
+                                panic!(
+                                    "unexpected type {:?}. This should not happen",
+                                    accessor.column_type
+                                )
+                            }
+                        };
+                        let bucket_index = (float_value / source.interval).floor() as i64;
+                        let bucket_value = i64::to_u64(bucket_index);
+                        if is_on_after_key {
+                            let should_skip = match current_level_source.order() {
+                                Order::Asc => current_level_accessors.after_key.gt(bucket_value),
+                                Order::Desc => current_level_accessors.after_key.lt(bucket_value),
+                            };
+                            if should_skip {
+                                continue;
+                            }
+                        }
+                        self.sub_level_values.push(InternalValueRepr::new_histogram(
+                            bucket_value,
+                            current_level_source.order(),
+                        ));
+                        let still_on_after_key =
+                            current_level_accessors.after_key.equals(bucket_value);
+                        self.visit(source_idx + 1, is_on_after_key && still_on_after_key)?;
+                        self.sub_level_values.pop();
+                    }
+                    CompositeAggregationSource::DateHistogram(_) => {
+                        let value_ns = match accessor.column_type {
+                            ColumnType::DateTime => i64::from_u64(value),
+                            _ => {
+                                panic!(
+                                    "unexpected type {:?}. This should not happen",
+                                    accessor.column_type
+                                )
+                            }
+                        };
+                        let bucket_index = match accessor.date_histogram_interval {
+                            PrecomputedDateInterval::FixedNanoseconds(fixed_interval_ns) => {
+                                (value_ns / fixed_interval_ns) * fixed_interval_ns
+                            }
+                            PrecomputedDateInterval::Calendar(CalendarInterval::Year) => {
+                                calendar_interval::try_year_bucket(value_ns)?
+                            }
+                            PrecomputedDateInterval::Calendar(CalendarInterval::Month) => {
+                                calendar_interval::try_month_bucket(value_ns)?
+                            }
+                            PrecomputedDateInterval::Calendar(CalendarInterval::Week) => {
+                                calendar_interval::week_bucket(value_ns)
+                            }
+                            PrecomputedDateInterval::NotApplicable => {
+                                panic!("interval not precomputed for date histogram source")
+                            }
+                        };
+                        let bucket_value = i64::to_u64(bucket_index);
+                        if is_on_after_key {
+                            let should_skip = match current_level_source.order() {
+                                Order::Asc => current_level_accessors.after_key.gt(bucket_value),
+                                Order::Desc => current_level_accessors.after_key.lt(bucket_value),
+                            };
+                            if should_skip {
+                                continue;
+                            }
+                        }
+                        self.sub_level_values.push(InternalValueRepr::new_histogram(
+                            bucket_value,
+                            current_level_source.order(),
+                        ));
+                        let still_on_after_key =
+                            current_level_accessors.after_key.equals(bucket_value);
+                        self.visit(source_idx + 1, is_on_after_key && still_on_after_key)?;
+                        self.sub_level_values.pop();
+                    }
+                };
+            }
+        }
+        if missing && current_level_source.missing_bucket() {
+            if is_on_after_key && current_level_accessors.skip_missing {
+                return Ok(());
+            }
+            self.sub_level_values.push(InternalValueRepr::new_missing(
+                current_level_source.order(),
+                current_level_source.missing_order(),
+            ));
+            self.visit(
+                source_idx + 1,
+                is_on_after_key && current_level_accessors.is_after_key_explicit_missing,
+            )?;
+            self.sub_level_values.pop();
+        }
+        Ok(())
+    }
+}
--- a/src/aggregation/bucket/composite/map.rs
+++ b/src/aggregation/bucket/composite/map.rs
@@ -0,0 +1,329 @@
+use std::collections::BinaryHeap;
+use std::fmt::Debug;
+use std::hash::Hash;
+
+use rustc_hash::FxHashMap;
+use smallvec::SmallVec;
+
+use crate::TantivyError;
+
+/// Map backed by a hash map for fast access and a binary heap to track the
+/// highest key. The key is an array of fixed size S.
+#[derive(Clone, Debug)]
+struct ArrayHeapMap<K: Ord, V, const S: usize> {
+    pub(crate) buckets: FxHashMap<[K; S], V>,
+    pub(crate) heap: BinaryHeap<[K; S]>,
+}
+
+impl<K: Ord, V, const S: usize> Default for ArrayHeapMap<K, V, S> {
+    fn default() -> Self {
+        ArrayHeapMap {
+            buckets: FxHashMap::default(),
+            heap: BinaryHeap::default(),
+        }
+    }
+}
+
+impl<K: Eq + Hash + Clone + Ord, V, const S: usize> ArrayHeapMap<K, V, S> {
+    /// Panics if the length of `key` is not S.
+    fn get_or_insert_with<F: FnOnce() -> V>(&mut self, key: &[K], f: F) -> &mut V {
+        let key_array: &[K; S] = key.try_into().expect("Key length mismatch");
+        self.buckets.entry(key_array.clone()).or_insert_with(|| {
+            self.heap.push(key_array.clone());
+            f()
+        })
+    }
+
+    /// Panics if the length of `key` is not S.
+    fn get_mut(&mut self, key: &[K]) -> Option<&mut V> {
+        let key_array: &[K; S] = key.try_into().expect("Key length mismatch");
+        self.buckets.get_mut(key_array)
+    }
+
+    fn peek_highest(&self) -> Option<&[K]> {
+        self.heap.peek().map(|k_array| k_array.as_slice())
+    }
+
+    fn evict_highest(&mut self) {
+        if let Some(highest) = self.heap.pop() {
+            self.buckets.remove(&highest);
+        }
+    }
+
+    fn memory_consumption(&self) -> u64 {
+        let key_size = std::mem::size_of::<[K; S]>();
+        let map_size = (key_size + std::mem::size_of::<V>()) * self.buckets.capacity();
+        let heap_size = key_size * self.heap.capacity();
+        (map_size + heap_size) as u64
+    }
+}
+
+impl<K: Copy + Ord + Clone + 'static, V: 'static, const S: usize> ArrayHeapMap<K, V, S> {
+    fn into_iter(self) -> Box<dyn Iterator<Item = (SmallVec<[K; MAX_DYN_ARRAY_SIZE]>, V)>> {
+        Box::new(
+            self.buckets
+                .into_iter()
+                .map(|(k, v)| (SmallVec::from_slice(&k), v)),
+        )
+    }
+}
+
+pub(super) const MAX_DYN_ARRAY_SIZE: usize = 16;
+const MAX_DYN_ARRAY_SIZE_PLUS_ONE: usize = MAX_DYN_ARRAY_SIZE + 1;
+
+/// A map optimized for memory footprint, fast access and efficient eviction of
+/// the highest key.
+///
+/// Keys are inlined arrays of size 1 to [MAX_DYN_ARRAY_SIZE] but for a given
+/// instance the key size is fixed. This allows to avoid heap allocations for the
+/// keys.
+#[derive(Clone, Debug)]
+pub(super) struct DynArrayHeapMap<K: Ord, V>(DynArrayHeapMapInner<K, V>);
+
+/// Wrapper around ArrayHeapMap to dynamically dispatch on the array size.
+#[derive(Clone, Debug)]
+enum DynArrayHeapMapInner<K: Ord, V> {
+    Dim1(ArrayHeapMap<K, V, 1>),
+    Dim2(ArrayHeapMap<K, V, 2>),
+    Dim3(ArrayHeapMap<K, V, 3>),
+    Dim4(ArrayHeapMap<K, V, 4>),
+    Dim5(ArrayHeapMap<K, V, 5>),
+    Dim6(ArrayHeapMap<K, V, 6>),
+    Dim7(ArrayHeapMap<K, V, 7>),
+    Dim8(ArrayHeapMap<K, V, 8>),
+    Dim9(ArrayHeapMap<K, V, 9>),
+    Dim10(ArrayHeapMap<K, V, 10>),
+    Dim11(ArrayHeapMap<K, V, 11>),
+    Dim12(ArrayHeapMap<K, V, 12>),
+    Dim13(ArrayHeapMap<K, V, 13>),
+    Dim14(ArrayHeapMap<K, V, 14>),
+    Dim15(ArrayHeapMap<K, V, 15>),
+    Dim16(ArrayHeapMap<K, V, 16>),
+}
+
+impl<K: Ord, V> DynArrayHeapMap<K, V> {
+    /// Creates a new heap map with dynamic array keys of size `key_dimension`.
+    pub(super) fn try_new(key_dimension: usize) -> crate::Result<Self> {
+        let inner = match key_dimension {
+            0 => {
+                return Err(TantivyError::InvalidArgument(
+                    "DynArrayHeapMap dimension must be at least 1".to_string(),
+                ))
+            }
+            1 => DynArrayHeapMapInner::Dim1(ArrayHeapMap::default()),
+            2 => DynArrayHeapMapInner::Dim2(ArrayHeapMap::default()),
+            3 => DynArrayHeapMapInner::Dim3(ArrayHeapMap::default()),
+            4 => DynArrayHeapMapInner::Dim4(ArrayHeapMap::default()),
+            5 => DynArrayHeapMapInner::Dim5(ArrayHeapMap::default()),
+            6 => DynArrayHeapMapInner::Dim6(ArrayHeapMap::default()),
+            7 => DynArrayHeapMapInner::Dim7(ArrayHeapMap::default()),
+            8 => DynArrayHeapMapInner::Dim8(ArrayHeapMap::default()),
+            9 => DynArrayHeapMapInner::Dim9(ArrayHeapMap::default()),
+            10 => DynArrayHeapMapInner::Dim10(ArrayHeapMap::default()),
+            11 => DynArrayHeapMapInner::Dim11(ArrayHeapMap::default()),
+            12 => DynArrayHeapMapInner::Dim12(ArrayHeapMap::default()),
+            13 => DynArrayHeapMapInner::Dim13(ArrayHeapMap::default()),
+            14 => DynArrayHeapMapInner::Dim14(ArrayHeapMap::default()),
+            15 => DynArrayHeapMapInner::Dim15(ArrayHeapMap::default()),
+            16 => DynArrayHeapMapInner::Dim16(ArrayHeapMap::default()),
+            MAX_DYN_ARRAY_SIZE_PLUS_ONE.. => {
+                return Err(TantivyError::InvalidArgument(format!(
+                    "DynArrayHeapMap supports maximum {MAX_DYN_ARRAY_SIZE} dimensions, got \
+                     {key_dimension}",
+                )))
+            }
+        };
+        Ok(DynArrayHeapMap(inner))
+    }
+
+    /// Number of elements in the map. This is not the dimension of the keys.
+    pub(super) fn size(&self) -> usize {
+        match &self.0 {
+            DynArrayHeapMapInner::Dim1(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim2(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim3(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim4(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim5(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim6(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim7(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim8(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim9(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim10(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim11(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim12(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim13(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim14(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim15(map) => map.buckets.len(),
+            DynArrayHeapMapInner::Dim16(map) => map.buckets.len(),
+        }
+    }
+}
+
+impl<K: Ord + Hash + Clone, V> DynArrayHeapMap<K, V> {
+    /// Get a mutable reference to the value corresponding to `key` or inserts a new
+    /// value created by calling `f`.
+    ///
+    /// Panics if the length of `key` does not match the key dimension of the map.
+    pub(super) fn get_or_insert_with<F: FnOnce() -> V>(&mut self, key: &[K], f: F) -> &mut V {
+        match &mut self.0 {
+            DynArrayHeapMapInner::Dim1(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim2(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim3(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim4(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim5(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim6(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim7(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim8(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim9(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim10(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim11(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim12(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim13(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim14(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim15(map) => map.get_or_insert_with(key, f),
+            DynArrayHeapMapInner::Dim16(map) => map.get_or_insert_with(key, f),
+        }
+    }
+
+    /// Returns a mutable reference to the value corresponding to `key`.
+    ///
+    /// Panics if the length of `key` does not match the key dimension of the map.
+    pub fn get_mut(&mut self, key: &[K]) -> Option<&mut V> {
+        match &mut self.0 {
+            DynArrayHeapMapInner::Dim1(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim2(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim3(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim4(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim5(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim6(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim7(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim8(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim9(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim10(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim11(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim12(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim13(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim14(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim15(map) => map.get_mut(key),
+            DynArrayHeapMapInner::Dim16(map) => map.get_mut(key),
+        }
+    }
+
+    /// Returns a reference to the highest key in the map.
+    pub(super) fn peek_highest(&self) -> Option<&[K]> {
+        match &self.0 {
+            DynArrayHeapMapInner::Dim1(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim2(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim3(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim4(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim5(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim6(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim7(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim8(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim9(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim10(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim11(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim12(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim13(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim14(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim15(map) => map.peek_highest(),
+            DynArrayHeapMapInner::Dim16(map) => map.peek_highest(),
+        }
+    }
+
+    /// Removes the entry with the highest key from the map.
+    pub(super) fn evict_highest(&mut self) {
+        match &mut self.0 {
+            DynArrayHeapMapInner::Dim1(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim2(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim3(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim4(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim5(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim6(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim7(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim8(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim9(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim10(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim11(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim12(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim13(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim14(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim15(map) => map.evict_highest(),
+            DynArrayHeapMapInner::Dim16(map) => map.evict_highest(),
+        }
+    }
+
+    pub(crate) fn memory_consumption(&self) -> u64 {
+        match &self.0 {
+            DynArrayHeapMapInner::Dim1(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim2(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim3(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim4(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim5(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim6(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim7(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim8(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim9(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim10(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim11(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim12(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim13(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim14(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim15(map) => map.memory_consumption(),
+            DynArrayHeapMapInner::Dim16(map) => map.memory_consumption(),
+        }
+    }
+}
+
+impl<K: Ord + Clone + Copy + 'static, V: 'static> DynArrayHeapMap<K, V> {
+    /// Turns this map into an iterator over key-value pairs.
+    pub fn into_iter(self) -> impl Iterator<Item = (SmallVec<[K; MAX_DYN_ARRAY_SIZE]>, V)> {
+        match self.0 {
+            DynArrayHeapMapInner::Dim1(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim2(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim3(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim4(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim5(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim6(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim7(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim8(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim9(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim10(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim11(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim12(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim13(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim14(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim15(map) => map.into_iter(),
+            DynArrayHeapMapInner::Dim16(map) => map.into_iter(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_dyn_array_heap_map() {
+        let mut map = DynArrayHeapMap::<u32, &str>::try_new(2).unwrap();
+        // insert
+        let key1 = [1u32, 2u32];
+        let key2 = [2u32, 1u32];
+        map.get_or_insert_with(&key1, || "a");
+        map.get_or_insert_with(&key2, || "b");
+        assert_eq!(map.size(), 2);
+
+        // evict highest
+        assert_eq!(map.peek_highest(), Some(&key2[..]));
+        map.evict_highest();
+        assert_eq!(map.size(), 1);
+        assert_eq!(map.peek_highest(), Some(&key1[..]));
+
+        // into_iter
+        let mut iter = map.into_iter();
+        let (k, v) = iter.next().unwrap();
+        assert_eq!(k.as_slice(), &key1);
+        assert_eq!(v, "a");
+        assert_eq!(iter.next(), None);
+    }
+}
--- a/src/aggregation/bucket/composite/mod.rs
+++ b/src/aggregation/bucket/composite/mod.rs
--- a/src/aggregation/bucket/composite/numeric_types.rs
+++ b/src/aggregation/bucket/composite/numeric_types.rs
@@ -0,0 +1,460 @@
+/// This module helps comparing numerical values of different types (i64, u64
+/// and f64).
+pub(super) mod num_cmp {
+    use std::cmp::Ordering;
+
+    use crate::TantivyError;
+
+    pub fn cmp_i64_f64(left_i: i64, right_f: f64) -> crate::Result<Ordering> {
+        if right_f.is_nan() {
+            return Err(TantivyError::InvalidArgument(
+                "NaN comparison is not supported".to_string(),
+            ));
+        }
+
+        // If right_f is < i64::MIN then left_i > right_f (i64::MIN=-2^63 can be
+        // exactly represented as f64)
+        if right_f < i64::MIN as f64 {
+            return Ok(Ordering::Greater);
+        }
+        // If right_f is >= i64::MAX then left_i < right_f (i64::MAX=2^63-1 cannot
+        // be exactly represented as f64)
+        if right_f >= i64::MAX as f64 {
+            return Ok(Ordering::Less);
+        }
+
+        // Now right_f is in (i64::MIN, i64::MAX), so `right_f as i64` is
+        // well-defined (truncation toward 0)
+        let right_as_i = right_f as i64;
+
+        let result = match left_i.cmp(&right_as_i) {
+            Ordering::Less => Ordering::Less,
+            Ordering::Greater => Ordering::Greater,
+            Ordering::Equal => {
+                // they have the same integer part, compare the fraction
+                let rem = right_f - (right_as_i as f64);
+                if rem == 0.0 {
+                    Ordering::Equal
+                } else if right_f > 0.0 {
+                    Ordering::Less
+                } else {
+                    Ordering::Greater
+                }
+            }
+        };
+        Ok(result)
+    }
+
+    pub fn cmp_u64_f64(left_u: u64, right_f: f64) -> crate::Result<Ordering> {
+        if right_f.is_nan() {
+            return Err(TantivyError::InvalidArgument(
+                "NaN comparison is not supported".to_string(),
+            ));
+        }
+
+        // Negative floats are always less than any u64 >= 0
+        if right_f < 0.0 {
+            return Ok(Ordering::Greater);
+        }
+
+        // If right_f is >= u64::MAX then left_u < right_f (u64::MAX=2^64-1 cannot be exactly)
+        let max_as_f = u64::MAX as f64;
+        if right_f > max_as_f {
+            return Ok(Ordering::Less);
+        }
+
+        // Now right_f is in (0, u64::MAX), so `right_f as u64` is well-defined
+        // (truncation toward 0)
+        let right_as_u = right_f as u64;
+
+        let result = match left_u.cmp(&right_as_u) {
+            Ordering::Less => Ordering::Less,
+            Ordering::Greater => Ordering::Greater,
+            Ordering::Equal => {
+                // they have the same integer part, compare the fraction
+                let rem = right_f - (right_as_u as f64);
+                if rem == 0.0 {
+                    Ordering::Equal
+                } else {
+                    Ordering::Less
+                }
+            }
+        };
+        Ok(result)
+    }
+
+    pub fn cmp_i64_u64(left_i: i64, right_u: u64) -> Ordering {
+        if left_i < 0 {
+            Ordering::Less
+        } else {
+            let left_as_u = left_i as u64;
+            left_as_u.cmp(&right_u)
+        }
+    }
+}
+
+/// This module helps projecting numerical values to other numerical types.
+/// When the target value space cannot exactly represent the source value, the
+/// next representable value is returned (or AfterLast if the source value is
+/// larger than the largest representable value).
+///
+/// All functions in this module assume that f64 values are not NaN.
+pub(super) mod num_proj {
+    #[derive(Debug, PartialEq)]
+    pub enum ProjectedNumber<T> {
+        Exact(T),
+        Next(T),
+        AfterLast,
+    }
+
+    pub fn i64_to_u64(value: i64) -> ProjectedNumber<u64> {
+        if value < 0 {
+            ProjectedNumber::Next(0)
+        } else {
+            ProjectedNumber::Exact(value as u64)
+        }
+    }
+
+    pub fn u64_to_i64(value: u64) -> ProjectedNumber<i64> {
+        if value > i64::MAX as u64 {
+            ProjectedNumber::AfterLast
+        } else {
+            ProjectedNumber::Exact(value as i64)
+        }
+    }
+
+    pub fn f64_to_u64(value: f64) -> ProjectedNumber<u64> {
+        if value < 0.0 {
+            ProjectedNumber::Next(0)
+        } else if value > u64::MAX as f64 {
+            ProjectedNumber::AfterLast
+        } else if value.fract() == 0.0 {
+            ProjectedNumber::Exact(value as u64)
+        } else {
+            // casting f64 to u64 truncates toward zero
+            ProjectedNumber::Next(value as u64 + 1)
+        }
+    }
+
+    pub fn f64_to_i64(value: f64) -> ProjectedNumber<i64> {
+        if value < (i64::MIN as f64) {
+            ProjectedNumber::Next(i64::MIN)
+        } else if value >= (i64::MAX as f64) {
+            ProjectedNumber::AfterLast
+        } else if value.fract() == 0.0 {
+            ProjectedNumber::Exact(value as i64)
+        } else if value > 0.0 {
+            // casting f64 to i64 truncates toward zero
+            ProjectedNumber::Next(value as i64 + 1)
+        } else {
+            ProjectedNumber::Next(value as i64)
+        }
+    }
+
+    pub fn i64_to_f64(value: i64) -> ProjectedNumber<f64> {
+        let value_f = value as f64;
+        let k_roundtrip = value_f as i64;
+        if k_roundtrip == value {
+            // between -2^53 and 2^53 all i64 are exactly represented as f64
+            ProjectedNumber::Exact(value_f)
+        } else {
+            // for very large/small i64 values, it is approximated to the closest f64
+            if k_roundtrip > value {
+                ProjectedNumber::Next(value_f)
+            } else {
+                ProjectedNumber::Next(value_f.next_up())
+            }
+        }
+    }
+
+    pub fn u64_to_f64(value: u64) -> ProjectedNumber<f64> {
+        let value_f = value as f64;
+        let k_roundtrip = value_f as u64;
+        if k_roundtrip == value {
+            // between 0 and 2^53 all u64 are exactly represented as f64
+            ProjectedNumber::Exact(value_f)
+        } else if k_roundtrip > value {
+            ProjectedNumber::Next(value_f)
+        } else {
+            ProjectedNumber::Next(value_f.next_up())
+        }
+    }
+}
+
+#[cfg(test)]
+mod num_cmp_tests {
+    use std::cmp::Ordering;
+
+    use super::num_cmp::*;
+
+    #[test]
+    fn test_cmp_u64_f64() {
+        // Basic comparisons
+        assert_eq!(cmp_u64_f64(5, 5.0).unwrap(), Ordering::Equal);
+        assert_eq!(cmp_u64_f64(5, 6.0).unwrap(), Ordering::Less);
+        assert_eq!(cmp_u64_f64(6, 5.0).unwrap(), Ordering::Greater);
+        assert_eq!(cmp_u64_f64(0, 0.0).unwrap(), Ordering::Equal);
+        assert_eq!(cmp_u64_f64(0, 0.1).unwrap(), Ordering::Less);
+
+        // Negative float values should always be less than any u64
+        assert_eq!(cmp_u64_f64(0, -0.1).unwrap(), Ordering::Greater);
+        assert_eq!(cmp_u64_f64(5, -5.0).unwrap(), Ordering::Greater);
+        assert_eq!(cmp_u64_f64(u64::MAX, -1e20).unwrap(), Ordering::Greater);
+
+        // Tests with extreme values
+        assert_eq!(cmp_u64_f64(u64::MAX, 1e20).unwrap(), Ordering::Less);
+
+        // Precision edge cases: large u64 that loses precision when converted to f64
+        // => 2^54, exactly represented as f64
+        let large_f64 = 18_014_398_509_481_984.0;
+        let large_u64 = 18_014_398_509_481_984;
+        // prove that large_u64 is exactly represented as f64
+        assert_eq!(large_u64 as f64, large_f64);
+        assert_eq!(cmp_u64_f64(large_u64, large_f64).unwrap(), Ordering::Equal);
+        // => (2^54 + 1) cannot be exactly represented in f64
+        let large_u64_plus_1 = 18_014_398_509_481_985;
+        // prove that it is represented as f64 by large_f64
+        assert_eq!(large_u64_plus_1 as f64, large_f64);
+        assert_eq!(
+            cmp_u64_f64(large_u64_plus_1, large_f64).unwrap(),
+            Ordering::Greater
+        );
+        // => (2^54 - 1) cannot be exactly represented in f64
+        let large_u64_minus_1 = 18_014_398_509_481_983;
+        // prove that it is also represented as f64 by large_f64
+        assert_eq!(large_u64_minus_1 as f64, large_f64);
+        assert_eq!(
+            cmp_u64_f64(large_u64_minus_1, large_f64).unwrap(),
+            Ordering::Less
+        );
+
+        // NaN comparison results in an error
+        assert!(cmp_u64_f64(0, f64::NAN).is_err());
+    }
+
+    #[test]
+    fn test_cmp_i64_f64() {
+        // Basic comparisons
+        assert_eq!(cmp_i64_f64(5, 5.0).unwrap(), Ordering::Equal);
+        assert_eq!(cmp_i64_f64(5, 6.0).unwrap(), Ordering::Less);
+        assert_eq!(cmp_i64_f64(6, 5.0).unwrap(), Ordering::Greater);
+        assert_eq!(cmp_i64_f64(-5, -5.0).unwrap(), Ordering::Equal);
+        assert_eq!(cmp_i64_f64(-5, -4.0).unwrap(), Ordering::Less);
+        assert_eq!(cmp_i64_f64(-4, -5.0).unwrap(), Ordering::Greater);
+        assert_eq!(cmp_i64_f64(-5, 5.0).unwrap(), Ordering::Less);
+        assert_eq!(cmp_i64_f64(5, -5.0).unwrap(), Ordering::Greater);
+        assert_eq!(cmp_i64_f64(0, -0.1).unwrap(), Ordering::Greater);
+        assert_eq!(cmp_i64_f64(0, 0.1).unwrap(), Ordering::Less);
+        assert_eq!(cmp_i64_f64(-1, -0.5).unwrap(), Ordering::Less);
+        assert_eq!(cmp_i64_f64(-1, 0.0).unwrap(), Ordering::Less);
+        assert_eq!(cmp_i64_f64(0, 0.0).unwrap(), Ordering::Equal);
+
+        // Tests with extreme values
+        assert_eq!(cmp_i64_f64(i64::MAX, 1e20).unwrap(), Ordering::Less);
+        assert_eq!(cmp_i64_f64(i64::MIN, -1e20).unwrap(), Ordering::Greater);
+
+        // Precision edge cases: large i64 that loses precision when converted to f64
+        // => 2^54, exactly represented as f64
+        let large_f64 = 18_014_398_509_481_984.0;
+        let large_i64 = 18_014_398_509_481_984;
+        // prove that large_i64 is exactly represented as f64
+        assert_eq!(large_i64 as f64, large_f64);
+        assert_eq!(cmp_i64_f64(large_i64, large_f64).unwrap(), Ordering::Equal);
+        // => (1_i64 << 54) + 1 cannot be exactly represented in f64
+        let large_i64_plus_1 = 18_014_398_509_481_985;
+        // prove that it is represented as f64 by large_f64
+        assert_eq!(large_i64_plus_1 as f64, large_f64);
+        assert_eq!(
+            cmp_i64_f64(large_i64_plus_1, large_f64).unwrap(),
+            Ordering::Greater
+        );
+        // => (1_i64 << 54) - 1 cannot be exactly represented in f64
+        let large_i64_minus_1 = 18_014_398_509_481_983;
+        // prove that it is also represented as f64 by large_f64
+        assert_eq!(large_i64_minus_1 as f64, large_f64);
+        assert_eq!(
+            cmp_i64_f64(large_i64_minus_1, large_f64).unwrap(),
+            Ordering::Less
+        );
+
+        // Same precision edge case but with negative values
+        // => -2^54, exactly represented as f64
+        let large_neg_f64 = -18_014_398_509_481_984.0;
+        let large_neg_i64 = -18_014_398_509_481_984;
+        // prove that large_neg_i64 is exactly represented as f64
+        assert_eq!(large_neg_i64 as f64, large_neg_f64);
+        assert_eq!(
+            cmp_i64_f64(large_neg_i64, large_neg_f64).unwrap(),
+            Ordering::Equal
+        );
+        // => (-2^54 + 1) cannot be exactly represented in f64
+        let large_neg_i64_plus_1 = -18_014_398_509_481_985;
+        // prove that it is represented as f64 by large_neg_f64
+        assert_eq!(large_neg_i64_plus_1 as f64, large_neg_f64);
+        assert_eq!(
+            cmp_i64_f64(large_neg_i64_plus_1, large_neg_f64).unwrap(),
+            Ordering::Less
+        );
+        // => (-2^54 - 1) cannot be exactly represented in f64
+        let large_neg_i64_minus_1 = -18_014_398_509_481_983;
+        // prove that it is also represented as f64 by large_neg_f64
+        assert_eq!(large_neg_i64_minus_1 as f64, large_neg_f64);
+        assert_eq!(
+            cmp_i64_f64(large_neg_i64_minus_1, large_neg_f64).unwrap(),
+            Ordering::Greater
+        );
+
+        // NaN comparison results in an error
+        assert!(cmp_i64_f64(0, f64::NAN).is_err());
+    }
+
+    #[test]
+    fn test_cmp_i64_u64() {
+        // Test with negative i64 values (should always be less than any u64)
+        assert_eq!(cmp_i64_u64(-1, 0), Ordering::Less);
+        assert_eq!(cmp_i64_u64(i64::MIN, 0), Ordering::Less);
+        assert_eq!(cmp_i64_u64(i64::MIN, u64::MAX), Ordering::Less);
+
+        // Test with positive i64 values
+        assert_eq!(cmp_i64_u64(0, 0), Ordering::Equal);
+        assert_eq!(cmp_i64_u64(1, 0), Ordering::Greater);
+        assert_eq!(cmp_i64_u64(1, 1), Ordering::Equal);
+        assert_eq!(cmp_i64_u64(0, 1), Ordering::Less);
+        assert_eq!(cmp_i64_u64(5, 10), Ordering::Less);
+        assert_eq!(cmp_i64_u64(10, 5), Ordering::Greater);
+
+        // Test with values near i64::MAX and u64 conversion
+        assert_eq!(cmp_i64_u64(i64::MAX, i64::MAX as u64), Ordering::Equal);
+        assert_eq!(cmp_i64_u64(i64::MAX, (i64::MAX as u64) + 1), Ordering::Less);
+        assert_eq!(cmp_i64_u64(i64::MAX, u64::MAX), Ordering::Less);
+    }
+}
+
+#[cfg(test)]
+mod num_proj_tests {
+    use super::num_proj::{self, ProjectedNumber};
+
+    #[test]
+    fn test_i64_to_u64() {
+        assert_eq!(num_proj::i64_to_u64(-1), ProjectedNumber::Next(0));
+        assert_eq!(num_proj::i64_to_u64(i64::MIN), ProjectedNumber::Next(0));
+        assert_eq!(num_proj::i64_to_u64(0), ProjectedNumber::Exact(0));
+        assert_eq!(num_proj::i64_to_u64(42), ProjectedNumber::Exact(42));
+        assert_eq!(
+            num_proj::i64_to_u64(i64::MAX),
+            ProjectedNumber::Exact(i64::MAX as u64)
+        );
+    }
+
+    #[test]
+    fn test_u64_to_i64() {
+        assert_eq!(num_proj::u64_to_i64(0), ProjectedNumber::Exact(0));
+        assert_eq!(num_proj::u64_to_i64(42), ProjectedNumber::Exact(42));
+        assert_eq!(
+            num_proj::u64_to_i64(i64::MAX as u64),
+            ProjectedNumber::Exact(i64::MAX)
+        );
+        assert_eq!(
+            num_proj::u64_to_i64((i64::MAX as u64) + 1),
+            ProjectedNumber::AfterLast
+        );
+        assert_eq!(num_proj::u64_to_i64(u64::MAX), ProjectedNumber::AfterLast);
+    }
+
+    #[test]
+    fn test_f64_to_u64() {
+        assert_eq!(num_proj::f64_to_u64(-1e25), ProjectedNumber::Next(0));
+        assert_eq!(num_proj::f64_to_u64(-0.1), ProjectedNumber::Next(0));
+        assert_eq!(num_proj::f64_to_u64(1e20), ProjectedNumber::AfterLast);
+        assert_eq!(
+            num_proj::f64_to_u64(f64::INFINITY),
+            ProjectedNumber::AfterLast
+        );
+        assert_eq!(num_proj::f64_to_u64(0.0), ProjectedNumber::Exact(0));
+        assert_eq!(num_proj::f64_to_u64(42.0), ProjectedNumber::Exact(42));
+        assert_eq!(num_proj::f64_to_u64(0.5), ProjectedNumber::Next(1));
+        assert_eq!(num_proj::f64_to_u64(42.1), ProjectedNumber::Next(43));
+    }
+
+    #[test]
+    fn test_f64_to_i64() {
+        assert_eq!(num_proj::f64_to_i64(-1e20), ProjectedNumber::Next(i64::MIN));
+        assert_eq!(
+            num_proj::f64_to_i64(f64::NEG_INFINITY),
+            ProjectedNumber::Next(i64::MIN)
+        );
+        assert_eq!(num_proj::f64_to_i64(1e20), ProjectedNumber::AfterLast);
+        assert_eq!(
+            num_proj::f64_to_i64(f64::INFINITY),
+            ProjectedNumber::AfterLast
+        );
+        assert_eq!(num_proj::f64_to_i64(0.0), ProjectedNumber::Exact(0));
+        assert_eq!(num_proj::f64_to_i64(42.0), ProjectedNumber::Exact(42));
+        assert_eq!(num_proj::f64_to_i64(-42.0), ProjectedNumber::Exact(-42));
+        assert_eq!(num_proj::f64_to_i64(0.5), ProjectedNumber::Next(1));
+        assert_eq!(num_proj::f64_to_i64(42.1), ProjectedNumber::Next(43));
+        assert_eq!(num_proj::f64_to_i64(-0.5), ProjectedNumber::Next(0));
+        assert_eq!(num_proj::f64_to_i64(-42.1), ProjectedNumber::Next(-42));
+    }
+
+    #[test]
+    fn test_i64_to_f64() {
+        assert_eq!(num_proj::i64_to_f64(0), ProjectedNumber::Exact(0.0));
+        assert_eq!(num_proj::i64_to_f64(42), ProjectedNumber::Exact(42.0));
+        assert_eq!(num_proj::i64_to_f64(-42), ProjectedNumber::Exact(-42.0));
+
+        let max_exact = 9_007_199_254_740_992; // 2^53
+        assert_eq!(
+            num_proj::i64_to_f64(max_exact),
+            ProjectedNumber::Exact(max_exact as f64)
+        );
+
+        // Test values that cannot be exactly represented as f64 (integers above 2^53)
+        let large_i64 = 9_007_199_254_740_993; // 2^53 + 1
+        let closest_f64 = 9_007_199_254_740_992.0;
+        assert_eq!(large_i64 as f64, closest_f64);
+        if let ProjectedNumber::Next(val) = num_proj::i64_to_f64(large_i64) {
+            // Verify that the returned float is different from the direct cast
+            assert!(val > closest_f64);
+            assert!(val - closest_f64 < 2. * f64::EPSILON * closest_f64);
+        } else {
+            panic!("Expected ProjectedNumber::Next for large_i64");
+        }
+
+        // Test with very large negative value
+        let large_neg_i64 = -9_007_199_254_740_993; // -(2^53 + 1)
+        let closest_neg_f64 = -9_007_199_254_740_992.0;
+        assert_eq!(large_neg_i64 as f64, closest_neg_f64);
+        if let ProjectedNumber::Next(val) = num_proj::i64_to_f64(large_neg_i64) {
+            // Verify that the returned float is the closest representable f64
+            assert_eq!(val, closest_neg_f64);
+        } else {
+            panic!("Expected ProjectedNumber::Next for large_neg_i64");
+        }
+    }
+
+    #[test]
+    fn test_u64_to_f64() {
+        assert_eq!(num_proj::u64_to_f64(0), ProjectedNumber::Exact(0.0));
+        assert_eq!(num_proj::u64_to_f64(42), ProjectedNumber::Exact(42.0));
+
+        // Test the largest u64 value that can be exactly represented as f64 (2^53)
+        let max_exact = 9_007_199_254_740_992; // 2^53
+        assert_eq!(
+            num_proj::u64_to_f64(max_exact),
+            ProjectedNumber::Exact(max_exact as f64)
+        );
+
+        // Test values that cannot be exactly represented as f64 (integers above 2^53)
+        let large_u64 = 9_007_199_254_740_993; // 2^53 + 1
+        let closest_f64 = 9_007_199_254_740_992.0;
+        assert_eq!(large_u64 as f64, closest_f64);
+        if let ProjectedNumber::Next(val) = num_proj::u64_to_f64(large_u64) {
+            // Verify that the returned float is different from the direct cast
+            assert!(val > closest_f64);
+            assert!(val - closest_f64 < 2. * f64::EPSILON * closest_f64);
+        } else {
+            panic!("Expected ProjectedNumber::Next for large_u64");
+        }
+    }
+}
--- a/src/aggregation/bucket/filter.rs
+++ b/src/aggregation/bucket/filter.rs
@@ -6,8 +6,8 @@ use serde::{Deserialize, Deserializer, Serialize, Serializer};
 use crate::aggregation::agg_data::{
    build_segment_agg_collectors, AggRefNode, AggregationsSegmentCtx,
 };
-use crate::aggregation::cached_sub_aggs::{
-    CachedSubAggs, HighCardSubAggCache, LowCardSubAggCache, SubAggCache,
+use crate::aggregation::buffered_sub_aggs::{
+    BufferedSubAggs, HighCardSubAggBuffer, LowCardSubAggBuffer, SubAggBuffer,
 };
 use crate::aggregation::intermediate_agg_result::{
    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateBucketResult,
@@ -503,17 +503,17 @@ struct DocCount {
 }

 /// Segment collector for filter aggregation
-pub struct SegmentFilterCollector<C: SubAggCache> {
+pub struct SegmentFilterCollector<B: SubAggBuffer> {
    /// Document counts per parent bucket
    parent_buckets: Vec<DocCount>,
    /// Sub-aggregation collectors
-    sub_aggregations: Option<CachedSubAggs<C>>,
+    sub_aggregations: Option<BufferedSubAggs<B>>,
    bucket_id_provider: BucketIdProvider,
    /// Accessor index for this filter aggregation (to access FilterAggReqData)
    accessor_idx: usize,
 }

-impl<C: SubAggCache> SegmentFilterCollector<C> {
+impl<B: SubAggBuffer> SegmentFilterCollector<B> {
    /// Create a new filter segment collector following the new agg_data pattern
    pub(crate) fn from_req_and_validate(
        req: &mut AggregationsSegmentCtx,
@@ -525,7 +525,7 @@ impl<C: SubAggCache> SegmentFilterCollector<C> {
        } else {
            None
        };
-        let sub_agg_collector = sub_agg_collector.map(CachedSubAggs::new);
+        let sub_agg_collector = sub_agg_collector.map(BufferedSubAggs::new);

        Ok(SegmentFilterCollector {
            parent_buckets: Vec::new(),
@@ -547,16 +547,16 @@ pub(crate) fn build_segment_filter_collector(

    if is_top_level {
        Ok(Box::new(
-            SegmentFilterCollector::<LowCardSubAggCache>::from_req_and_validate(req, node)?,
+            SegmentFilterCollector::<LowCardSubAggBuffer>::from_req_and_validate(req, node)?,
        ))
    } else {
        Ok(Box::new(
-            SegmentFilterCollector::<HighCardSubAggCache>::from_req_and_validate(req, node)?,
+            SegmentFilterCollector::<HighCardSubAggBuffer>::from_req_and_validate(req, node)?,
        ))
    }
 }

-impl<C: SubAggCache> Debug for SegmentFilterCollector<C> {
+impl<B: SubAggBuffer> Debug for SegmentFilterCollector<B> {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("SegmentFilterCollector")
            .field("buckets", &self.parent_buckets)
@@ -566,7 +566,7 @@ impl<C: SubAggCache> Debug for SegmentFilterCollector<C> {
    }
 }

-impl<C: SubAggCache> SegmentAggregationCollector for SegmentFilterCollector<C> {
+impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentFilterCollector<B> {
    fn add_intermediate_aggregation_result(
        &mut self,
        agg_data: &AggregationsSegmentCtx,
@@ -674,6 +674,17 @@ impl<C: SubAggCache> SegmentAggregationCollector for SegmentFilterCollector<C> {
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        // TODO: forward into the inner `sub_agg` for nested order paths (`filter.metric`).
+        None
+    }
 }

 /// Intermediate result for filter aggregation
--- a/src/aggregation/bucket/histogram/date_histogram.rs
+++ b/src/aggregation/bucket/histogram/date_histogram.rs
@@ -207,7 +207,7 @@ fn parse_offset_into_milliseconds(input: &str) -> Result<i64, AggregationError>
    }
 }

-fn parse_into_milliseconds(input: &str) -> Result<i64, AggregationError> {
+pub(crate) fn parse_into_milliseconds(input: &str) -> Result<i64, AggregationError> {
    let split_boundary = input
        .as_bytes()
        .iter()
--- a/src/aggregation/bucket/histogram/histogram.rs
+++ b/src/aggregation/bucket/histogram/histogram.rs
@@ -10,7 +10,7 @@ use crate::aggregation::agg_data::{
 };
 use crate::aggregation::agg_req::Aggregations;
 use crate::aggregation::agg_result::BucketEntry;
-use crate::aggregation::cached_sub_aggs::{CachedSubAggs, HighCardCachedSubAggs};
+use crate::aggregation::buffered_sub_aggs::{BufferedSubAggs, HighCardBufferedSubAggs};
 use crate::aggregation::intermediate_agg_result::{
    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateBucketResult,
    IntermediateHistogramBucketEntry,
@@ -258,7 +258,7 @@ pub(crate) struct SegmentHistogramBucketEntry {
 impl SegmentHistogramBucketEntry {
    pub(crate) fn into_intermediate_bucket_entry(
        self,
-        sub_aggregation: &mut Option<HighCardCachedSubAggs>,
+        sub_aggregation: &mut Option<HighCardBufferedSubAggs>,
        agg_data: &AggregationsSegmentCtx,
    ) -> crate::Result<IntermediateHistogramBucketEntry> {
        let mut sub_aggregation_res = IntermediateAggregationResults::default();
@@ -283,6 +283,11 @@ impl SegmentHistogramBucketEntry {
 struct HistogramBuckets {
    pub buckets: FxHashMap<i64, SegmentHistogramBucketEntry>,
 }
+impl HistogramBuckets {
+    fn memory_consumption(&self) -> u64 {
+        self.buckets.capacity() as u64 * std::mem::size_of::<SegmentHistogramBucketEntry>() as u64
+    }
+}

 /// The collector puts values from the fast field into the correct buckets and does a conversion to
 /// the correct datatype.
@@ -291,7 +296,7 @@ pub struct SegmentHistogramCollector {
    /// The buckets containing the aggregation data.
    /// One Histogram bucket per parent bucket id.
    parent_buckets: Vec<HistogramBuckets>,
-    sub_agg: Option<HighCardCachedSubAggs>,
+    sub_agg: Option<HighCardBufferedSubAggs>,
    accessor_idx: usize,
    bucket_id_provider: BucketIdProvider,
 }
@@ -324,7 +329,7 @@ impl SegmentAggregationCollector for SegmentHistogramCollector {
        agg_data: &mut AggregationsSegmentCtx,
    ) -> crate::Result<()> {
        let req = agg_data.take_histogram_req_data(self.accessor_idx);
-        let mem_pre = self.get_memory_consumption();
+        let mem_pre = self.get_memory_consumption(parent_bucket_id);
        let buckets = &mut self.parent_buckets[parent_bucket_id as usize].buckets;

        let bounds = req.bounds;
@@ -358,12 +363,9 @@ impl SegmentAggregationCollector for SegmentHistogramCollector {
        }
        agg_data.put_back_histogram_req_data(self.accessor_idx, req);

-        let mem_delta = self.get_memory_consumption() - mem_pre;
+        let mem_delta = self.get_memory_consumption(parent_bucket_id) - mem_pre;
        if mem_delta > 0 {
-            agg_data
-                .context
-                .limits
-                .add_memory_consumed(mem_delta as u64)?;
+            agg_data.context.limits.add_memory_consumed(mem_delta)?;
        }

        if let Some(sub_agg) = &mut self.sub_agg {
@@ -392,14 +394,24 @@ impl SegmentAggregationCollector for SegmentHistogramCollector {
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        // Histogram is a multi-bucket agg with no single value to extract.
+        None
+    }
 }

 impl SegmentHistogramCollector {
-    fn get_memory_consumption(&self) -> usize {
-        let self_mem = std::mem::size_of::<Self>();
-        let buckets_mem = self.parent_buckets.len() * std::mem::size_of::<HistogramBuckets>();
-        self_mem + buckets_mem
+    fn get_memory_consumption(&self, parent_bucket_id: BucketId) -> u64 {
+        self.parent_buckets[parent_bucket_id as usize].memory_consumption()
    }
+
    /// Converts the collector result into a intermediate bucket result.
    fn add_intermediate_bucket_result(
        &mut self,
@@ -444,7 +456,7 @@ impl SegmentHistogramCollector {
            max: f64::MAX,
        });
        req_data.offset = req_data.req.offset.unwrap_or(0.0);
-        let sub_agg = sub_agg.map(CachedSubAggs::new);
+        let sub_agg = sub_agg.map(BufferedSubAggs::new);

        Ok(Self {
            parent_buckets: Default::default(),
--- a/src/aggregation/bucket/mod.rs
+++ b/src/aggregation/bucket/mod.rs
@@ -22,6 +22,7 @@
 //! - [Range](RangeAggregation)
 //! - [Terms](TermsAggregation)

+mod composite;
 mod filter;
 mod histogram;
 mod range;
@@ -31,6 +32,7 @@ mod term_missing_agg;
 use std::collections::HashMap;
 use std::fmt;

+pub use composite::*;
 pub use filter::*;
 pub use histogram::*;
 pub use range::*;
--- a/src/aggregation/bucket/range.rs
+++ b/src/aggregation/bucket/range.rs
@@ -9,8 +9,9 @@ use crate::aggregation::agg_data::{
    build_segment_agg_collectors, AggRefNode, AggregationsSegmentCtx,
 };
 use crate::aggregation::agg_limits::AggregationLimitsGuard;
-use crate::aggregation::cached_sub_aggs::{
-    CachedSubAggs, HighCardSubAggCache, LowCardCachedSubAggs, LowCardSubAggCache, SubAggCache,
+use crate::aggregation::buffered_sub_aggs::{
+    BufferedSubAggs, HighCardSubAggBuffer, LowCardBufferedSubAggs, LowCardSubAggBuffer,
+    SubAggBuffer,
 };
 use crate::aggregation::intermediate_agg_result::{
    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateBucketResult,
@@ -155,13 +156,13 @@ pub(crate) struct SegmentRangeAndBucketEntry {

 /// The collector puts values from the fast field into the correct buckets and does a conversion to
 /// the correct datatype.
-pub struct SegmentRangeCollector<C: SubAggCache> {
+pub struct SegmentRangeCollector<B: SubAggBuffer> {
    /// The buckets containing the aggregation data.
    /// One for each ParentBucketId
    parent_buckets: Vec<Vec<SegmentRangeAndBucketEntry>>,
    column_type: ColumnType,
    pub(crate) accessor_idx: usize,
-    sub_agg: Option<CachedSubAggs<C>>,
+    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.
    /// This allows a kind of flattening of the bucket ids across all parent buckets.
@@ -178,7 +179,7 @@ pub struct SegmentRangeCollector<C: SubAggCache> {
    limits: AggregationLimitsGuard,
 }

-impl<C: SubAggCache> Debug for SegmentRangeCollector<C> {
+impl<B: SubAggBuffer> Debug for SegmentRangeCollector<B> {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("SegmentRangeCollector")
            .field("parent_buckets_len", &self.parent_buckets.len())
@@ -229,7 +230,7 @@ impl SegmentRangeBucketEntry {
    }
 }

-impl<C: SubAggCache> SegmentAggregationCollector for SegmentRangeCollector<C> {
+impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
    fn add_intermediate_aggregation_result(
        &mut self,
        agg_data: &AggregationsSegmentCtx,
@@ -327,6 +328,17 @@ impl<C: SubAggCache> SegmentAggregationCollector for SegmentRangeCollector<C> {

        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        // Range is a multi-bucket agg with no single value to extract.
+        None
+    }
 }
 /// Build a concrete `SegmentRangeCollector` with either a Vec- or HashMap-backed
 /// bucket storage, depending on the column type and aggregation level.
@@ -350,8 +362,8 @@ pub(crate) fn build_segment_range_collector(
    };

    if is_low_card {
-        Ok(Box::new(SegmentRangeCollector::<LowCardSubAggCache> {
-            sub_agg: sub_agg.map(LowCardCachedSubAggs::new),
+        Ok(Box::new(SegmentRangeCollector::<LowCardSubAggBuffer> {
+            sub_agg: sub_agg.map(LowCardBufferedSubAggs::new),
            column_type: field_type,
            accessor_idx,
            parent_buckets: Vec::new(),
@@ -359,8 +371,8 @@ pub(crate) fn build_segment_range_collector(
            limits: agg_data.context.limits.clone(),
        }))
    } else {
-        Ok(Box::new(SegmentRangeCollector::<HighCardSubAggCache> {
-            sub_agg: sub_agg.map(CachedSubAggs::new),
+        Ok(Box::new(SegmentRangeCollector::<HighCardSubAggBuffer> {
+            sub_agg: sub_agg.map(BufferedSubAggs::new),
            column_type: field_type,
            accessor_idx,
            parent_buckets: Vec::new(),
@@ -370,7 +382,7 @@ pub(crate) fn build_segment_range_collector(
    }
 }

-impl<C: SubAggCache> SegmentRangeCollector<C> {
+impl<B: SubAggBuffer> SegmentRangeCollector<B> {
    pub(crate) fn create_new_buckets(
        &mut self,
        agg_data: &AggregationsSegmentCtx,
@@ -554,7 +566,7 @@ mod tests {
    pub fn get_collector_from_ranges(
        ranges: Vec<RangeAggregationRange>,
        field_type: ColumnType,
-    ) -> SegmentRangeCollector<HighCardSubAggCache> {
+    ) -> SegmentRangeCollector<HighCardSubAggBuffer> {
        let req = RangeAggregation {
            field: "dummy".to_string(),
            ranges,
--- a/src/aggregation/bucket/term_agg.rs
+++ b/src/aggregation/bucket/term_agg.rs
@@ -1,5 +1,4 @@
 use std::fmt::Debug;
-use std::io;
 use std::net::Ipv6Addr;

 use columnar::column_values::CompactSpaceU64Accessor;
@@ -17,8 +16,9 @@ use crate::aggregation::agg_data::{
 };
 use crate::aggregation::agg_limits::MemoryConsumption;
 use crate::aggregation::agg_req::Aggregations;
-use crate::aggregation::cached_sub_aggs::{
-    CachedSubAggs, HighCardSubAggCache, LowCardCachedSubAggs, LowCardSubAggCache, SubAggCache,
+use crate::aggregation::buffered_sub_aggs::{
+    BufferedSubAggs, HighCardSubAggBuffer, LowCardBufferedSubAggs, LowCardSubAggBuffer,
+    SubAggBuffer,
 };
 use crate::aggregation::intermediate_agg_result::{
    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateBucketResult,
@@ -352,19 +352,15 @@ pub(crate) fn build_segment_term_collector(
        )));
    }

-    // Validate sub aggregation exists when ordering by sub-aggregation.
-    {
-        if let OrderTarget::SubAggregation(sub_agg_name) = &terms_req_data.req.order.target {
-            let (agg_name, _agg_property) = get_agg_name_and_property(sub_agg_name);
-
-            node.get_sub_agg(agg_name, &req_data.per_request)
-                .ok_or_else(|| {
-                    TantivyError::InvalidArgument(format!(
-                        "could not find aggregation with name {agg_name} in metric \
-                         sub_aggregations"
-                    ))
-                })?;
-        }
+    // Validate that the referenced sub-aggregation exists when ordering by one.
+    if let OrderTarget::SubAggregation(sub_agg_name) = &terms_req_data.req.order.target {
+        let (agg_name, _agg_property) = get_agg_name_and_property(sub_agg_name);
+        node.get_sub_agg(agg_name, &req_data.per_request)
+            .ok_or_else(|| {
+                TantivyError::InvalidArgument(format!(
+                    "could not find aggregation with name {agg_name} in metric sub_aggregations"
+                ))
+            })?;
    }

    // Build sub-aggregation blueprint if there are children.
@@ -391,7 +387,7 @@ pub(crate) fn build_segment_term_collector(
    // 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);
-        let collector: SegmentTermCollector<_, HighCardSubAggCache> = SegmentTermCollector {
+        let collector: SegmentTermCollector<_, HighCardSubAggBuffer> = SegmentTermCollector {
            parent_buckets: vec![term_buckets],
            sub_agg: None,
            bucket_id_provider,
@@ -401,8 +397,8 @@ pub(crate) fn build_segment_term_collector(
        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);
-        let sub_agg = sub_agg_collector.map(LowCardCachedSubAggs::new);
-        let collector: SegmentTermCollector<_, LowCardSubAggCache> = SegmentTermCollector {
+        let sub_agg = sub_agg_collector.map(LowCardBufferedSubAggs::new);
+        let collector: SegmentTermCollector<_, LowCardSubAggBuffer> = SegmentTermCollector {
            parent_buckets: vec![term_buckets],
            sub_agg,
            bucket_id_provider,
@@ -414,8 +410,8 @@ pub(crate) fn build_segment_term_collector(
        let term_buckets: PagedTermMap =
            PagedTermMap::new(max_term_id + 1, &mut bucket_id_provider);
        // Build sub-aggregation blueprint (flat pairs)
-        let sub_agg = sub_agg_collector.map(CachedSubAggs::new);
-        let collector: SegmentTermCollector<PagedTermMap, HighCardSubAggCache> =
+        let sub_agg = sub_agg_collector.map(BufferedSubAggs::new);
+        let collector: SegmentTermCollector<PagedTermMap, HighCardSubAggBuffer> =
            SegmentTermCollector {
                parent_buckets: vec![term_buckets],
                sub_agg,
@@ -427,8 +423,8 @@ pub(crate) fn build_segment_term_collector(
    } else {
        let term_buckets: HashMapTermBuckets = HashMapTermBuckets::default();
        // Build sub-aggregation blueprint (flat pairs)
-        let sub_agg = sub_agg_collector.map(CachedSubAggs::new);
-        let collector: SegmentTermCollector<HashMapTermBuckets, HighCardSubAggCache> =
+        let sub_agg = sub_agg_collector.map(BufferedSubAggs::new);
+        let collector: SegmentTermCollector<HashMapTermBuckets, HighCardSubAggBuffer> =
            SegmentTermCollector {
                parent_buckets: vec![term_buckets],
                sub_agg,
@@ -758,10 +754,10 @@ impl TermAggregationMap for VecTermBuckets {
 /// The collector puts values from the fast field into the correct buckets and does a conversion to
 /// the correct datatype.
 #[derive(Debug)]
-struct SegmentTermCollector<TermMap: TermAggregationMap, C: SubAggCache> {
+struct SegmentTermCollector<TermMap: TermAggregationMap, B: SubAggBuffer> {
    /// The buckets containing the aggregation data.
    parent_buckets: Vec<TermMap>,
-    sub_agg: Option<CachedSubAggs<C>>,
+    sub_agg: Option<BufferedSubAggs<B>>,
    bucket_id_provider: BucketIdProvider,
    max_term_id: u64,
    terms_req_data: TermsAggReqData,
@@ -772,8 +768,8 @@ pub(crate) fn get_agg_name_and_property(name: &str) -> (&str, &str) {
    (agg_name, agg_property)
 }

-impl<TermMap: TermAggregationMap, C: SubAggCache> SegmentAggregationCollector
-    for SegmentTermCollector<TermMap, C>
+impl<TermMap: TermAggregationMap, B: SubAggBuffer> SegmentAggregationCollector
+    for SegmentTermCollector<TermMap, B>
 {
    fn add_intermediate_aggregation_result(
        &mut self,
@@ -790,8 +786,14 @@ impl<TermMap: TermAggregationMap, C: SubAggCache> SegmentAggregationCollector
        let term_req = &self.terms_req_data;
        let name = term_req.name.clone();

-        let bucket =
-            Self::into_intermediate_bucket_result(term_req, &mut self.sub_agg, bucket, agg_data)?;
+        let bucket = Self::into_intermediate_bucket_result(
+            term_req,
+            self.sub_agg
+                .as_mut()
+                .map(BufferedSubAggs::get_sub_agg_collector),
+            bucket,
+            agg_data,
+        )?;
        results.push(name, IntermediateAggregationResult::Bucket(bucket))?;
        Ok(())
    }
@@ -803,15 +805,17 @@ impl<TermMap: TermAggregationMap, C: SubAggCache> SegmentAggregationCollector
        docs: &[crate::DocId],
        agg_data: &mut AggregationsSegmentCtx,
    ) -> crate::Result<()> {
-        let mem_pre = self.get_memory_consumption();
+        let mem_pre = self.get_memory_consumption(parent_bucket_id);

        let req_data = &mut self.terms_req_data;

-        agg_data.column_block_accessor.fetch_block_with_missing(
-            docs,
-            &req_data.accessor,
-            req_data.missing_value_for_accessor,
-        );
+        agg_data
+            .column_block_accessor
+            .fetch_block_with_missing_unique_per_doc(
+                docs,
+                &req_data.accessor,
+                req_data.missing_value_for_accessor,
+            );

        if let Some(sub_agg) = &mut self.sub_agg {
            let term_buckets = &mut self.parent_buckets[parent_bucket_id as usize];
@@ -845,7 +849,7 @@ impl<TermMap: TermAggregationMap, C: SubAggCache> SegmentAggregationCollector
            }
        }

-        let mem_delta = self.get_memory_consumption() - mem_pre;
+        let mem_delta = self.get_memory_consumption(parent_bucket_id) - mem_pre;
        if mem_delta > 0 {
            agg_data
                .context
@@ -879,6 +883,17 @@ impl<TermMap: TermAggregationMap, C: SubAggCache> SegmentAggregationCollector
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        // Terms is a multi-bucket agg with no single value to extract.
+        None
+    }
 }

 /// Missing value are represented as a sentinel value in the column.
@@ -905,30 +920,53 @@ fn extract_missing_value<T>(
    Some((key, bucket))
 }

-impl<TermMap, C> SegmentTermCollector<TermMap, C>
+fn reborrow_opt_collector<'a>(
+    opt: &'a mut Option<&mut dyn SegmentAggregationCollector>,
+) -> Option<&'a mut dyn SegmentAggregationCollector> {
+    match opt {
+        Some(inner) => Some(*inner),
+        None => None,
+    }
+}
+
+fn into_intermediate_bucket_entry(
+    bucket: Bucket,
+    sub_agg_collector: Option<&mut dyn SegmentAggregationCollector>,
+    agg_data: &AggregationsSegmentCtx,
+) -> crate::Result<IntermediateTermBucketEntry> {
+    let mut sub_aggregation_res = IntermediateAggregationResults::default();
+    if let Some(sub_agg_collector) = sub_agg_collector {
+        sub_agg_collector.add_intermediate_aggregation_result(
+            agg_data,
+            &mut sub_aggregation_res,
+            bucket.bucket_id,
+        )?;
+    }
+    Ok(IntermediateTermBucketEntry {
+        doc_count: bucket.count,
+        sub_aggregation: sub_aggregation_res,
+    })
+}
+
+impl<TermMap, B> SegmentTermCollector<TermMap, B>
 where
    TermMap: TermAggregationMap,
-    C: SubAggCache,
+    B: SubAggBuffer,
 {
-    fn get_memory_consumption(&self) -> usize {
-        self.parent_buckets
-            .iter()
-            .map(|b| b.get_memory_consumption())
-            .sum()
+    #[inline]
+    fn get_memory_consumption(&self, parent_bucket_id: BucketId) -> usize {
+        self.parent_buckets[parent_bucket_id as usize].get_memory_consumption()
    }

    #[inline]
    pub(crate) fn into_intermediate_bucket_result(
        term_req: &TermsAggReqData,
-        sub_agg: &mut Option<CachedSubAggs<C>>,
+        mut sub_agg_collector: Option<&mut dyn SegmentAggregationCollector>,
        term_buckets: TermMap,
        agg_data: &AggregationsSegmentCtx,
    ) -> crate::Result<IntermediateBucketResult> {
        let mut entries: Vec<(u64, Bucket)> = term_buckets.into_vec();

-        let order_by_sub_aggregation =
-            matches!(term_req.req.order.target, OrderTarget::SubAggregation(_));
-
        match &term_req.req.order.target {
            OrderTarget::Key => {
                // We rely on the fact, that term ordinals match the order of the strings
@@ -940,10 +978,37 @@ where
                    entries.sort_unstable_by_key(|bucket| bucket.0);
                }
            }
-            OrderTarget::SubAggregation(_name) => {
-                // don't sort and cut off since it's hard to make assumptions on the quality of the
-                // results when cutting off du to unknown nature of the sub_aggregation (possible
-                // to check).
+            OrderTarget::SubAggregation(sub_agg_path) => {
+                // Peek segment-level metric values, sort, then fall through to
+                // `cut_off_buckets`. Like Elasticsearch, we always cut off when ordering
+                // by a sub-agg: top-K results are approximate and may differ from the
+                // global ordering, especially for non-monotonic metrics like avg/min.
+                let coll = sub_agg_collector.as_deref().ok_or_else(|| {
+                    TantivyError::InvalidArgument(format!(
+                        "Could not find sub-aggregation collector for path {sub_agg_path}"
+                    ))
+                })?;
+                let (agg_name, agg_prop) = get_agg_name_and_property(sub_agg_path);
+                // Fetch values up-front; otherwise sort would re-compute per comparison
+                let mut keyed: Vec<(f64, (u64, Bucket))> = entries
+                    .into_iter()
+                    .map(|bucket| {
+                        let metric_value = coll
+                            .compute_metric_value(bucket.1.bucket_id, agg_name, agg_prop, agg_data)
+                            .unwrap_or(0.0);
+                        (metric_value, bucket)
+                    })
+                    .collect();
+                if term_req.req.order.order == Order::Desc {
+                    keyed.sort_unstable_by(|a, b| {
+                        b.0.partial_cmp(&a.0).unwrap_or(std::cmp::Ordering::Equal)
+                    });
+                } else {
+                    keyed.sort_unstable_by(|a, b| {
+                        a.0.partial_cmp(&b.0).unwrap_or(std::cmp::Ordering::Equal)
+                    });
+                }
+                entries = keyed.into_iter().map(|(_, e)| e).collect();
            }
            OrderTarget::Count => {
                if term_req.req.order.order == Order::Desc {
@@ -954,40 +1019,12 @@ where
            }
        }

-        let (term_doc_count_before_cutoff, sum_other_doc_count) = if order_by_sub_aggregation {
-            (0, 0)
-        } else {
-            cut_off_buckets(&mut entries, term_req.req.segment_size as usize)
-        };
+        let (term_doc_count_before_cutoff, sum_other_doc_count) =
+            cut_off_buckets(&mut entries, term_req.req.segment_size as usize);

        let mut dict: FxHashMap<IntermediateKey, IntermediateTermBucketEntry> = Default::default();
        dict.reserve(entries.len());

-        let into_intermediate_bucket_entry =
-            |bucket: Bucket,
-             sub_agg: &mut Option<CachedSubAggs<C>>|
-             -> crate::Result<IntermediateTermBucketEntry> {
-                if let Some(sub_agg) = sub_agg {
-                    let mut sub_aggregation_res = IntermediateAggregationResults::default();
-                    sub_agg
-                        .get_sub_agg_collector()
-                        .add_intermediate_aggregation_result(
-                            agg_data,
-                            &mut sub_aggregation_res,
-                            bucket.bucket_id,
-                        )?;
-                    Ok(IntermediateTermBucketEntry {
-                        doc_count: bucket.count,
-                        sub_aggregation: sub_aggregation_res,
-                    })
-                } else {
-                    Ok(IntermediateTermBucketEntry {
-                        doc_count: bucket.count,
-                        sub_aggregation: Default::default(),
-                    })
-                }
-            };
-
        if term_req.column_type == ColumnType::Str {
            let fallback_dict = Dictionary::empty();
            let term_dict = term_req
@@ -998,7 +1035,11 @@ where

            if let Some((intermediate_key, bucket)) = extract_missing_value(&mut entries, term_req)
            {
-                let intermediate_entry = into_intermediate_bucket_entry(bucket, sub_agg)?;
+                let intermediate_entry = into_intermediate_bucket_entry(
+                    bucket,
+                    reborrow_opt_collector(&mut sub_agg_collector),
+                    agg_data,
+                )?;
                dict.insert(intermediate_key, intermediate_entry);
            }

@@ -1006,19 +1047,28 @@ where
            entries.sort_unstable_by_key(|bucket| bucket.0);

            let (term_ids, buckets): (Vec<u64>, Vec<Bucket>) = entries.into_iter().unzip();
-            let mut buckets_it = buckets.into_iter();

-            term_dict.sorted_ords_to_term_cb(term_ids.into_iter(), |term| {
-                let bucket = buckets_it.next().unwrap();
-                let intermediate_entry =
-                    into_intermediate_bucket_entry(bucket, sub_agg).map_err(io::Error::other)?;
+            let intermediate_entries: Vec<IntermediateTermBucketEntry> = buckets
+                .into_iter()
+                .map(|bucket| {
+                    into_intermediate_bucket_entry(
+                        bucket,
+                        reborrow_opt_collector(&mut sub_agg_collector),
+                        agg_data,
+                    )
+                })
+                .collect::<crate::Result<_>>()?;
+
+            let mut intermediate_entry_it = intermediate_entries.into_iter();
+
+            term_dict.sorted_ords_to_term_cb(&term_ids[..], |term| {
+                let intermediate_entry = intermediate_entry_it.next().unwrap();
                dict.insert(
                    IntermediateKey::Str(
                        String::from_utf8(term.to_vec()).expect("could not convert to String"),
                    ),
                    intermediate_entry,
                );
-                Ok(())
            })?;

            if term_req.req.min_doc_count == 0 {
@@ -1053,14 +1103,22 @@ where
            }
        } else if term_req.column_type == ColumnType::DateTime {
            for (val, doc_count) in entries {
-                let intermediate_entry = into_intermediate_bucket_entry(doc_count, sub_agg)?;
+                let intermediate_entry = into_intermediate_bucket_entry(
+                    doc_count,
+                    reborrow_opt_collector(&mut sub_agg_collector),
+                    agg_data,
+                )?;
                let val = i64::from_u64(val);
                let date = format_date(val)?;
                dict.insert(IntermediateKey::Str(date), intermediate_entry);
            }
        } else if term_req.column_type == ColumnType::Bool {
            for (val, doc_count) in entries {
-                let intermediate_entry = into_intermediate_bucket_entry(doc_count, sub_agg)?;
+                let intermediate_entry = into_intermediate_bucket_entry(
+                    doc_count,
+                    reborrow_opt_collector(&mut sub_agg_collector),
+                    agg_data,
+                )?;
                let val = bool::from_u64(val);
                dict.insert(IntermediateKey::Bool(val), intermediate_entry);
            }
@@ -1080,14 +1138,22 @@ where
                })?;

            for (val, doc_count) in entries {
-                let intermediate_entry = into_intermediate_bucket_entry(doc_count, sub_agg)?;
+                let intermediate_entry = into_intermediate_bucket_entry(
+                    doc_count,
+                    reborrow_opt_collector(&mut sub_agg_collector),
+                    agg_data,
+                )?;
                let val: u128 = compact_space_accessor.compact_to_u128(val as u32);
                let val = Ipv6Addr::from_u128(val);
                dict.insert(IntermediateKey::IpAddr(val), intermediate_entry);
            }
        } else {
            for (val, doc_count) in entries {
-                let intermediate_entry = into_intermediate_bucket_entry(doc_count, sub_agg)?;
+                let intermediate_entry = into_intermediate_bucket_entry(
+                    doc_count,
+                    reborrow_opt_collector(&mut sub_agg_collector),
+                    agg_data,
+                )?;
                if term_req.column_type == ColumnType::U64 {
                    dict.insert(IntermediateKey::U64(val), intermediate_entry);
                } else if term_req.column_type == ColumnType::I64 {
@@ -1121,13 +1187,13 @@ where
    }
 }

-impl<TermMap: TermAggregationMap, C: SubAggCache> SegmentTermCollector<TermMap, C> {
+impl<TermMap: TermAggregationMap, B: SubAggBuffer> SegmentTermCollector<TermMap, B> {
    #[inline]
    fn collect_terms_with_docs(
        iter: impl Iterator<Item = (crate::DocId, u64)>,
        term_buckets: &mut TermMap,
        bucket_id_provider: &mut BucketIdProvider,
-        sub_agg: &mut CachedSubAggs<C>,
+        sub_agg: &mut BufferedSubAggs<B>,
    ) {
        for (doc, term_id) in iter {
            let bucket_id = term_buckets.term_entry(term_id, bucket_id_provider);
@@ -1200,7 +1266,7 @@ mod tests {
    use crate::aggregation::{AggregationLimitsGuard, DistributedAggregationCollector};
    use crate::indexer::NoMergePolicy;
    use crate::query::AllQuery;
-    use crate::schema::{IntoIpv6Addr, Schema, FAST, STRING};
+    use crate::schema::{IntoIpv6Addr, Schema, FAST, INDEXED, STRING, TEXT};
    use crate::{Index, IndexWriter};

    #[test]
@@ -1729,6 +1795,263 @@ mod tests {
        Ok(())
    }

+    #[test]
+    fn terms_aggregation_order_by_cardinality_desc_single_segment() -> crate::Result<()> {
+        terms_aggregation_order_by_cardinality_desc(true)
+    }
+    #[test]
+    fn terms_aggregation_order_by_cardinality_desc_multi_segment() -> crate::Result<()> {
+        terms_aggregation_order_by_cardinality_desc(false)
+    }
+    fn terms_aggregation_order_by_cardinality_desc(merge_segments: bool) -> crate::Result<()> {
+        // Distinct score values per bucket key: A→5, B→1, C→3.
+        // Order by cardinality desc must yield A, C, B.
+        let segment_and_terms = vec![vec![
+            (1.0, "A".to_string()),
+            (2.0, "A".to_string()),
+            (3.0, "A".to_string()),
+            (4.0, "A".to_string()),
+            (5.0, "A".to_string()),
+            (1.0, "B".to_string()),
+            (1.0, "B".to_string()),
+            (1.0, "B".to_string()),
+            (1.0, "C".to_string()),
+            (2.0, "C".to_string()),
+            (3.0, "C".to_string()),
+        ]];
+        let index = get_test_index_from_values_and_terms(merge_segments, &segment_and_terms)?;
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "order": { "card": "desc" }
+                },
+                "aggs": {
+                    "card": { "cardinality": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "A");
+        assert_eq!(res["my_texts"]["buckets"][0]["card"]["value"], 5.0);
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"][1]["card"]["value"], 3.0);
+        assert_eq!(res["my_texts"]["buckets"][2]["key"], "B");
+        assert_eq!(res["my_texts"]["buckets"][2]["card"]["value"], 1.0);
+
+        // Asc engages the segment-cutoff path too (monotonic-safe: discarded buckets had
+        // local card >= cutoff, so merged card >= cutoff and they cannot be globally smallest).
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "order": { "card": "asc" }
+                },
+                "aggs": {
+                    "card": { "cardinality": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "B");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"][2]["key"], "A");
+
+        // size=2 with desc engages the segment cutoff: must keep top-2 by cardinality (A, C),
+        // and `sum_other_doc_count` reflects the dropped B (3 docs).
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "size": 2,
+                    "order": { "card": "desc" }
+                },
+                "aggs": {
+                    "card": { "cardinality": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "A");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"].as_array().unwrap().len(), 2);
+
+        // size=2 with asc engages the segment cutoff: must keep bottom-2 by cardinality (B, C).
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "size": 2,
+                    "order": { "card": "asc" }
+                },
+                "aggs": {
+                    "card": { "cardinality": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "B");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"].as_array().unwrap().len(), 2);
+
+        Ok(())
+    }
+
+    #[test]
+    fn terms_aggregation_order_by_sum_single_segment() -> crate::Result<()> {
+        terms_aggregation_order_by_sum(true)
+    }
+    #[test]
+    fn terms_aggregation_order_by_sum_multi_segment() -> crate::Result<()> {
+        terms_aggregation_order_by_sum(false)
+    }
+    fn terms_aggregation_order_by_sum(merge_segments: bool) -> crate::Result<()> {
+        // Per-bucket sums on the U64 `score` column (non-negative => sum is monotonic):
+        //   A → 1+2+3+4+5 = 15, B → 1+1+1 = 3, C → 1+2+3 = 6.
+        let segment_and_terms = vec![
+            vec![
+                (1.0, "A".to_string()),
+                (2.0, "A".to_string()),
+                (3.0, "A".to_string()),
+                (1.0, "B".to_string()),
+                (1.0, "C".to_string()),
+            ],
+            vec![
+                (4.0, "A".to_string()),
+                (5.0, "A".to_string()),
+                (1.0, "B".to_string()),
+                (1.0, "B".to_string()),
+                (2.0, "C".to_string()),
+                (3.0, "C".to_string()),
+            ],
+        ];
+        let index = get_test_index_from_values_and_terms(merge_segments, &segment_and_terms)?;
+
+        // Desc on a Sum metric engages the fast path (column is U64).
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "order": { "total": "desc" }
+                },
+                "aggs": {
+                    "total": { "sum": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "A");
+        assert_eq!(res["my_texts"]["buckets"][0]["total"]["value"], 15.0);
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"][1]["total"]["value"], 6.0);
+        assert_eq!(res["my_texts"]["buckets"][2]["key"], "B");
+        assert_eq!(res["my_texts"]["buckets"][2]["total"]["value"], 3.0);
+
+        // Asc engages the fast path too — discarded buckets had local sum >= cutoff,
+        // and merged sum >= local (non-negative addends), so they cannot be globally smallest.
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "order": { "total": "asc" }
+                },
+                "aggs": {
+                    "total": { "sum": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "B");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"][2]["key"], "A");
+
+        // size=2 desc with cutoff: top-2 by sum (A, C).
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "size": 2,
+                    "order": { "total": "desc" }
+                },
+                "aggs": {
+                    "total": { "sum": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "A");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"].as_array().unwrap().len(), 2);
+
+        // Stats sub-property: ordering by `mystats.sum` on a U64 column also engages.
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "order": { "mystats.sum": "desc" }
+                },
+                "aggs": {
+                    "mystats": { "stats": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "A");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"][2]["key"], "B");
+
+        // Sum on a signed column (I64) takes the same cutoff path. Results may be
+        // approximate near the boundary on adversarial data, but for this dataset the
+        // top-K is unambiguous.
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "order": { "total": "desc" }
+                },
+                "aggs": {
+                    "total": { "sum": { "field": "score_i64" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "A");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"][2]["key"], "B");
+
+        // Order by extended_stats sub-property exercises compute_metric_value on the
+        // ExtendedStats collector. A→max=5, B→max=1, C→max=3, so desc by max → A, C, B.
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "my_texts": {
+                "terms": {
+                    "field": "string_id",
+                    "order": { "ext.max": "desc" }
+                },
+                "aggs": {
+                    "ext": { "extended_stats": { "field": "score" } }
+                }
+            }
+        }))
+        .unwrap();
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "A");
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "C");
+        assert_eq!(res["my_texts"]["buckets"][2]["key"], "B");
+
+        Ok(())
+    }
+
    #[test]
    fn terms_aggregation_test_order_key_single_segment() -> crate::Result<()> {
        terms_aggregation_test_order_key_merge_segment(true)
@@ -2347,7 +2670,7 @@ mod tests {

        // text field
        assert_eq!(res["my_texts"]["buckets"][0]["key"], "Hello Hello");
-        assert_eq!(res["my_texts"]["buckets"][0]["doc_count"], 5);
+        assert_eq!(res["my_texts"]["buckets"][0]["doc_count"], 4);
        assert_eq!(res["my_texts"]["buckets"][1]["key"], "Empty");
        assert_eq!(res["my_texts"]["buckets"][1]["doc_count"], 2);
        assert_eq!(
@@ -2356,7 +2679,7 @@ mod tests {
        );
        // text field with number as missing fallback
        assert_eq!(res["my_texts2"]["buckets"][0]["key"], "Hello Hello");
-        assert_eq!(res["my_texts2"]["buckets"][0]["doc_count"], 5);
+        assert_eq!(res["my_texts2"]["buckets"][0]["doc_count"], 4);
        assert_eq!(res["my_texts2"]["buckets"][1]["key"], 1337.0);
        assert_eq!(res["my_texts2"]["buckets"][1]["doc_count"], 2);
        assert_eq!(
@@ -2370,7 +2693,7 @@ mod tests {
        assert_eq!(res["my_ids"]["buckets"][0]["key"], 1337.0);
        assert_eq!(res["my_ids"]["buckets"][0]["doc_count"], 4);
        assert_eq!(res["my_ids"]["buckets"][1]["key"], 1.0);
-        assert_eq!(res["my_ids"]["buckets"][1]["doc_count"], 3);
+        assert_eq!(res["my_ids"]["buckets"][1]["doc_count"], 2);
        assert_eq!(res["my_ids"]["buckets"][2]["key"], serde_json::Value::Null);

        Ok(())
@@ -2894,4 +3217,101 @@ mod tests {

        Ok(())
    }
+
+    fn prep_index_with_n_unique_terms_plus_one_null(n: u64) -> crate::Result<Index> {
+        let mut schema_builder = Schema::builder();
+        let id_field = schema_builder.add_u64_field("id", INDEXED);
+        let title_field = schema_builder.add_text_field("title", TEXT | FAST);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema.clone());
+        // set to one thread to guarantee all docs end up in the same segment
+        let mut writer = index.writer_with_num_threads(1, 50_000_000)?;
+
+        writer.add_document(doc!(
+            id_field => 0u64,
+        ))?;
+        for i in 1u64..=n {
+            let title = format!("foo{i}");
+            writer.add_document(doc!(
+                id_field => i,
+                title_field => title,
+            ))?;
+        }
+
+        writer.commit()?;
+
+        Ok(index)
+    }
+
+    #[test]
+    fn null_bitset_bounds_check_regression() -> crate::Result<()> {
+        // include cases
+        for i in 0..=4 {
+            let index = prep_index_with_n_unique_terms_plus_one_null(i * 64)?;
+            let normal_req: Aggregations = serde_json::from_value(json!({
+                "my_bool": {
+                    "terms": {
+                        "field": "title",
+                        "missing": "__NULL__",
+                        "size": 1000,
+                    }
+                }
+            }))?;
+            let include_req: Aggregations = serde_json::from_value(json!({
+                "my_bool": {
+                    "terms": {
+                        "field": "title",
+                        "include": "foo(.*)",
+                        "missing": "__NULL__",
+                        "size": 1000,
+                    }
+                }
+            }))?;
+            let exclude_req: Aggregations = serde_json::from_value(json!({
+                "my_bool": {
+                    "terms": {
+                        "field": "title",
+                        "exclude": "foo(.*)",
+                        "missing": "__NULL__",
+                        "size": 1000,
+                    }
+                }
+            }))?;
+
+            let normal_res = exec_request(normal_req, &index)?;
+            let normal_buckets = normal_res["my_bool"]["buckets"].as_array().unwrap();
+            assert_eq!(
+                normal_buckets.len(),
+                (i * 64) as usize + 1,
+                "The normal request should return all 'foo' buckets, plus the missing term bucket",
+            );
+
+            let include_res = exec_request(include_req, &index)?;
+            eprintln!("include_res: {include_res:?}");
+            let include_buckets = include_res["my_bool"]["buckets"].as_array().unwrap();
+            assert_eq!(
+                include_buckets.len(),
+                (i * 64) as usize,
+                "The include request should return all 'foo' buckets, and not the missing term \
+                 bucket",
+            );
+            assert!(include_buckets
+                .iter()
+                .all(|b| b["key"].as_str().unwrap().starts_with("foo")));
+
+            let exclude_res = exec_request(exclude_req, &index)?;
+            let exclude_buckets = exclude_res["my_bool"]["buckets"].as_array().unwrap();
+            if i != 0 {
+                // TODO: Remove this if after fixing exclude + missing bug
+                assert_eq!(
+                    exclude_buckets.len(),
+                    1,
+                    "The exclude request should exclude all 'foo' buckets, and only the missing \
+                     term bucket",
+                );
+                assert_eq!(exclude_buckets[0]["key"], "__NULL__");
+            }
+        }
+        Ok(())
+    }
 }
--- a/src/aggregation/bucket/term_missing_agg.rs
+++ b/src/aggregation/bucket/term_missing_agg.rs
@@ -5,7 +5,7 @@ use crate::aggregation::agg_data::{
    build_segment_agg_collectors, AggRefNode, AggregationsSegmentCtx,
 };
 use crate::aggregation::bucket::term_agg::TermsAggregation;
-use crate::aggregation::cached_sub_aggs::{CachedSubAggs, HighCardCachedSubAggs};
+use crate::aggregation::buffered_sub_aggs::{BufferedSubAggs, HighCardBufferedSubAggs};
 use crate::aggregation::intermediate_agg_result::{
    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateBucketResult,
    IntermediateKey, IntermediateTermBucketEntry, IntermediateTermBucketResult,
@@ -47,7 +47,7 @@ struct MissingCount {
 #[derive(Default, Debug)]
 pub struct TermMissingAgg {
    accessor_idx: usize,
-    sub_agg: Option<HighCardCachedSubAggs>,
+    sub_agg: Option<HighCardBufferedSubAggs>,
    /// Idx = parent bucket id, Value = missing count for that bucket
    missing_count_per_bucket: Vec<MissingCount>,
    bucket_id_provider: BucketIdProvider,
@@ -66,7 +66,7 @@ impl TermMissingAgg {
            None
        };

-        let sub_agg = sub_agg.map(CachedSubAggs::new);
+        let sub_agg = sub_agg.map(BufferedSubAggs::new);
        let bucket_id_provider = BucketIdProvider::default();

        Ok(Self {
@@ -177,6 +177,17 @@ impl SegmentAggregationCollector for TermMissingAgg {
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        // TODO: forward to `sub_agg` for nested order paths (`missing_agg>metric`).
+        None
+    }
 }

 #[cfg(test)]
--- a/src/aggregation/buffered_sub_aggs.rs
+++ b/src/aggregation/buffered_sub_aggs.rs
@@ -6,7 +6,7 @@ use crate::aggregation::bucket::MAX_NUM_TERMS_FOR_VEC;
 use crate::aggregation::BucketId;
 use crate::DocId;

-/// A cache for sub-aggregations, storing doc ids per bucket id.
+/// A buffer for sub-aggregations, storing doc ids per bucket id.
 /// Depending on the cardinality of the parent aggregation, we use different
 /// storage strategies.
 ///
@@ -24,21 +24,21 @@ use crate::DocId;
 /// aggregations.
 /// What this datastructure does in general is to group docs by bucket id.
 #[derive(Debug)]
-pub(crate) struct CachedSubAggs<C: SubAggCache> {
-    cache: C,
+pub(crate) struct BufferedSubAggs<B: SubAggBuffer> {
+    buffer: B,
    sub_agg_collector: Box<dyn SegmentAggregationCollector>,
    num_docs: usize,
 }

-pub type LowCardCachedSubAggs = CachedSubAggs<LowCardSubAggCache>;
-pub type HighCardCachedSubAggs = CachedSubAggs<HighCardSubAggCache>;
+pub type LowCardBufferedSubAggs = BufferedSubAggs<LowCardSubAggBuffer>;
+pub type HighCardBufferedSubAggs = BufferedSubAggs<HighCardSubAggBuffer>;

 const FLUSH_THRESHOLD: usize = 2048;

-/// A trait for caching sub-aggregation doc ids per bucket id.
+/// A trait for buffering sub-aggregation doc ids per bucket id.
 /// Different implementations can be used depending on the cardinality
 /// of the parent aggregation.
-pub trait SubAggCache: Debug {
+pub trait SubAggBuffer: Debug {
    fn new() -> Self;
    fn push(&mut self, bucket_id: BucketId, doc_id: DocId);
    fn flush_local(
@@ -49,22 +49,22 @@ pub trait SubAggCache: Debug {
    ) -> crate::Result<()>;
 }

-impl<Backend: SubAggCache + Debug> CachedSubAggs<Backend> {
+impl<Backend: SubAggBuffer + Debug> BufferedSubAggs<Backend> {
    pub fn new(sub_agg: Box<dyn SegmentAggregationCollector>) -> Self {
        Self {
-            cache: Backend::new(),
+            buffer: Backend::new(),
            sub_agg_collector: sub_agg,
            num_docs: 0,
        }
    }

-    pub fn get_sub_agg_collector(&mut self) -> &mut Box<dyn SegmentAggregationCollector> {
-        &mut self.sub_agg_collector
+    pub fn get_sub_agg_collector(&mut self) -> &mut dyn SegmentAggregationCollector {
+        &mut *self.sub_agg_collector
    }

    #[inline]
    pub fn push(&mut self, bucket_id: BucketId, doc_id: DocId) {
-        self.cache.push(bucket_id, doc_id);
+        self.buffer.push(bucket_id, doc_id);
        self.num_docs += 1;
    }

@@ -75,7 +75,7 @@ impl<Backend: SubAggCache + Debug> CachedSubAggs<Backend> {
        agg_data: &mut AggregationsSegmentCtx,
    ) -> crate::Result<()> {
        if self.num_docs >= FLUSH_THRESHOLD {
-            self.cache
+            self.buffer
                .flush_local(&mut self.sub_agg_collector, agg_data, false)?;
            self.num_docs = 0;
        }
@@ -85,7 +85,7 @@ impl<Backend: SubAggCache + Debug> CachedSubAggs<Backend> {
    /// Note: this _does_ flush the sub aggregations.
    pub fn flush(&mut self, agg_data: &mut AggregationsSegmentCtx) -> crate::Result<()> {
        if self.num_docs != 0 {
-            self.cache
+            self.buffer
                .flush_local(&mut self.sub_agg_collector, agg_data, true)?;
            self.num_docs = 0;
        }
@@ -94,11 +94,11 @@ impl<Backend: SubAggCache + Debug> CachedSubAggs<Backend> {
    }
 }

-/// Number of partitions for high cardinality sub-aggregation cache.
+/// Number of partitions for high cardinality sub-aggregation buffer.
 const NUM_PARTITIONS: usize = 16;

 #[derive(Debug)]
-pub(crate) struct HighCardSubAggCache {
+pub(crate) struct HighCardSubAggBuffer {
    /// This weird partitioning is used to do some cheap grouping on the bucket ids.
    /// bucket ids are dense, e.g. when we don't detect the cardinality as low cardinality,
    /// but there are just 16 bucket ids, each bucket id will go to its own partition.
@@ -108,7 +108,7 @@ pub(crate) struct HighCardSubAggCache {
    partitions: Box<[PartitionEntry; NUM_PARTITIONS]>,
 }

-impl HighCardSubAggCache {
+impl HighCardSubAggBuffer {
    #[inline]
    fn clear(&mut self) {
        for partition in self.partitions.iter_mut() {
@@ -131,7 +131,7 @@ impl PartitionEntry {
    }
 }

-impl SubAggCache for HighCardSubAggCache {
+impl SubAggBuffer for HighCardSubAggBuffer {
    fn new() -> Self {
        Self {
            partitions: Box::new(core::array::from_fn(|_| PartitionEntry::default())),
@@ -173,14 +173,14 @@ impl SubAggCache for HighCardSubAggCache {
 }

 #[derive(Debug)]
-pub(crate) struct LowCardSubAggCache {
-    /// Cache doc ids per bucket for sub-aggregations.
+pub(crate) struct LowCardSubAggBuffer {
+    /// Buffer doc ids per bucket for sub-aggregations.
    ///
    /// The outer Vec is indexed by BucketId.
    per_bucket_docs: Vec<Vec<DocId>>,
 }

-impl LowCardSubAggCache {
+impl LowCardSubAggBuffer {
    #[inline]
    fn clear(&mut self) {
        for v in &mut self.per_bucket_docs {
@@ -189,7 +189,7 @@ impl LowCardSubAggCache {
    }
 }

-impl SubAggCache for LowCardSubAggCache {
+impl SubAggBuffer for LowCardSubAggBuffer {
    fn new() -> Self {
        Self {
            per_bucket_docs: Vec::new(),
--- a/src/aggregation/collector.rs
+++ b/src/aggregation/collector.rs
@@ -1,6 +1,6 @@
 use super::agg_req::Aggregations;
 use super::agg_result::AggregationResults;
-use super::cached_sub_aggs::LowCardCachedSubAggs;
+use super::buffered_sub_aggs::LowCardBufferedSubAggs;
 use super::intermediate_agg_result::IntermediateAggregationResults;
 use super::AggContextParams;
 // group buffering strategy is chosen explicitly by callers; no need to hash-group on the fly.
@@ -136,7 +136,7 @@ fn merge_fruits(
 /// `AggregationSegmentCollector` does the aggregation collection on a segment.
 pub struct AggregationSegmentCollector {
    aggs_with_accessor: AggregationsSegmentCtx,
-    agg_collector: LowCardCachedSubAggs,
+    agg_collector: LowCardBufferedSubAggs,
    error: Option<TantivyError>,
 }

@@ -152,7 +152,7 @@ impl AggregationSegmentCollector {
        let mut agg_data =
            build_aggregations_data_from_req(agg, reader, segment_ordinal, context.clone())?;
        let mut result =
-            LowCardCachedSubAggs::new(build_segment_agg_collectors_root(&mut agg_data)?);
+            LowCardBufferedSubAggs::new(build_segment_agg_collectors_root(&mut agg_data)?);
        result
            .get_sub_agg_collector()
            .prepare_max_bucket(0, &agg_data)?; // prepare for bucket zero
--- a/src/aggregation/intermediate_agg_result.rs
+++ b/src/aggregation/intermediate_agg_result.rs
@@ -15,8 +15,9 @@ use serde::{Deserialize, Serialize};
 use super::agg_req::{Aggregation, AggregationVariants, Aggregations};
 use super::agg_result::{AggregationResult, BucketResult, MetricResult, RangeBucketEntry};
 use super::bucket::{
-    cut_off_buckets, get_agg_name_and_property, intermediate_histogram_buckets_to_final_buckets,
-    GetDocCount, Order, OrderTarget, RangeAggregation, TermsAggregation,
+    composite_intermediate_key_ordering, cut_off_buckets, get_agg_name_and_property,
+    intermediate_histogram_buckets_to_final_buckets, CompositeAggregation, GetDocCount,
+    MissingOrder, Order, OrderTarget, RangeAggregation, TermsAggregation,
 };
 use super::metric::{
    IntermediateAverage, IntermediateCount, IntermediateExtendedStats, IntermediateMax,
@@ -25,7 +26,7 @@ use super::metric::{
 use super::segment_agg_result::AggregationLimitsGuard;
 use super::{format_date, AggregationError, Key, SerializedKey};
 use crate::aggregation::agg_result::{
-    AggregationResults, BucketEntries, BucketEntry, FilterBucketResult,
+    AggregationResults, BucketEntries, BucketEntry, CompositeBucketEntry, FilterBucketResult,
 };
 use crate::aggregation::bucket::TermsAggregationInternal;
 use crate::aggregation::metric::CardinalityCollector;
@@ -280,6 +281,11 @@ pub(crate) fn empty_from_req(req: &Aggregation) -> IntermediateAggregationResult
            doc_count: 0,
            sub_aggregations: IntermediateAggregationResults::default(),
        }),
+        Composite(_) => {
+            IntermediateAggregationResult::Bucket(IntermediateBucketResult::Composite {
+                buckets: IntermediateCompositeBucketResult::default(),
+            })
+        }
    }
 }

@@ -473,6 +479,11 @@ pub enum IntermediateBucketResult {
        /// Sub-aggregation results
        sub_aggregations: IntermediateAggregationResults,
    },
+    /// Composite aggregation
+    Composite {
+        /// The composite buckets
+        buckets: IntermediateCompositeBucketResult,
+    },
 }

 impl IntermediateBucketResult {
@@ -568,6 +579,13 @@ impl IntermediateBucketResult {
                    sub_aggregations: final_sub_aggregations,
                }))
            }
+            IntermediateBucketResult::Composite { buckets } => {
+                let composite_req = req
+                    .agg
+                    .as_composite()
+                    .expect("unexpected aggregation, expected composite aggregation");
+                buckets.into_final_result(composite_req, req.sub_aggregation(), limits)
+            }
        }
    }

@@ -634,6 +652,16 @@ impl IntermediateBucketResult {
                *doc_count_left += doc_count_right;
                sub_aggs_left.merge_fruits(sub_aggs_right)?;
            }
+            (
+                IntermediateBucketResult::Composite {
+                    buckets: composite_left,
+                },
+                IntermediateBucketResult::Composite {
+                    buckets: composite_right,
+                },
+            ) => {
+                composite_left.merge_fruits(composite_right)?;
+            }
            (IntermediateBucketResult::Range(_), _) => {
                panic!("try merge on different types")
            }
@@ -646,6 +674,9 @@ impl IntermediateBucketResult {
            (IntermediateBucketResult::Filter { .. }, _) => {
                panic!("try merge on different types")
            }
+            (IntermediateBucketResult::Composite { .. }, _) => {
+                panic!("try merge on different types")
+            }
        }
        Ok(())
    }
@@ -914,6 +945,172 @@ impl MergeFruits for IntermediateHistogramBucketEntry {
    }
 }

+/// Entry for the composite bucket.
+pub type IntermediateCompositeBucketEntry = IntermediateTermBucketEntry;
+
+/// The fully typed key for composite aggregation
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub enum CompositeIntermediateKey {
+    /// Bool key
+    Bool(bool),
+    /// String key
+    Str(String),
+    /// Float key
+    F64(f64),
+    /// Signed integer key
+    I64(i64),
+    /// Unsigned integer key
+    U64(u64),
+    /// DateTime key, nanoseconds since epoch
+    DateTime(i64),
+    /// IP Address key
+    IpAddr(Ipv6Addr),
+    /// Missing value key
+    Null,
+}
+
+impl Eq for CompositeIntermediateKey {}
+
+impl std::hash::Hash for CompositeIntermediateKey {
+    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
+        core::mem::discriminant(self).hash(state);
+        match self {
+            CompositeIntermediateKey::Bool(val) => val.hash(state),
+            CompositeIntermediateKey::Str(text) => text.hash(state),
+            CompositeIntermediateKey::F64(val) => val.to_bits().hash(state),
+            CompositeIntermediateKey::U64(val) => val.hash(state),
+            CompositeIntermediateKey::I64(val) => val.hash(state),
+            CompositeIntermediateKey::DateTime(val) => val.hash(state),
+            CompositeIntermediateKey::IpAddr(val) => val.hash(state),
+            CompositeIntermediateKey::Null => {}
+        }
+    }
+}
+
+/// Composite aggregation page.
+#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct IntermediateCompositeBucketResult {
+    pub(crate) entries: FxHashMap<Vec<CompositeIntermediateKey>, IntermediateCompositeBucketEntry>,
+    pub(crate) target_size: u32,
+    pub(crate) orders: Vec<(Order, MissingOrder)>,
+}
+
+impl IntermediateCompositeBucketResult {
+    pub(crate) fn into_final_result(
+        self,
+        req: &CompositeAggregation,
+        sub_aggregation_req: &Aggregations,
+        limits: &mut AggregationLimitsGuard,
+    ) -> crate::Result<BucketResult> {
+        let trimmed_entry_vec =
+            trim_composite_buckets(self.entries, &self.orders, self.target_size)?;
+        let after_key = trimmed_entry_vec
+            .last()
+            .map(|bucket| {
+                let (intermediate_key, _entry) = bucket;
+                intermediate_key
+                    .iter()
+                    .enumerate()
+                    .map(|(idx, intermediate_key)| {
+                        let source = &req.sources[idx];
+                        (source.name().to_string(), intermediate_key.clone().into())
+                    })
+                    .collect()
+            })
+            .unwrap_or_default();
+
+        let buckets = trimmed_entry_vec
+            .into_iter()
+            .map(|(intermediate_key, entry)| {
+                let key = intermediate_key
+                    .into_iter()
+                    .enumerate()
+                    .map(|(idx, intermediate_key)| {
+                        let source = &req.sources[idx];
+                        (source.name().to_string(), intermediate_key.into())
+                    })
+                    .collect();
+                Ok(CompositeBucketEntry {
+                    key,
+                    doc_count: entry.doc_count as u64,
+                    sub_aggregation: entry
+                        .sub_aggregation
+                        .into_final_result_internal(sub_aggregation_req, limits)?,
+                })
+            })
+            .collect::<crate::Result<Vec<_>>>()?;
+
+        Ok(BucketResult::Composite { after_key, buckets })
+    }
+
+    fn merge_fruits(&mut self, other: IntermediateCompositeBucketResult) -> crate::Result<()> {
+        merge_maps(&mut self.entries, other.entries)?;
+        if self.entries.len() as u32 > 2 * self.target_size {
+            self.trim()?;
+        }
+        Ok(())
+    }
+
+    /// Trim the composite buckets to the target size, according to the ordering.
+    pub(crate) fn trim(&mut self) -> crate::Result<()> {
+        if self.entries.len() as u32 <= self.target_size {
+            return Ok(());
+        }
+
+        let sorted_entries = trim_composite_buckets(
+            std::mem::take(&mut self.entries),
+            &self.orders,
+            self.target_size,
+        )?;
+
+        self.entries = sorted_entries.into_iter().collect();
+        Ok(())
+    }
+}
+
+fn trim_composite_buckets(
+    entries: FxHashMap<Vec<CompositeIntermediateKey>, IntermediateCompositeBucketEntry>,
+    orders: &[(Order, MissingOrder)],
+    target_size: u32,
+) -> crate::Result<
+    Vec<(
+        Vec<CompositeIntermediateKey>,
+        IntermediateCompositeBucketEntry,
+    )>,
+> {
+    let mut entries: Vec<_> = entries.into_iter().collect();
+    let mut sort_error: Option<TantivyError> = None;
+    entries.sort_by(|(left_key, _), (right_key, _)| {
+        if sort_error.is_some() {
+            return Ordering::Equal;
+        }
+
+        for idx in 0..orders.len() {
+            match composite_intermediate_key_ordering(
+                &left_key[idx],
+                &right_key[idx],
+                orders[idx].0,
+                orders[idx].1,
+            ) {
+                Ok(ordering) if ordering != Ordering::Equal => return ordering,
+                Ok(_) => continue,
+                Err(err) => {
+                    sort_error = Some(err);
+                    break;
+                }
+            }
+        }
+        Ordering::Equal
+    });
+
+    if let Some(err) = sort_error {
+        return Err(err);
+    }
+
+    entries.truncate(target_size as usize);
+    Ok(entries)
+}
+
 #[cfg(test)]
 mod tests {
    use std::collections::HashMap;
--- a/src/aggregation/metric/cardinality.rs
+++ b/src/aggregation/metric/cardinality.rs
@@ -1,10 +1,11 @@
+use std::fmt::Debug;
 use std::hash::Hash;
+use std::io;

 use columnar::column_values::CompactSpaceU64Accessor;
 use columnar::{Column, ColumnType, Dictionary, StrColumn};
-use common::f64_to_u64;
-use datasketches::hll::{HllSketch, HllType, HllUnion};
-use rustc_hash::FxHashSet;
+use datasketches::hll::{Coupon, HllSketch, HllType, HllUnion};
+use rustc_hash::{FxBuildHasher, FxHashMap, FxHashSet};
 use serde::{Deserialize, Deserializer, Serialize, Serializer};

 use crate::aggregation::agg_data::AggregationsSegmentCtx;
@@ -120,9 +121,65 @@ impl CardinalityAggregationReq {
    }
 }

-#[derive(Clone, Debug)]
+/// A CouponCache is here to cache the mapping term ordinal -> coupon (see above).
+/// The idea is that we do not want to fetch terms associated to several term ordinals,
+/// several times due to the fact that we have several buckets.
+enum CouponCache {
+    Dense {
+        coupon_map: Vec<Coupon>,
+        missing_coupon_opt: Option<Coupon>,
+    },
+    Sparse {
+        coupon_map: FxHashMap<u64, Coupon>,
+        missing_coupon_opt: Option<Coupon>,
+    },
+}
+
+impl CouponCache {
+    fn new(
+        term_ords: Vec<u64>,
+        coupons: Vec<Coupon>,
+        missing_coupon_opt: Option<Coupon>,
+    ) -> CouponCache {
+        let num_terms = term_ords.len();
+        assert_eq!(num_terms, coupons.len());
+        if term_ords.is_empty() {
+            return CouponCache::Dense {
+                coupon_map: Vec::new(),
+                missing_coupon_opt,
+            };
+        }
+        let highest_term_ord = term_ords.last().copied().unwrap_or(0u64);
+        // We prefer the dense implementation, if it is not too wasteful.
+        // There are two cases for which we can use it.
+        // 1- if the data is small.
+        // 2- if the data is not necessarily small, but due to a high occupancy ratio, the RAM usage
+        // is not that much bigger than if we had used a HashSet. (occupancy ratio + extra
+        // metadata ~ x2.25)
+        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()) {
+                coupon_map[term_ord as usize] = coupon;
+            }
+            CouponCache::Dense {
+                coupon_map,
+                missing_coupon_opt,
+            }
+        } else {
+            let coupon_map: FxHashMap<u64, Coupon> = term_ords.into_iter().zip(coupons).collect();
+            CouponCache::Sparse {
+                coupon_map,
+                missing_coupon_opt,
+            }
+        }
+    }
+}
+
 pub(crate) struct SegmentCardinalityCollector {
-    buckets: Vec<SegmentCardinalityCollectorBucket>,
+    /// Buckets are Some(_) until they get consumed by into_intermediate_results().
+    buckets: Vec<Option<SegmentCardinalityCollectorBucket>>,
    accessor_idx: usize,
    /// The column accessor to access the fast field values.
    accessor: Column<u64>,
@@ -130,75 +187,133 @@ pub(crate) struct SegmentCardinalityCollector {
    column_type: ColumnType,
    /// The missing value normalized to the internal u64 representation of the field type.
    missing_value_for_accessor: Option<u64>,
+    coupon_cache: Option<CouponCache>,
+}
+
+impl Debug for SegmentCardinalityCollector {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        f.debug_struct("SegmentCardinalityCollector")
+            .field("column_type", &self.column_type)
+            .field(
+                "missing_value_for_accessor",
+                &self.missing_value_for_accessor,
+            )
+            .finish()
+    }
 }

-#[derive(Clone, Debug, PartialEq, Default)]
 pub(crate) struct SegmentCardinalityCollectorBucket {
    cardinality: CardinalityCollector,
    entries: FxHashSet<u64>,
 }
 impl SegmentCardinalityCollectorBucket {
+    #[inline(always)]
    pub fn new(column_type: ColumnType) -> Self {
        Self {
            cardinality: CardinalityCollector::new(column_type as u8),
            entries: FxHashSet::default(),
        }
    }
+
+    // Returns a intermediate metric result.
+    //
+    // If the column is not str, the values have been added to the
+    // sketch during collection.
+    //
+    // If the column is str, then the values are dictionary encoded
+    // and have not been added to the sketch yet.
+    // We need to resolves the term ords accumulated in self.entries
+    // with the coupon cache, and append the results to the sketch.
    fn into_intermediate_metric_result(
        mut self,
-        req_data: &CardinalityAggReqData,
+        coupon_cache_opt: Option<&CouponCache>,
    ) -> crate::Result<IntermediateMetricResult> {
-        if req_data.column_type == ColumnType::Str {
-            let fallback_dict = Dictionary::empty();
-            let dict = req_data
-                .str_dict_column
-                .as_ref()
-                .map(|el| el.dictionary())
-                .unwrap_or_else(|| &fallback_dict);
-            let mut has_missing = false;
+        if let Some(coupon_cache) = coupon_cache_opt {
+            assert!(self.cardinality.sketch.is_empty());
+            append_to_sketch(&self.entries, coupon_cache, &mut self.cardinality);
+        }
+        Ok(IntermediateMetricResult::Cardinality(self.cardinality))
+    }
+}

-            // TODO: replace FxHashSet with something that allows iterating in order
-            // (e.g. sparse bitvec)
-            let mut term_ids = Vec::new();
-            for term_ord in self.entries.into_iter() {
-                if term_ord == u64::MAX {
-                    has_missing = true;
-                } else {
-                    // we can reasonably exclude values above u32::MAX
-                    term_ids.push(term_ord as u32);
-                }
-            }
+/// Builds a coupon cache from the given buckets, dictionary, and optional missing value.
+/// Returns a mapping from term_ord to the hash (coupon) of the associated term.
+fn build_coupon_cache(
+    buckets: &[Option<SegmentCardinalityCollectorBucket>],
+    dictionary: &Dictionary,
+    missing_value_opt: Option<&Key>,
+) -> io::Result<CouponCache> {
+    let term_ords_capacity: usize = buckets
+        .iter()
+        .flatten()
+        .map(|bucket| bucket.entries.len())
+        .max()
+        .unwrap_or(0)
+        * 2;
+    let mut term_ords_set = FxHashSet::with_capacity_and_hasher(term_ords_capacity, FxBuildHasher);
+    for bucket in buckets.iter().flatten() {
+        term_ords_set.extend(bucket.entries.iter().copied());
+    }
+    let mut term_ords: Vec<u64> = term_ords_set.into_iter().collect();
+    term_ords.sort_unstable();

-            term_ids.sort_unstable();
-            dict.sorted_ords_to_term_cb(term_ids.iter().map(|term| *term as u64), |term| {
-                self.cardinality.insert(term);
-                Ok(())
-            })?;
-            if has_missing {
-                // Replace missing with the actual value provided
-                let missing_key =
-                    req_data.req.missing.as_ref().expect(
-                        "Found sentinel value u64::MAX for term_ord but `missing` is not set",
-                    );
-                match missing_key {
-                    Key::Str(missing) => {
-                        self.cardinality.insert(missing.as_str());
-                    }
-                    Key::F64(val) => {
-                        let val = f64_to_u64(*val);
-                        self.cardinality.insert(val);
-                    }
-                    Key::U64(val) => {
-                        self.cardinality.insert(*val);
-                    }
-                    Key::I64(val) => {
-                        self.cardinality.insert(*val);
-                    }
+    term_ords.pop_if(|highest_term_ord| *highest_term_ord >= dictionary.num_terms() as u64);
+
+    let mut coupons: Vec<Coupon> = Vec::with_capacity(term_ords.len());
+    let all_term_ords_found: bool =
+        dictionary.sorted_ords_to_term_cb(&term_ords, |term_bytes| {
+            let coupon: Coupon = Coupon::from_hash(term_bytes);
+            coupons.push(coupon);
+        })?;
+    assert!(all_term_ords_found);
+
+    // Regardless of whether or not there is effectively a missing value in one of the buckets,
+    // we populate the cache with the missing key too (if any).
+    let missing_coupon_opt: Option<Coupon> = missing_value_opt.map(|missing_key| {
+        if let Key::Str(missing_value_str) = missing_key {
+            Coupon::from_hash(missing_value_str.as_bytes())
+        } else {
+            // See https://github.com/quickwit-oss/tantivy/issues/2891
+            // A missing key with a type different from Str will not work as intended
+            // for the moment.
+            //
+            // Right now this is just a partial workaround.
+            Coupon::from_hash("__tantivy_missing_non_str__".as_bytes())
+        }
+    });
+    Ok(CouponCache::new(term_ords, coupons, missing_coupon_opt))
+}
+
+fn append_to_sketch(
+    term_ords: &FxHashSet<u64>,
+    coupon_cache: &CouponCache,
+    sketch: &mut CardinalityCollector,
+) {
+    match coupon_cache {
+        CouponCache::Dense {
+            coupon_map,
+            missing_coupon_opt,
+        } => {
+            for &term_ord in term_ords {
+                if let Some(coupon) = coupon_map
+                    .get(term_ord as usize)
+                    .copied()
+                    .or(*missing_coupon_opt)
+                {
+                    sketch.insert_coupon(coupon);
+                }
+            }
+        }
+        CouponCache::Sparse {
+            coupon_map,
+            missing_coupon_opt,
+        } => {
+            for term_ord in term_ords {
+                if let Some(coupon) = coupon_map.get(term_ord).copied().or(*missing_coupon_opt) {
+                    sketch.insert_coupon(coupon);
                }
            }
        }
-
-        Ok(IntermediateMetricResult::Cardinality(self.cardinality))
    }
 }

@@ -210,11 +325,12 @@ impl SegmentCardinalityCollector {
        missing_value_for_accessor: Option<u64>,
    ) -> Self {
        Self {
-            buckets: vec![SegmentCardinalityCollectorBucket::new(column_type); 1],
+            buckets: Vec::new(),
            column_type,
            accessor_idx,
            accessor,
            missing_value_for_accessor,
+            coupon_cache: None,
        }
    }

@@ -236,15 +352,35 @@ impl SegmentAggregationCollector for SegmentCardinalityCollector {
        &mut self,
        agg_data: &AggregationsSegmentCtx,
        results: &mut IntermediateAggregationResults,
-        parent_bucket_id: BucketId,
+        bucket_id: BucketId,
    ) -> crate::Result<()> {
-        self.prepare_max_bucket(parent_bucket_id, agg_data)?;
+        self.prepare_max_bucket(bucket_id, agg_data)?;
        let req_data = &agg_data.get_cardinality_req_data(self.accessor_idx);
+        // Strings are dictionary encoded. Fetching the terms associated to strings
+        // is expensive. For this reason, we do that once for all buckets and cache the results
+        // here.
+        if let Some(str_dict_column) = &req_data.str_dict_column {
+            // Ensure the coupon cache is populated.
+            // A mapping from term_ord to the hash of the associated term.
+            // The missing value sentinel will be associated to the hash of the missing value if
+            // any.
+            if self.coupon_cache.is_none() {
+                self.coupon_cache = Some(build_coupon_cache(
+                    &self.buckets,
+                    str_dict_column.dictionary(),
+                    req_data.req.missing.as_ref(),
+                )?);
+            }
+        }
        let name = req_data.name.to_string();
        // take the bucket in buckets and replace it with a new empty one
-        let bucket = std::mem::take(&mut self.buckets[parent_bucket_id as usize]);
-
-        let intermediate_result = bucket.into_intermediate_metric_result(req_data)?;
+        let Some(bucket) = self.buckets[bucket_id as usize].take() else {
+            return Err(crate::TantivyError::InternalError(
+                "the same bucket should not be finalized twice.".to_string(),
+            ));
+        };
+        let intermediate_result =
+            bucket.into_intermediate_metric_result(self.coupon_cache.as_ref())?;
        results.push(
            name,
            IntermediateAggregationResult::Metric(intermediate_result),
@@ -260,8 +396,11 @@ impl SegmentAggregationCollector for SegmentCardinalityCollector {
        agg_data: &mut AggregationsSegmentCtx,
    ) -> crate::Result<()> {
        self.fetch_block_with_field(docs, agg_data);
-        let bucket = &mut self.buckets[parent_bucket_id as usize];
-
+        let Some(bucket) = &mut self.buckets[parent_bucket_id as usize].as_mut() else {
+            return Err(crate::TantivyError::InternalError(
+                "collection should not happen after finalization".to_string(),
+            ));
+        };
        let col_block_accessor = &agg_data.column_block_accessor;
        if self.column_type == ColumnType::Str {
            for term_ord in col_block_accessor.iter_vals() {
@@ -301,11 +440,33 @@ impl SegmentAggregationCollector for SegmentCardinalityCollector {
    ) -> crate::Result<()> {
        if max_bucket as usize >= self.buckets.len() {
            self.buckets.resize_with(max_bucket as usize + 1, || {
-                SegmentCardinalityCollectorBucket::new(self.column_type)
+                Some(SegmentCardinalityCollectorBucket::new(self.column_type))
            });
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        let req_data = &agg_data.get_cardinality_req_data(self.accessor_idx);
+        if req_data.name != sub_agg_name || !sub_agg_property.is_empty() {
+            return None;
+        }
+        let bucket = self.buckets.get(bucket_id as usize)?.as_ref()?;
+        // For string columns the HLL sketch is empty until materialization; entries holds
+        // the deduplicated term ordinals seen, which is the exact distinct count.
+        // For numeric columns the sketch is populated during collect.
+        if self.column_type == ColumnType::Str {
+            Some(bucket.entries.len() as f64)
+        } else {
+            Some(bucket.cardinality.sketch.estimate().trunc())
+        }
+    }
 }

 #[derive(Clone, Debug)]
@@ -358,10 +519,14 @@ impl CardinalityCollector {
    /// Insert a value into the HLL sketch, salted by the column type.
    /// The salt ensures that identical u64 values from different column types
    /// (e.g. bool `false` vs i64 `0`) are counted as distinct.
-    pub(crate) fn insert<T: Hash>(&mut self, value: T) {
+    fn insert<T: Hash>(&mut self, value: T) {
        self.sketch.update((self.salt, value));
    }

+    fn insert_coupon(&mut self, coupon: Coupon) {
+        self.sketch.update_with_coupon(coupon);
+    }
+
    /// Compute the final cardinality estimate.
    pub fn finalize(self) -> Option<f64> {
        Some(self.sketch.estimate().trunc())
@@ -377,7 +542,7 @@ impl CardinalityCollector {
        let mut union = HllUnion::new(LG_K);
        union.update(&self.sketch);
        union.update(&right.sketch);
-        self.sketch = union.get_result(HllType::Hll4);
+        self.sketch = union.to_sketch(HllType::Hll4);
        Ok(())
    }
 }
@@ -392,7 +557,7 @@ mod tests {

    use crate::aggregation::agg_req::Aggregations;
    use crate::aggregation::tests::{exec_request, get_test_index_from_terms};
-    use crate::schema::{IntoIpv6Addr, Schema, FAST};
+    use crate::schema::{IntoIpv6Addr, Schema, FAST, STRING};
    use crate::Index;

    #[test]
@@ -575,6 +740,30 @@ mod tests {
        assert_eq!(estimate, 3.0);
    }

+    /// Verifies that merging two small sketches (both in List/Set coupon mode)
+    /// produces an exact result — i.e. the HllUnion does not unnecessarily
+    /// promote to the full HLL array when the combined cardinality is small.
+    #[test]
+    fn cardinality_collector_merge_stays_exact_for_small_sets() {
+        use super::CardinalityCollector;
+
+        let mut left = CardinalityCollector::default();
+        for i in 0u64..50 {
+            left.insert(i);
+        }
+
+        let mut right = CardinalityCollector::default();
+        for i in 30u64..100 {
+            right.insert(i);
+        }
+
+        left.merge_fruits(right).unwrap();
+        let estimate = left.finalize().unwrap();
+        // 100 distinct values (0..100). Both sketches are in Set mode (< 192 coupons),
+        // so the union should stay in coupon mode and give an exact count.
+        assert_eq!(estimate, 100.0);
+    }
+
    #[test]
    fn cardinality_collector_serialize_deserialize_binary() {
        use datasketches::hll::HllSketch;
@@ -591,6 +780,98 @@ mod tests {
        assert!((deserialized.estimate() - 3.0).abs() < 0.01);
    }

+    /// Tests that the `missing` parameter correctly counts a single empty document
+    /// for both u64 and str columns.
+    #[test]
+    fn cardinality_aggregation_missing_value_single_empty_doc() {
+        let mut schema_builder = Schema::builder();
+        let id_field = schema_builder.add_u64_field("id", FAST);
+        let name_field = schema_builder.add_text_field("name", STRING | FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut writer = index.writer_for_tests().unwrap();
+        writer
+            .add_document(doc!(id_field=>1u64,name_field=>"some_name"))
+            .unwrap();
+        writer.add_document(doc!()).unwrap();
+        writer.commit().unwrap();
+
+        {
+            // int colum with missing value non redundant
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "id",
+                        "missing": 42u64
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 2.0);
+        }
+
+        {
+            // int colum with missing value redundant
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "id",
+                        "missing": 1u64
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 1.0);
+        }
+
+        {
+            // str colum with missing value non redundant
+            // With more than one segment, this is not well handled.
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "name",
+                        "missing": "other_name"
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 2.0);
+        }
+
+        {
+            // str colum with missing value redundant
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "name",
+                        "missing": "some_name"
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 1.0);
+        }
+
+        {
+            // str column with missing value with a number type.
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "name",
+                        "missing": 3,
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 2.0);
+        }
+    }
+
    #[test]
    fn cardinality_collector_salt_differentiates_types() {
        use super::CardinalityCollector;
--- a/src/aggregation/metric/extended_stats.rs
+++ b/src/aggregation/metric/extended_stats.rs
@@ -399,6 +399,26 @@ impl SegmentAggregationCollector for SegmentExtendedStatsCollector {
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        if self.name != sub_agg_name {
+            return None;
+        }
+        let extended = self.buckets.get(bucket_id as usize)?;
+        // Finalize is a pure read of accumulators — calling it here for the cutoff sort
+        // doesn't disturb the eventual intermediate result.
+        extended
+            .finalize()
+            .get_value(sub_agg_property)
+            .ok()
+            .flatten()
+    }
 }

 #[cfg(test)]
--- a/src/aggregation/metric/mod.rs
+++ b/src/aggregation/metric/mod.rs
@@ -107,10 +107,9 @@ pub enum PercentileValues {
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 /// The entry when requesting percentiles with keyed: false
 pub struct PercentileValuesVecEntry {
-    /// Percentile
+    /// The percentile key (e.g. 1.0, 5.0, 25.0).
    pub key: f64,
-
-    /// Value at the percentile
+    /// The percentile value. `NaN` when there are no values.
    pub value: f64,
 }

--- a/src/aggregation/metric/percentiles.rs
+++ b/src/aggregation/metric/percentiles.rs
@@ -312,6 +312,26 @@ impl SegmentAggregationCollector for SegmentPercentilesCollector {
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        if agg_data.get_metric_req_data(self.accessor_idx).name != sub_agg_name {
+            return None;
+        }
+        let percentile: f64 = sub_agg_property.parse().ok()?;
+        if !(0.0..=100.0).contains(&percentile) {
+            return None;
+        }
+        let bucket = self.buckets.get(bucket_id as usize)?;
+        // DDSketch.quantile is a pure read; calling it here for the cutoff sort does
+        // not affect the intermediate state used for the final result.
+        bucket.sketch.quantile(percentile / 100.0).ok().flatten()
+    }
 }

 #[cfg(test)]
--- a/src/aggregation/metric/stats.rs
+++ b/src/aggregation/metric/stats.rs
@@ -321,6 +321,40 @@ impl<const COLUMN_TYPE_ID: u8> SegmentAggregationCollector
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        if self.name != sub_agg_name {
+            return None;
+        }
+        let stats = self.buckets.get(bucket_id as usize)?;
+        // The property depends on what we're collecting:
+        //   - StatsType::Stats exposes count/sum/min/max/avg via dotted property.
+        //   - Single-value kinds (Sum/Count/Min/Max/Average) expect an empty property and return
+        //     the value they were configured to collect.
+        let prop = match self.collecting_for {
+            StatsType::Stats if !sub_agg_property.is_empty() => sub_agg_property,
+            StatsType::Sum if sub_agg_property.is_empty() => "sum",
+            StatsType::Count if sub_agg_property.is_empty() => "count",
+            StatsType::Max if sub_agg_property.is_empty() => "max",
+            StatsType::Min if sub_agg_property.is_empty() => "min",
+            StatsType::Average if sub_agg_property.is_empty() => "avg",
+            _ => return None,
+        };
+        match prop {
+            "count" => Some(stats.count as f64),
+            "sum" => Some(stats.sum),
+            "min" if stats.count > 0 => Some(stats.min),
+            "max" if stats.count > 0 => Some(stats.max),
+            "avg" if stats.count > 0 => Some(stats.sum / stats.count as f64),
+            _ => None,
+        }
+    }
 }

 #[inline]
--- a/src/aggregation/metric/top_hits.rs
+++ b/src/aggregation/metric/top_hits.rs
@@ -644,6 +644,17 @@ impl SegmentAggregationCollector for TopHitsSegmentCollector {
        );
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        _bucket_id: BucketId,
+        _sub_agg_name: &str,
+        _sub_agg_property: &str,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        // top_hits is not a numeric metric and cannot be used as an order target.
+        None
+    }
 }

 #[cfg(test)]
--- a/src/aggregation/mod.rs
+++ b/src/aggregation/mod.rs
@@ -133,7 +133,7 @@ mod agg_limits;
 pub mod agg_req;
 pub mod agg_result;
 pub mod bucket;
-pub(crate) mod cached_sub_aggs;
+pub(crate) mod buffered_sub_aggs;
 mod collector;
 mod date;
 mod error;
--- a/src/aggregation/segment_agg_result.rs
+++ b/src/aggregation/segment_agg_result.rs
@@ -76,6 +76,31 @@ pub trait SegmentAggregationCollector: Debug {
    fn flush(&mut self, _agg_data: &mut AggregationsSegmentCtx) -> crate::Result<()> {
        Ok(())
    }
+
+    /// Compute the segment-level metric value of the named direct-child metric for `bucket_id`.
+    ///
+    /// Used by parent term aggs that order by a sub-aggregation: the parent sorts on
+    /// this value and cuts off at segment time, matching the approximation tradeoff
+    /// Elasticsearch makes for any sub-agg ordering.
+    ///
+    /// `sub_agg_property` is the dotted suffix (e.g. `"sum"` in `mystats.sum`); empty when
+    /// the metric is a single-value kind such as cardinality.
+    ///
+    /// Returns `None` only on name mismatch, unknown property, or empty bucket. Implementations
+    /// may finalize their per-bucket state (e.g. compute a percentile from a sketch); calls
+    /// must be idempotent so the final intermediate result is unaffected.
+    ///
+    /// No default impl on purpose: every collector must decide explicitly whether it
+    /// produces a metric value, forwards into children (single-bucket aggs), or rejects
+    /// the lookup. A silent `None` default would let a parent term agg's cutoff sort all
+    /// buckets to the same key and drop arbitrary winners.
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64>;
 }

 #[derive(Default)]
@@ -137,4 +162,21 @@ impl SegmentAggregationCollector for GenericSegmentAggregationResultsCollector {
        }
        Ok(())
    }
+
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        for agg in &self.aggs {
+            if let Some(value) =
+                agg.compute_metric_value(bucket_id, sub_agg_name, sub_agg_property, agg_data)
+            {
+                return Some(value);
+            }
+        }
+        None
+    }
 }
--- a/src/collector/count_collector.rs
+++ b/src/collector/count_collector.rs
@@ -1,5 +1,6 @@
 use super::Collector;
 use crate::collector::SegmentCollector;
+use crate::query::Weight;
 use crate::{DocId, Score, SegmentOrdinal, SegmentReader};

 /// `CountCollector` collector only counts how many
@@ -55,6 +56,15 @@ impl Collector for Count {
    fn merge_fruits(&self, segment_counts: Vec<usize>) -> crate::Result<usize> {
        Ok(segment_counts.into_iter().sum())
    }
+
+    fn collect_segment(
+        &self,
+        weight: &dyn Weight,
+        _segment_ord: u32,
+        reader: &SegmentReader,
+    ) -> crate::Result<usize> {
+        Ok(weight.count(reader)? as usize)
+    }
 }

 #[derive(Default)]
--- a/src/collector/facet_collector.rs
+++ b/src/collector/facet_collector.rs
@@ -389,6 +389,13 @@ impl SegmentCollector for FacetSegmentCollector {
            }
            let mut facet = vec![];
            let (facet_ord, facet_depth) = self.unique_facet_ords[collapsed_facet_ord];
+            // u64::MAX is used as a sentinel for unmapped ordinals (e.g. when a
+            // document has the exact registered facet, not a child of it).
+            // Passing it to ord_to_term would resolve to the last dictionary
+            // entry and produce a spurious facet from an unrelated branch.
+            if facet_ord == u64::MAX {
+                continue;
+            }
            // TODO handle errors.
            if facet_dict.ord_to_term(facet_ord, &mut facet).is_ok() {
                if let Some((end_collapsed_facet, _)) = facet
@@ -814,6 +821,63 @@ mod tests {
        assert!(!super::is_child_facet(&b"foo\0bar"[..], &b"foo"[..]));
        assert!(!super::is_child_facet(&b"foo"[..], &b"foobar\0baz"[..]));
    }
+
+    // Regression test for https://github.com/quickwit-oss/tantivy/issues/2494
+    // When a document has the exact registered facet path (not just a child),
+    // harvest() must not turn the unmapped sentinel into a spurious root entry.
+    #[test]
+    fn test_facet_collector_wrong_root() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let facet_field = schema_builder.add_facet_field("facet", FacetOptions::default());
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+
+        let mut index_writer: IndexWriter = index.writer_for_tests()?;
+        let facets: Vec<&str> = vec![
+            "/science-fiction/asimov",
+            "/science-fiction/clarke",
+            "/science-fiction/dick",
+            "/science-fiction/herbert",
+            "/science-fiction/orwell",
+            // This exact match on the registered facet is the bug trigger:
+            // its ordinal maps to the sentinel (u64::MAX, 0) in the collapse
+            // mapping, which without the fix resolves to an unrelated term.
+            "/fantasy/epic-fantasy",
+            "/fantasy/epic-fantasy/tolkien",
+            "/fantasy/epic-fantasy/martin",
+        ];
+        for facet_str in &facets {
+            index_writer.add_document(doc!(
+                facet_field => Facet::from(*facet_str)
+            ))?;
+        }
+        index_writer.commit()?;
+
+        let reader = index.reader()?;
+        let searcher = reader.searcher();
+
+        let term = Term::from_facet(facet_field, &Facet::from("/fantasy/epic-fantasy"));
+        let query = TermQuery::new(term, IndexRecordOption::Basic);
+
+        let mut facet_collector = FacetCollector::for_field("facet");
+        facet_collector.add_facet("/fantasy/epic-fantasy");
+        let counts: FacetCounts = searcher.search(&query, &facet_collector)?;
+
+        let result: Vec<(String, u64)> = counts
+            .get("/")
+            .map(|(facet, count)| (facet.to_string(), count))
+            .collect();
+
+        // Only children of /fantasy/epic-fantasy should appear, not /science-fiction
+        assert_eq!(
+            result,
+            vec![
+                ("/fantasy/epic-fantasy/martin".to_string(), 1),
+                ("/fantasy/epic-fantasy/tolkien".to_string(), 1),
+            ]
+        );
+        Ok(())
+    }
 }

 #[cfg(all(test, feature = "unstable"))]
--- a/src/collector/sort_key/sort_by_score.rs
+++ b/src/collector/sort_key/sort_by_score.rs
@@ -1,5 +1,8 @@
+use std::cmp::{Ordering, Reverse};
+use std::collections::BinaryHeap;
+
 use crate::collector::sort_key::NaturalComparator;
-use crate::collector::{SegmentSortKeyComputer, SortKeyComputer, TopNComputer};
+use crate::collector::{SegmentSortKeyComputer, SortKeyComputer};
 use crate::{DocAddress, DocId, Score};

 /// Sort by similarity score.
@@ -25,6 +28,10 @@ impl SortKeyComputer for SortBySimilarityScore {
    }

    // Sorting by score is special in that it allows for the Block-Wand optimization.
+    //
+    // We use a BinaryHeap (TopNHeap) instead of TopNComputer here so that the
+    // threshold is always the exact K-th best score. TopNComputer only updates its
+    // threshold every K docs (at truncation), giving Block-WAND a stale bound.
    fn collect_segment_top_k(
        &self,
        k: usize,
@@ -32,12 +39,10 @@ impl SortKeyComputer for SortBySimilarityScore {
        reader: &crate::SegmentReader,
        segment_ord: u32,
    ) -> crate::Result<Vec<(Self::SortKey, DocAddress)>> {
-        let mut top_n: TopNComputer<Score, DocId, Self::Comparator> =
-            TopNComputer::new_with_comparator(k, self.comparator());
+        let mut top_n = TopNHeap::new(k);

        if let Some(alive_bitset) = reader.alive_bitset() {
            let mut threshold = Score::MIN;
-            top_n.threshold = Some(threshold);
            weight.for_each_pruning(Score::MIN, reader, &mut |doc, score| {
                if alive_bitset.is_deleted(doc) {
                    return threshold;
@@ -56,7 +61,7 @@ impl SortKeyComputer for SortBySimilarityScore {
        Ok(top_n
            .into_vec()
            .into_iter()
-            .map(|cid| (cid.sort_key, DocAddress::new(segment_ord, cid.doc)))
+            .map(|(score, doc)| (score, DocAddress::new(segment_ord, doc)))
            .collect())
    }
 }
@@ -75,3 +80,204 @@ impl SegmentSortKeyComputer for SortBySimilarityScore {
        score
    }
 }
+
+/// Min-heap entry: higher score = greater, lower doc wins ties.
+struct ScoreHeapEntry {
+    score: Score,
+    doc: DocId,
+}
+
+impl Eq for ScoreHeapEntry {}
+
+impl PartialEq for ScoreHeapEntry {
+    fn eq(&self, other: &Self) -> bool {
+        self.cmp(other) == Ordering::Equal
+    }
+}
+
+impl PartialOrd for ScoreHeapEntry {
+    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl Ord for ScoreHeapEntry {
+    fn cmp(&self, other: &Self) -> Ordering {
+        self.score
+            .partial_cmp(&other.score)
+            .unwrap_or(Ordering::Equal)
+            .then_with(|| other.doc.cmp(&self.doc))
+    }
+}
+
+/// Heap-based top-K for score collection. O(log K) per insert, but the threshold
+/// is always tight, so Block-WAND prunes better than with [`TopNComputer`]'s
+/// buffer/median approach.
+///
+/// Like [`TopNComputer`], items must arrive in ascending doc order, and equal
+/// scores are rejected (strict `>`) so that lower doc IDs win ties.
+///
+/// [`TopNComputer`]: crate::collector::TopNComputer
+struct TopNHeap {
+    heap: BinaryHeap<Reverse<ScoreHeapEntry>>,
+    top_n: usize,
+    threshold: Option<Score>,
+}
+
+impl TopNHeap {
+    fn new(top_n: usize) -> Self {
+        TopNHeap {
+            heap: BinaryHeap::with_capacity(top_n),
+            top_n,
+            threshold: None,
+        }
+    }
+
+    #[inline]
+    fn push(&mut self, score: Score, doc: DocId) {
+        if self.heap.len() < self.top_n {
+            self.heap.push(Reverse(ScoreHeapEntry { score, doc }));
+            if self.heap.len() == self.top_n {
+                self.threshold = self.heap.peek().map(|Reverse(entry)| entry.score);
+            }
+        } else if let Some(threshold) = self.threshold {
+            if score > threshold {
+                // peek_mut + assign is a single sift-down, vs pop + push = two sifts.
+                if let Some(mut min) = self.heap.peek_mut() {
+                    *min = Reverse(ScoreHeapEntry { score, doc });
+                }
+                self.threshold = self.heap.peek().map(|Reverse(entry)| entry.score);
+            }
+        }
+    }
+
+    fn into_vec(self) -> Vec<(Score, DocId)> {
+        self.heap
+            .into_vec()
+            .into_iter()
+            .map(|Reverse(entry)| (entry.score, entry.doc))
+            .collect()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use proptest::prelude::*;
+
+    use super::*;
+    use crate::collector::sort_key::NaturalComparator;
+    use crate::collector::TopNComputer;
+
+    #[test]
+    fn test_top_n_heap_zero_capacity() {
+        let mut heap = TopNHeap::new(0);
+        heap.push(1.0, 0);
+        heap.push(2.0, 1);
+        assert!(heap.into_vec().is_empty());
+    }
+
+    #[test]
+    fn test_top_n_heap_basic() {
+        let mut heap = TopNHeap::new(2);
+        heap.push(1.0, 0);
+        heap.push(3.0, 1);
+        heap.push(2.0, 2);
+
+        let mut results = heap.into_vec();
+        results.sort_by(|a, b| b.0.partial_cmp(&a.0).unwrap().then_with(|| a.1.cmp(&b.1)));
+        assert_eq!(results, vec![(3.0, 1), (2.0, 2)]);
+    }
+
+    #[test]
+    fn test_top_n_heap_threshold_always_accurate() {
+        let mut heap = TopNHeap::new(2);
+        assert_eq!(heap.threshold, None);
+
+        heap.push(1.0, 0);
+        assert_eq!(heap.threshold, None);
+
+        heap.push(3.0, 1);
+        assert_eq!(heap.threshold, Some(1.0));
+
+        heap.push(2.0, 2); // evicts 1.0
+        assert_eq!(heap.threshold, Some(2.0));
+
+        heap.push(4.0, 3); // evicts 2.0
+        assert_eq!(heap.threshold, Some(3.0));
+    }
+
+    #[test]
+    fn test_top_n_heap_tiebreaking_lower_doc_wins() {
+        let mut heap = TopNHeap::new(2);
+        heap.push(5.0, 0);
+        heap.push(5.0, 1);
+        heap.push(5.0, 2); // rejected: not strictly > threshold
+
+        let mut results = heap.into_vec();
+        results.sort_by_key(|&(_, doc)| doc);
+        assert_eq!(results, vec![(5.0, 0), (5.0, 1)]);
+    }
+
+    #[test]
+    fn test_top_n_heap_single_element() {
+        let mut heap = TopNHeap::new(1);
+        heap.push(1.0, 0);
+        assert_eq!(heap.threshold, Some(1.0));
+
+        heap.push(0.5, 1); // rejected
+        heap.push(2.0, 2); // accepted
+        assert_eq!(heap.threshold, Some(2.0));
+
+        let results = heap.into_vec();
+        assert_eq!(results, vec![(2.0, 2)]);
+    }
+
+    #[test]
+    fn test_top_n_heap_under_capacity() {
+        let mut heap = TopNHeap::new(5);
+        heap.push(3.0, 0);
+        heap.push(1.0, 1);
+        heap.push(2.0, 2);
+        // Only 3 elements, capacity is 5 — all should be kept
+        assert_eq!(heap.threshold, None);
+
+        let mut results = heap.into_vec();
+        results.sort_by(|a, b| b.0.partial_cmp(&a.0).unwrap().then_with(|| a.1.cmp(&b.1)));
+        assert_eq!(results, vec![(3.0, 0), (2.0, 2), (1.0, 1)]);
+    }
+
+    proptest! {
+        #[test]
+        fn test_top_n_heap_matches_top_n_computer(
+            limit in 0..20_usize,
+            mut docs in proptest::collection::vec((0..1000_u32, 0..1000_u32), 0..200_usize),
+        ) {
+            // Both require ascending doc order.
+            docs.sort_by_key(|(_, doc_id)| *doc_id);
+            docs.dedup_by_key(|(_, doc_id)| *doc_id);
+
+            let mut heap = TopNHeap::new(limit);
+            let mut computer: TopNComputer<Score, DocId, NaturalComparator> =
+                TopNComputer::new_with_comparator(limit, NaturalComparator);
+
+            for &(score_u32, doc) in &docs {
+                let score = score_u32 as Score;
+                heap.push(score, doc);
+                computer.push(score, doc);
+            }
+
+            let mut heap_results = heap.into_vec();
+            heap_results.sort_by(|a, b| {
+                b.0.partial_cmp(&a.0).unwrap().then_with(|| a.1.cmp(&b.1))
+            });
+
+            let computer_results: Vec<(Score, DocId)> = computer
+                .into_sorted_vec()
+                .into_iter()
+                .map(|cd| (cd.sort_key, cd.doc))
+                .collect();
+
+            prop_assert_eq!(heap_results, computer_results);
+        }
+    }
+}
--- a/src/collector/top_score_collector.rs
+++ b/src/collector/top_score_collector.rs
@@ -513,7 +513,9 @@ pub struct TopNComputer<Score, D, C> {
    /// The buffer reverses sort order to get top-semantics instead of bottom-semantics
    buffer: Vec<ComparableDoc<Score, D>>,
    top_n: usize,
-    pub(crate) threshold: Option<Score>,
+    /// The current threshold for pruning. Documents with scores at or below
+    /// this value are skipped by `push()`. Updated when the buffer is truncated.
+    pub threshold: Option<Score>,
    comparator: C,
 }

--- a/src/directory/composite_file.rs
+++ b/src/directory/composite_file.rs
@@ -167,6 +167,7 @@ impl CompositeFile {
            .map(|byte_range| self.data.slice(byte_range.clone()))
    }

+    /// Returns the space usage per field in this composite file.
    pub fn space_usage(&self, schema: &Schema) -> PerFieldSpaceUsage {
        let mut fields = Vec::new();
        for (&field_addr, byte_range) in &self.offsets_index {
--- a/src/docset.rs
+++ b/src/docset.rs
@@ -1,5 +1,7 @@
 use std::borrow::{Borrow, BorrowMut};

+use common::TinySet;
+
 use crate::fastfield::AliveBitSet;
 use crate::DocId;

@@ -14,6 +16,12 @@ pub const TERMINATED: DocId = i32::MAX as u32;
 /// exactly this size as long as we can fill the buffer.
 pub const COLLECT_BLOCK_BUFFER_LEN: usize = 64;

+/// Number of `TinySet` (64-bit) buckets in a block used by [`DocSet::fill_bitset_block`].
+pub const BLOCK_NUM_TINYBITSETS: usize = 16;
+
+/// Number of doc IDs covered by one block: `BLOCK_NUM_TINYBITSETS * 64 = 1024`.
+pub const BLOCK_WINDOW: u32 = BLOCK_NUM_TINYBITSETS as u32 * 64;
+
 /// Represents an iterable set of sorted doc ids.
 pub trait DocSet: Send {
    /// Goes to the next element.
@@ -160,6 +168,31 @@ pub trait DocSet: Send {
        self.size_hint() as u64
    }

+    /// Fills a bitmask representing which documents in `[min_doc, min_doc + BLOCK_WINDOW)` are
+    /// present in this docset.
+    ///
+    /// The window is divided into `BLOCK_NUM_TINYBITSETS` buckets of 64 docs each.
+    /// Returns the next doc `>= min_doc + BLOCK_WINDOW`, or `TERMINATED` if exhausted.
+    fn fill_bitset_block(
+        &mut self,
+        min_doc: DocId,
+        mask: &mut [TinySet; BLOCK_NUM_TINYBITSETS],
+    ) -> DocId {
+        self.seek(min_doc);
+        let horizon = min_doc + BLOCK_WINDOW;
+        loop {
+            let doc = self.doc();
+            if doc >= horizon {
+                return doc;
+            }
+            let delta = doc - min_doc;
+            mask[(delta / 64) as usize].insert_mut(delta % 64);
+            if self.advance() == TERMINATED {
+                return TERMINATED;
+            }
+        }
+    }
+
    /// Returns the number documents matching.
    /// Calling this method consumes the `DocSet`.
    fn count(&mut self, alive_bitset: &AliveBitSet) -> u32 {
@@ -214,6 +247,18 @@ impl DocSet for &mut dyn DocSet {
        (**self).seek_danger(target)
    }

+    fn fill_buffer(&mut self, buffer: &mut [DocId; COLLECT_BLOCK_BUFFER_LEN]) -> usize {
+        (**self).fill_buffer(buffer)
+    }
+
+    fn fill_bitset_block(
+        &mut self,
+        min_doc: DocId,
+        mask: &mut [TinySet; BLOCK_NUM_TINYBITSETS],
+    ) -> DocId {
+        (**self).fill_bitset_block(min_doc, mask)
+    }
+
    fn doc(&self) -> u32 {
        (**self).doc()
    }
@@ -256,6 +301,15 @@ impl<TDocSet: DocSet + ?Sized> DocSet for Box<TDocSet> {
        unboxed.fill_buffer(buffer)
    }

+    fn fill_bitset_block(
+        &mut self,
+        min_doc: DocId,
+        mask: &mut [TinySet; BLOCK_NUM_TINYBITSETS],
+    ) -> DocId {
+        let unboxed: &mut TDocSet = self.borrow_mut();
+        unboxed.fill_bitset_block(min_doc, mask)
+    }
+
    fn doc(&self) -> DocId {
        let unboxed: &TDocSet = self.borrow();
        unboxed.doc()
--- a/src/indexer/segment_updater.rs
+++ b/src/indexer/segment_updater.rs
@@ -403,7 +403,8 @@ impl SegmentUpdater {
            // from the different drives.
            //
            // Segment 1 from disk 1, Segment 1 from disk 2, etc.
-            committed_segment_metas.sort_by_key(|segment_meta| std::cmp::Reverse(segment_meta.max_doc()));
+            committed_segment_metas
+                .sort_by_key(|segment_meta| std::cmp::Reverse(segment_meta.max_doc()));
            let index_meta = IndexMeta {
                index_settings: index.settings().clone(),
                segments: committed_segment_metas,
@@ -648,9 +649,6 @@ impl SegmentUpdater {
                                    merge_operation.segment_ids(),
                                    advance_deletes_err
                                );
-                                assert!(!cfg!(test), "Merge failed.");
-
-                                // ... cancel merge
                                // `merge_operations` are tracked. As it is dropped, the
                                // the segment_ids will be available again for merge.
                                return Err(advance_deletes_err);
@@ -705,12 +703,12 @@ mod tests {
    use crate::collector::TopDocs;
    use crate::directory::RamDirectory;
    use crate::fastfield::AliveBitSet;
+    use crate::index::{SegmentId, SegmentMetaInventory};
    use crate::indexer::merge_policy::tests::MergeWheneverPossible;
    use crate::indexer::merger::IndexMerger;
    use crate::indexer::segment_updater::merge_filtered_segments;
    use crate::query::QueryParser;
    use crate::schema::*;
-    use crate::index::{SegmentId, SegmentMetaInventory};
    use crate::{Directory, DocAddress, Index, Segment};

    #[test]
@@ -718,7 +716,7 @@ mod tests {
        // Regression test: -(max_doc as i32) overflows for max_doc >= 2^31.
        // Using std::cmp::Reverse avoids this.
        let inventory = SegmentMetaInventory::default();
-        let mut metas = vec![
+        let mut metas = [
            inventory.new_segment_meta(SegmentId::generate_random(), 100),
            inventory.new_segment_meta(SegmentId::generate_random(), (1u32 << 31) - 1),
            inventory.new_segment_meta(SegmentId::generate_random(), 50_000),
--- a/src/postings/block_segment_postings.rs
+++ b/src/postings/block_segment_postings.rs
@@ -249,6 +249,12 @@ impl BlockSegmentPostings {

    /// Returns the length of the current block.
    ///
+    /// Returns the decoded term-frequency buffer for the current block.
+    #[inline]
+    pub(crate) fn freq_output_array(&self) -> &[u32] {
+        self.freq_decoder.output_array()
+    }
+
    /// All blocks have a length of `NUM_DOCS_PER_BLOCK`,
    /// except the last block that may have a length
    /// of any number between 1 and `NUM_DOCS_PER_BLOCK - 1`
@@ -298,6 +304,11 @@ impl BlockSegmentPostings {
        }
    }

+    #[inline]
+    pub(crate) fn has_remaining_docs(&self) -> bool {
+        self.skip_reader.has_remaining_docs()
+    }
+
    pub(crate) fn block_is_loaded(&self) -> bool {
        self.block_loaded
    }
--- a/src/postings/skip.rs
+++ b/src/postings/skip.rs
@@ -146,6 +146,11 @@ impl SkipReader {
        skip_reader
    }

+    #[inline(always)]
+    pub fn has_remaining_docs(&self) -> bool {
+        self.remaining_docs != 0
+    }
+
    pub fn reset(&mut self, data: OwnedBytes, doc_freq: u32) {
        self.last_doc_in_block = if doc_freq >= COMPRESSION_BLOCK_SIZE as u32 {
            0
--- a/src/query/boolean_query/block_wand_intersection.rs
+++ b/src/query/boolean_query/block_wand_intersection.rs
@@ -0,0 +1,464 @@
+use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
+use crate::query::term_query::TermScorer;
+use crate::query::Scorer;
+use crate::{DocId, DocSet, Score, TERMINATED};
+
+/// Block-max pruning for top-K over intersection of term scorers.
+///
+/// Uses the least-frequent term as "leader" to define 128-doc processing windows.
+/// For each window, the sum of block_max_scores is compared to the current threshold;
+/// if the block can't beat it, the entire block is skipped.
+///
+/// Within non-skipped blocks, individual documents are pruned by checking whether
+/// leader_score + sum(secondary block_max_scores) can exceed the threshold before
+/// performing the expensive intersection membership check (seeking into secondary scorers).
+///
+/// # Preconditions
+/// - `scorers` has at least 2 elements
+/// - All scorers read frequencies (`FreqReadingOption::ReadFreq`)
+pub(crate) fn block_wand_intersection(
+    mut scorers: Vec<TermScorer>,
+    mut threshold: Score,
+    callback: &mut dyn FnMut(DocId, Score) -> Score,
+) {
+    assert!(scorers.len() >= 2);
+
+    // Sort by cost (ascending). scorers[0] becomes the "leader" (rarest term).
+    scorers.sort_by_key(TermScorer::size_hint);
+
+    let (leader, secondaries) = scorers.split_first_mut().unwrap();
+
+    // Precompute global max scores for early termination checks.
+    let leader_max_score: Score = leader.max_score();
+    let secondaries_global_max_sum: Score = secondaries.iter().map(TermScorer::max_score).sum();
+
+    // Early exit: no document can possibly beat the threshold.
+    if leader_max_score + secondaries_global_max_sum <= threshold {
+        return;
+    }
+
+    // Borrow fieldnorm reader and BM25 weight before the main loop.
+    // These are immutable references to disjoint fields from block_cursor,
+    // but Rust's borrow checker can't see through method calls, so we
+    // extract them once upfront.
+    let fieldnorm_reader = leader.fieldnorm_reader().clone();
+    let bm25_weight = leader.bm25_weight().clone();
+
+    let mut doc = leader.doc();
+
+    let mut secondary_block_max_scores: Box<[f32]> =
+        vec![0.0f32; secondaries.len()].into_boxed_slice();
+    let mut secondary_suffix_block_max: Box<[f32]> =
+        vec![0.0f32; secondaries.len()].into_boxed_slice();
+
+    while doc < TERMINATED {
+        // --- Phase 1: Block-level pruning ---
+        //
+        // Position all skip readers on the block containing `doc`.
+        // seek_block is cheap: it only advances the skip reader, no block decompression.
+        leader.seek_block(doc);
+        let leader_block_max: Score = leader.block_max_score();
+
+        // Compute the window end as the minimum last_doc_in_block across all scorers.
+        // This ensures the block_max values are valid for all docs in [doc, window_end].
+        // Different scorers have independently aligned blocks, so we must use the
+        // smallest window where all block_max values hold.
+        let mut window_end: DocId = leader.last_doc_in_block();
+
+        let mut secondary_block_max_sum: Score = 0.0;
+        let num_secondaries = secondaries.len();
+        for (idx, secondary) in secondaries.iter_mut().enumerate() {
+            secondary.block_cursor().seek_block(doc);
+            if !secondary.block_cursor().has_remaining_docs() {
+                return;
+            }
+            window_end = window_end.min(secondary.last_doc_in_block());
+            let bms = secondary.block_max_score();
+            secondary_block_max_scores[idx] = bms;
+            secondary_block_max_sum += bms;
+        }
+
+        if leader_block_max + secondary_block_max_sum <= threshold {
+            // The entire window cannot beat the threshold. Skip past it.
+            doc = window_end + 1;
+            continue;
+        }
+
+        // --- Phase 2: Batch processing within the window ---
+        //
+        // Score-first approach: decode the leader's block, filter by threshold,
+        // then check intersection membership only for survivors. This avoids expensive
+        // secondary seeks for docs that can't beat the threshold.
+        let block_cursor = leader.block_cursor();
+        // seek loads the block and returns the in-block index of the first doc >= `doc`.
+        let start_idx = block_cursor.seek(doc);
+
+        // Use the branchless binary search on the doc decoder to find the first
+        // index past window_end.
+        let end_idx = block_cursor
+            .doc_decoder
+            .seek_within_block(window_end + 1)
+            .min(block_cursor.block_len());
+
+        let block_docs = &block_cursor.doc_decoder.output_array()[start_idx..end_idx];
+        let block_freqs = &block_cursor.freq_output_array()[start_idx..end_idx];
+
+        // Pass 1: Batch-compute leader BM25 scores and branchlessly filter
+        // candidates that can't beat the threshold.
+        //
+        // The trick: always write to the buffer at `num_candidates`, then
+        // conditionally advance the count. The compiler can turn this into
+        // a cmov instead of a branch, avoiding misprediction costs.
+        let score_threshold = threshold - secondary_block_max_sum;
+        let mut candidate_doc_ids = [0u32; COMPRESSION_BLOCK_SIZE];
+        let mut candidate_scores = [0.0f32; COMPRESSION_BLOCK_SIZE];
+        let mut num_candidates = 0usize;
+
+        for (candidate_doc, term_freq) in
+            block_docs.iter().copied().zip(block_freqs.iter().copied())
+        {
+            let fieldnorm_id = fieldnorm_reader.fieldnorm_id(candidate_doc);
+            let leader_score = bm25_weight.score(fieldnorm_id, term_freq);
+            candidate_doc_ids[num_candidates] = candidate_doc;
+            candidate_scores[num_candidates] = leader_score;
+            num_candidates += (leader_score > score_threshold) as usize;
+        }
+
+        // Precompute suffix sums: suffix[i] = sum of block_max for secondaries[i+1..].
+        // Used in Phase 2 to prune candidates that can't beat threshold even with
+        // remaining secondaries contributing their block_max.
+        if num_candidates == 0 {
+            doc = window_end + 1;
+            continue;
+        }
+
+        let mut running = 0.0f32;
+        for idx in (0..num_secondaries).rev() {
+            secondary_suffix_block_max[idx] = running;
+            running += secondary_block_max_scores[idx];
+        }
+
+        // Pass 2: Check intersection membership only for survivors.
+        // score_threshold may be stale (threshold can increase from callbacks),
+        // but that's conservative — we may check a few extra candidates, never miss one.
+        'next_candidate: for candidate_idx in 0..num_candidates {
+            let candidate_doc = candidate_doc_ids[candidate_idx];
+            let mut total_score: Score = candidate_scores[candidate_idx];
+
+            for (secondary_idx, secondary) in secondaries.iter_mut().enumerate() {
+                // If a previous candidate already advanced this secondary past
+                // candidate_doc, the candidate can't be in the intersection.
+                if secondary.doc() > candidate_doc {
+                    continue 'next_candidate;
+                }
+                let seek_result = secondary.seek(candidate_doc);
+                if seek_result != candidate_doc {
+                    continue 'next_candidate;
+                }
+                total_score += secondary.score();
+
+                // Prune: even if all remaining secondaries score at their block max,
+                // can we still beat the threshold?
+                if total_score + secondary_suffix_block_max[secondary_idx] <= threshold {
+                    continue 'next_candidate;
+                }
+            }
+
+            // All secondaries matched.
+            if total_score > threshold {
+                threshold = callback(candidate_doc, total_score);
+
+                if leader_max_score + secondaries_global_max_sum <= threshold {
+                    return;
+                }
+            }
+        }
+
+        doc = window_end + 1;
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::cmp::Ordering;
+    use std::collections::BinaryHeap;
+
+    use proptest::prelude::*;
+
+    use crate::query::term_query::TermScorer;
+    use crate::query::{Bm25Weight, Scorer};
+    use crate::{DocId, DocSet, Score, TERMINATED};
+
+    struct Float(Score);
+
+    impl Eq for Float {}
+
+    impl PartialEq for Float {
+        fn eq(&self, other: &Self) -> bool {
+            self.cmp(other) == Ordering::Equal
+        }
+    }
+
+    impl PartialOrd for Float {
+        fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+            Some(self.cmp(other))
+        }
+    }
+
+    impl Ord for Float {
+        fn cmp(&self, other: &Self) -> Ordering {
+            other.0.partial_cmp(&self.0).unwrap_or(Ordering::Equal)
+        }
+    }
+
+    fn nearly_equals(left: Score, right: Score) -> bool {
+        (left - right).abs() < 0.0001 * (left + right).abs()
+    }
+
+    /// Run block_wand_intersection and collect (doc, score) pairs above threshold.
+    fn compute_checkpoints_block_wand_intersection(
+        term_scorers: Vec<TermScorer>,
+        top_k: usize,
+    ) -> Vec<(DocId, Score)> {
+        let mut heap: BinaryHeap<Float> = BinaryHeap::with_capacity(top_k);
+        let mut checkpoints: Vec<(DocId, Score)> = Vec::new();
+        let mut limit: Score = 0.0;
+
+        let callback = &mut |doc, score| {
+            heap.push(Float(score));
+            if heap.len() > top_k {
+                heap.pop().unwrap();
+            }
+            if heap.len() == top_k {
+                limit = heap.peek().unwrap().0;
+            }
+            if !nearly_equals(score, limit) {
+                checkpoints.push((doc, score));
+            }
+            limit
+        };
+
+        super::block_wand_intersection(term_scorers, Score::MIN, callback);
+        checkpoints
+    }
+
+    /// Naive baseline: intersect by iterating all docs.
+    fn compute_checkpoints_naive_intersection(
+        mut term_scorers: Vec<TermScorer>,
+        top_k: usize,
+    ) -> Vec<(DocId, Score)> {
+        let mut heap: BinaryHeap<Float> = BinaryHeap::with_capacity(top_k);
+        let mut checkpoints: Vec<(DocId, Score)> = Vec::new();
+        let mut limit = Score::MIN;
+
+        // Sort by cost to use the cheapest as driver.
+        term_scorers.sort_by_key(|s| s.cost());
+
+        let (leader, secondaries) = term_scorers.split_first_mut().unwrap();
+
+        let mut doc = leader.doc();
+        while doc != TERMINATED {
+            let mut all_match = true;
+            for secondary in secondaries.iter_mut() {
+                let secondary_doc = secondary.doc();
+                let seek_result = if secondary_doc <= doc {
+                    secondary.seek(doc)
+                } else {
+                    secondary_doc
+                };
+                if seek_result != doc {
+                    all_match = false;
+                    break;
+                }
+            }
+
+            if all_match {
+                let score: Score =
+                    leader.score() + secondaries.iter_mut().map(|s| s.score()).sum::<Score>();
+
+                if score > limit {
+                    heap.push(Float(score));
+                    if heap.len() > top_k {
+                        heap.pop().unwrap();
+                    }
+                    if heap.len() == top_k {
+                        limit = heap.peek().unwrap().0;
+                    }
+                    if !nearly_equals(score, limit) {
+                        checkpoints.push((doc, score));
+                    }
+                }
+            }
+            doc = leader.advance();
+        }
+        checkpoints
+    }
+
+    const MAX_TERM_FREQ: u32 = 100u32;
+
+    fn posting_list(max_doc: u32) -> BoxedStrategy<Vec<(DocId, u32)>> {
+        (1..max_doc + 1)
+            .prop_flat_map(move |doc_freq| {
+                (
+                    proptest::bits::bitset::sampled(doc_freq as usize, 0..max_doc as usize),
+                    proptest::collection::vec(1u32..MAX_TERM_FREQ, doc_freq as usize),
+                )
+            })
+            .prop_map(|(docset, term_freqs)| {
+                docset
+                    .iter()
+                    .map(|doc| doc as u32)
+                    .zip(term_freqs.iter().cloned())
+                    .collect::<Vec<_>>()
+            })
+            .boxed()
+    }
+
+    #[expect(clippy::type_complexity)]
+    fn gen_term_scorers(num_scorers: usize) -> BoxedStrategy<(Vec<Vec<(DocId, u32)>>, Vec<u32>)> {
+        (1u32..100u32)
+            .prop_flat_map(move |max_doc: u32| {
+                (
+                    proptest::collection::vec(posting_list(max_doc), num_scorers),
+                    proptest::collection::vec(2u32..10u32 * MAX_TERM_FREQ, max_doc as usize),
+                )
+            })
+            .boxed()
+    }
+
+    fn test_block_wand_intersection_aux(posting_lists: &[Vec<(DocId, u32)>], fieldnorms: &[u32]) {
+        // Repeat docs 64 times to create multi-block scenarios, matching block_wand.rs test
+        // strategy.
+        const REPEAT: usize = 64;
+        let fieldnorms_expanded: Vec<u32> = fieldnorms
+            .iter()
+            .cloned()
+            .flat_map(|fieldnorm| std::iter::repeat_n(fieldnorm, REPEAT))
+            .collect();
+
+        let postings_lists_expanded: Vec<Vec<(DocId, u32)>> = posting_lists
+            .iter()
+            .map(|posting_list| {
+                posting_list
+                    .iter()
+                    .cloned()
+                    .flat_map(|(doc, term_freq)| {
+                        (0_u32..REPEAT as u32).map(move |offset| {
+                            (
+                                doc * (REPEAT as u32) + offset,
+                                if offset == 0 { term_freq } else { 1 },
+                            )
+                        })
+                    })
+                    .collect::<Vec<(DocId, u32)>>()
+            })
+            .collect();
+
+        let total_fieldnorms: u64 = fieldnorms_expanded
+            .iter()
+            .cloned()
+            .map(|fieldnorm| fieldnorm as u64)
+            .sum();
+        let average_fieldnorm = (total_fieldnorms as Score) / (fieldnorms_expanded.len() as Score);
+        let max_doc = fieldnorms_expanded.len();
+
+        let make_scorers = || -> Vec<TermScorer> {
+            postings_lists_expanded
+                .iter()
+                .map(|postings| {
+                    let bm25_weight = Bm25Weight::for_one_term(
+                        postings.len() as u64,
+                        max_doc as u64,
+                        average_fieldnorm,
+                    );
+                    TermScorer::create_for_test(postings, &fieldnorms_expanded[..], bm25_weight)
+                })
+                .collect()
+        };
+
+        for top_k in 1..4 {
+            let checkpoints_optimized =
+                compute_checkpoints_block_wand_intersection(make_scorers(), top_k);
+            let checkpoints_naive = compute_checkpoints_naive_intersection(make_scorers(), top_k);
+            assert_eq!(
+                checkpoints_optimized.len(),
+                checkpoints_naive.len(),
+                "Mismatch in checkpoint count for top_k={top_k}"
+            );
+            for (&(left_doc, left_score), &(right_doc, right_score)) in
+                checkpoints_optimized.iter().zip(checkpoints_naive.iter())
+            {
+                assert_eq!(left_doc, right_doc);
+                assert!(
+                    nearly_equals(left_score, right_score),
+                    "Score mismatch for doc {left_doc}: {left_score} vs {right_score}"
+                );
+            }
+        }
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(500))]
+        #[test]
+        fn test_block_wand_intersection_two_scorers(
+            (posting_lists, fieldnorms) in gen_term_scorers(2)
+        ) {
+            test_block_wand_intersection_aux(&posting_lists[..], &fieldnorms[..]);
+        }
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(500))]
+        #[test]
+        fn test_block_wand_intersection_three_scorers(
+            (posting_lists, fieldnorms) in gen_term_scorers(3)
+        ) {
+            test_block_wand_intersection_aux(&posting_lists[..], &fieldnorms[..]);
+        }
+    }
+
+    #[test]
+    fn test_block_wand_intersection_disjoint() {
+        // Two posting lists with no overlap — intersection is empty.
+        let fieldnorms: Vec<u32> = vec![10; 200];
+        let average_fieldnorm = 10.0;
+        let postings_a: Vec<(DocId, u32)> = (0..100).map(|d| (d, 1)).collect();
+        let postings_b: Vec<(DocId, u32)> = (100..200).map(|d| (d, 1)).collect();
+
+        let scorer_a = TermScorer::create_for_test(
+            &postings_a,
+            &fieldnorms,
+            Bm25Weight::for_one_term(100, 200, average_fieldnorm),
+        );
+        let scorer_b = TermScorer::create_for_test(
+            &postings_b,
+            &fieldnorms,
+            Bm25Weight::for_one_term(100, 200, average_fieldnorm),
+        );
+
+        let checkpoints = compute_checkpoints_block_wand_intersection(vec![scorer_a, scorer_b], 10);
+        assert!(checkpoints.is_empty());
+    }
+
+    #[test]
+    fn test_block_wand_intersection_all_overlap() {
+        // Two posting lists with full overlap.
+        let fieldnorms: Vec<u32> = vec![10; 50];
+        let average_fieldnorm = 10.0;
+        let postings: Vec<(DocId, u32)> = (0..50).map(|d| (d, 3)).collect();
+
+        let make_scorer = || {
+            TermScorer::create_for_test(
+                &postings,
+                &fieldnorms,
+                Bm25Weight::for_one_term(50, 50, average_fieldnorm),
+            )
+        };
+
+        let checkpoints_opt =
+            compute_checkpoints_block_wand_intersection(vec![make_scorer(), make_scorer()], 5);
+        let checkpoints_naive =
+            compute_checkpoints_naive_intersection(vec![make_scorer(), make_scorer()], 5);
+        assert_eq!(checkpoints_opt.len(), checkpoints_naive.len());
+    }
+}
--- a/src/query/boolean_query/block_wand_union.rs
+++ b/src/query/boolean_query/block_wand_union.rs
@@ -50,7 +50,7 @@ fn block_max_was_too_low_advance_one_scorer(
    scorers: &mut [TermScorerWithMaxScore],
    pivot_len: usize,
 ) {
-    debug_assert!(is_sorted(scorers.iter().map(|scorer| scorer.doc())));
+    debug_assert!(scorers.iter().map(|scorer| scorer.doc()).is_sorted());
    let mut scorer_to_seek = pivot_len - 1;
    let mut global_max_score = scorers[scorer_to_seek].max_score;
    let mut doc_to_seek_after = scorers[scorer_to_seek].last_doc_in_block();
@@ -76,7 +76,7 @@ fn block_max_was_too_low_advance_one_scorer(
    scorers[scorer_to_seek].seek(doc_to_seek_after);

    restore_ordering(scorers, scorer_to_seek);
-    debug_assert!(is_sorted(scorers.iter().map(|scorer| scorer.doc())));
+    debug_assert!(scorers.iter().map(|scorer| scorer.doc()).is_sorted());
 }

 // Given a list of term_scorers and a `ord` and assuming that `term_scorers[ord]` is sorted
@@ -90,7 +90,7 @@ fn restore_ordering(term_scorers: &mut [TermScorerWithMaxScore], ord: usize) {
        }
        term_scorers.swap(i, i - 1);
    }
-    debug_assert!(is_sorted(term_scorers.iter().map(|scorer| scorer.doc())));
+    debug_assert!(term_scorers.iter().map(|scorer| scorer.doc()).is_sorted());
 }

 // Attempts to advance all term_scorers between `&term_scorers[0..before_len]` to the pivot.
@@ -150,17 +150,21 @@ pub fn block_wand(
    mut threshold: Score,
    callback: &mut dyn FnMut(u32, Score) -> Score,
 ) {
+    scorers.retain(|scorer| scorer.doc() < TERMINATED);
+    if scorers.len() == 1 {
+        let scorer = scorers.pop().unwrap();
+        return block_wand_single_scorer(scorer, threshold, callback);
+    }
    let mut scorers: Vec<TermScorerWithMaxScore> = scorers
        .iter_mut()
        .map(TermScorerWithMaxScore::from)
        .collect();
-    scorers.sort_by_key(|scorer| scorer.doc());
    // At this point we need to ensure that the scorers are sorted!
-    debug_assert!(is_sorted(scorers.iter().map(|scorer| scorer.doc())));
+    scorers.sort_by_key(|scorer| scorer.doc());
    while let Some((before_pivot_len, pivot_len, pivot_doc)) =
        find_pivot_doc(&scorers[..], threshold)
    {
-        debug_assert!(is_sorted(scorers.iter().map(|scorer| scorer.doc())));
+        debug_assert!(scorers.iter().map(|scorer| scorer.doc()).is_sorted());
        debug_assert_ne!(pivot_doc, TERMINATED);
        debug_assert!(before_pivot_len < pivot_len);

@@ -228,7 +232,7 @@ pub fn block_wand_single_scorer(
    loop {
        // We position the scorer on a block that can reach
        // the threshold.
-        while scorer.block_max_score() < threshold {
+        while scorer.block_max_score() <= threshold {
            let last_doc_in_block = scorer.last_doc_in_block();
            if last_doc_in_block == TERMINATED {
                return;
@@ -286,18 +290,6 @@ impl DerefMut for TermScorerWithMaxScore<'_> {
    }
 }

-fn is_sorted<I: Iterator<Item = DocId>>(mut it: I) -> bool {
-    if let Some(first) = it.next() {
-        let mut prev = first;
-        for doc in it {
-            if doc < prev {
-                return false;
-            }
-            prev = doc;
-        }
-    }
-    true
-}
 #[cfg(test)]
 mod tests {
    use std::cmp::Ordering;
--- a/src/query/boolean_query/boolean_weight.rs
+++ b/src/query/boolean_query/boolean_weight.rs
@@ -16,6 +16,7 @@ use crate::{DocId, Score};

 enum SpecializedScorer {
    TermUnion(Vec<TermScorer>),
+    TermIntersection(Vec<TermScorer>),
    Other(Box<dyn Scorer>),
 }

@@ -49,10 +50,9 @@ where
    TScoreCombiner: ScoreCombiner,
 {
    assert!(!scorers.is_empty());
-    if scorers.len() == 1 {
+    if scorers.len() == 1 && !scorers[0].is::<TermScorer>() {
        return SpecializedScorer::Other(scorers.into_iter().next().unwrap()); //< we checked the size beforehand
    }
-
    {
        let is_all_term_queries = scorers.iter().all(|scorer| scorer.is::<TermScorer>());
        if is_all_term_queries {
@@ -66,6 +66,9 @@ where
            {
                // Block wand is only available if we read frequencies.
                return SpecializedScorer::TermUnion(scorers);
+            } else if scorers.len() == 1 {
+                // Single TermScorer without freq reading — unwrap directly.
+                return SpecializedScorer::Other(Box::new(scorers.into_iter().next().unwrap()));
            } else {
                return SpecializedScorer::Other(Box::new(BufferedUnionScorer::build(
                    scorers,
@@ -93,6 +96,13 @@ fn into_box_scorer<TScoreCombiner: ScoreCombiner>(
                BufferedUnionScorer::build(term_scorers, score_combiner_fn, num_docs);
            Box::new(union_scorer)
        }
+        SpecializedScorer::TermIntersection(term_scorers) => {
+            let boxed_scorers: Vec<Box<dyn Scorer>> = term_scorers
+                .into_iter()
+                .map(|s| Box::new(s) as Box<dyn Scorer>)
+                .collect();
+            intersect_scorers(boxed_scorers, num_docs)
+        }
        SpecializedScorer::Other(scorer) => scorer,
    }
 }
@@ -297,14 +307,43 @@ impl<TScoreCombiner: ScoreCombiner> BooleanWeight<TScoreCombiner> {
                // Result depends entirely on MUST + any removed AllScorers.
                let combined_all_scorer_count = must_special_scorer_counts.num_all_scorers
                    + should_special_scorer_counts.num_all_scorers;
-                let boxed_scorer: Box<dyn Scorer> = effective_must_scorer(
-                    must_scorers,
-                    combined_all_scorer_count,
-                    reader.max_doc(),
-                    num_docs,
-                )
-                .unwrap_or_else(|| Box::new(EmptyScorer));
-                SpecializedScorer::Other(boxed_scorer)
+
+                // Try to detect a pure TermScorer intersection for block-max optimization.
+                // Preconditions: no removed AllScorers, at least 2 scorers, all TermScorer
+                // with frequency reading enabled.
+                if combined_all_scorer_count == 0
+                    && must_scorers.len() >= 2
+                    && must_scorers.iter().all(|s| s.is::<TermScorer>())
+                {
+                    let term_scorers: Vec<TermScorer> = must_scorers
+                        .into_iter()
+                        .map(|s| *(s.downcast::<TermScorer>().map_err(|_| ()).unwrap()))
+                        .collect();
+                    if term_scorers
+                        .iter()
+                        .all(|s| s.freq_reading_option() == FreqReadingOption::ReadFreq)
+                    {
+                        SpecializedScorer::TermIntersection(term_scorers)
+                    } else {
+                        let must_scorers: Vec<Box<dyn Scorer>> = term_scorers
+                            .into_iter()
+                            .map(|s| Box::new(s) as Box<dyn Scorer>)
+                            .collect();
+                        let boxed_scorer: Box<dyn Scorer> =
+                            effective_must_scorer(must_scorers, 0, reader.max_doc(), num_docs)
+                                .unwrap_or_else(|| Box::new(EmptyScorer));
+                        SpecializedScorer::Other(boxed_scorer)
+                    }
+                } else {
+                    let boxed_scorer: Box<dyn Scorer> = effective_must_scorer(
+                        must_scorers,
+                        combined_all_scorer_count,
+                        reader.max_doc(),
+                        num_docs,
+                    )
+                    .unwrap_or_else(|| Box::new(EmptyScorer));
+                    SpecializedScorer::Other(boxed_scorer)
+                }
            }
            (ShouldScorersCombinationMethod::Optional(should_scorer), must_scorers) => {
                // Optional SHOULD: contributes to scoring but not required for matching.
@@ -463,15 +502,21 @@ impl<TScoreCombiner: ScoreCombiner + Sync> Weight for BooleanWeight<TScoreCombin
        callback: &mut dyn FnMut(DocId, Score),
    ) -> crate::Result<()> {
        let scorer = self.complex_scorer(reader, 1.0, &self.score_combiner_fn)?;
+        let num_docs = reader.num_docs();
        match scorer {
            SpecializedScorer::TermUnion(term_scorers) => {
-                let mut union_scorer = BufferedUnionScorer::build(
-                    term_scorers,
-                    &self.score_combiner_fn,
-                    reader.num_docs(),
-                );
+                let mut union_scorer =
+                    BufferedUnionScorer::build(term_scorers, &self.score_combiner_fn, num_docs);
                for_each_scorer(&mut union_scorer, callback);
            }
+            SpecializedScorer::TermIntersection(term_scorers) => {
+                let boxed_scorers: Vec<Box<dyn Scorer>> = term_scorers
+                    .into_iter()
+                    .map(|term_scorer| Box::new(term_scorer) as Box<dyn Scorer>)
+                    .collect();
+                let mut intersection = intersect_scorers(boxed_scorers, num_docs);
+                for_each_scorer(intersection.as_mut(), callback);
+            }
            SpecializedScorer::Other(mut scorer) => {
                for_each_scorer(scorer.as_mut(), callback);
            }
@@ -485,17 +530,23 @@ impl<TScoreCombiner: ScoreCombiner + Sync> Weight for BooleanWeight<TScoreCombin
        callback: &mut dyn FnMut(&[DocId]),
    ) -> crate::Result<()> {
        let scorer = self.complex_scorer(reader, 1.0, || DoNothingCombiner)?;
+        let num_docs = reader.num_docs();
        let mut buffer = [0u32; COLLECT_BLOCK_BUFFER_LEN];

        match scorer {
            SpecializedScorer::TermUnion(term_scorers) => {
-                let mut union_scorer = BufferedUnionScorer::build(
-                    term_scorers,
-                    &self.score_combiner_fn,
-                    reader.num_docs(),
-                );
+                let mut union_scorer =
+                    BufferedUnionScorer::build(term_scorers, &self.score_combiner_fn, num_docs);
                for_each_docset_buffered(&mut union_scorer, &mut buffer, callback);
            }
+            SpecializedScorer::TermIntersection(term_scorers) => {
+                let boxed_scorers: Vec<Box<dyn Scorer>> = term_scorers
+                    .into_iter()
+                    .map(|term_scorer| Box::new(term_scorer) as Box<dyn Scorer>)
+                    .collect();
+                let mut intersection = intersect_scorers(boxed_scorers, num_docs);
+                for_each_docset_buffered(intersection.as_mut(), &mut buffer, callback);
+            }
            SpecializedScorer::Other(mut scorer) => {
                for_each_docset_buffered(scorer.as_mut(), &mut buffer, callback);
            }
@@ -524,6 +575,9 @@ impl<TScoreCombiner: ScoreCombiner + Sync> Weight for BooleanWeight<TScoreCombin
            SpecializedScorer::TermUnion(term_scorers) => {
                super::block_wand(term_scorers, threshold, callback);
            }
+            SpecializedScorer::TermIntersection(term_scorers) => {
+                super::block_wand_intersection(term_scorers, threshold, callback);
+            }
            SpecializedScorer::Other(mut scorer) => {
                for_each_pruning_scorer(scorer.as_mut(), threshold, callback);
            }
--- a/src/query/boolean_query/mod.rs
+++ b/src/query/boolean_query/mod.rs
@@ -1,8 +1,10 @@
-mod block_wand;
+mod block_wand_intersection;
+mod block_wand_union;
 mod boolean_query;
 mod boolean_weight;

-pub(crate) use self::block_wand::{block_wand, block_wand_single_scorer};
+pub(crate) use self::block_wand_intersection::block_wand_intersection;
+pub(crate) use self::block_wand_union::{block_wand, block_wand_single_scorer};
 pub use self::boolean_query::BooleanQuery;
 pub use self::boolean_weight::BooleanWeight;

--- a/src/query/intersection.rs
+++ b/src/query/intersection.rs
@@ -1,5 +1,7 @@
+use common::TinySet;
+
 use super::size_hint::estimate_intersection;
-use crate::docset::{DocSet, SeekDangerResult, TERMINATED};
+use crate::docset::{DocSet, SeekDangerResult, BLOCK_NUM_TINYBITSETS, TERMINATED};
 use crate::query::term_query::TermScorer;
 use crate::query::{EmptyScorer, Scorer};
 use crate::{DocId, Score};
@@ -17,7 +19,7 @@ use crate::{DocId, Score};
 /// `size_hint` of the intersection.
 pub fn intersect_scorers(
    mut scorers: Vec<Box<dyn Scorer>>,
-    num_docs_segment: u32,
+    segment_num_docs: u32,
 ) -> Box<dyn Scorer> {
    if scorers.is_empty() {
        return Box::new(EmptyScorer);
@@ -42,14 +44,14 @@ pub fn intersect_scorers(
            left: *(left.downcast::<TermScorer>().map_err(|_| ()).unwrap()),
            right: *(right.downcast::<TermScorer>().map_err(|_| ()).unwrap()),
            others: scorers,
-            num_docs: num_docs_segment,
+            segment_num_docs,
        });
    }
    Box::new(Intersection {
        left,
        right,
        others: scorers,
-        num_docs: num_docs_segment,
+        segment_num_docs,
    })
 }

@@ -58,7 +60,7 @@ pub struct Intersection<TDocSet: DocSet, TOtherDocSet: DocSet = Box<dyn Scorer>>
    left: TDocSet,
    right: TDocSet,
    others: Vec<TOtherDocSet>,
-    num_docs: u32,
+    segment_num_docs: u32,
 }

 fn go_to_first_doc<TDocSet: DocSet>(docsets: &mut [TDocSet]) -> DocId {
@@ -78,7 +80,10 @@ fn go_to_first_doc<TDocSet: DocSet>(docsets: &mut [TDocSet]) -> DocId {

 impl<TDocSet: DocSet> Intersection<TDocSet, TDocSet> {
    /// num_docs is the number of documents in the segment.
-    pub(crate) fn new(mut docsets: Vec<TDocSet>, num_docs: u32) -> Intersection<TDocSet, TDocSet> {
+    pub(crate) fn new(
+        mut docsets: Vec<TDocSet>,
+        segment_num_docs: u32,
+    ) -> Intersection<TDocSet, TDocSet> {
        let num_docsets = docsets.len();
        assert!(num_docsets >= 2);
        docsets.sort_by_key(|docset| docset.cost());
@@ -97,7 +102,7 @@ impl<TDocSet: DocSet> Intersection<TDocSet, TDocSet> {
            left,
            right,
            others: docsets,
-            num_docs,
+            segment_num_docs,
        }
    }
 }
@@ -214,7 +219,7 @@ impl<TDocSet: DocSet, TOtherDocSet: DocSet> DocSet for Intersection<TDocSet, TOt
            [self.left.size_hint(), self.right.size_hint()]
                .into_iter()
                .chain(self.others.iter().map(DocSet::size_hint)),
-            self.num_docs,
+            self.segment_num_docs,
        )
    }

@@ -224,6 +229,91 @@ impl<TDocSet: DocSet, TOtherDocSet: DocSet> DocSet for Intersection<TDocSet, TOt
        // If there are docsets that are bad at skipping, they should also influence the cost.
        self.left.cost()
    }
+
+    fn count_including_deleted(&mut self) -> u32 {
+        const DENSITY_THRESHOLD_INVERSE: u32 = 32;
+        if self
+            .left
+            .size_hint()
+            .saturating_mul(DENSITY_THRESHOLD_INVERSE)
+            < self.segment_num_docs
+        {
+            // Sparse path: if the lead iterator covers less than ~3% of docs,
+            // the block approach wastes time on mostly-empty blocks.
+            self.count_including_deleted_sparse()
+        } else {
+            // Dense approach. We push documents into a block bitset to then
+            // perform count using popcount.
+            self.count_including_deleted_dense()
+        }
+    }
+}
+
+const EMPTY_BLOCK: [TinySet; BLOCK_NUM_TINYBITSETS] = [TinySet::EMPTY; BLOCK_NUM_TINYBITSETS];
+
+/// ANDs `other` into `mask` in-place. Returns `true` if the result is all zeros.
+#[inline]
+fn and_blocks_and_return_is_empty(
+    mask: &mut [TinySet; BLOCK_NUM_TINYBITSETS],
+    update: &[TinySet; BLOCK_NUM_TINYBITSETS],
+) -> bool {
+    let mut all_empty = true;
+    for (mask_tinyset, update_tinyset) in mask.iter_mut().zip(update.iter()) {
+        *mask_tinyset = mask_tinyset.intersect(*update_tinyset);
+        all_empty &= mask_tinyset.is_empty();
+    }
+    all_empty
+}
+
+impl<TDocSet: DocSet, TOtherDocSet: DocSet> Intersection<TDocSet, TOtherDocSet> {
+    fn count_including_deleted_sparse(&mut self) -> u32 {
+        let mut count = 0u32;
+        let mut doc = self.doc();
+        while doc != TERMINATED {
+            count += 1;
+            doc = self.advance();
+        }
+        count
+    }
+
+    /// Dense block-wise bitmask intersection count.
+    ///
+    /// Fills a 1024-doc window from each iterator, ANDs the bitmasks together,
+    /// and popcounts the result. `fill_bitset_block` handles seeking tails forward
+    /// when they lag behind the current block.
+    fn count_including_deleted_dense(&mut self) -> u32 {
+        let mut count = 0u32;
+        let mut next_base = self.left.doc();
+
+        while next_base < TERMINATED {
+            let base = next_base;
+
+            // Fill lead bitmask.
+            let mut mask = EMPTY_BLOCK;
+            next_base = next_base.max(self.left.fill_bitset_block(base, &mut mask));
+
+            let mut tail_mask = EMPTY_BLOCK;
+            next_base = next_base.max(self.right.fill_bitset_block(base, &mut tail_mask));
+
+            if and_blocks_and_return_is_empty(&mut mask, &tail_mask) {
+                continue;
+            }
+            // AND with each additional tail.
+            for other in &mut self.others {
+                let mut other_mask = EMPTY_BLOCK;
+                next_base = next_base.max(other.fill_bitset_block(base, &mut other_mask));
+                if and_blocks_and_return_is_empty(&mut mask, &other_mask) {
+                    continue;
+                }
+            }
+
+            for tinyset in &mask {
+                count += tinyset.len();
+            }
+        }
+
+        count
+    }
 }

 impl<TScorer, TOtherScorer> Scorer for Intersection<TScorer, TOtherScorer>
@@ -421,6 +511,82 @@ mod tests {
        }
    }

+    proptest! {
+        #[test]
+        fn prop_test_count_including_deleted_matches_default(
+            a in sorted_deduped_vec(1200, 400),
+            b in sorted_deduped_vec(1200, 400),
+            c in sorted_deduped_vec(1200, 400),
+            num_docs in 1200u32..2000u32,
+        ) {
+            // Compute expected count via set intersection.
+            let expected: u32 = a.iter()
+                .filter(|doc| b.contains(doc) && c.contains(doc))
+                .count() as u32;
+
+            // Test count_including_deleted (dense path).
+            let make_intersection = || {
+                Intersection::new(
+                    vec![
+                        VecDocSet::from(a.clone()),
+                        VecDocSet::from(b.clone()),
+                        VecDocSet::from(c.clone()),
+                    ],
+                    num_docs,
+                )
+            };
+
+            let mut intersection = make_intersection();
+            let count = intersection.count_including_deleted();
+            prop_assert_eq!(count, expected,
+                "count_including_deleted mismatch: a={:?}, b={:?}, c={:?}", a, b, c);
+        }
+    }
+
+    #[test]
+    fn test_count_including_deleted_two_way() {
+        let left = VecDocSet::from(vec![1, 3, 9]);
+        let right = VecDocSet::from(vec![3, 4, 9, 18]);
+        let mut intersection = Intersection::new(vec![left, right], 100);
+        assert_eq!(intersection.count_including_deleted(), 2);
+    }
+
+    #[test]
+    fn test_count_including_deleted_empty() {
+        let a = VecDocSet::from(vec![1, 3]);
+        let b = VecDocSet::from(vec![1, 4]);
+        let c = VecDocSet::from(vec![3, 9]);
+        let mut intersection = Intersection::new(vec![a, b, c], 100);
+        assert_eq!(intersection.count_including_deleted(), 0);
+    }
+
+    /// Test with enough documents to exercise the dense path (>= num_docs/32).
+    #[test]
+    fn test_count_including_deleted_dense_path() {
+        // Create dense docsets: many docs relative to segment size.
+        let docs_a: Vec<u32> = (0..2000).step_by(2).collect(); // even numbers 0..2000
+        let docs_b: Vec<u32> = (0..2000).step_by(3).collect(); // multiples of 3
+        let expected = docs_a.iter().filter(|d| *d % 3 == 0).count() as u32;
+
+        let a = VecDocSet::from(docs_a);
+        let b = VecDocSet::from(docs_b);
+        let mut intersection = Intersection::new(vec![a, b], 2000);
+        assert_eq!(intersection.count_including_deleted(), expected);
+    }
+
+    /// Test that spans multiple blocks (>1024 docs).
+    #[test]
+    fn test_count_including_deleted_multi_block() {
+        let docs_a: Vec<u32> = (0..5000).collect();
+        let docs_b: Vec<u32> = (0..5000).step_by(7).collect();
+        let expected = docs_b.len() as u32; // all of b is in a
+
+        let a = VecDocSet::from(docs_a);
+        let b = VecDocSet::from(docs_b);
+        let mut intersection = Intersection::new(vec![a, b], 5000);
+        assert_eq!(intersection.count_including_deleted(), expected);
+    }
+
    #[test]
    fn test_bug_2811_intersection_candidate_should_increase() {
        let mut schema_builder = Schema::builder();
--- a/src/query/term_query/term_scorer.rs
+++ b/src/query/term_query/term_scorer.rs
@@ -1,6 +1,6 @@
 use crate::docset::DocSet;
 use crate::fieldnorm::FieldNormReader;
-use crate::postings::{FreqReadingOption, Postings, SegmentPostings};
+use crate::postings::{BlockSegmentPostings, FreqReadingOption, Postings, SegmentPostings};
 use crate::query::bm25::Bm25Weight;
 use crate::query::{Explanation, Scorer};
 use crate::{DocId, Score};
@@ -95,6 +95,21 @@ impl TermScorer {
    pub fn last_doc_in_block(&self) -> DocId {
        self.postings.block_cursor.skip_reader().last_doc_in_block()
    }
+
+    /// Returns a mutable reference to the underlying block cursor.
+    pub(crate) fn block_cursor(&mut self) -> &mut BlockSegmentPostings {
+        &mut self.postings.block_cursor
+    }
+
+    /// Returns a reference to the fieldnorm reader for batch lookups.
+    pub(crate) fn fieldnorm_reader(&self) -> &FieldNormReader {
+        &self.fieldnorm_reader
+    }
+
+    /// Returns a reference to the BM25 weight for batch score computation.
+    pub(crate) fn bm25_weight(&self) -> &Bm25Weight {
+        &self.similarity_weight
+    }
 }

 impl DocSet for TermScorer {
@@ -117,6 +132,12 @@ impl DocSet for TermScorer {
    fn size_hint(&self) -> u32 {
        self.postings.size_hint()
    }
+
+    // TODO
+    // It is probably possible to optimize fill_bitset_block for TermScorer,
+    // working directly with the blocks, enabling vectorization.
+    // I did not manage to get a performance improvement on Mac ARM,
+    // and do not have access to x86 to investigate.
 }

 impl Scorer for TermScorer {
--- a/src/query/union/buffered_union.rs
+++ b/src/query/union/buffered_union.rs
@@ -1,6 +1,6 @@
 use common::TinySet;

-use crate::docset::{DocSet, SeekDangerResult, TERMINATED};
+use crate::docset::{DocSet, SeekDangerResult, COLLECT_BLOCK_BUFFER_LEN, TERMINATED};
 use crate::query::score_combiner::{DoNothingCombiner, ScoreCombiner};
 use crate::query::size_hint::estimate_union;
 use crate::query::Scorer;
@@ -172,6 +172,46 @@ where
        self.doc
    }

+    fn fill_buffer(&mut self, buffer: &mut [DocId; COLLECT_BLOCK_BUFFER_LEN]) -> usize {
+        if self.doc == TERMINATED {
+            return 0;
+        }
+        // The current doc (self.doc) has already been popped from the bitsets,
+        // so the loop below won't yield it. Emit it here first.
+        buffer[0] = self.doc;
+        let mut count = 1;
+
+        loop {
+            // Drain docs directly from the pre-computed bitsets.
+            while self.bucket_idx < HORIZON_NUM_TINYBITSETS {
+                // Move bitset to a local variable to avoid read/store on self.bitsets while
+                // iterating through the bits.
+                let mut tinyset: TinySet = self.bitsets[self.bucket_idx];
+
+                while let Some(val) = tinyset.pop_lowest() {
+                    let delta = val + (self.bucket_idx as u32) * 64;
+                    self.doc = self.window_start_doc + delta;
+
+                    if count >= COLLECT_BLOCK_BUFFER_LEN {
+                        // Buffer full; put remaining bits back.
+                        self.bitsets[self.bucket_idx] = tinyset;
+                        return COLLECT_BLOCK_BUFFER_LEN;
+                    }
+                    buffer[count] = self.doc;
+                    count += 1;
+                }
+                self.bitsets[self.bucket_idx] = TinySet::empty();
+                self.bucket_idx += 1;
+            }
+
+            // Current window exhausted, refill.
+            if !self.refill() {
+                self.doc = TERMINATED;
+                return count;
+            }
+        }
+    }
+
    fn seek(&mut self, target: DocId) -> DocId {
        if self.doc >= target {
            return self.doc;
--- a/src/termdict/fst_termdict/term_info_store.rs
+++ b/src/termdict/fst_termdict/term_info_store.rs
@@ -48,8 +48,7 @@ impl BinarySerializable for TermInfoBlockMeta {
 }

 impl FixedSize for TermInfoBlockMeta {
-    const SIZE_IN_BYTES: usize =
-        u64::SIZE_IN_BYTES + TermInfo::SIZE_IN_BYTES + 3 * u8::SIZE_IN_BYTES;
+    const SIZE_IN_BYTES: usize = u64::SIZE_IN_BYTES + TermInfo::SIZE_IN_BYTES + 3;
 }

 impl TermInfoBlockMeta {
--- a/sstable/Cargo.toml
+++ b/sstable/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-sstable"
-version = "0.6.0"
+version = "0.7.0"
 edition = "2024"
 license = "MIT"
 homepage = "https://github.com/quickwit-oss/tantivy"
@@ -10,10 +10,10 @@ categories = ["database-implementations", "data-structures", "compression"]
 description = "sstables for tantivy"

 [dependencies]
-common = {version= "0.10", path="../common", package="tantivy-common"}
+common = {version= "0.11", path="../common", package="tantivy-common"}
 futures-util = "0.3.30"
 itertools = "0.14.0"
-tantivy-bitpacker = { version= "0.9", path="../bitpacker" }
+tantivy-bitpacker = { version= "0.10", path="../bitpacker" }
 tantivy-fst = "0.5"
 # experimental gives us access to Decompressor::upper_bound
 zstd = { version = "0.13", optional = true, features = ["experimental"] }
@@ -23,7 +23,7 @@ zstd-compression = ["zstd"]

 [dev-dependencies]
 proptest = "1"
-criterion = { version = "0.5", default-features = false }
+criterion = { version = "0.8", default-features = false }
 names = "0.14"
 rand = "0.9"

--- a/sstable/src/dictionary.rs
+++ b/sstable/src/dictionary.rs
@@ -512,11 +512,13 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
    /// Returns the terms for a _sorted_ list of term ordinals.
    ///
    /// Returns true if and only if all terms have been found.
-    pub fn sorted_ords_to_term_cb<F: FnMut(&[u8]) -> io::Result<()>>(
+    pub fn sorted_ords_to_term_cb(
        &self,
-        mut ords: impl Iterator<Item = TermOrdinal>,
-        mut cb: F,
+        ords: &[TermOrdinal],
+        mut cb: impl FnMut(&[u8]),
    ) -> io::Result<bool> {
+        assert!(ords.is_sorted());
+        let mut ords = ords.iter().copied();
        let Some(mut ord) = ords.next() else {
            return Ok(true);
        };
@@ -538,33 +540,36 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
                bytes.extend_from_slice(current_sstable_delta_reader.suffix());
                current_block_ordinal += 1;
            }
-            cb(&bytes)?;
+            cb(&bytes);

            // fetch the next ordinal
-            let Some(next_ord) = ords.next() else {
-                return Ok(true);
+            let next_ord = loop {
+                let Some(next_ord) = ords.next() else {
+                    return Ok(true);
+                };
+                if next_ord == ord {
+                    // This is the same ordinal, let's just call the callback directly.
+                    cb(&bytes);
+                } else {
+                    // we checked it was sorted beforehands
+                    debug_assert!(next_ord > ord);
+                    break next_ord;
+                }
            };

-            // advance forward if the new ord is different than the one we just processed
+            // TODO optimization: it is silly to do a binary search to get the block every single
+            // time.
            //
-            // this allows the input TermOrdinal iterator to contain duplicates, so long as it's
-            // still sorted
-            if next_ord < ord {
-                panic!("Ordinals were not sorted: received {next_ord} after {ord}");
-            } else if next_ord > ord {
-                // 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 {
-                    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();
-                }
-                ord = next_ord;
-            } else {
-                // The next ord is equal to the previous ord: no need to seek or advance.
+            // 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 {
+                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();
            }
+            ord = next_ord;
        }
    }

@@ -671,8 +676,8 @@ mod tests {
    use common::OwnedBytes;

    use super::Dictionary;
-    use crate::MonotonicU64SSTable;
    use crate::dictionary::TermOrdHit;
+    use crate::{MonotonicU64SSTable, TermOrdinal};

    #[derive(Debug)]
    struct PermissionedHandle {
@@ -935,25 +940,24 @@ mod tests {
    }

    #[test]
-    fn test_ords_term() {
+    fn test_sorted_ords_to_term() {
        let (dic, _slice) = make_test_sstable();

        // Single term
        let mut terms = Vec::new();
        assert!(
-            dic.sorted_ords_to_term_cb(100_000..100_001, |term| {
+            dic.sorted_ords_to_term_cb(&[100_000], |term| {
                terms.push(term.to_vec());
-                Ok(())
            })
            .unwrap()
        );
        assert_eq!(terms, vec![format!("{:05X}", 100_000).into_bytes(),]);
        // Single term
        let mut terms = Vec::new();
+        let ords: Vec<TermOrdinal> = (100_001..100_002).collect();
        assert!(
-            dic.sorted_ords_to_term_cb(100_001..100_002, |term| {
+            dic.sorted_ords_to_term_cb(&ords, |term| {
                terms.push(term.to_vec());
-                Ok(())
            })
            .unwrap()
        );
@@ -961,9 +965,8 @@ mod tests {
        // both terms
        let mut terms = Vec::new();
        assert!(
-            dic.sorted_ords_to_term_cb(100_000..100_002, |term| {
+            dic.sorted_ords_to_term_cb(&[100_000, 100_001], |term| {
                terms.push(term.to_vec());
-                Ok(())
            })
            .unwrap()
        );
@@ -976,10 +979,10 @@ mod tests {
        );
        // Test cross block
        let mut terms = Vec::new();
+        let ords: Vec<TermOrdinal> = (98653..=98655).collect();
        assert!(
-            dic.sorted_ords_to_term_cb(98653..=98655, |term| {
+            dic.sorted_ords_to_term_cb(&ords, |term| {
                terms.push(term.to_vec());
-                Ok(())
            })
            .unwrap()
        );
@@ -991,6 +994,43 @@ mod tests {
                format!("{:05X}", 98655).into_bytes(),
            ]
        );
+        // redundant
+        let mut terms = Vec::new();
+        let ords: Vec<TermOrdinal> = vec![1, 1, 2];
+        assert!(
+            dic.sorted_ords_to_term_cb(&ords, |term| {
+                terms.push(term.to_vec());
+            })
+            .unwrap()
+        );
+        assert_eq!(
+            terms,
+            vec![
+                format!("{:05X}", 1).into_bytes(),
+                format!("{:05X}", 1).into_bytes(),
+                format!("{:05X}", 2).into_bytes(),
+            ]
+        );
+        // redundant cross block
+        let mut terms = Vec::new();
+        let ords: Vec<TermOrdinal> = vec![98653, 98653, 98654, 98654, 98655, 98655];
+        assert!(
+            dic.sorted_ords_to_term_cb(&ords, |term| {
+                terms.push(term.to_vec());
+            })
+            .unwrap()
+        );
+        assert_eq!(
+            terms,
+            vec![
+                format!("{:05X}", 98_653).into_bytes(),
+                format!("{:05X}", 98_653).into_bytes(),
+                format!("{:05X}", 98_654).into_bytes(),
+                format!("{:05X}", 98_654).into_bytes(),
+                format!("{:05X}", 98_655).into_bytes(),
+                format!("{:05X}", 98_655).into_bytes(),
+            ]
+        );
    }

    #[test]
--- a/sstable/src/lib.rs
+++ b/sstable/src/lib.rs
@@ -51,7 +51,7 @@ mod sstable_index_v3;
 pub use sstable_index_v3::{BlockAddr, SSTableIndex, SSTableIndexBuilder, SSTableIndexV3};
 mod sstable_index_v2;
 pub(crate) mod vint;
-pub use dictionary::Dictionary;
+pub use dictionary::{Dictionary, TermOrdHit};
 pub use streamer::{Streamer, StreamerBuilder};

 mod block_reader;
--- a/sstable/src/sstable_index_v3.rs
+++ b/sstable/src/sstable_index_v3.rs
@@ -553,7 +553,7 @@ impl FixedSize for BlockAddrBlockMetadata {
    const SIZE_IN_BYTES: usize = u64::SIZE_IN_BYTES
        + BlockStartAddr::SIZE_IN_BYTES
        + 2 * u32::SIZE_IN_BYTES
-        + 2 * u8::SIZE_IN_BYTES
+        + 2
        + u16::SIZE_IN_BYTES;
 }

--- a/stacker/Cargo.toml
+++ b/stacker/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-stacker"
-version = "0.6.0"
+version = "0.7.0"
 edition = "2024"
 license = "MIT"
 homepage = "https://github.com/quickwit-oss/tantivy"
@@ -9,7 +9,7 @@ description = "term hashmap used for indexing"

 [dependencies]
 murmurhash32 = "0.3"
-common = { version = "0.10", path = "../common/", package = "tantivy-common" }
+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 = "0.14.0" }
+binggan = { version = "0.16.1" }
 rand_distr = "0.5"

 [features]
--- a/tokenizer-api/Cargo.toml
+++ b/tokenizer-api/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-tokenizer-api"
-version = "0.6.0"
+version = "0.7.0"
 license = "MIT"
 edition = "2021"
 description = "Tokenizer API of tantivy"