attempt to optimize sorted_ords_to_term_cb

Co-authored-by: Pascal Seitz <pascal.seitz@datadoghq.com>

.gitattributes

@@ -1 +0,0 @@
-cpp/* linguist-vendored

.github/FUNDING.yml

@@ -0,0 +1,12 @@
+# These are supported funding model platforms
+
+github: fulmicoton
+patreon: # Replace with a single Patreon username
+open_collective: # Replace with a single Open Collective username
+ko_fi: # Replace with a single Ko-fi username
+tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
+community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
+liberapay: # Replace with a single Liberapay username
+issuehunt: # Replace with a single IssueHunt username
+otechie: # Replace with a single Otechie username
+custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']

.github/ISSUE_TEMPLATE/actions.md

@@ -0,0 +1,13 @@
+---
+name: Actions
+about: Actions not directly related to producing code.
+
+---
+
+# Actions title
+
+Action description. 
+e.g. 
+- benchmark
+- investigate and report
+- etc.

.github/dependabot.yml

@@ -0,0 +1,15 @@
+version: 2
+updates:
+- package-ecosystem: cargo
+  directory: "/"
+  schedule:
+    interval: daily
+    time: "20:00"
+  open-pull-requests-limit: 10
+
+- package-ecosystem: "github-actions"
+  directory: "/"
+  schedule:
+    interval: daily
+    time: "20:00"
+  open-pull-requests-limit: 10

.github/workflows/coverage.yml

@@ -0,0 +1,29 @@
+name: Coverage
+
+on:
+  push:
+    branches: [main]
+
+# Ensures that we cancel running jobs for the same PR / same workflow.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  coverage:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install Rust
+        run: rustup toolchain install nightly-2024-07-01 --profile minimal --component llvm-tools-preview
+      - uses: Swatinem/rust-cache@v2
+      - uses: taiki-e/install-action@cargo-llvm-cov
+      - name: Generate code coverage
+        run: cargo +nightly-2024-07-01 llvm-cov --all-features --workspace --doctests --lcov --output-path lcov.info
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v3
+        continue-on-error: true
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }} # not required for public repos
+          files: lcov.info
+          fail_ci_if_error: true

.github/workflows/long_running.yml

@@ -0,0 +1,33 @@
+name: Long running tests
+
+on:
+  push:
+    branches: [ main ]
+
+env:
+  CARGO_TERM_COLOR: always
+  NUM_FUNCTIONAL_TEST_ITERATIONS: 20000
+
+# Ensures that we cancel running jobs for the same PR / same workflow.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+    - name: Install stable
+      uses: actions-rs/toolchain@v1
+      with:
+          toolchain: stable
+          profile: minimal
+          override: true
+
+    - name: Run indexing_unsorted
+      run: cargo test indexing_unsorted -- --ignored
+    - name: Run indexing_sorted
+      run: cargo test indexing_sorted -- --ignored

.github/workflows/test.yml

@@ -0,0 +1,86 @@
+name: Unit tests
+
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+
+env:
+  CARGO_TERM_COLOR: always
+
+# Ensures that we cancel running jobs for the same PR / same workflow.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  check:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install nightly
+      uses: actions-rs/toolchain@v1
+      with:
+            toolchain: nightly
+            profile: minimal
+            components: rustfmt
+    - name: Install stable
+      uses: actions-rs/toolchain@v1
+      with:
+            toolchain: stable
+            profile: minimal
+            components: clippy
+
+    - uses: Swatinem/rust-cache@v2
+
+    - name: Check Formatting
+      run: cargo +nightly fmt --all -- --check
+    
+    - name: Check Stable Compilation
+      run: cargo build --all-features
+
+    
+    - name: Check Bench Compilation
+      run: cargo +nightly bench --no-run --profile=dev --all-features
+
+    - uses: actions-rs/clippy-check@v1
+      with:
+        toolchain: stable
+        token: ${{ secrets.GITHUB_TOKEN }}
+        args: --tests
+
+  test:
+
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        features: [
+            { label: "all", flags: "mmap,stopwords,lz4-compression,zstd-compression,failpoints" },
+            { label: "quickwit", flags: "mmap,quickwit,failpoints" }
+        ]
+
+    name: test-${{ matrix.features.label}}
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install stable
+      uses: actions-rs/toolchain@v1
+      with:
+            toolchain: stable
+            profile: minimal
+            override: true
+
+    - uses: taiki-e/install-action@nextest
+    - uses: Swatinem/rust-cache@v2
+
+    - name: Run tests
+      run: cargo +stable nextest run --features ${{ matrix.features.flags }} --verbose --workspace
+
+    - name: Run doctests
+      run: cargo +stable test --doc --features ${{ matrix.features.flags }} --verbose --workspace

.gitignore

@@ -1,4 +1,6 @@
 tantivy.iml
+.cargo
+proptest-regressions
 *.swp
 target
 target/debug
@@ -7,7 +9,9 @@ target/release
 Cargo.lock
 benchmark
 .DS_Store
-cpp/simdcomp/bitpackingbenchmark
 *.bk
 .idea
 trace.dat
+cargo-timing*
+control
+variable

.travis.yml

@@ -1,90 +0,0 @@
-# Based on the "trust" template v0.1.2
-# https://github.com/japaric/trust/tree/v0.1.2
-
-dist: trusty
-language: rust
-services: docker
-sudo: required
-
-env:
-  global:
-    - CRATE_NAME=tantivy
-    - TRAVIS_CARGO_NIGHTLY_FEATURE=""
-    # - secure: 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
-
-addons:
-  apt:
-    sources:
-      - ubuntu-toolchain-r-test
-      - kalakris-cmake
-    packages:
-      - gcc-4.8
-      - g++-4.8
-      - libcurl4-openssl-dev
-      - libelf-dev
-      - libdw-dev
-      - binutils-dev
-      - cmake
-
-matrix:
-  include:
-    # Android
-    - env: TARGET=aarch64-linux-android DISABLE_TESTS=1
-    #- env: TARGET=arm-linux-androideabi DISABLE_TESTS=1
-    #- env: TARGET=armv7-linux-androideabi DISABLE_TESTS=1
-    #- env: TARGET=i686-linux-android DISABLE_TESTS=1
-    #- env: TARGET=x86_64-linux-android DISABLE_TESTS=1
-
-    # Linux
-    #- env: TARGET=aarch64-unknown-linux-gnu
-    #- env: TARGET=i686-unknown-linux-gnu
-    - env: TARGET=x86_64-unknown-linux-gnu CODECOV=1 #UPLOAD_DOCS=1
-    # - env: TARGET=x86_64-unknown-linux-musl CODECOV=1
-    # OSX
-    #- env: TARGET=x86_64-apple-darwin
-    #  os: osx
-
-before_install:
-  - set -e
-  - rustup self update
-
-install:
-  - sh ci/install.sh
-  - source ~/.cargo/env || true
-  - env | grep "TRAVIS"
-
-before_script:
-  - export PATH=$HOME/.cargo/bin:$PATH
-  - cargo install cargo-update || echo "cargo-update already installed"
-  - cargo install cargo-travis || echo "cargo-travis already installed"
-
-script:
-  - bash ci/script.sh
-
-before_deploy:
-  - sh ci/before_deploy.sh
-
-after_success:
-  # Needs GH_TOKEN env var to be set in travis settings
-  - if [[ -v GH_TOKEN ]]; then echo "GH TOKEN IS SET"; else echo "GH TOKEN NOT SET"; fi
-  - if [[ -v UPLOAD_DOCS ]]; then cargo doc; cargo doc-upload; else echo "doc upload disabled."; fi
-
-#cache: cargo
-#before_cache:
-#  # Travis can't cache files that are not readable by "others"
-#  - chmod -R a+r $HOME/.cargo
-#  - find ./target/debug -type f -maxdepth 1 -delete
-#  - rm -f  ./target/.rustc_info.json
-#  - rm -fr ./target/debug/{deps,.fingerprint}/tantivy*
-#  - rm -r target/debug/examples/
-#  - ls -1 examples/ | sed -e 's/\.rs$//' | xargs -I "{}" find target/* -name "*{}*" -type f -delete
-
-#branches:
-#  only:
-#    # release tags
-#    - /^v\d+\.\d+\.\d+.*$/
-#    - master
-
-notifications:
-  email:
-    on_success: never

ARCHITECTURE.md

@@ -0,0 +1,295 @@
+# Tantivy
+
+## What is tantivy?
+
+Tantivy is a library that is meant to build search engines. Although it is by no means a port of Lucene, its architecture is strongly inspired by it. If you are familiar with Lucene, you may be struck by the overlapping vocabulary.
+This is not fortuitous.
+
+Tantivy's bread and butter is to address the problem of full-text search :
+
+Given a large set of textual documents, and a text query, return the K-most relevant documents in a very efficient way. To execute these queries rapidly, the tantivy needs to build an index beforehand. The relevance score implemented in the tantivy is not configurable. Tantivy uses the same score as the default similarity used in Lucene / Elasticsearch, called [BM25](https://en.wikipedia.org/wiki/Okapi_BM25).
+
+But tantivy's scope does not stop there. Numerous features are required to power rich-search applications. For instance, one may want to:
+
+- compute the count of documents matching a query in the different section of an e-commerce website,
+- display an average price per meter square for a real estate search engine,
+- take into account historical user data to rank documents in a specific way,
+- or even use tantivy to power an OLAP database.
+
+A more abstract description of the problem space tantivy is trying to address is the following.
+
+Ingest a large set of documents, create an index that makes it possible to
+rapidly select all documents matching a given predicate (also known as a query) and
+collect some information about them ([See collector](#collector-define-what-to-do-with-matched-documents)).
+
+Roughly speaking the design is following these guiding principles:
+
+- Search should be O(1) in memory.
+- Indexing should be O(1) in memory. (In practice it is just sublinear)
+- Search should be as fast as possible
+
+This comes at the cost of the dynamicity of the index: while it is possible to add, and delete documents from our corpus, the tantivy is designed to handle these updates in large batches.
+
+## [core/](src/core): Index, segments, searchers
+
+Core contains all of the high-level code to make it possible to create an index, add documents, delete documents and commit.
+
+This is both the most high-level part of tantivy, the least performance-sensitive one, the seemingly most mundane code... And paradoxically the most complicated part.
+
+### Index and Segments
+
+A tantivy index is a collection of smaller independent immutable segments.
+Each segment contains its own independent set of data structures.
+
+A segment is identified by a segment id that is in fact a UUID.
+The file of a segment has the format
+
+ ```segment-id . ext```
+
+The extension signals which data structure (or [`SegmentComponent`](src/index/segment_component.rs)) is stored in the file.
+
+A small `meta.json` file is in charge of keeping track of the list of segments, as well as the schema.
+
+On commit, one segment per indexing thread is written to disk, and the `meta.json` is then updated atomically.
+
+For a better idea of how indexing works, you may read the [following blog post](https://fulmicoton.com/posts/behold-tantivy-part2/).
+
+### Deletes
+
+Deletes happen by deleting a "term". Tantivy does not offer any notion of primary id, so it is up to the user to use a field in their schema as if it was a primary id, and delete the associated term if they want to delete only one specific document.
+
+On commit, tantivy will find all of the segments with documents matching this existing term and remove from [alive bitset file](src/fastfield/alive_bitset.rs) that represents the bitset of the alive document ids.
+Like all segment files, this file is immutable. Because it is possible to have more than one alive bitset file at a given instant, the alive bitset filename has the format ```segment_id . commit_opstamp . del```.
+
+An opstamp is simply an incremental id that identifies any operation applied to the index. For instance, performing a commit or adding a document.
+
+### DocId
+
+Within a segment, all documents are identified by a DocId that ranges within `[0, max_doc)`.
+where `max_doc` is the number of documents in the segment, (deleted or not). Having such a compact `DocId` space is key to the compression of our data structures.
+
+The DocIds are simply allocated in the order documents are added to the index.
+
+### Merges
+
+In separate threads, tantivy's index writer search for opportunities to merge segments.
+The point of segment merge is to:
+
+- eventually get rid of tombstoned documents
+- reduce the otherwise ever-growing number of segments.
+
+Indeed, while having several segments instead of one does not hurt search too much, having hundreds can have a measurable impact on the search performance.
+
+### Searcher
+
+The user of the library usually does not need to know about the existence of Segments.
+Searching is done through an object called a [`Searcher`](src/core/searcher.rs), that captures a
+snapshot of the index at one point of time, by holding a list of [SegmentReader](src/core/segment_reader.rs).
+
+In other words, regardless of commits, file garbage collection, or segment merge that might happen, as long as the user holds and reuse the same [Searcher](src/core/searcher.rs), search will happen on an immutable snapshot of the index.
+
+## [directory/](src/directory): Where should the data be stored?
+
+Tantivy, like Lucene, abstracts the place where the data should be stored in a key-trait
+called [`Directory`](src/directory/directory.rs).
+Contrary to Lucene however, "files" are quite different from some kind of `io::Read` object.
+Check out [`src/directory/directory.rs`](src/directory/directory.rs) trait for more details.
+
+Tantivy ships two main directory implementation: the `MmapDirectory` and the `RamDirectory`,
+but users can extend tantivy with their own implementation.
+
+## [schema/](src/schema): What are documents?
+
+Tantivy's document follows a very strict schema, decided before building any index.
+
+The schema defines all of the fields that the indexes [`Document`](src/schema/document/mod.rs) may and should contain, their types (`text`, `i64`, `u64`, `Date`, ...) as well as how it should be indexed / represented in tantivy.
+
+Depending on the type of the field, you can decide to
+
+- put it in the docstore
+- store it as a fast field
+- index it
+
+Practically, tantivy will push values associated with this type to up to 3 respective
+data structures.
+
+*Limitations*
+
+As of today, tantivy's schema imposes a 1:1 relationship between a field that is being ingested and a field represented in the search index. In sophisticated search application, it is fairly common to want to index a field twice using different tokenizers, or to index the concatenation of several fields together into one field.
+
+This is not something tantivy supports, and it is up to the user to duplicate field / concatenate fields before feeding them to tantivy.
+
+## General information about these data structures
+
+All data structures in tantivy, have:
+
+- a writer
+- a serializer
+- a reader
+
+The writer builds an in-memory representation of a batch of documents. This representation is not searchable. It is just meant as an intermediary mutable representation, to which we can sequentially add
+the document of a batch. At the end of the batch (or if a memory limit is reached), this representation
+is then converted into an on-disk immutable representation, that is extremely compact.
+This conversion is done by the serializer.
+
+Finally, the reader is in charge of offering an API to read on this on-disk read-only representation.
+In tantivy, readers are designed to require very little anonymous memory. The data is read straight from an mmapped file, and loading an index is as fast as mmapping its files.
+
+## [store/](src/store): Here is my DocId, Gimme my document
+
+The docstore is a row-oriented storage that, for each document, stores a subset of the fields
+that are marked as stored in the schema. The docstore is compressed using a general-purpose algorithm
+like LZ4.
+
+**Useful for**
+
+In search engines, it is often used to display search results.
+Once the top 10 documents have been identified, we fetch them from the store, and display them or their snippet on the search result page (aka SERP).
+
+**Not useful for**
+
+Fetching a document from the store is typically a "slow" operation. It usually consists in
+
+- searching into a compact tree-like data structure to find the position of the right block.
+- decompressing a small block
+- returning the document from this block.
+
+It is NOT meant to be called for every document matching a query.
+
+As a rule of thumb, if you hit the docstore more than 100 times per search query, you are probably misusing tantivy.
+
+## [fastfield/](src/fastfield): Here is my DocId, Gimme my value
+
+Fast fields are stored in a column-oriented storage that allows for random access.
+The only compression applied is bitpacking. The column comes with two meta data.
+The minimum value in the column and the number of bits per doc.
+
+Fetching a value for a `DocId` is then as simple as computing
+
+```rust
+min_value + fetch_bits(num_bits * doc_id..num_bits * (doc_id+1))
+```
+
+This operation just requires one memory fetch.
+Because, DocSets are scanned through in order (DocId are iterated in a sorted manner) which
+also help locality.
+
+In Lucene's jargon, fast fields are called DocValues.
+
+**Useful for**
+
+They are typically integer values that are useful to either rank or compute aggregate over
+all of the documents matching a query (aka [DocSet](src/docset.rs)).
+
+For instance, one could define a function to combine upvotes with tantivy's internal relevancy score.
+This can be done by fetching a fast field during scoring.
+One could also compute the mean price of the items matching a query in an e-commerce website.
+This can be done by fetching a fast field in a collector.
+Finally one could decide to post-filter a docset to remove docset with a price within a specific range.
+If the ratio of filtered out documents is not too low, an efficient way to do this is to fetch the price and apply the filter on the collector side.
+
+Aside from integer values, it is also possible to store an actual byte payload.
+For advanced search engine, it is possible to store all of the features required for learning-to-rank in a byte payload, access it during search, and apply the learning-to-rank model.
+
+Finally facets are a specific kind of fast field, and the associated source code is in [`fastfield/facet_reader.rs`](src/fastfield/facet_reader.rs).
+
+# The inverted search index
+
+The inverted index is the core part of full-text search.
+When presented a new document with the text field "Hello, happy tax payer!", tantivy breaks it into a list of so-called tokens. In addition to just splitting these strings into tokens, it might also do different kinds of operations like dropping the punctuation, converting the character to lowercase, apply stemming, etc. Tantivy makes it possible to configure the operations to be applied in the schema (tokenizer/ is the place where these operations are implemented).
+
+For instance, the default tokenizer of tantivy would break our text into: `[hello, happy, tax, payer]`.
+The document will therefore be registered in the inverted index as containing the terms
+`[text:hello, text:happy, text:tax, text:payer]`.
+
+The role of the inverted index is, when given a term, gives us in return a very fast iterator over the sorted doc ids that match the term.
+
+Such an iterator is called a posting list. In addition to giving us `DocId`, they can also give us optionally the number of occurrence of the term for each document, also called term frequency or TF.
+
+These iterators being sorted by DocId, one can create an iterator over the document containing `text:tax AND text:payer`, `(text:tax AND text:payer) OR (text:contribuable)` or any boolean expression.
+
+In order to represent the function
+```Term ⟶ Posting```
+
+The inverted index actually consists of two data structures chained together.
+
+- [Term](src/schema/term.rs) ⟶ [TermInfo](src/postings/term_info.rs) is addressed by the term dictionary.
+- [TermInfo](src/postings/term_info.rs) ⟶ [Posting](src/postings/postings.rs) is addressed by the posting lists.
+
+Where [TermInfo](src/postings/term_info.rs) is an object containing some meta data about a term.
+
+## [termdict/](src/termdict): Here is a term, give me the [TermInfo](src/postings/term_info.rs)
+
+Tantivy's term dictionary is mainly in charge of supplying the function
+
+[Term](src/schema/term.rs) ⟶ [TermInfo](src/postings/term_info.rs)
+
+It is itself broken into two parts.
+
+- [Term](src/schema/term.rs) ⟶ [TermOrdinal](src/termdict/mod.rs) is addressed by a finite state transducer, implemented by the fst crate.
+- [TermOrdinal](src/termdict/mod.rs) ⟶ [TermInfo](src/postings/term_info.rs) is addressed by the term info store.
+
+## [postings/](src/postings): Iterate over documents... very fast
+
+A posting list makes it possible to store a sorted list of doc ids and for each doc store
+a term frequency as well.
+
+The posting lists are stored in a separate file. The [TermInfo](src/postings/term_info.rs) contains an offset into that file and a number of documents for the given posting list. Both are required and sufficient to read the posting list.
+
+The posting list is organized in block of 128 documents.
+One block of doc ids is followed by one block of term frequencies.
+
+The doc ids are delta encoded and bitpacked.
+The term frequencies are bitpacked.
+
+Because the number of docs is rarely a multiple of 128, the last block may contain an arbitrary number of docs between 1 and 127 documents. We then use variable int encoding instead of bitpacking.
+
+## [positions/](src/positions): Where are my terms within the documents?
+
+Phrase queries make it possible to search for documents containing a specific sequence of terms.
+For instance, when the phrase query "the art of war" does not match "the war of art".
+To make it possible, it is possible to specify in the schema that a field should store positions in addition to being indexed.
+
+The token positions of all of the terms are then stored in a separate file with the extension `.pos`.
+The [TermInfo](src/postings/term_info.rs) gives an offset (expressed in position this time) in this file. As we iterate through the docset,
+we advance the position reader by the number of term frequencies of the current document.
+
+## [fieldnorm/](src/fieldnorm): Here is my doc, how many tokens in this field?
+
+The [BM25](https://en.wikipedia.org/wiki/Okapi_BM25) formula also requires to know the number of tokens stored in a specific field for a given document. We store this information on one byte per document in the fieldnorm.
+The fieldnorm is therefore compressed. Values up to 40 are encoded unchanged.
+
+## [tokenizer/](src/tokenizer): How should we process text?
+
+Text processing is key to a good search experience.
+Splits or normalize your text too much, and the search results will have a less precision and a higher recall.
+Do not normalize, or under split your text, you will end up with a higher precision and a lesser recall.
+
+Text processing can be configured by selecting an off-the-shelf [`Tokenizer`](./src/tokenizer/tokenizer.rs) or implementing your own to first split the text into tokens, and then chain different [`TokenFilter`](src/tokenizer/tokenizer.rs)'s to it.
+
+Tantivy's comes with few tokenizers, but external crates are offering advanced tokenizers, such as [Lindera](https://crates.io/crates/lindera) for Japanese.
+
+## [query/](src/query): Define and compose queries
+
+The [Query](src/query/query.rs) trait defines what a query is.
+Due to the necessity for some queries to compute some statistics over the entire index, and because the
+index is composed of several `SegmentReader`, the path from transforming a `Query` to an iterator over documents is slightly convoluted, but fundamentally, this is what a Query is.
+
+The iterator over a document comes with some scoring function. The resulting trait is called a
+[Scorer](src/query/scorer.rs) and is specific to a segment.
+
+Different queries can be combined using the [BooleanQuery](src/query/boolean_query/).
+Tantivy comes with different types of queries and can be extended by implementing
+the `Query`, `Weight`, and `Scorer` traits.
+
+## [collector](src/collector): Define what to do with matched documents
+
+Collectors define how to aggregate the documents matching a query, in the broadest sense possible.
+The search will push matched documents one by one, calling their
+`fn collect(doc: DocId, score: Score);` method.
+
+Users may implement their own collectors by implementing the [Collector](src/collector/mod.rs) trait.
+
+## [query-grammar](query-grammar): Defines the grammar of the query parser
+
+While the [QueryParser](src/query/query_parser/query_parser.rs) struct is located in the `query/` directory, the actual parser combinator used to convert user queries into an AST is in an external crate called `query-grammar`. This part was externalized to lighten the work of the compiler.

CHANGELOG.md

@@ -1,7 +1,520 @@
+Tantivy 0.24
+================================
+Tantivy 0.24 will be backwards compatible with indices created with v0.22 and v0.21. The new minimum rust version will be 1.75. Tantivy 0.23 will be skipped.
+
+#### Bugfixes
+- fix potential endless loop in merge [#2457](https://github.com/quickwit-oss/tantivy/pull/2457)(@PSeitz)
+- fix bug that causes out-of-order sstable key. [#2445](https://github.com/quickwit-oss/tantivy/pull/2445)(@fulmicoton)
+- fix ReferenceValue API flaw [#2372](https://github.com/quickwit-oss/tantivy/pull/2372)(@PSeitz)
+- fix `OwnedBytes` debug panic [#2512](https://github.com/quickwit-oss/tantivy/pull/2512)(@b41sh)
+- catch panics during merges [#2582](https://github.com/quickwit-oss/tantivy/pull/2582)(@rdettai)
+- switch from u32 to usize in bitpacker. This enables multivalued columns larger than 4GB, which crashed during merge before. [#2581](https://github.com/quickwit-oss/tantivy/pull/2581) [#2586](https://github.com/quickwit-oss/tantivy/pull/2586)(@fulmicoton-dd @PSeitz)
+
+#### Breaking API Changes
+- remove index sorting [#2434](https://github.com/quickwit-oss/tantivy/pull/2434)(@PSeitz)
+
+#### Features/Improvements
+- **Aggregation**
+    - Support for cardinality aggregation [#2337](https://github.com/quickwit-oss/tantivy/pull/2337) [#2446](https://github.com/quickwit-oss/tantivy/pull/2446) (@raphaelcoeffic @PSeitz)
+    - Support for extended stats aggregation [#2247](https://github.com/quickwit-oss/tantivy/pull/2247)(@giovannicuccu)
+    - Add Key::I64 and Key::U64 variants in aggregation to avoid f64 precision issues [#2468](https://github.com/quickwit-oss/tantivy/pull/2468)(@PSeitz)
+    - Faster term aggregation fetch terms [#2447](https://github.com/quickwit-oss/tantivy/pull/2447)(@PSeitz)
+    - Improve custom order deserialization [#2451](https://github.com/quickwit-oss/tantivy/pull/2451)(@PSeitz)
+    - Change AggregationLimits behavior [#2495](https://github.com/quickwit-oss/tantivy/pull/2495)(@PSeitz)
+    - lower contention on AggregationLimits [#2394](https://github.com/quickwit-oss/tantivy/pull/2394)(@PSeitz)
+    - fix postcard compatibility for top_hits, add postcard test [#2346](https://github.com/quickwit-oss/tantivy/pull/2346)(@PSeitz)
+    - reduce top hits memory consumption [#2426](https://github.com/quickwit-oss/tantivy/pull/2426)(@PSeitz)
+    - check unsupported parameters top_hits [#2351](https://github.com/quickwit-oss/tantivy/pull/2351)(@PSeitz)
+    - Change AggregationLimits to AggregationLimitsGuard [#2495](https://github.com/quickwit-oss/tantivy/pull/2495)(@PSeitz)
+    - add support for counting non integer in aggregation [#2547](https://github.com/quickwit-oss/tantivy/pull/2547)(@trinity-1686a)
+- **Range Queries**
+    - Support fast field range queries on json fields [#2456](https://github.com/quickwit-oss/tantivy/pull/2456)(@PSeitz)
+    - Add support for str fast field range query [#2460](https://github.com/quickwit-oss/tantivy/pull/2460) [#2452](https://github.com/quickwit-oss/tantivy/pull/2452) [#2453](https://github.com/quickwit-oss/tantivy/pull/2453)(@PSeitz)
+    - modify fastfield range query heuristic [#2375](https://github.com/quickwit-oss/tantivy/pull/2375)(@trinity-1686a)
+    - add FastFieldRangeQuery for explicit range queries on fast field (for `RangeQuery` it is autodetected) [#2477](https://github.com/quickwit-oss/tantivy/pull/2477)(@PSeitz)
+
+- add format backwards-compatibility tests [#2485](https://github.com/quickwit-oss/tantivy/pull/2485)(@PSeitz)
+- add columnar format compatibility tests [#2433](https://github.com/quickwit-oss/tantivy/pull/2433)(@PSeitz)
+- Improved snippet ranges algorithm [#2474](https://github.com/quickwit-oss/tantivy/pull/2474)(@gezihuzi)
+- make find_field_with_default return json fields without path [#2476](https://github.com/quickwit-oss/tantivy/pull/2476)(@trinity-1686a)
+- Make `BooleanQuery` support `minimum_number_should_match` [#2405](https://github.com/quickwit-oss/tantivy/pull/2405)(@LebranceBW)
+- Make `NUM_MERGE_THREADS` configurable [#2535](https://github.com/quickwit-oss/tantivy/pull/2535)(@Barre)
+
+- **RegexPhraseQuery** 
+`RegexPhraseQuery` supports phrase queries with regex. E.g. query "b.* b.* wolf" matches "big bad wolf". Slop is supported as well: "b.* wolf"~2 matches "big bad wolf" [#2516](https://github.com/quickwit-oss/tantivy/pull/2516)(@PSeitz)
+
+- **Optional Index in Multivalue Columnar Index** 
+For mostly empty multivalued indices there was a large overhead during creation when iterating all docids (merge case). 
+This is alleviated by placing an optional index in the multivalued index to mark documents that have values. 
+This will slightly increase space and access time. [#2439](https://github.com/quickwit-oss/tantivy/pull/2439)(@PSeitz)
+
+- **Store DateTime as nanoseconds in doc store** DateTime in the doc store was truncated to microseconds previously. This removes this truncation, while still keeping backwards compatibility. [#2486](https://github.com/quickwit-oss/tantivy/pull/2486)(@PSeitz)
+
+- **Performace/Memory**
+    - lift clauses in LogicalAst for optimized ast during execution [#2449](https://github.com/quickwit-oss/tantivy/pull/2449)(@PSeitz)
+    - Use Vec instead of BTreeMap to back OwnedValue object [#2364](https://github.com/quickwit-oss/tantivy/pull/2364)(@fulmicoton)
+    - Replace TantivyDocument with CompactDoc. CompactDoc is much smaller and provides similar performance. [#2402](https://github.com/quickwit-oss/tantivy/pull/2402)(@PSeitz)
+    - Recycling buffer in PrefixPhraseScorer [#2443](https://github.com/quickwit-oss/tantivy/pull/2443)(@fulmicoton)
+
+- **Json Type**
+    - JSON supports now all values on the root level. Previously an object was required. This enables support for flat mixed types. allow more JSON values, fix i64 special case [#2383](https://github.com/quickwit-oss/tantivy/pull/2383)(@PSeitz)
+    - add json path constructor to term [#2367](https://github.com/quickwit-oss/tantivy/pull/2367)(@PSeitz)
+
+- **QueryParser**
+    - fix de-escaping too much in query parser [#2427](https://github.com/quickwit-oss/tantivy/pull/2427)(@trinity-1686a)
+    - improve query parser [#2416](https://github.com/quickwit-oss/tantivy/pull/2416)(@trinity-1686a)
+    - Support field grouping `title:(return AND "pink panther")` [#2333](https://github.com/quickwit-oss/tantivy/pull/2333)(@trinity-1686a)
+    - allow term starting with wildcard [#2568](https://github.com/quickwit-oss/tantivy/pull/2568)(@trinity-1686a)
+
+- Exist queries match subpath fields [#2558](https://github.com/quickwit-oss/tantivy/pull/2558)(@rdettai)
+- add access benchmark for columnar [#2432](https://github.com/quickwit-oss/tantivy/pull/2432)(@PSeitz)
+- extend indexwriter proptests [#2342](https://github.com/quickwit-oss/tantivy/pull/2342)(@PSeitz)
+- add bench & test for columnar merging [#2428](https://github.com/quickwit-oss/tantivy/pull/2428)(@PSeitz)
+- Change in Executor API [#2391](https://github.com/quickwit-oss/tantivy/pull/2391)(@fulmicoton)
+- Removed usage of num_cpus [#2387](https://github.com/quickwit-oss/tantivy/pull/2387)(@fulmicoton)
+- use bingang for agg and stacker benchmark [#2378](https://github.com/quickwit-oss/tantivy/pull/2378)[#2492](https://github.com/quickwit-oss/tantivy/pull/2492)(@PSeitz) 
+- cleanup top level exports [#2382](https://github.com/quickwit-oss/tantivy/pull/2382)(@PSeitz)
+- make convert_to_fast_value_and_append_to_json_term pub [#2370](https://github.com/quickwit-oss/tantivy/pull/2370)(@PSeitz)
+- remove JsonTermWriter [#2238](https://github.com/quickwit-oss/tantivy/pull/2238)(@PSeitz)
+- validate sort by field type [#2336](https://github.com/quickwit-oss/tantivy/pull/2336)(@PSeitz)
+- Fix trait bound of StoreReader::iter [#2360](https://github.com/quickwit-oss/tantivy/pull/2360)(@adamreichold)
+- remove read_postings_no_deletes [#2526](https://github.com/quickwit-oss/tantivy/pull/2526)(@PSeitz)
+
+Tantivy 0.22
+================================
+
+Tantivy 0.22 will be able to read indices created with Tantivy 0.21.
+
+#### Bugfixes
+- Fix null byte handling in JSON paths (null bytes in json keys caused panic during indexing) [#2345](https://github.com/quickwit-oss/tantivy/pull/2345)(@PSeitz)
+- Fix bug that can cause `get_docids_for_value_range` to panic. [#2295](https://github.com/quickwit-oss/tantivy/pull/2295)(@fulmicoton)
+- Avoid 1 document indices by increase min memory to 15MB for indexing [#2176](https://github.com/quickwit-oss/tantivy/pull/2176)(@PSeitz)
+- Fix merge panic for JSON fields [#2284](https://github.com/quickwit-oss/tantivy/pull/2284)(@PSeitz)
+- Fix bug occurring when merging JSON object indexed with positions. [#2253](https://github.com/quickwit-oss/tantivy/pull/2253)(@fulmicoton)
+- Fix empty DateHistogram gap bug [#2183](https://github.com/quickwit-oss/tantivy/pull/2183)(@PSeitz)
+- Fix range query end check (fields with less than 1 value per doc are affected) [#2226](https://github.com/quickwit-oss/tantivy/pull/2226)(@PSeitz)
+- Handle exclusive out of bounds ranges on fastfield range queries [#2174](https://github.com/quickwit-oss/tantivy/pull/2174)(@PSeitz)
+
+#### Breaking API Changes
+- rename ReloadPolicy onCommit to onCommitWithDelay [#2235](https://github.com/quickwit-oss/tantivy/pull/2235)(@giovannicuccu)
+- Move exports from the root into modules [#2220](https://github.com/quickwit-oss/tantivy/pull/2220)(@PSeitz)
+- Accept field name instead of `Field` in FilterCollector [#2196](https://github.com/quickwit-oss/tantivy/pull/2196)(@PSeitz)
+- remove deprecated IntOptions and DateTime [#2353](https://github.com/quickwit-oss/tantivy/pull/2353)(@PSeitz)
+
+#### Features/Improvements
+- Tantivy documents as a trait: Index data directly without converting to tantivy types first [#2071](https://github.com/quickwit-oss/tantivy/pull/2071)(@ChillFish8)
+- encode some part of posting list as -1 instead of direct values (smaller inverted indices) [#2185](https://github.com/quickwit-oss/tantivy/pull/2185)(@trinity-1686a)
+- **Aggregation**
+  - Support to deserialize f64 from string [#2311](https://github.com/quickwit-oss/tantivy/pull/2311)(@PSeitz)
+  - Add a top_hits aggregator [#2198](https://github.com/quickwit-oss/tantivy/pull/2198)(@ditsuke)
+  - Support bool type in term aggregation [#2318](https://github.com/quickwit-oss/tantivy/pull/2318)(@PSeitz)
+  - Support ip addresses in term aggregation [#2319](https://github.com/quickwit-oss/tantivy/pull/2319)(@PSeitz)
+  - Support date type in term aggregation [#2172](https://github.com/quickwit-oss/tantivy/pull/2172)(@PSeitz)
+  - Support escaped dot when addressing field [#2250](https://github.com/quickwit-oss/tantivy/pull/2250)(@PSeitz)
+
+- Add ExistsQuery to check documents that have a value [#2160](https://github.com/quickwit-oss/tantivy/pull/2160)(@imotov)
+- Expose TopDocs::order_by_u64_field again [#2282](https://github.com/quickwit-oss/tantivy/pull/2282)(@ditsuke)
+
+- **Memory/Performance**
+  - Faster TopN: replace BinaryHeap with TopNComputer [#2186](https://github.com/quickwit-oss/tantivy/pull/2186)(@PSeitz)
+  - reduce number of allocations during indexing [#2257](https://github.com/quickwit-oss/tantivy/pull/2257)(@PSeitz)
+  - Less Memory while indexing: docid deltas while indexing [#2249](https://github.com/quickwit-oss/tantivy/pull/2249)(@PSeitz)
+  - Faster indexing: use term hashmap in fastfield [#2243](https://github.com/quickwit-oss/tantivy/pull/2243)(@PSeitz)
+  - term hashmap remove copy in is_empty, unused unordered_id [#2229](https://github.com/quickwit-oss/tantivy/pull/2229)(@PSeitz)
+  - add method to fetch block of first values in columnar [#2330](https://github.com/quickwit-oss/tantivy/pull/2330)(@PSeitz)
+  - Faster aggregations: add fast path for full columns in fetch_block [#2328](https://github.com/quickwit-oss/tantivy/pull/2328)(@PSeitz)
+  - Faster sstable loading: use fst for sstable index [#2268](https://github.com/quickwit-oss/tantivy/pull/2268)(@trinity-1686a)
+
+- **QueryParser**
+  - allow newline where we allow space in query parser [#2302](https://github.com/quickwit-oss/tantivy/pull/2302)(@trinity-1686a)
+  - allow some mixing of occur and bool in strict query parser [#2323](https://github.com/quickwit-oss/tantivy/pull/2323)(@trinity-1686a)
+  - handle * inside term in lenient query parser [#2228](https://github.com/quickwit-oss/tantivy/pull/2228)(@trinity-1686a)
+  - add support for exists query syntax in query parser [#2170](https://github.com/quickwit-oss/tantivy/pull/2170)(@trinity-1686a)
+- Add shared search executor [#2312](https://github.com/quickwit-oss/tantivy/pull/2312)(@MochiXu)
+- Truncate keys to u16::MAX in term hashmap [#2299](https://github.com/quickwit-oss/tantivy/pull/2299)(@PSeitz)
+- report if a term matched when warming up posting list [#2309](https://github.com/quickwit-oss/tantivy/pull/2309)(@trinity-1686a)
+- Support json fields in FuzzyTermQuery [#2173](https://github.com/quickwit-oss/tantivy/pull/2173)(@PingXia-at)
+- Read list of fields encoded in term dictionary for JSON fields [#2184](https://github.com/quickwit-oss/tantivy/pull/2184)(@PSeitz)
+- add collect_block to BoxableSegmentCollector [#2331](https://github.com/quickwit-oss/tantivy/pull/2331)(@PSeitz)
+- expose collect_block buffer size [#2326](https://github.com/quickwit-oss/tantivy/pull/2326)(@PSeitz)
+- Forward regex parser errors [#2288](https://github.com/quickwit-oss/tantivy/pull/2288)(@adamreichold)
+- Make FacetCounts defaultable and cloneable. [#2322](https://github.com/quickwit-oss/tantivy/pull/2322)(@adamreichold)
+- Derive Debug for SchemaBuilder [#2254](https://github.com/quickwit-oss/tantivy/pull/2254)(@GodTamIt)
+- add missing inlines to tantivy options [#2245](https://github.com/quickwit-oss/tantivy/pull/2245)(@PSeitz)
+
+Tantivy 0.21.1
+================================
+#### Bugfixes
+- Range queries on fast fields with less values on that field than documents had an invalid end condition, leading to missing results. [#2226](https://github.com/quickwit-oss/tantivy/issues/2226)(@appaquet @PSeitz)
+- Increase the minimum memory budget from 3MB to 15MB to avoid single doc segments (API fix). [#2176](https://github.com/quickwit-oss/tantivy/issues/2176)(@PSeitz)
+
+Tantivy 0.21
+================================
+#### Bugfixes
+- Fix track fast field memory consumption, which led to higher memory consumption than the budget allowed during indexing [#2148](https://github.com/quickwit-oss/tantivy/issues/2148)[#2147](https://github.com/quickwit-oss/tantivy/issues/2147)(@PSeitz)
+- Fix a regression from 0.20 where sort index by date wasn't working anymore [#2124](https://github.com/quickwit-oss/tantivy/issues/2124)(@PSeitz)
+- Fix getting the root facet on the `FacetCollector`. [#2086](https://github.com/quickwit-oss/tantivy/issues/2086)(@adamreichold)
+- Align numerical type priority order of columnar and query. [#2088](https://github.com/quickwit-oss/tantivy/issues/2088)(@fmassot)
+#### Breaking Changes
+- Remove support for Brotli and Snappy compression [#2123](https://github.com/quickwit-oss/tantivy/issues/2123)(@adamreichold)
+#### Features/Improvements
+- Implement lenient query parser [#2129](https://github.com/quickwit-oss/tantivy/pull/2129)(@trinity-1686a)
+- order_by_u64_field and order_by_fast_field allow sorting in ascending and descending order [#2111](https://github.com/quickwit-oss/tantivy/issues/2111)(@naveenann)
+- Allow dynamic filters in text analyzer builder [#2110](https://github.com/quickwit-oss/tantivy/issues/2110)(@fulmicoton @fmassot)
+- **Aggregation**
+  - Add missing parameter for term aggregation [#2149](https://github.com/quickwit-oss/tantivy/issues/2149)[#2103](https://github.com/quickwit-oss/tantivy/issues/2103)(@PSeitz)
+  - Add missing parameter for percentiles [#2157](https://github.com/quickwit-oss/tantivy/issues/2157)(@PSeitz)
+  - Add missing parameter for stats,min,max,count,sum,avg [#2151](https://github.com/quickwit-oss/tantivy/issues/2151)(@PSeitz)
+  - Improve aggregation deserialization error message [#2150](https://github.com/quickwit-oss/tantivy/issues/2150)(@PSeitz)
+  - Add validation for type Bytes to term_agg [#2077](https://github.com/quickwit-oss/tantivy/issues/2077)(@PSeitz)
+  - Alternative mixed field collection [#2135](https://github.com/quickwit-oss/tantivy/issues/2135)(@PSeitz)
+- Add missing query_terms impl for TermSetQuery. [#2120](https://github.com/quickwit-oss/tantivy/issues/2120)(@adamreichold)
+- Minor improvements to OwnedBytes [#2134](https://github.com/quickwit-oss/tantivy/issues/2134)(@adamreichold)
+- Remove allocations in split compound words [#2080](https://github.com/quickwit-oss/tantivy/issues/2080)(@PSeitz)
+- Ngram tokenizer now returns an error with invalid arguments [#2102](https://github.com/quickwit-oss/tantivy/issues/2102)(@fmassot)
+- Make TextAnalyzerBuilder public [#2097](https://github.com/quickwit-oss/tantivy/issues/2097)(@adamreichold)
+- Return an error when tokenizer is not found while indexing [#2093](https://github.com/quickwit-oss/tantivy/issues/2093)(@naveenann)
+- Delayed column opening during merge [#2132](https://github.com/quickwit-oss/tantivy/issues/2132)(@PSeitz)
+
+Tantivy 0.20.2
+================================
+- Align numerical type priority order on the search side.  [#2088](https://github.com/quickwit-oss/tantivy/issues/2088) (@fmassot)
+- Fix is_child_of function not considering the root facet. [#2086](https://github.com/quickwit-oss/tantivy/issues/2086) (@adamreichhold)
+
+Tantivy 0.20.1
+================================
+- Fix building on windows with mmap [#2070](https://github.com/quickwit-oss/tantivy/issues/2070) (@ChillFish8)
+
+Tantivy 0.20
+================================
+#### Bugfixes
+- Fix phrase queries with slop (slop supports now transpositions, algorithm that carries slop so far for num terms > 2) [#2031](https://github.com/quickwit-oss/tantivy/issues/2031)[#2020](https://github.com/quickwit-oss/tantivy/issues/2020)(@PSeitz)
+- Handle error for exists on MMapDirectory [#1988](https://github.com/quickwit-oss/tantivy/issues/1988) (@PSeitz)
+- Aggregation
+  - Fix min doc_count empty merge bug [#2057](https://github.com/quickwit-oss/tantivy/issues/2057) (@PSeitz)
+  - Fix: Sort order for term aggregations (sort order on key was inverted) [#1858](https://github.com/quickwit-oss/tantivy/issues/1858) (@PSeitz)
+
+#### Features/Improvements
+- Add PhrasePrefixQuery [#1842](https://github.com/quickwit-oss/tantivy/issues/1842) (@trinity-1686a)
+- Add `coerce` option for text and numbers types (convert the value instead of returning an error during indexing) [#1904](https://github.com/quickwit-oss/tantivy/issues/1904) (@PSeitz)
+- Add regex tokenizer [#1759](https://github.com/quickwit-oss/tantivy/issues/1759)(@mkleen)
+- Move tokenizer API to separate crate. Having a separate crate with a stable API will allow us to use tokenizers with different tantivy versions. [#1767](https://github.com/quickwit-oss/tantivy/issues/1767) (@PSeitz)
+- **Columnar crate**: New fast field handling (@fulmicoton @PSeitz) [#1806](https://github.com/quickwit-oss/tantivy/issues/1806)[#1809](https://github.com/quickwit-oss/tantivy/issues/1809)
+  - Support for fast fields with optional values. Previously tantivy supported only single-valued and multi-value fast fields. The encoding of optional fast fields is now very compact.
+  - Fast field Support for JSON (schemaless fast fields). Support multiple types on the same column. [#1876](https://github.com/quickwit-oss/tantivy/issues/1876) (@fulmicoton)
+  - Unified access for fast fields over different cardinalities.
+  - Unified storage for typed and untyped fields.
+  - Move fastfield codecs into columnar. [#1782](https://github.com/quickwit-oss/tantivy/issues/1782) (@fulmicoton)
+  - Sparse dense index for optional values [#1716](https://github.com/quickwit-oss/tantivy/issues/1716) (@PSeitz)
+  - Switch to nanosecond precision in DateTime fastfield [#2016](https://github.com/quickwit-oss/tantivy/issues/2016) (@PSeitz)
+- **Aggregation**
+  - Add `date_histogram` aggregation (only `fixed_interval` for now) [#1900](https://github.com/quickwit-oss/tantivy/issues/1900) (@PSeitz)
+  - Add `percentiles` aggregations [#1984](https://github.com/quickwit-oss/tantivy/issues/1984) (@PSeitz)
+  - [**breaking**] Drop JSON support on intermediate agg result (we use postcard as format in `quickwit` to send intermediate results) [#1992](https://github.com/quickwit-oss/tantivy/issues/1992) (@PSeitz)
+  - Set memory limit in bytes for aggregations after which they abort (Previously there was only the bucket limit) [#1942](https://github.com/quickwit-oss/tantivy/issues/1942)[#1957](https://github.com/quickwit-oss/tantivy/issues/1957)(@PSeitz)
+  - Add support for u64,i64,f64 fields in term aggregation [#1883](https://github.com/quickwit-oss/tantivy/issues/1883) (@PSeitz)
+  - Allow histogram bounds to be passed as Rfc3339 [#2076](https://github.com/quickwit-oss/tantivy/issues/2076) (@PSeitz)
+  - Add count, min, max, and sum aggregations [#1794](https://github.com/quickwit-oss/tantivy/issues/1794) (@guilload)
+  - Switch to Aggregation without serde_untagged => better deserialization errors. [#2003](https://github.com/quickwit-oss/tantivy/issues/2003) (@PSeitz)
+  - Switch to ms in histogram for date type (ES compatibility) [#2045](https://github.com/quickwit-oss/tantivy/issues/2045) (@PSeitz)
+  - Reduce term aggregation memory consumption [#2013](https://github.com/quickwit-oss/tantivy/issues/2013) (@PSeitz)
+  - Reduce agg memory consumption: Replace generic aggregation collector (which has a high memory requirement per instance) in aggregation tree with optimized versions behind a trait.
+  - Split term collection count and sub_agg (Faster term agg with less memory consumption for cases without sub-aggs) [#1921](https://github.com/quickwit-oss/tantivy/issues/1921) (@PSeitz)
+  - Schemaless aggregations: In combination with stacker tantivy supports now schemaless aggregations via the JSON type.
+    - Add aggregation support for JSON type [#1888](https://github.com/quickwit-oss/tantivy/issues/1888) (@PSeitz)
+    - Mixed types support on JSON fields in aggs [#1971](https://github.com/quickwit-oss/tantivy/issues/1971) (@PSeitz)
+  - Perf: Fetch blocks of vals in aggregation for all cardinality [#1950](https://github.com/quickwit-oss/tantivy/issues/1950) (@PSeitz)
+  - Allow histogram bounds to be passed as Rfc3339 [#2076](https://github.com/quickwit-oss/tantivy/issues/2076) (@PSeitz)
+- `Searcher` with disabled scoring via `EnableScoring::Disabled` [#1780](https://github.com/quickwit-oss/tantivy/issues/1780) (@shikhar)
+- Enable tokenizer on json fields [#2053](https://github.com/quickwit-oss/tantivy/issues/2053) (@PSeitz)
+- Enforcing "NOT" and "-" queries consistency in UserInputAst [#1609](https://github.com/quickwit-oss/tantivy/issues/1609) (@bazhenov)
+- Faster indexing
+  - Refactor tokenization pipeline to use GATs [#1924](https://github.com/quickwit-oss/tantivy/issues/1924) (@trinity-1686a)
+  - Faster term hash map [#2058](https://github.com/quickwit-oss/tantivy/issues/2058)[#1940](https://github.com/quickwit-oss/tantivy/issues/1940) (@PSeitz)
+  - tokenizer-api: reduce Tokenizer allocation overhead [#2062](https://github.com/quickwit-oss/tantivy/issues/2062) (@PSeitz)
+  - Refactor vint [#2010](https://github.com/quickwit-oss/tantivy/issues/2010) (@PSeitz)
+- Faster search
+  - Work in batches of docs on the SegmentCollector (Only for cases without score for now) [#1937](https://github.com/quickwit-oss/tantivy/issues/1937) (@PSeitz)
+  - Faster fast field range queries using SIMD [#1954](https://github.com/quickwit-oss/tantivy/issues/1954) (@fulmicoton)
+  - Improve fast field range query performance [#1864](https://github.com/quickwit-oss/tantivy/issues/1864) (@PSeitz)
+- Make BM25 scoring more flexible [#1855](https://github.com/quickwit-oss/tantivy/issues/1855) (@alexcole)
+- Switch fs2 to fs4 as it is now unmaintained and does not support illumos [#1944](https://github.com/quickwit-oss/tantivy/issues/1944) (@Toasterson)
+- Made BooleanWeight and BoostWeight public [#1991](https://github.com/quickwit-oss/tantivy/issues/1991) (@fulmicoton)
+- Make index compatible with virtual drives on Windows [#1843](https://github.com/quickwit-oss/tantivy/issues/1843) (@gyk)
+- Add stop words for Hungarian language [#2069](https://github.com/quickwit-oss/tantivy/issues/2069) (@tnxbutno)
+- Auto downgrade index record option, instead of vint error [#1857](https://github.com/quickwit-oss/tantivy/issues/1857) (@PSeitz)
+- Enable range query on fast field for u64 compatible types [#1762](https://github.com/quickwit-oss/tantivy/issues/1762) (@PSeitz) [#1876]
+- sstable
+  - Isolating sstable and stacker in independent crates. [#1718](https://github.com/quickwit-oss/tantivy/issues/1718) (@fulmicoton)
+  - New sstable format [#1943](https://github.com/quickwit-oss/tantivy/issues/1943)[#1953](https://github.com/quickwit-oss/tantivy/issues/1953) (@trinity-1686a)
+  - Use DeltaReader directly to implement Dictionary::ord_to_term [#1928](https://github.com/quickwit-oss/tantivy/issues/1928) (@trinity-1686a)
+  - Use DeltaReader directly to implement Dictionary::term_ord [#1925](https://github.com/quickwit-oss/tantivy/issues/1925) (@trinity-1686a)
+- Add separate tokenizer manager for fast fields [#2019](https://github.com/quickwit-oss/tantivy/issues/2019) (@PSeitz)
+- Make construction of LevenshteinAutomatonBuilder for FuzzyTermQuery instances lazy. [#1756](https://github.com/quickwit-oss/tantivy/issues/1756) (@adamreichold)
+- Added support for madvise when opening an mmapped Index [#2036](https://github.com/quickwit-oss/tantivy/issues/2036) (@fulmicoton)
+- Rename `DatePrecision` to `DateTimePrecision` [#2051](https://github.com/quickwit-oss/tantivy/issues/2051) (@guilload)
+- Query Parser
+  - Quotation mark can now be used for phrase queries. [#2050](https://github.com/quickwit-oss/tantivy/issues/2050) (@fulmicoton)
+  - PhrasePrefixQuery is supported in the query parser via: `field:"phrase ter"*` [#2044](https://github.com/quickwit-oss/tantivy/issues/2044) (@adamreichold)
+- Docs
+  - Update examples for literate docs [#1880](https://github.com/quickwit-oss/tantivy/issues/1880) (@PSeitz)
+  - Add ip field example [#1775](https://github.com/quickwit-oss/tantivy/issues/1775) (@PSeitz)
+  - Fix doc store cache documentation [#1821](https://github.com/quickwit-oss/tantivy/issues/1821) (@PSeitz)
+  - Fix BooleanQuery document [#1999](https://github.com/quickwit-oss/tantivy/issues/1999) (@RT_Enzyme)
+  - Update comments in the faceted search example [#1737](https://github.com/quickwit-oss/tantivy/issues/1737) (@DawChihLiou)
+
+
+Tantivy 0.19
+================================
+#### Bugfixes
+- Fix missing fieldnorms for u64, i64, f64, bool, bytes and date [#1620](https://github.com/quickwit-oss/tantivy/pull/1620) (@PSeitz)
+- Fix interpolation overflow in linear interpolation fastfield codec [#1480](https://github.com/quickwit-oss/tantivy/pull/1480) (@PSeitz @fulmicoton)
+
+#### Features/Improvements
+- Add support for `IN` in queryparser , e.g. `field: IN [val1 val2 val3]` [#1683](https://github.com/quickwit-oss/tantivy/pull/1683) (@trinity-1686a)
+- Skip score calculation, when no scoring is required [#1646](https://github.com/quickwit-oss/tantivy/pull/1646) (@PSeitz)
+- Limit fast fields to u32 (`get_val(u32)`) [#1644](https://github.com/quickwit-oss/tantivy/pull/1644) (@PSeitz)
+- The `DateTime` type has been updated to hold timestamps with microseconds precision.
+  `DateOptions` and `DatePrecision` have been added to configure Date fields. The precision is used to hint on fast values compression. Otherwise, seconds precision is used everywhere else (i.e terms, indexing) [#1396](https://github.com/quickwit-oss/tantivy/pull/1396) (@evanxg852000)
+- Add IP address field type [#1553](https://github.com/quickwit-oss/tantivy/pull/1553) (@PSeitz)
+- Add boolean field type [#1382](https://github.com/quickwit-oss/tantivy/pull/1382) (@boraarslan)
+- Remove Searcher pool and make `Searcher` cloneable. (@PSeitz)
+- Validate settings on create [#1570](https://github.com/quickwit-oss/tantivy/pull/1570) (@PSeitz)
+- Detect and apply gcd on fastfield codecs [#1418](https://github.com/quickwit-oss/tantivy/pull/1418) (@PSeitz)
+- Doc store
+  - use separate thread to compress block store [#1389](https://github.com/quickwit-oss/tantivy/pull/1389) [#1510](https://github.com/quickwit-oss/tantivy/pull/1510) (@PSeitz @fulmicoton)
+  - Expose doc store cache size [#1403](https://github.com/quickwit-oss/tantivy/pull/1403) (@PSeitz)
+  - Enable compression levels for doc store [#1378](https://github.com/quickwit-oss/tantivy/pull/1378) (@PSeitz)
+  - Make block size configurable [#1374](https://github.com/quickwit-oss/tantivy/pull/1374) (@kryesh)
+- Make `tantivy::TantivyError` cloneable [#1402](https://github.com/quickwit-oss/tantivy/pull/1402) (@PSeitz)
+- Add support for phrase slop in query language [#1393](https://github.com/quickwit-oss/tantivy/pull/1393) (@saroh)
+- Aggregation
+  - Add aggregation support for date type [#1693](https://github.com/quickwit-oss/tantivy/pull/1693)(@PSeitz)
+  - Add support for keyed parameter in range and histogram aggregations [#1424](https://github.com/quickwit-oss/tantivy/pull/1424) (@k-yomo)
+  - Add aggregation bucket limit [#1363](https://github.com/quickwit-oss/tantivy/pull/1363) (@PSeitz)
+- Faster indexing
+  - [#1610](https://github.com/quickwit-oss/tantivy/pull/1610) (@PSeitz)
+  - [#1594](https://github.com/quickwit-oss/tantivy/pull/1594) (@PSeitz)
+  - [#1582](https://github.com/quickwit-oss/tantivy/pull/1582) (@PSeitz)
+  - [#1611](https://github.com/quickwit-oss/tantivy/pull/1611) (@PSeitz)
+  - Added a pre-configured stop word filter for various language [#1666](https://github.com/quickwit-oss/tantivy/pull/1666) (@adamreichold)
+
+Tantivy 0.18
+================================
+
+- For date values `chrono` has been replaced with `time` (@uklotzde) #1304 :
+  - The `time` crate is re-exported as `tantivy::time` instead of `tantivy::chrono`.
+  - The type alias `tantivy::DateTime` has been removed.
+  - `Value::Date` wraps `time::PrimitiveDateTime` without time zone information.
+  - Internally date/time values are stored as seconds since UNIX epoch in UTC.
+  - Converting a `time::OffsetDateTime` to `Value::Date` implicitly converts the value into UTC.
+    If this is not desired do the time zone conversion yourself and use `time::PrimitiveDateTime`
+    directly instead.
+- Add [histogram](https://github.com/quickwit-oss/tantivy/pull/1306) aggregation (@PSeitz)
+- Add support for fastfield on text fields (@PSeitz)
+- Add terms aggregation (@PSeitz)
+- Add support for zstd compression (@kryesh)
+
+Tantivy 0.18.1
+================================
+- Hotfix: positions computation.  #1629 (@fmassot, @fulmicoton, @PSeitz)
+
+Tantivy 0.17
+================================
+
+- LogMergePolicy now triggers merges if the ratio of deleted documents reaches a threshold (@shikhar @fulmicoton) [#115](https://github.com/quickwit-oss/tantivy/issues/115)
+- Adds a searcher Warmer API (@shikhar @fulmicoton)
+- Change to non-strict schema. Ignore fields in data which are not defined in schema. Previously this returned an error. #1211
+- Facets are necessarily indexed. Existing index with indexed facets should work out of the box. Index without facets that are marked with index: false should be broken (but they were already broken in a sense). (@fulmicoton) #1195 .
+- Bugfix that could in theory impact durability in theory on some filesystems [#1224](https://github.com/quickwit-oss/tantivy/issues/1224)
+- Schema now offers not indexing fieldnorms (@lpouget) [#922](https://github.com/quickwit-oss/tantivy/issues/922)
+- Reduce the number of fsync calls [#1225](https://github.com/quickwit-oss/tantivy/issues/1225)
+- Fix opening bytes index with dynamic codec (@PSeitz) [#1278](https://github.com/quickwit-oss/tantivy/issues/1278)
+- Added an aggregation collector for range, average and stats compatible with Elasticsearch. (@PSeitz)
+- Added a JSON schema type @fulmicoton [#1251](https://github.com/quickwit-oss/tantivy/issues/1251)
+- Added support for slop in phrase queries @halvorboe [#1068](https://github.com/quickwit-oss/tantivy/issues/1068)
+
+Tantivy 0.16.2
+================================
+
+- Bugfix in FuzzyTermQuery. (transposition_cost_one was not doing anything)
+
+Tantivy 0.16.1
+========================
+
+- Major Bugfix on multivalued fastfield.  #1151
+- Demux operation (@PSeitz)
+
+Tantivy 0.16.0
+=========================
+
+- Bugfix in the filesum check. (@evanxg852000) #1127
+- Bugfix in positions when the index is sorted by a field. (@appaquet) #1125
+
+Tantivy 0.15.3
+=========================
+
+- Major bugfix. Deleting documents was broken when the index was sorted by a field. (@appaquet, @fulmicoton) #1101
+
+Tantivy 0.15.2
+========================
+
+- Major bugfix. DocStore still panics when a deleted doc is at the beginning of a block. (@appaquet) #1088
+
+Tantivy 0.15.1
+=========================
+
+- Major bugfix. DocStore panics when first block is deleted. (@appaquet) #1077
+
+Tantivy 0.15.0
+=========================
+
+- API Changes. Using Range instead of (start, end) in the API and internals (`FileSlice`, `OwnedBytes`, `Snippets`, ...)
+  This change is breaking but migration is trivial.
+- Added an Histogram collector. (@fulmicoton) #994
+- Added support for Option<TCollector>.  (@fulmicoton)
+- DocAddress is now a struct (@scampi) #987
+- Bugfix consistent tie break handling in facet's topk (@hardikpnsp) #357
+- Date field support for range queries (@rihardsk) #516
+- Added lz4-flex as the default compression scheme in tantivy (@PSeitz) #1009
+- Renamed a lot of symbols to avoid all uppercasing on acronyms, as per new clippy recommendation. For instance, RAMDirectory -> RamDirectory. (@fulmicoton)
+- Simplified positions index format (@fulmicoton) #1022
+- Moved bitpacking to bitpacker subcrate and add BlockedBitpacker, which bitpacks blocks of 128 elements (@PSeitz) #1030
+- Added support for more-like-this query in tantivy (@evanxg852000) #1011
+- Added support for sorting an index, e.g presorting documents in an index by a timestamp field. This can heavily improve performance for certain scenarios, by utilizing the sorted data (Top-n optimizations)(@PSeitz). #1026
+- Add iterator over documents in doc store (@PSeitz). #1044
+- Fix log merge policy (@PSeitz). #1043
+- Add detection to avoid small doc store blocks on merge (@PSeitz). #1054
+- Make doc store compression dynamic (@PSeitz). #1060
+- Switch to json for footer version handling (@PSeitz). #1060
+- Updated TermMerger implementation to rely on the union feature of the FST (@scampi) #469
+- Add boolean marking whether position is required in the query_terms API call (@fulmicoton). #1070
+
+Tantivy 0.14.0
+=========================
+
+- Remove dependency to atomicwrites #833 .Implemented by @fulmicoton upon suggestion and research from @asafigan).
+- Migrated tantivy error from the now deprecated `failure` crate to `thiserror` #760. (@hirevo)
+- API Change. Accessing the typed value off a `Schema::Value` now returns an Option instead of panicking if the type does not match.
+- Large API Change in the Directory API. Tantivy used to assume that all files could be somehow memory mapped. After this change, Directory return a `FileSlice` that can be reduced and eventually read into an `OwnedBytes` object. Long and blocking io operation are still required by they do not span over the entire file.
+- Added support for Brotli compression in the DocStore. (@ppodolsky)
+- Added helper for building intersections and unions in BooleanQuery (@guilload)
+- Bugfix in `Query::explain`
+- Removed dependency on `notify` #924. Replaced with `FileWatcher` struct that polls meta file every 500ms in background thread. (@halvorboe @guilload)
+- Added `FilterCollector`, which wraps another collector and filters docs using a predicate over a fast field (@barrotsteindev)
+- Simplified the encoding of the skip reader struct. BlockWAND max tf is now encoded over a single byte. (@fulmicoton)
+- `FilterCollector` now supports all Fast Field value types (@barrotsteindev)
+- FastField are not all loaded when opening the segment reader. (@fulmicoton)
+- Added an API to merge segments, see `tantivy::merge_segments` #1005. (@evanxg852000)
+
+This version breaks compatibility and requires users to reindex everything.
+
+Tantivy 0.13.2
+===================
+
+Bugfix. Acquiring a facet reader on a segment that does not contain any
+doc with this facet returns `None`. (#896)
+
+Tantivy 0.13.1
+===================
+
+Made `Query` and `Collector` `Send + Sync`.
+Updated misc dependency versions.
+
+Tantivy 0.13.0
+======================
+
+Tantivy 0.13 introduce a change in the index format that will require
+you to reindex your index (BlockWAND information are added in the skiplist).
+The index size increase is minor as this information is only added for
+full blocks.
+If you have a massive index for which reindexing is not an option, please contact me
+so that we can discuss possible solutions.
+
+- Bugfix in `FuzzyTermQuery` not matching terms by prefix when it should (@Peachball)
+- Relaxed constraints on the custom/tweak score functions. At the segment level, they can be mut, and they are not required to be Sync + Send.
+- `MMapDirectory::open` does not return a `Result` anymore.
+- Change in the DocSet and Scorer API. (@fulmicoton).
+A freshly created DocSet point directly to their first doc. A sentinel value called TERMINATED marks the end of a DocSet.
+`.advance()` returns the new DocId. `Scorer::skip(target)` has been replaced by `Scorer::seek(target)` and returns the resulting DocId.
+As a result, iterating through DocSet now looks as follows
+
+```rust
+let mut doc = docset.doc();
+while doc != TERMINATED {
+   // ...
+   doc = docset.advance();
+}
+```
+
+The change made it possible to greatly simplify a lot of the docset's code.
+
+- Misc internal optimization and introduction of the `Scorer::for_each_pruning` function. (@fulmicoton)
+- Added an offset option to the Top(.*)Collectors. (@robyoung)
+- Added Block WAND. Performance on TOP-K on term-unions should be greatly increased. (@fulmicoton, and special thanks
+to the PISA team for answering all my questions!)
+
+Tantivy 0.12.0
+======================
+
+- Removing static dispatch in tokenizers for simplicity. (#762)
+- Added backward iteration for `TermDictionary` stream. (@halvorboe)
+- Fixed a performance issue when searching for the posting lists of a missing term (@audunhalland)
+- Added a configurable maximum number of docs (10M by default) for a segment to be considered for merge (@hntd187, landed by @halvorboe #713)
+- Important Bugfix #777, causing tantivy to retain memory mapping. (diagnosed by @poljar)
+- Added support for field boosting. (#547, @fulmicoton)
+
+## How to update?
+
+Crates relying on custom tokenizer, or registering tokenizer in the manager will require some
+minor changes. Check <https://github.com/quickwit-oss/tantivy/blob/main/examples/custom_tokenizer.rs>
+to check for some code sample.
+
+Tantivy 0.11.3
+=======================
+
+- Fixed DateTime as a fast field (#735)
+
+Tantivy 0.11.2
+=======================
+
+- The future returned by `IndexWriter::merge` does not borrow `self` mutably anymore (#732)
+- Exposing a constructor for `WatchHandle` (#731)
+
+Tantivy 0.11.1
+=====================
+
+- Bug fix #729
+
 Tantivy 0.11.0
 =====================
 
 - Added f64 field. Internally reuse u64 code the same way i64 does (@fdb-hiroshima)
+- Various bugfixes in the query parser.
+  - Better handling of hyphens in query parser. (#609)
+  - Better handling of whitespaces.
+- Closes #498 - add support for Elastic-style unbounded range queries for alphanumeric types eg. "title:>hello", "weight:>=70.5", "height:<200" (@petr-tik)
+- API change around `Box<BoxableTokenizer>`. See detail in #629
+- Avoid rebuilding Regex automaton whenever a regex query is reused. #639 (@brainlock)
+- Add footer with some metadata to index files. #605 (@fdb-hiroshima)
+- Add a method to check the compatibility of the footer in the index with the running version of tantivy (@petr-tik)
+- TopDocs collector: ensure stable sorting on equal score. #671 (@brainlock)
+- Added handling of pre-tokenized text fields (#642), which will enable users to
+  load tokens created outside tantivy. See usage in examples/pre_tokenized_text. (@kkoziara)
+- Fix crash when committing multiple times with deleted documents. #681 (@brainlock)
+
+## How to update?
+
+- The index format is changed. You are required to reindex your data to use tantivy 0.11.
+- `Box<dyn BoxableTokenizer>` has been replaced by a `BoxedTokenizer` struct.
+- Regex are now compiled when the `RegexQuery` instance is built. As a result, it can now return
+an error and handling the `Result` is required.
+- `tantivy::version()` now returns a `Version` object. This object implements `ToString()`
+
+Tantivy 0.10.2
+=====================
+
+- Closes #656. Solving memory leak.
 
 Tantivy 0.10.1
 =====================
@@ -10,32 +523,31 @@ Tantivy 0.10.1
 Avoid watching the mmap directory until someone effectively creates a reader that uses
 this functionality.
 
-
 Tantivy 0.10.0
 =====================
 
 *Tantivy 0.10.0 index format is compatible with the index format in 0.9.0.*
 
-- Added an API to easily tweak or entirely replace the 
- default score. See `TopDocs::tweak_score`and `TopScore::custom_score` (@pmasurel)
+- Added an API to easily tweak or entirely replace the
+ default score. See `TopDocs::tweak_score`and `TopScore::custom_score` (@fulmicoton)
 - Added an ASCII folding filter (@drusellers)
-- Bugfix in `query.count` in presence of deletes (@pmasurel)
-- Added `.explain(...)` in `Query` and `Weight` to (@pmasurel)
-- Added an efficient way to `delete_all_documents` in `IndexWriter` (@petr-tik). 
+- Bugfix in `query.count` in presence of deletes (@fulmicoton)
+- Added `.explain(...)` in `Query` and `Weight` to (@fulmicoton)
+- Added an efficient way to `delete_all_documents` in `IndexWriter` (@petr-tik).
   All segments are simply removed.
 
 Minor
 ---------
+
 - Switched to Rust 2018 (@uvd)
-- Small simplification of the code. 
+- Small simplification of the code.
 Calling .freq() or .doc() when .advance() has never been called
 on segment postings should panic from now on.
 - Tokens exceeding `u16::max_value() - 4` chars are discarded silently instead of panicking.
 - Fast fields are now preloaded when the `SegmentReader` is created.
 - `IndexMeta` is now public.  (@hntd187)
-- `IndexWriter` `add_document`, `delete_term`. `IndexWriter` is `Sync`, making it possible to use it with a `
-Arc<RwLock<IndexWriter>>`. `add_document` and `delete_term` can 
-only require a read lock. (@pmasurel)
+- `IndexWriter` `add_document`, `delete_term`. `IndexWriter` is `Sync`, making it possible to use it with a `Arc<RwLock<IndexWriter>>`. `add_document` and `delete_term` can
+only require a read lock. (@fulmicoton)
 - Introducing `Opstamp` as an expressive type alias for `u64`. (@petr-tik)
 - Stamper now relies on `AtomicU64` on all platforms (@petr-tik)
 - Bugfix - Files get deleted slightly earlier
@@ -49,25 +561,26 @@ Your program should be usable as is.
 
 Fast fields used to be accessed directly from the `SegmentReader`.
 The API changed, you are now required to acquire your fast field reader via the
-`segment_reader.fast_fields()`, and use one of the typed method: 
+`segment_reader.fast_fields()`, and use one of the typed method:
+
 - `.u64()`, `.i64()` if your field is single-valued ;
 - `.u64s()`, `.i64s()` if your field is multi-valued ;
 - `.bytes()` if your field is bytes fast field.
 
-
-
 Tantivy 0.9.0
 =====================
-*0.9.0 index format is not compatible with the 
+
+*0.9.0 index format is not compatible with the
 previous index format.*
-- MAJOR BUGFIX : 
+
+- MAJOR BUGFIX :
   Some `Mmap` objects were being leaked, and would never get released. (@fulmicoton)
 - Removed most unsafe (@fulmicoton)
 - Indexer memory footprint improved. (VInt comp, inlining the first block. (@fulmicoton)
 - Stemming in other language possible (@pentlander)
 - Segments with no docs are deleted earlier (@barrotsteindev)
-- Added grouped add and delete operations. 
-  They are guaranteed to happen together (i.e. they cannot be split by a commit). 
+- Added grouped add and delete operations.
+  They are guaranteed to happen together (i.e. they cannot be split by a commit).
   In addition, adds are guaranteed to happen on the same segment. (@elbow-jason)
 - Removed `INT_STORED` and `INT_INDEXED`. It is now possible to use `STORED` and `INDEXED`
   for int fields. (@fulmicoton)
@@ -81,59 +594,62 @@ tantivy 0.9 brought some API breaking change.
 To update from tantivy 0.8, you will need to go through the following steps.
 
 - `schema::INT_INDEXED` and `schema::INT_STORED`  should be replaced by `schema::INDEXED` and `schema::INT_STORED`.
-- The index now does not hold the pool of searcher anymore. You are required to create an intermediary object called 
-`IndexReader` for this. 
-    
+- The index now does not hold the pool of searcher anymore. You are required to create an intermediary object called
+`IndexReader` for this.
+
     ```rust
     // create the reader. You typically need to create 1 reader for the entire
     // lifetime of you program.
     let reader = index.reader()?;
-    
+
     // Acquire a searcher (previously `index.searcher()`) is now written:
     let searcher = reader.searcher();
-    
-    // With the default setting of the reader, you are not required to 
+
+    // With the default setting of the reader, you are not required to
     // call `index.load_searchers()` anymore.
     //
     // The IndexReader will pick up that change automatically, regardless
     // of whether the update was done in a different process or not.
-    // If this behavior is not wanted, you can create your reader with 
+    // If this behavior is not wanted, you can create your reader with
     // the `ReloadPolicy::Manual`, and manually decide when to reload the index
     // by calling `reader.reload()?`.
-  
-    ```
 
+    ```
 
 Tantivy 0.8.2
 =====================
+
 Fixing build for x86_64 platforms. (#496)
 No need to update from 0.8.1 if tantivy
 is building on your platform.
 
-
 Tantivy 0.8.1
 =====================
+
 Hotfix of #476.
 
-Merge was reflecting deletes before commit was passed. 
+Merge was reflecting deletes before commit was passed.
 Thanks @barrotsteindev  for reporting the bug.
 
-
 Tantivy 0.8.0
 =====================
-*No change in the index format*
-- API Breaking change in the collector API. (@jwolfe, @fulmicoton)
-- Multithreaded search (@jwolfe, @fulmicoton) 
 
+*No change in the index format*
+
+- API Breaking change in the collector API. (@jwolfe, @fulmicoton)
+- Multithreaded search (@jwolfe, @fulmicoton)
 
 Tantivy 0.7.1
 =====================
+
 *No change in the index format*
+
 - Bugfix: NGramTokenizer panics on non ascii chars
 - Added a space usage API
 
 Tantivy 0.7
 =====================
+
 - Skip data for doc ids and positions (@fulmicoton),
   greatly improving performance
 - Tantivy error now rely on the failure crate (@drusellers)
@@ -143,15 +659,15 @@ Tantivy 0.7
 
 Tantivy 0.6.1
 =========================
+
 - Bugfix #324. GC removing was removing file that were still in useful
 - Added support for parsing AllQuery and RangeQuery via QueryParser
-    - AllQuery: `*`
-    - RangeQuery:
-        - Inclusive `field:[startIncl to endIncl]`
-        - Exclusive `field:{startExcl to endExcl}`
-        - Mixed `field:[startIncl to endExcl}` and vice versa
-        - Unbounded `field:[start to *]`, `field:[* to end]`
- 
+  - AllQuery: `*`
+  - RangeQuery:
+    - Inclusive `field:[startIncl to endIncl]`
+    - Exclusive `field:{startExcl to endExcl}`
+    - Mixed `field:[startIncl to endExcl}` and vice versa
+    - Unbounded `field:[start to *]`, `field:[* to end]`
 
 Tantivy 0.6
 ==========================
@@ -159,70 +675,65 @@ Tantivy 0.6
 Special thanks to @drusellers and @jason-wolfe for their contributions
 to this release!
 
-- Removed C code. Tantivy is now pure Rust. (@pmasurel)
-- BM25 (@pmasurel)
-- Approximate field norms encoded over 1 byte. (@pmasurel)
-- Compiles on stable rust (@pmasurel)
+- Removed C code. Tantivy is now pure Rust. (@fulmicoton)
+- BM25 (@fulmicoton)
+- Approximate field norms encoded over 1 byte. (@fulmicoton)
+- Compiles on stable rust (@fulmicoton)
 - Add &[u8] fastfield for associating arbitrary bytes to each document (@jason-wolfe) (#270)
-    - Completely uncompressed
-    - Internally: One u64 fast field for indexes, one fast field for the bytes themselves.
+  - Completely uncompressed
+  - Internally: One u64 fast field for indexes, one fast field for the bytes themselves.
 - Add NGram token support (@drusellers)
 - Add Stopword Filter support (@drusellers)
 - Add a FuzzyTermQuery (@drusellers)
 - Add a RegexQuery (@drusellers)
-- Various performance improvements (@pmasurel)_
-
+- Various performance improvements (@fulmicoton)_
 
 Tantivy 0.5.2
 ===========================
+
 - bugfix #274
 - bugfix #280
 - bugfix #289
 
-
 Tantivy 0.5.1
 ==========================
-- bugfix #254 : tantivy failed if no documents in a segment contained a specific field.
 
+- bugfix #254 : tantivy failed if no documents in a segment contained a specific field.
 
 Tantivy 0.5
 ==========================
+
 - Faceting
 - RangeQuery
 - Configurable tokenization pipeline
 - Bugfix in PhraseQuery
 - Various query optimisation
 - Allowing very large indexes
-    - 64 bits file address
-    - Smarter encoding of the `TermInfo` objects
-
-
+  - 64 bits file address
+  - Smarter encoding of the `TermInfo` objects
 
 Tantivy 0.4.3
 ==========================
 
 - Bugfix race condition when deleting files. (#198)
 
-
 Tantivy 0.4.2
 ==========================
 
 - Prevent usage of AVX2 instructions (#201)
 
-
 Tantivy 0.4.1
 ==========================
 
 - Bugfix for non-indexed fields. (#199)
 
-
 Tantivy 0.4.0
 ==========================
 
 - Raise the limit of number of fields (previously 256 fields) (@fulmicoton)
 - Removed u32 fields. They are replaced by u64 and i64 fields (#65) (@fulmicoton)
 - Optimized skip in SegmentPostings (#130) (@lnicola)
-- Replacing rustc_serialize by serde. Kudos to @KodrAus and @lnicola
+- Replacing rustc_serialize by serde. Kudos to  benchmark@KodrAus and @lnicola
 - Using error-chain (@KodrAus)
 - QueryParser: (@fulmicoton)
   - Explicit error returned when searched for a term that is not indexed
@@ -230,37 +741,31 @@ Tantivy 0.4.0
   - Searching for a non-indexed field returns an explicit Error
   - Phrase query for non-tokenized field are not tokenized by the query parser.
 - Faster/Better indexing (@fulmicoton)
-    - using murmurhash2
-    - faster merging
-    - more memory efficient fast field writer (@lnicola )
-    - better handling of collisions
-    - lesser memory usage
+  - using murmurhash2
+  - faster merging
+  - more memory efficient fast field writer (@lnicola )
+  - better handling of collisions
+  - lesser memory usage
 - Added API, most notably to iterate over ranges of terms (@fulmicoton)
 - Bugfix that was preventing to unmap segment files, on index drop (@fulmicoton)
 - Made the doc! macro public (@fulmicoton)
 - Added an alternative implementation of the streaming dictionary (@fulmicoton)
 
-
-
 Tantivy 0.3.1
 ==========================
 
 - Expose a method to trigger files garbage collection
 
-
-
 Tantivy 0.3
 ==========================
 
-
 Special thanks to @Kodraus @lnicola @Ameobea @manuel-woelker @celaus
 for their contribution to this release.
 
 Thanks also to everyone in tantivy gitter chat
 for their advise and company :)
 
-https://gitter.im/tantivy-search/tantivy
-
+<https://gitter.im/tantivy-search/tantivy>
 
 Warning:
 
@@ -269,19 +774,16 @@ code and index format.
 You should not expect backward compatibility before
 tantivy 1.0.
 
-
-
 New Features
 ------------
 
 - Delete. You can now delete documents from an index.
 - Support for windows (Thanks to @lnicola)
 
-
 Various Bugfixes & small improvements
 ----------------------------------------
 
-- Added CI for Windows (https://ci.appveyor.com/project/fulmicoton/tantivy)
+- Added CI for Windows (<https://ci.appveyor.com/project/fulmicoton/tantivy>)
 Thanks to @KodrAus ! (#108)
 - Various dependy version update (Thanks to @Ameobea) #76
 - Fixed several race conditions in `Index.wait_merge_threads`
@@ -293,7 +795,3 @@ Thanks to @KodrAus ! (#108)
 - Building binary targets for tantivy-cli (Thanks to @KodrAus)
 - Misc invisible bug fixes, and code cleanup.
 - Use
-
-
-
-

CITATION.cff

@@ -0,0 +1,10 @@
+cff-version: 1.2.0
+message: "If you use this software, please cite it as below."
+authors:
+  - alias: Quickwit Inc.
+    website: "https://quickwit.io"
+title: "tantivy"
+version: 0.22.0
+doi: 10.5281/zenodo.13942948
+date-released: 2024-10-17
+url: "https://github.com/quickwit-oss/tantivy"

Cargo.toml

@@ -1,92 +1,148 @@
 [package]
 name = "tantivy"
-version = "0.10.1"
+version = "0.24.0"
 authors = ["Paul Masurel <paul.masurel@gmail.com>"]
 license = "MIT"
 categories = ["database-implementations", "data-structures"]
 description = """Search engine library"""
-documentation = "https://tantivy-search.github.io/tantivy/tantivy/index.html"
-homepage = "https://github.com/tantivy-search/tantivy"
-repository = "https://github.com/tantivy-search/tantivy"
+documentation = "https://docs.rs/tantivy/"
+homepage = "https://github.com/quickwit-oss/tantivy"
+repository = "https://github.com/quickwit-oss/tantivy"
 readme = "README.md"
 keywords = ["search", "information", "retrieval"]
-edition = "2018"
+edition = "2021"
+rust-version = "1.85"
+exclude = ["benches/*.json", "benches/*.txt"]
 
 [dependencies]
-base64 = "0.10.0"
-byteorder = "1.0"
-once_cell = "0.2"
-regex = "1.0"
-tantivy-fst = "0.1"
-memmap = {version = "0.7", optional=true}
-lz4 = {version="1.20", optional=true}
-snap = {version="0.2"}
-atomicwrites = {version="0.2.2", optional=true}
-tempfile = "3.0"
-log = "0.4"
-combine = ">=3.6.0,<4.0.0"
-tempdir = "0.3"
-serde = "1.0"
-serde_derive = "1.0"
-serde_json = "1.0"
-num_cpus = "1.2"
-fs2={version="0.4", optional=true}
-itertools = "0.8"
-levenshtein_automata = {version="0.1", features=["fst_automaton"]}
-notify = {version="4", optional=true}
-bit-set = "0.5"
-uuid = { version = "0.7.2", features = ["v4", "serde"] }
-crossbeam = "0.5"
-futures = "0.1"
-futures-cpupool = "0.1"
-owning_ref = "0.4"
-stable_deref_trait = "1.0.0"
-rust-stemmers = "1.1"
-downcast-rs = { version="1.0" }
-bitpacking = {version="0.8", default-features = false, features=["bitpacker4x"]}
-census = "0.2"
-fnv = "1.0.6"
-owned-read = "0.4"
-failure = "0.1"
+oneshot = "0.1.7"
+base64 = "0.22.0"
+byteorder = "1.4.3"
+crc32fast = "1.3.2"
+once_cell = "1.10.0"
+regex = { version = "1.5.5", default-features = false, features = [
+    "std",
+    "unicode",
+] }
+aho-corasick = "1.0"
+tantivy-fst = "0.5"
+memmap2 = { version = "0.9.0", optional = true }
+lz4_flex = { version = "0.11", 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 = "0.8.0", optional = true }
+levenshtein_automata = "0.2.1"
+uuid = { version = "1.0.0", features = ["v4", "serde"] }
+crossbeam-channel = "0.5.4"
+rust-stemmers = "1.2.0"
+downcast-rs = "2.0.1"
+bitpacking = { version = "0.9.2", default-features = false, features = [
+    "bitpacker4x",
+] }
+census = "0.4.2"
+rustc-hash = "2.0.0"
+thiserror = "2.0.1"
 htmlescape = "0.3.1"
-fail = "0.3"
-scoped-pool = "1.0"
-murmurhash32 = "0.2"
-chrono = "0.4"
-smallvec = "0.6"
+fail = { version = "0.5.0", optional = true }
+time = { version = "0.3.35", features = ["serde-well-known"] }
+smallvec = "1.8.0"
+rayon = "1.5.2"
+lru = "0.12.0"
+fastdivide = "0.4.0"
+itertools = "0.14.0"
+measure_time = "0.9.0"
+arc-swap = "1.5.0"
+bon = "3.3.1"
+
+columnar = { version = "0.5", path = "./columnar", package = "tantivy-columnar" }
+sstable = { version = "0.5", path = "./sstable", package = "tantivy-sstable", optional = true }
+stacker = { version = "0.5", path = "./stacker", package = "tantivy-stacker" }
+query-grammar = { version = "0.24.0", path = "./query-grammar", package = "tantivy-query-grammar" }
+tantivy-bitpacker = { version = "0.8", path = "./bitpacker" }
+common = { version = "0.9", path = "./common/", package = "tantivy-common" }
+tokenizer-api = { version = "0.5", path = "./tokenizer-api", package = "tantivy-tokenizer-api" }
+sketches-ddsketch = { version = "0.3.0", features = ["use_serde"] }
+hyperloglogplus = { version = "0.4.1", features = ["const-loop"] }
+futures-util = { version = "0.3.28", optional = true }
+futures-channel = { version = "0.3.28", optional = true }
+fnv = "1.0.7"
 
 [target.'cfg(windows)'.dependencies]
-winapi = "0.3"
+winapi = "0.3.9"
 
 [dev-dependencies]
-rand = "0.7"
-maplit = "1"
-matches = "0.1.8"
-time = "0.1.42"
+binggan = "0.14.0"
+rand = "0.8.5"
+maplit = "1.0.2"
+matches = "0.1.9"
+pretty_assertions = "1.2.1"
+proptest = "1.0.0"
+test-log = "0.2.10"
+futures = "0.3.21"
+paste = "1.0.11"
+more-asserts = "0.3.1"
+rand_distr = "0.4.3"
+time = { version = "0.3.10", features = ["serde-well-known", "macros"] }
+postcard = { version = "1.0.4", features = [
+  "use-std",
+], default-features = false }
+
+[target.'cfg(not(windows))'.dev-dependencies]
+criterion = { version = "0.5", default-features = false }
+
+[dev-dependencies.fail]
+version = "0.5.0"
+features = ["failpoints"]
 
 [profile.release]
 opt-level = 3
 debug = false
 debug-assertions = false
 
+[profile.bench]
+opt-level = 3
+debug = true
+debug-assertions = false
+
 [profile.test]
 debug-assertions = true
 overflow-checks = true
 
 [features]
-default = ["mmap"]
-mmap = ["atomicwrites", "fs2", "memmap", "notify"]
-lz4-compression = ["lz4"]
-failpoints = ["fail/failpoints"]
-unstable = [] # useful for benches.
-wasm-bindgen = ["uuid/wasm-bindgen"]
+default = ["mmap", "stopwords", "lz4-compression", "columnar-zstd-compression"]
+mmap = ["fs4", "tempfile", "memmap2"]
+stopwords = []
 
-[badges]
-travis-ci = { repository = "tantivy-search/tantivy" }
+lz4-compression = ["lz4_flex"]
+zstd-compression = ["zstd"]
 
-[dev-dependencies.fail]
-features = ["failpoints"]
+# enable zstd-compression in columnar (and sstable)
+columnar-zstd-compression = ["columnar/zstd-compression"]
 
+failpoints = ["fail", "fail/failpoints"]
+unstable = []                            # useful for benches.
+
+quickwit = ["sstable", "futures-util", "futures-channel"]
+
+# Compares only the hash of a string when indexing data.
+# Increases indexing speed, but may lead to extremely rare missing terms, when there's a hash collision.
+# Uses 64bit ahash.
+compare_hash_only = ["stacker/compare_hash_only"]
+
+[workspace]
+members = [
+    "query-grammar",
+    "bitpacker",
+    "common",
+    "ownedbytes",
+    "stacker",
+    "sstable",
+    "tokenizer-api",
+    "columnar",
+]
 
 # Following the "fail" crate best practises, we isolate
 # tests that define specific behavior in fail check points
@@ -98,4 +154,16 @@ features = ["failpoints"]
 [[test]]
 name = "failpoints"
 path = "tests/failpoints/mod.rs"
-required-features = ["fail/failpoints"]
+required-features = ["failpoints"]
+
+[[bench]]
+name = "analyzer"
+harness = false
+
+[[bench]]
+name = "index-bench"
+harness = false
+
+[[bench]]
+name = "agg_bench"
+harness = false

Makefile

@@ -0,0 +1,6 @@
+test:
+	@echo "Run test only... No examples."
+	cargo test --tests --lib
+
+fmt:
+	cargo +nightly fmt --all

README.md

@@ -1,138 +1,144 @@
-
-[![Build Status](https://travis-ci.org/tantivy-search/tantivy.svg?branch=master)](https://travis-ci.org/tantivy-search/tantivy)
-[![codecov](https://codecov.io/gh/tantivy-search/tantivy/branch/master/graph/badge.svg)](https://codecov.io/gh/tantivy-search/tantivy)
-[![Join the chat at https://gitter.im/tantivy-search/tantivy](https://badges.gitter.im/tantivy-search/tantivy.svg)](https://gitter.im/tantivy-search/tantivy?utm_source=badge&utm_medium=badge&utm_campaign=pr-badge&utm_content=badge)
+[![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)
+[![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)
-[![Build status](https://ci.appveyor.com/api/projects/status/r7nb13kj23u8m9pj/branch/master?svg=true)](https://ci.appveyor.com/project/fulmicoton/tantivy/branch/master)
 [![Crates.io](https://img.shields.io/crates/v/tantivy.svg)](https://crates.io/crates/tantivy)
-[![Say Thanks!](https://img.shields.io/badge/Say%20Thanks-!-1EAEDB.svg)](https://saythanks.io/to/fulmicoton)
 
-![Tantivy](https://tantivy-search.github.io/logo/tantivy-logo.png)
+<img src="https://tantivy-search.github.io/logo/tantivy-logo.png" alt="Tantivy, the fastest full-text search engine library written in Rust" height="250">
 
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/0)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/0)
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/1)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/1)
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/2)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/2)
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/3)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/3)
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/4)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/4)
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/5)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/5)
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/6)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/6)
-[![](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/images/7)](https://sourcerer.io/fame/fulmicoton/tantivy-search/tantivy/links/7)
+## Fast full-text search engine library written in Rust
 
-[![Become a patron](https://c5.patreon.com/external/logo/become_a_patron_button.png)](https://www.patreon.com/fulmicoton)
+**If you are looking for an alternative to Elasticsearch or Apache Solr, check out [Quickwit](https://github.com/quickwit-oss/quickwit), our distributed search engine built on top of Tantivy.**
 
-
-**Tantivy** is a **full text search engine library** written in rust.
-
-It is closer to [Apache Lucene](https://lucene.apache.org/) than to [Elasticsearch](https://www.elastic.co/products/elasticsearch) and [Apache Solr](https://lucene.apache.org/solr/) in the sense it is not
-an off-the-shelf search engine server, but rather a crate that can be used
-to build such a search engine.
+Tantivy is closer to [Apache Lucene](https://lucene.apache.org/) than to [Elasticsearch](https://www.elastic.co/products/elasticsearch) or [Apache Solr](https://lucene.apache.org/solr/) in the sense it is not
+an off-the-shelf search engine server, but rather a crate that can be used to build such a search engine.
 
 Tantivy is, in fact, strongly inspired by Lucene's design.
 
-# Benchmark
+## Benchmark
 
-Tantivy is typically faster than Lucene, but the results will depend on 
-the nature of the queries in your workload.
+The following [benchmark](https://tantivy-search.github.io/bench/) breaks down the
+performance for different types of queries/collections.
 
-The following [benchmark](https://tantivy-search.github.io/bench/) break downs 
-performance for different type of queries / collection.
+Your mileage WILL vary depending on the nature of queries and their load.
 
-# Features
+<img src="doc/assets/images/searchbenchmark.png">
+
+Details about the benchmark can be found at this [repository](https://github.com/quickwit-oss/search-benchmark-game).
+
+## Features
 
 - Full-text search
-- Configurable tokenizer. (stemming available for 17 latin languages. Third party support for Chinese ([tantivy-jieba](https://crates.io/crates/tantivy-jieba) and [cang-jie](https://crates.io/crates/cang-jie)) and [Japanese](https://crates.io/crates/tantivy-tokenizer-tiny-segmenter)
+- Configurable tokenizer (stemming available for 17 Latin languages) with third party support for Chinese ([tantivy-jieba](https://crates.io/crates/tantivy-jieba) and [cang-jie](https://crates.io/crates/cang-jie)), Japanese ([lindera](https://github.com/lindera-morphology/lindera-tantivy), [Vaporetto](https://crates.io/crates/vaporetto_tantivy), and [tantivy-tokenizer-tiny-segmenter](https://crates.io/crates/tantivy-tokenizer-tiny-segmenter)) and Korean ([lindera](https://github.com/lindera-morphology/lindera-tantivy) + [lindera-ko-dic-builder](https://github.com/lindera-morphology/lindera-ko-dic-builder))
 - Fast (check out the :racehorse: :sparkles: [benchmark](https://tantivy-search.github.io/bench/) :sparkles: :racehorse:)
-- Tiny startup time (<10ms), perfect for command line tools
-- BM25 scoring (the same as lucene)
-- Natural query language `(michael AND jackson) OR "king of pop"`
-- Phrase queries search (`"michael jackson"`)
+- Tiny startup time (<10ms), perfect for command-line tools
+- BM25 scoring (the same as Lucene)
+- Natural query language (e.g. `(michael AND jackson) OR "king of pop"`)
+- Phrase queries search (e.g. `"michael jackson"`)
 - Incremental indexing
 - Multithreaded indexing (indexing English Wikipedia takes < 3 minutes on my desktop)
 - Mmap directory
-- SIMD integer compression when the platform/CPU includes the SSE2 instruction set.
-- Single valued and multivalued u64, i64 and f64 fast fields (equivalent of doc values in Lucene)
+- SIMD integer compression when the platform/CPU includes the SSE2 instruction set
+- Single valued and multivalued u64, i64, and f64 fast fields (equivalent of doc values in Lucene)
 - `&[u8]` fast fields
-- Text, i64, u64, f64, dates and hierarchical facet fields
-- LZ4 compressed document store
+- Text, i64, u64, f64, dates, ip, bool, and hierarchical facet fields
+- Compressed document store (LZ4, Zstd, None)
 - Range queries
 - Faceted search
 - Configurable indexing (optional term frequency and position indexing)
+- JSON Field
+- Aggregation Collector: histogram, range buckets, average, and stats metrics
+- LogMergePolicy with deletes
+- Searcher Warmer API
 - Cheesy logo with a horse
 
-# Non-features
+### Non-features
 
-- Distributed search is out of the scope of tantivy. That being said, tantivy is meant as a
-library upon which one could build a distributed search. Serializable/mergeable collector state for instance, 
-are within the scope of tantivy.
+Distributed search is out of the scope of Tantivy, but if you are looking for this feature, check out [Quickwit](https://github.com/quickwit-oss/quickwit/).
 
-# Supported OS and compiler
+## Getting started
 
-Tantivy works on stable rust (>= 1.27) and supports Linux, MacOS and Windows.
+Tantivy works on stable Rust and supports Linux, macOS, and Windows.
 
-# Getting started
+- [Tantivy's simple search example](https://tantivy-search.github.io/examples/basic_search.html)
+- [tantivy-cli and its tutorial](https://github.com/quickwit-oss/tantivy-cli) - `tantivy-cli` is an actual command-line interface that makes it easy for you to create a search engine,
+index documents, and search via the CLI or a small server with a REST API.
+It walks you through getting a Wikipedia search engine up and running in a few minutes.
+- [Reference doc for the last released version](https://docs.rs/tantivy/)
 
-- [tantivy's simple search example](https://tantivy-search.github.io/examples/basic_search.html)
-- [tantivy-cli and its tutorial](https://github.com/tantivy-search/tantivy-cli).
-`tantivy-cli` is an actual command line interface that makes it easy for you to create a search engine,
-index documents and search via the CLI or a small server with a REST API.
-It will walk you through getting a wikipedia search engine up and running in a few minutes.
-- [reference doc for the last released version](https://docs.rs/tantivy/)
+## How can I support this project?
 
-# How can I support this project?
+There are many ways to support this project.
 
-There are many ways to support this project. 
-
-- Use tantivy and tell us about your experience on [gitter](https://gitter.im/tantivy-search/tantivy) or by email (paul.masurel@gmail.com)
+- Use Tantivy and tell us about your experience on [Discord](https://discord.gg/MT27AG5EVE) or by email (paul.masurel@gmail.com)
 - Report bugs
 - Write a blog post
 - Help with documentation by asking questions or submitting PRs
-- Contribute code (you can join [our gitter](https://gitter.im/tantivy-search/tantivy) )
-- Talk about tantivy around you
-- Drop a word on on [![Say Thanks!](https://img.shields.io/badge/Say%20Thanks-!-1EAEDB.svg)](https://saythanks.io/to/fulmicoton) or even [![Become a patron](https://c5.patreon.com/external/logo/become_a_patron_button.png)](https://www.patreon.com/fulmicoton)
+- Contribute code (you can join [our Discord server](https://discord.gg/MT27AG5EVE))
+- Talk about Tantivy around you
 
-# Contributing code
+## Contributing code
 
-We use the GitHub Pull Request workflow - reference a GitHub ticket and/or include a comprehensive commit message when opening a PR.
+We use the GitHub Pull Request workflow: reference a GitHub ticket and/or include a comprehensive commit message when opening a PR.
+Feel free to update CHANGELOG.md with your contribution.
 
-## Clone and build locally
+### Tokenizer
 
-Tantivy compiles on stable rust but requires `Rust >= 1.27`.
-To check out and run tests, you can simply run :
+When implementing a tokenizer for tantivy depend on the `tantivy-tokenizer-api` crate.
+
+### Clone and build locally
+
+Tantivy compiles on stable Rust.
+To check out and run tests, you can simply run:
 
 ```bash
-    git clone https://github.com/tantivy-search/tantivy.git
-    cd tantivy
-    cargo build
+git clone https://github.com/quickwit-oss/tantivy.git
+cd tantivy
+cargo test
 ```
 
-## Run tests
+## Companies Using Tantivy
 
-Some tests will not run with just `cargo test` because of `fail-rs`.
-To run the tests exhaustively, run `./run-tests.sh`
+<p align="left">
+<img align="center" src="doc/assets/images/etsy.png" alt="Etsy" height="25" width="auto" /> &nbsp;
+<img align="center" src="doc/assets/images/paradedb.png" alt="ParadeDB" height="25" width="auto" /> &nbsp;
+<img align="center" src="doc/assets/images/Nuclia.png#gh-light-mode-only" alt="Nuclia" height="25" width="auto" /> &nbsp;
+<img align="center" src="doc/assets/images/humanfirst.png#gh-light-mode-only" alt="Humanfirst.ai" height="30" width="auto" />
+<img align="center" src="doc/assets/images/element.io.svg#gh-light-mode-only" alt="Element.io" height="25" width="auto" />
+<img align="center" src="doc/assets/images/nuclia-dark-theme.png#gh-dark-mode-only" alt="Nuclia" height="35" width="auto" /> &nbsp;
+<img align="center" src="doc/assets/images/humanfirst.ai-dark-theme.png#gh-dark-mode-only" alt="Humanfirst.ai" height="25" width="auto" />&nbsp; &nbsp;
+<img align="center" src="doc/assets/images/element-dark-theme.png#gh-dark-mode-only" alt="Element.io" height="25" width="auto" />
+</p>
 
-## Debug
+## FAQ
 
-You might find it useful to step through the programme with a debugger.
+### Can I use Tantivy in other languages?
 
-### A failing test
+- Python → [tantivy-py](https://github.com/quickwit-oss/tantivy-py)
+- Ruby → [tantiny](https://github.com/baygeldin/tantiny)
 
-Make sure you haven't run `cargo clean` after the most recent `cargo test` or `cargo build` to guarantee that `target/` dir exists. Use this bash script to find the most name of the most recent debug build of tantivy and run it under rust-gdb.
+You can also find other bindings on [GitHub](https://github.com/search?q=tantivy) but they may be less maintained.
 
-```bash
-find target/debug/ -maxdepth 1 -executable -type f -name "tantivy*" -printf '%TY-%Tm-%Td %TT %p\n' | sort -r | cut -d " " -f 3 | xargs -I RECENT_DBG_TANTIVY rust-gdb RECENT_DBG_TANTIVY
-```
+### What are some examples of Tantivy use?
 
-Now that you are in rust-gdb, you can set breakpoints on lines and methods that match your source-code and run the debug executable with flags that you normally pass to `cargo test` to like this
+- [seshat](https://github.com/matrix-org/seshat/): A matrix message database/indexer
+- [tantiny](https://github.com/baygeldin/tantiny): Tiny full-text search for Ruby
+- [lnx](https://github.com/lnx-search/lnx): adaptable, typo tolerant search engine with a REST API
+- and [more](https://github.com/search?q=tantivy)!
 
-```bash
-$gdb run --test-threads 1 --test $NAME_OF_TEST
-```
+### On average, how much faster is Tantivy compared to Lucene?
 
-### An example
+- According to our [search latency benchmark](https://tantivy-search.github.io/bench/), Tantivy is approximately 2x faster than Lucene.
 
-By default, rustc compiles everything in the `examples/` dir in debug mode. This makes it easy for you to make examples to reproduce bugs.
+### Does tantivy support incremental indexing?
 
-```bash
-rust-gdb target/debug/examples/$EXAMPLE_NAME
-$ gdb run
-```
+- Yes.
+
+### How can I edit documents?
+
+- Data in tantivy is immutable. To edit a document, the document needs to be deleted and reindexed.
+
+### When will my documents be searchable during indexing?
+
+- Documents will be searchable after a `commit` is called on an `IndexWriter`. Existing `IndexReader`s will also need to be reloaded in order to reflect the changes. Finally, changes are only visible to newly acquired `Searcher`.

RELEASE.md

@@ -0,0 +1,21 @@
+# Release a new Tantivy Version
+
+## Steps
+
+1. Identify new packages in workspace since last release
+2. Identify changed packages in workspace since last release
+3. Bump version in `Cargo.toml` and their dependents for all changed packages
+4. Update version of root `Cargo.toml`
+5. Publish version starting with leaf nodes
+6. Set git tag with new version
+
+
+In conjucation with `cargo-release` Steps 1-4 (I'm not sure if the change detection works):
+Set new packages to version 0.0.0
+
+Replace prev-tag-name
+```bash
+cargo release --workspace --no-publish -v --prev-tag-name 0.19 --push-remote origin minor --no-tag --execute
+```
+
+no-tag or it will create tags for all the subpackages

TODO.txt

@@ -0,0 +1,18 @@
+Make schema_builder API fluent.
+fix doc serialization and prevent compression problems
+
+u64 , etc. should return Result<Option> now that we support optional missing a column is really not an error
+remove fastfield codecs
+ditch the first_or_default trick. if it is still useful, improve its implementation.
+rename FastFieldReaders::open to load
+
+
+remove fast field reader
+
+find a way to unify the two DateTime.
+readd type check in the filter wrapper
+
+add unit test on columnar list columns.
+
+make sure sort works
+

appveyor.yml

@@ -1,22 +0,0 @@
-# Appveyor configuration template for Rust using rustup for Rust installation
-# https://github.com/starkat99/appveyor-rust
-
-os: Visual Studio 2015
-environment:
-  matrix:
-    - channel: stable
-      target: x86_64-pc-windows-msvc
-
-install:
-  - appveyor DownloadFile https://win.rustup.rs/ -FileName rustup-init.exe
-  - rustup-init -yv --default-toolchain %channel% --default-host %target%
-  - set PATH=%PATH%;%USERPROFILE%\.cargo\bin
-  - if defined msys_bits set PATH=%PATH%;C:\msys64\mingw%msys_bits%\bin
-  - rustc -vV
-  - cargo -vV
-
-build: false
-
-test_script:
-  - REM SET RUST_LOG=tantivy,test & cargo test --verbose --no-default-features --features mmap
-  - REM SET RUST_BACKTRACE=1 & cargo build --examples

benches/agg_bench.rs

@@ -0,0 +1,462 @@
+use binggan::plugins::PeakMemAllocPlugin;
+use binggan::{black_box, InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM};
+use rand::prelude::SliceRandom;
+use rand::rngs::StdRng;
+use rand::{Rng, SeedableRng};
+use rand_distr::Distribution;
+use serde_json::json;
+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};
+
+#[global_allocator]
+pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
+
+/// Mini macro to register a function via its name
+/// runner.register("average_u64", move |index| average_u64(index));
+macro_rules! register {
+    ($runner:expr, $func:ident) => {
+        $runner.register(stringify!($func), move |index| {
+            $func(index);
+        })
+    };
+}
+
+fn main() {
+    let inputs = vec![
+        ("full", get_test_index_bench(Cardinality::Full).unwrap()),
+        (
+            "dense",
+            get_test_index_bench(Cardinality::OptionalDense).unwrap(),
+        ),
+        (
+            "sparse",
+            get_test_index_bench(Cardinality::OptionalSparse).unwrap(),
+        ),
+        (
+            "multivalue",
+            get_test_index_bench(Cardinality::Multivalued).unwrap(),
+        ),
+    ];
+
+    bench_agg(InputGroup::new_with_inputs(inputs));
+}
+
+fn bench_agg(mut group: InputGroup<Index>) {
+    group.add_plugin(PeakMemAllocPlugin::new(GLOBAL));
+
+    register!(group, average_u64);
+    register!(group, average_f64);
+    register!(group, average_f64_u64);
+    register!(group, stats_f64);
+    register!(group, extendedstats_f64);
+    register!(group, percentiles_f64);
+    register!(group, terms_few);
+    register!(group, terms_many);
+    register!(group, terms_many_top_1000);
+    register!(group, terms_many_order_by_term);
+    register!(group, terms_many_with_top_hits);
+    register!(group, terms_many_with_avg_sub_agg);
+    register!(group, terms_many_json_mixed_type_with_avg_sub_agg);
+
+    register!(group, cardinality_agg);
+    register!(group, terms_few_with_cardinality_agg);
+
+    register!(group, range_agg);
+    register!(group, range_agg_with_avg_sub_agg);
+    register!(group, range_agg_with_term_agg_few);
+    register!(group, range_agg_with_term_agg_many);
+    register!(group, histogram);
+    register!(group, histogram_hard_bounds);
+    register!(group, histogram_with_avg_sub_agg);
+    register!(group, avg_and_range_with_avg_sub_agg);
+
+    group.run();
+}
+
+fn exec_term_with_agg(index: &Index, agg_req: serde_json::Value) {
+    let agg_req: Aggregations = serde_json::from_value(agg_req).unwrap();
+
+    let reader = index.reader().unwrap();
+    let text_field = reader.searcher().schema().get_field("text").unwrap();
+    let term_query = TermQuery::new(
+        Term::from_field_text(text_field, "cool"),
+        IndexRecordOption::Basic,
+    );
+    let collector = get_collector(agg_req);
+    let searcher = reader.searcher();
+    black_box(searcher.search(&term_query, &collector).unwrap());
+}
+
+fn average_u64(index: &Index) {
+    let agg_req = json!({
+        "average": { "avg": { "field": "score", } }
+    });
+    exec_term_with_agg(index, agg_req)
+}
+fn average_f64(index: &Index) {
+    let agg_req = json!({
+        "average": { "avg": { "field": "score_f64", } }
+    });
+    exec_term_with_agg(index, agg_req)
+}
+fn average_f64_u64(index: &Index) {
+    let agg_req = json!({
+        "average_f64": { "avg": { "field": "score_f64" } },
+        "average": { "avg": { "field": "score" } },
+    });
+    exec_term_with_agg(index, agg_req)
+}
+fn stats_f64(index: &Index) {
+    let agg_req = json!({
+        "average_f64": { "stats": { "field": "score_f64", } }
+    });
+    exec_term_with_agg(index, agg_req)
+}
+fn extendedstats_f64(index: &Index) {
+    let agg_req = json!({
+        "extendedstats_f64": { "extended_stats": { "field": "score_f64", } }
+    });
+    exec_term_with_agg(index, agg_req)
+}
+fn percentiles_f64(index: &Index) {
+    let agg_req = json!({
+      "mypercentiles": {
+        "percentiles": {
+          "field": "score_f64",
+          "percents": [ 95, 99, 99.9 ]
+        }
+      }
+    });
+    execute_agg(index, agg_req);
+}
+
+fn cardinality_agg(index: &Index) {
+    let agg_req = json!({
+        "cardinality": {
+            "cardinality": {
+                "field": "text_many_terms"
+            },
+        }
+    });
+    execute_agg(index, agg_req);
+}
+fn terms_few_with_cardinality_agg(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_few_terms" },
+            "aggs": {
+                "cardinality": {
+                    "cardinality": {
+                        "field": "text_many_terms"
+                    },
+                }
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+
+fn terms_few(index: &Index) {
+    let agg_req = json!({
+        "my_texts": { "terms": { "field": "text_few_terms" } },
+    });
+    execute_agg(index, agg_req);
+}
+fn terms_many(index: &Index) {
+    let agg_req = json!({
+        "my_texts": { "terms": { "field": "text_many_terms" } },
+    });
+    execute_agg(index, agg_req);
+}
+fn terms_many_top_1000(index: &Index) {
+    let agg_req = json!({
+        "my_texts": { "terms": { "field": "text_many_terms", "size": 1000 } },
+    });
+    execute_agg(index, agg_req);
+}
+fn terms_many_order_by_term(index: &Index) {
+    let agg_req = json!({
+        "my_texts": { "terms": { "field": "text_many_terms", "order": { "_key": "desc" } } },
+    });
+    execute_agg(index, agg_req);
+}
+fn terms_many_with_top_hits(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_many_terms" },
+            "aggs": {
+                "top_hits": { "top_hits":
+                    {
+                        "sort": [
+                            { "score": "desc" }
+                        ],
+                        "size": 2,
+                        "doc_value_fields": ["score_f64"]
+                    }
+                }
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+fn terms_many_with_avg_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "text_many_terms" },
+            "aggs": {
+                "average_f64": { "avg": { "field": "score_f64" } }
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+fn terms_many_json_mixed_type_with_avg_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "my_texts": {
+            "terms": { "field": "json.mixed_type" },
+            "aggs": {
+                "average_f64": { "avg": { "field": "score_f64" } }
+            }
+        },
+    });
+    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);
+
+    let reader = index.reader().unwrap();
+    let searcher = reader.searcher();
+    black_box(searcher.search(&AllQuery, &collector).unwrap());
+}
+fn range_agg(index: &Index) {
+    let agg_req = json!({
+        "range_f64": { "range": { "field": "score_f64", "ranges": [
+            { "from": 3, "to": 7000 },
+            { "from": 7000, "to": 20000 },
+            { "from": 20000, "to": 30000 },
+            { "from": 30000, "to": 40000 },
+            { "from": 40000, "to": 50000 },
+            { "from": 50000, "to": 60000 }
+        ] } },
+    });
+    execute_agg(index, agg_req);
+}
+fn range_agg_with_avg_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "rangef64": {
+            "range": {
+                "field": "score_f64",
+                "ranges": [
+                    { "from": 3, "to": 7000 },
+                    { "from": 7000, "to": 20000 },
+                    { "from": 20000, "to": 30000 },
+                    { "from": 30000, "to": 40000 },
+                    { "from": 40000, "to": 50000 },
+                    { "from": 50000, "to": 60000 }
+                ]
+            },
+            "aggs": {
+                "average_f64": { "avg": { "field": "score_f64" } }
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+
+fn range_agg_with_term_agg_few(index: &Index) {
+    let agg_req = json!({
+        "rangef64": {
+            "range": {
+                "field": "score_f64",
+                "ranges": [
+                    { "from": 3, "to": 7000 },
+                    { "from": 7000, "to": 20000 },
+                    { "from": 20000, "to": 30000 },
+                    { "from": 30000, "to": 40000 },
+                    { "from": 40000, "to": 50000 },
+                    { "from": 50000, "to": 60000 }
+                ]
+            },
+            "aggs": {
+                "my_texts": { "terms": { "field": "text_few_terms" } },
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+fn range_agg_with_term_agg_many(index: &Index) {
+    let agg_req = json!({
+        "rangef64": {
+            "range": {
+                "field": "score_f64",
+                "ranges": [
+                    { "from": 3, "to": 7000 },
+                    { "from": 7000, "to": 20000 },
+                    { "from": 20000, "to": 30000 },
+                    { "from": 30000, "to": 40000 },
+                    { "from": 40000, "to": 50000 },
+                    { "from": 50000, "to": 60000 }
+                ]
+            },
+            "aggs": {
+                "my_texts": { "terms": { "field": "text_many_terms" } },
+            }
+        },
+    });
+    execute_agg(index, agg_req);
+}
+
+fn histogram(index: &Index) {
+    let agg_req = json!({
+        "rangef64": {
+            "histogram": {
+                "field": "score_f64",
+                "interval": 100 // 1000 buckets
+            },
+        }
+    });
+    execute_agg(index, agg_req);
+}
+fn histogram_hard_bounds(index: &Index) {
+    let agg_req = json!({
+        "rangef64": { "histogram": { "field": "score_f64", "interval": 100, "hard_bounds": { "min": 1000, "max": 300000 } } },
+    });
+    execute_agg(index, agg_req);
+}
+fn histogram_with_avg_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "rangef64": {
+            "histogram": { "field": "score_f64", "interval": 100 },
+            "aggs": {
+                "average_f64": { "avg": { "field": "score_f64" } }
+            }
+        }
+    });
+    execute_agg(index, agg_req);
+}
+fn avg_and_range_with_avg_sub_agg(index: &Index) {
+    let agg_req = json!({
+        "rangef64": {
+            "range": {
+                "field": "score_f64",
+                "ranges": [
+                    { "from": 3, "to": 7000 },
+                    { "from": 7000, "to": 20000 },
+                    { "from": 20000, "to": 60000 }
+                ]
+            },
+            "aggs": {
+                "average_in_range": { "avg": { "field": "score" } }
+            }
+        },
+        "average": { "avg": { "field": "score" } }
+    });
+    execute_agg(index, agg_req);
+}
+
+#[derive(Clone, Copy, Hash, Default, Debug, PartialEq, Eq, PartialOrd, Ord)]
+enum Cardinality {
+    /// All documents contain exactly one value.
+    /// `Full` is the default for auto-detecting the Cardinality, since it is the most strict.
+    #[default]
+    Full = 0,
+    /// All documents contain at most one value.
+    OptionalDense = 1,
+    /// All documents may contain any number of values.
+    Multivalued = 2,
+    /// 1 / 20 documents has a value
+    OptionalSparse = 3,
+}
+
+fn get_collector(agg_req: Aggregations) -> AggregationCollector {
+    AggregationCollector::from_aggs(agg_req, Default::default())
+}
+
+fn get_test_index_bench(cardinality: Cardinality) -> tantivy::Result<Index> {
+    let mut schema_builder = Schema::builder();
+    let text_fieldtype = tantivy::schema::TextOptions::default()
+        .set_indexing_options(
+            TextFieldIndexing::default().set_index_option(IndexRecordOption::WithFreqs),
+        )
+        .set_stored();
+    let text_field = schema_builder.add_text_field("text", text_fieldtype);
+    let json_field = schema_builder.add_json_field("json", 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 score_fieldtype = tantivy::schema::NumericOptions::default().set_fast();
+    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 index = Index::create_from_tempdir(schema_builder.build())?;
+    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)
+        .map(|num| format!("author{num}"))
+        .collect::<Vec<_>>();
+    {
+        let mut rng = StdRng::from_seed([1u8; 32]);
+        let mut index_writer = index.writer_with_num_threads(1, 200_000_000)?;
+        // To make the different test cases comparable we just change one doc to force the
+        // cardinality
+        if cardinality == Cardinality::OptionalDense {
+            index_writer.add_document(doc!())?;
+        }
+        if cardinality == Cardinality::Multivalued {
+            index_writer.add_document(doc!(
+                json_field => json!({"mixed_type": 10.0}),
+                json_field => json!({"mixed_type": 10.0}),
+                text_field => "cool",
+                text_field => "cool",
+                text_field_many_terms => "cool",
+                text_field_many_terms => "cool",
+                text_field_few_terms => "cool",
+                text_field_few_terms => "cool",
+                score_field => 1u64,
+                score_field => 1u64,
+                score_field_f64 => lg_norm.sample(&mut rng),
+                score_field_f64 => lg_norm.sample(&mut rng),
+                score_field_i64 => 1i64,
+                score_field_i64 => 1i64,
+            ))?;
+        }
+        let mut doc_with_value = 1_000_000;
+        if cardinality == Cardinality::OptionalSparse {
+            doc_with_value /= 20;
+        }
+        let _val_max = 1_000_000.0;
+        for _ in 0..doc_with_value {
+            let val: f64 = rng.gen_range(0.0..1_000_000.0);
+            let json = if rng.gen_bool(0.1) {
+                // 10% are numeric values
+                json!({ "mixed_type": val })
+            } else {
+                json!({"mixed_type": many_terms_data.choose(&mut rng).unwrap().to_string()})
+            };
+            index_writer.add_document(doc!(
+                text_field => "cool",
+                json_field => json,
+                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(),
+                score_field => val as u64,
+                score_field_f64 => lg_norm.sample(&mut rng),
+                score_field_i64 => val as i64,
+            ))?;
+            if cardinality == Cardinality::OptionalSparse {
+                for _ in 0..20 {
+                    index_writer.add_document(doc!(text_field => "cool"))?;
+                }
+            }
+        }
+        // writing the segment
+        index_writer.commit()?;
+    }
+
+    Ok(index)
+}

benches/analyzer.rs

@@ -0,0 +1,43 @@
+use criterion::{criterion_group, criterion_main, Criterion};
+use tantivy::tokenizer::{
+    LowerCaser, RemoveLongFilter, SimpleTokenizer, TextAnalyzer, TokenizerManager,
+};
+
+const ALICE_TXT: &str = include_str!("alice.txt");
+
+pub fn criterion_benchmark(c: &mut Criterion) {
+    let tokenizer_manager = TokenizerManager::default();
+    let mut tokenizer = tokenizer_manager.get("default").unwrap();
+    c.bench_function("default-tokenize-alice", |b| {
+        b.iter(|| {
+            let mut word_count = 0;
+            let mut token_stream = tokenizer.token_stream(ALICE_TXT);
+            while token_stream.advance() {
+                word_count += 1;
+            }
+            assert_eq!(word_count, 30_731);
+        })
+    });
+    let mut dynamic_analyzer = TextAnalyzer::builder(SimpleTokenizer::default())
+        .dynamic()
+        .filter_dynamic(RemoveLongFilter::limit(40))
+        .filter_dynamic(LowerCaser)
+        .build();
+    c.bench_function("dynamic-tokenize-alice", |b| {
+        b.iter(|| {
+            let mut word_count = 0;
+            let mut token_stream = dynamic_analyzer.token_stream(ALICE_TXT);
+            while token_stream.advance() {
+                word_count += 1;
+            }
+            assert_eq!(word_count, 30_731);
+        })
+    });
+}
+
+criterion_group! {
+    name = benches;
+    config = Criterion::default().sample_size(200);
+    targets = criterion_benchmark
+}
+criterion_main!(benches);

benches/index-bench.rs

@@ -0,0 +1,218 @@
+use criterion::{criterion_group, criterion_main, BatchSize, Bencher, Criterion, Throughput};
+use tantivy::schema::{TantivyDocument, FAST, INDEXED, STORED, STRING, TEXT};
+use tantivy::{tokenizer, Index, IndexWriter};
+
+const HDFS_LOGS: &str = include_str!("hdfs.json");
+const GH_LOGS: &str = include_str!("gh.json");
+const WIKI: &str = include_str!("wiki.json");
+
+fn benchmark(
+    b: &mut Bencher,
+    input: &str,
+    schema: tantivy::schema::Schema,
+    commit: bool,
+    parse_json: bool,
+    is_dynamic: bool,
+) {
+    if is_dynamic {
+        benchmark_dynamic_json(b, input, schema, commit, parse_json)
+    } else {
+        _benchmark(b, input, schema, commit, parse_json, |schema, doc_json| {
+            TantivyDocument::parse_json(schema, doc_json).unwrap()
+        })
+    }
+}
+
+fn get_index(schema: tantivy::schema::Schema) -> Index {
+    let mut index = Index::create_in_ram(schema.clone());
+    let ff_tokenizer_manager = tokenizer::TokenizerManager::default();
+    ff_tokenizer_manager.register(
+        "raw",
+        tokenizer::TextAnalyzer::builder(tokenizer::RawTokenizer::default())
+            .filter(tokenizer::RemoveLongFilter::limit(255))
+            .build(),
+    );
+    index.set_fast_field_tokenizers(ff_tokenizer_manager.clone());
+    index
+}
+
+fn _benchmark(
+    b: &mut Bencher,
+    input: &str,
+    schema: tantivy::schema::Schema,
+    commit: bool,
+    include_json_parsing: bool,
+    create_doc: impl Fn(&tantivy::schema::Schema, &str) -> TantivyDocument,
+) {
+    if include_json_parsing {
+        let lines: Vec<&str> = input.trim().split('\n').collect();
+        b.iter(|| {
+            let index = get_index(schema.clone());
+            let mut index_writer: IndexWriter =
+                index.writer_with_num_threads(1, 100_000_000).unwrap();
+            for doc_json in &lines {
+                let doc = create_doc(&schema, doc_json);
+                index_writer.add_document(doc).unwrap();
+            }
+            if commit {
+                index_writer.commit().unwrap();
+            }
+        })
+    } else {
+        let docs: Vec<_> = input
+            .trim()
+            .split('\n')
+            .map(|doc_json| create_doc(&schema, doc_json))
+            .collect();
+        b.iter_batched(
+            || docs.clone(),
+            |docs| {
+                let index = get_index(schema.clone());
+                let mut index_writer: IndexWriter =
+                    index.writer_with_num_threads(1, 100_000_000).unwrap();
+                for doc in docs {
+                    index_writer.add_document(doc).unwrap();
+                }
+                if commit {
+                    index_writer.commit().unwrap();
+                }
+            },
+            BatchSize::SmallInput,
+        )
+    }
+}
+fn benchmark_dynamic_json(
+    b: &mut Bencher,
+    input: &str,
+    schema: tantivy::schema::Schema,
+    commit: bool,
+    parse_json: bool,
+) {
+    let json_field = schema.get_field("json").unwrap();
+    _benchmark(b, input, schema, commit, parse_json, |_schema, doc_json| {
+        let json_val: serde_json::Value = serde_json::from_str(doc_json).unwrap();
+        tantivy::doc!(json_field=>json_val)
+    })
+}
+
+pub fn hdfs_index_benchmark(c: &mut Criterion) {
+    let schema = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_u64_field("timestamp", INDEXED);
+        schema_builder.add_text_field("body", TEXT);
+        schema_builder.add_text_field("severity", STRING);
+        schema_builder.build()
+    };
+    let schema_only_fast = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_u64_field("timestamp", FAST);
+        schema_builder.add_text_field("body", FAST);
+        schema_builder.add_text_field("severity", FAST);
+        schema_builder.build()
+    };
+    let _schema_with_store = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_u64_field("timestamp", INDEXED | STORED);
+        schema_builder.add_text_field("body", TEXT | STORED);
+        schema_builder.add_text_field("severity", STRING | STORED);
+        schema_builder.build()
+    };
+    let dynamic_schema = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_json_field("json", TEXT | FAST);
+        schema_builder.build()
+    };
+
+    let mut group = c.benchmark_group("index-hdfs");
+    group.throughput(Throughput::Bytes(HDFS_LOGS.len() as u64));
+    group.sample_size(20);
+
+    let benches = [
+        ("only-indexed-".to_string(), schema, false),
+        //("stored-".to_string(), _schema_with_store, false),
+        ("only-fast-".to_string(), schema_only_fast, false),
+        ("dynamic-".to_string(), dynamic_schema, true),
+    ];
+
+    for (prefix, schema, is_dynamic) in benches {
+        for commit in [false, true] {
+            let suffix = if commit { "with-commit" } else { "no-commit" };
+            {
+                let parse_json = false;
+                // for parse_json in [false, true] {
+                let suffix = if parse_json {
+                    format!("{suffix}-with-json-parsing")
+                } else {
+                    suffix.to_string()
+                };
+
+                let bench_name = format!("{prefix}{suffix}");
+                group.bench_function(bench_name, |b| {
+                    benchmark(b, HDFS_LOGS, schema.clone(), commit, parse_json, is_dynamic)
+                });
+            }
+        }
+    }
+}
+
+pub fn gh_index_benchmark(c: &mut Criterion) {
+    let dynamic_schema = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_json_field("json", TEXT | FAST);
+        schema_builder.build()
+    };
+    let dynamic_schema_fast = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_json_field("json", FAST);
+        schema_builder.build()
+    };
+
+    let mut group = c.benchmark_group("index-gh");
+    group.throughput(Throughput::Bytes(GH_LOGS.len() as u64));
+
+    group.bench_function("index-gh-no-commit", |b| {
+        benchmark_dynamic_json(b, GH_LOGS, dynamic_schema.clone(), false, false)
+    });
+    group.bench_function("index-gh-fast", |b| {
+        benchmark_dynamic_json(b, GH_LOGS, dynamic_schema_fast.clone(), false, false)
+    });
+
+    group.bench_function("index-gh-fast-with-commit", |b| {
+        benchmark_dynamic_json(b, GH_LOGS, dynamic_schema_fast.clone(), true, false)
+    });
+}
+
+pub fn wiki_index_benchmark(c: &mut Criterion) {
+    let dynamic_schema = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_json_field("json", TEXT | FAST);
+        schema_builder.build()
+    };
+
+    let mut group = c.benchmark_group("index-wiki");
+    group.throughput(Throughput::Bytes(WIKI.len() as u64));
+
+    group.bench_function("index-wiki-no-commit", |b| {
+        benchmark_dynamic_json(b, WIKI, dynamic_schema.clone(), false, false)
+    });
+    group.bench_function("index-wiki-with-commit", |b| {
+        benchmark_dynamic_json(b, WIKI, dynamic_schema.clone(), true, false)
+    });
+}
+
+criterion_group! {
+    name = benches;
+    config = Criterion::default();
+    targets = hdfs_index_benchmark
+}
+criterion_group! {
+    name = gh_benches;
+    config = Criterion::default();
+    targets = gh_index_benchmark
+}
+criterion_group! {
+    name = wiki_benches;
+    config = Criterion::default();
+    targets = wiki_index_benchmark
+}
+criterion_main!(benches, gh_benches, wiki_benches);

bitpacker/Cargo.toml

@@ -0,0 +1,22 @@
+[package]
+name = "tantivy-bitpacker"
+version = "0.8.0"
+edition = "2024"
+authors = ["Paul Masurel <paul.masurel@gmail.com>"]
+license = "MIT"
+categories = []
+description = """Tantivy-sub crate: bitpacking"""
+repository = "https://github.com/quickwit-oss/tantivy"
+keywords = []
+documentation = "https://docs.rs/tantivy-bitpacker/latest/tantivy_bitpacker"
+homepage = "https://github.com/quickwit-oss/tantivy"
+
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+bitpacking = { version = "0.9.2", default-features = false, features = ["bitpacker1x"] }
+
+[dev-dependencies]
+rand = "0.8"
+proptest = "1"

bitpacker/benches/bench.rs

@@ -0,0 +1,65 @@
+#![feature(test)]
+
+extern crate test;
+
+#[cfg(test)]
+mod tests {
+    use rand::seq::IteratorRandom;
+    use rand::thread_rng;
+    use tantivy_bitpacker::{BitPacker, BitUnpacker, BlockedBitpacker};
+    use test::Bencher;
+
+    #[inline(never)]
+    fn create_bitpacked_data(bit_width: u8, num_els: u32) -> Vec<u8> {
+        let mut bitpacker = BitPacker::new();
+        let mut buffer = Vec::new();
+        for _ in 0..num_els {
+            // the values do not matter.
+            bitpacker.write(0u64, bit_width, &mut buffer).unwrap();
+            bitpacker.flush(&mut buffer).unwrap();
+        }
+        buffer
+    }
+
+    #[bench]
+    fn bench_bitpacking_read(b: &mut Bencher) {
+        let bit_width = 3;
+        let num_els = 1_000_000u32;
+        let bit_unpacker = BitUnpacker::new(bit_width);
+        let data = create_bitpacked_data(bit_width, num_els);
+        let idxs: Vec<u32> = (0..num_els).choose_multiple(&mut thread_rng(), 100_000);
+        b.iter(|| {
+            let mut out = 0u64;
+            for &idx in &idxs {
+                out = out.wrapping_add(bit_unpacker.get(idx, &data[..]));
+            }
+            out
+        });
+    }
+
+    #[bench]
+    fn bench_blockedbitp_read(b: &mut Bencher) {
+        let mut blocked_bitpacker = BlockedBitpacker::new();
+        for val in 0..=21500 {
+            blocked_bitpacker.add(val * val);
+        }
+        b.iter(|| {
+            let mut out = 0u64;
+            for val in 0..=21500 {
+                out = out.wrapping_add(blocked_bitpacker.get(val));
+            }
+            out
+        });
+    }
+
+    #[bench]
+    fn bench_blockedbitp_create(b: &mut Bencher) {
+        b.iter(|| {
+            let mut blocked_bitpacker = BlockedBitpacker::new();
+            for val in 0..=21500 {
+                blocked_bitpacker.add(val * val);
+            }
+            blocked_bitpacker
+        });
+    }
+}

bitpacker/src/bitpacker.rs

@@ -0,0 +1,380 @@
+use std::io;
+use std::ops::{Range, RangeInclusive};
+
+use bitpacking::{BitPacker as ExternalBitPackerTrait, BitPacker1x};
+
+pub struct BitPacker {
+    mini_buffer: u64,
+    mini_buffer_written: usize,
+}
+
+impl Default for BitPacker {
+    fn default() -> Self {
+        BitPacker::new()
+    }
+}
+impl BitPacker {
+    pub fn new() -> BitPacker {
+        BitPacker {
+            mini_buffer: 0u64,
+            mini_buffer_written: 0,
+        }
+    }
+
+    #[inline]
+    pub fn write<TWrite: io::Write + ?Sized>(
+        &mut self,
+        val: u64,
+        num_bits: u8,
+        output: &mut TWrite,
+    ) -> io::Result<()> {
+        let num_bits = num_bits as usize;
+        if self.mini_buffer_written + num_bits > 64 {
+            self.mini_buffer |= val.wrapping_shl(self.mini_buffer_written as u32);
+            output.write_all(self.mini_buffer.to_le_bytes().as_ref())?;
+            self.mini_buffer = val.wrapping_shr((64 - self.mini_buffer_written) as u32);
+            self.mini_buffer_written = self.mini_buffer_written + num_bits - 64;
+        } else {
+            self.mini_buffer |= val << self.mini_buffer_written;
+            self.mini_buffer_written += num_bits;
+            if self.mini_buffer_written == 64 {
+                output.write_all(self.mini_buffer.to_le_bytes().as_ref())?;
+                self.mini_buffer_written = 0;
+                self.mini_buffer = 0u64;
+            }
+        }
+        Ok(())
+    }
+
+    pub fn flush<TWrite: io::Write + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
+        if self.mini_buffer_written > 0 {
+            let num_bytes = (self.mini_buffer_written + 7) / 8;
+            let bytes = self.mini_buffer.to_le_bytes();
+            output.write_all(&bytes[..num_bytes])?;
+            self.mini_buffer_written = 0;
+            self.mini_buffer = 0;
+        }
+        Ok(())
+    }
+
+    pub fn close<TWrite: io::Write + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
+        self.flush(output)?;
+        Ok(())
+    }
+}
+
+#[derive(Clone, Debug, Default, Copy)]
+pub struct BitUnpacker {
+    num_bits: usize,
+    mask: u64,
+}
+
+impl BitUnpacker {
+    /// Creates a bit unpacker, that assumes the same bitwidth for all values.
+    ///
+    /// The bitunpacker works by doing an unaligned read of 8 bytes.
+    /// For this reason, values of `num_bits` between
+    /// [57..63] are forbidden.
+    pub fn new(num_bits: u8) -> BitUnpacker {
+        assert!(num_bits <= 7 * 8 || num_bits == 64);
+        let mask: u64 = if num_bits == 64 {
+            !0u64
+        } else {
+            (1u64 << num_bits) - 1u64
+        };
+        BitUnpacker {
+            num_bits: usize::from(num_bits),
+            mask,
+        }
+    }
+
+    pub fn bit_width(&self) -> u8 {
+        self.num_bits as u8
+    }
+
+    #[inline]
+    pub fn get(&self, idx: u32, data: &[u8]) -> u64 {
+        let addr_in_bits = idx as usize * self.num_bits;
+        let addr = addr_in_bits >> 3;
+        if addr + 8 > data.len() {
+            if self.num_bits == 0 {
+                return 0;
+            }
+            let bit_shift = addr_in_bits & 7;
+            return self.get_slow_path(addr, bit_shift as u32, data);
+        }
+        let bit_shift = addr_in_bits & 7;
+        let bytes: [u8; 8] = (&data[addr..addr + 8]).try_into().unwrap();
+        let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
+        let val_shifted = val_unshifted_unmasked >> bit_shift;
+        val_shifted & self.mask
+    }
+
+    #[inline(never)]
+    fn get_slow_path(&self, addr: usize, bit_shift: u32, data: &[u8]) -> u64 {
+        let mut bytes: [u8; 8] = [0u8; 8];
+        let available_bytes = data.len() - addr;
+        // This function is meant to only be called if we did not have 8 bytes to load.
+        debug_assert!(available_bytes < 8);
+        bytes[..available_bytes].copy_from_slice(&data[addr..]);
+        let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
+        let val_shifted = val_unshifted_unmasked >> bit_shift;
+        val_shifted & self.mask
+    }
+
+    // Decodes the range of bitpacked `u32` values with idx
+    // in [start_idx, start_idx + output.len()).
+    //
+    // #Panics
+    //
+    // This methods panics if `num_bits` is > 32.
+    fn get_batch_u32s(&self, start_idx: u32, data: &[u8], output: &mut [u32]) {
+        assert!(
+            self.bit_width() <= 32,
+            "Bitwidth must be <= 32 to use this method."
+        );
+
+        let end_idx: u32 = start_idx + output.len() as u32;
+
+        // We use `usize` here to avoid overflow issues.
+        let end_bit_read = (end_idx as usize) * self.num_bits;
+        let end_byte_read = (end_bit_read + 7) / 8;
+        assert!(
+            end_byte_read <= data.len(),
+            "Requested index is out of bounds."
+        );
+
+        // Simple slow implementation of get_batch_u32s, to deal with our ramps.
+        let get_batch_ramp = |start_idx: u32, output: &mut [u32]| {
+            for (out, idx) in output.iter_mut().zip(start_idx..) {
+                *out = self.get(idx, data) as u32;
+            }
+        };
+
+        // We use an unrolled routine to decode 32 values at once.
+        // We therefore decompose our range of values to decode into three ranges:
+        // - Entrance ramp: [start_idx, fast_track_start) (up to 31 values)
+        // - Highway: [fast_track_start, fast_track_end) (a length multiple of 32s)
+        // - Exit ramp: [fast_track_end, start_idx + output.len()) (up to 31 values)
+
+        // We want the start of the fast track to start align with bytes.
+        // A sufficient condition is to start with an idx that is a multiple of 8,
+        // so highway start is the closest multiple of 8 that is >= start_idx.
+        let entrance_ramp_len: u32 = 8 - (start_idx % 8) % 8;
+
+        let highway_start: u32 = start_idx + entrance_ramp_len;
+
+        if highway_start + (BitPacker1x::BLOCK_LEN as u32) > end_idx {
+            // We don't have enough values to have even a single block of highway.
+            // Let's just supply the values the simple way.
+            get_batch_ramp(start_idx, output);
+            return;
+        }
+
+        let num_blocks: usize = (end_idx - highway_start) as usize / BitPacker1x::BLOCK_LEN;
+
+        // Entrance ramp
+        get_batch_ramp(start_idx, &mut output[..entrance_ramp_len as usize]);
+
+        // Highway
+        let mut offset = (highway_start as usize * self.num_bits) / 8;
+        let mut output_cursor = (highway_start - start_idx) as usize;
+        for _ in 0..num_blocks {
+            offset += BitPacker1x.decompress(
+                &data[offset..],
+                &mut output[output_cursor..],
+                self.num_bits as u8,
+            );
+            output_cursor += 32;
+        }
+
+        // Exit ramp
+        let highway_end: u32 = highway_start + (num_blocks * BitPacker1x::BLOCK_LEN) as u32;
+        get_batch_ramp(highway_end, &mut output[output_cursor..]);
+    }
+
+    pub fn get_ids_for_value_range(
+        &self,
+        range: RangeInclusive<u64>,
+        id_range: Range<u32>,
+        data: &[u8],
+        positions: &mut Vec<u32>,
+    ) {
+        if self.bit_width() > 32 {
+            self.get_ids_for_value_range_slow(range, id_range, data, positions)
+        } else {
+            if *range.start() > u32::MAX as u64 {
+                positions.clear();
+                return;
+            }
+            let range_u32 = (*range.start() as u32)..=(*range.end()).min(u32::MAX as u64) as u32;
+            self.get_ids_for_value_range_fast(range_u32, id_range, data, positions)
+        }
+    }
+
+    fn get_ids_for_value_range_slow(
+        &self,
+        range: RangeInclusive<u64>,
+        id_range: Range<u32>,
+        data: &[u8],
+        positions: &mut Vec<u32>,
+    ) {
+        positions.clear();
+        for i in id_range {
+            // If we cared we could make this branchless, but the slow implementation should rarely
+            // kick in.
+            let val = self.get(i, data);
+            if range.contains(&val) {
+                positions.push(i);
+            }
+        }
+    }
+
+    fn get_ids_for_value_range_fast(
+        &self,
+        value_range: RangeInclusive<u32>,
+        id_range: Range<u32>,
+        data: &[u8],
+        positions: &mut Vec<u32>,
+    ) {
+        positions.resize(id_range.len(), 0u32);
+        self.get_batch_u32s(id_range.start, data, positions);
+        crate::filter_vec::filter_vec_in_place(value_range, id_range.start, positions)
+    }
+}
+
+#[cfg(test)]
+mod test {
+    use super::{BitPacker, BitUnpacker};
+
+    fn create_bitpacker(len: usize, num_bits: u8) -> (BitUnpacker, Vec<u64>, Vec<u8>) {
+        let mut data = Vec::new();
+        let mut bitpacker = BitPacker::new();
+        let max_val: u64 = (1u64 << num_bits as u64) - 1u64;
+        let vals: Vec<u64> = (0u64..len as u64)
+            .map(|i| if max_val == 0 { 0 } else { i % max_val })
+            .collect();
+        for &val in &vals {
+            bitpacker.write(val, num_bits, &mut data).unwrap();
+        }
+        bitpacker.close(&mut data).unwrap();
+        assert_eq!(data.len(), ((num_bits as usize) * len + 7) / 8);
+        let bitunpacker = BitUnpacker::new(num_bits);
+        (bitunpacker, vals, data)
+    }
+
+    fn test_bitpacker_util(len: usize, num_bits: u8) {
+        let (bitunpacker, vals, data) = create_bitpacker(len, num_bits);
+        for (i, val) in vals.iter().enumerate() {
+            assert_eq!(bitunpacker.get(i as u32, &data), *val);
+        }
+    }
+
+    #[test]
+    fn test_bitpacker() {
+        test_bitpacker_util(10, 3);
+        test_bitpacker_util(10, 0);
+        test_bitpacker_util(10, 1);
+        test_bitpacker_util(6, 14);
+        test_bitpacker_util(1000, 14);
+    }
+
+    use proptest::prelude::*;
+
+    fn num_bits_strategy() -> impl Strategy<Value = u8> {
+        prop_oneof!(Just(0), Just(1), 2u8..56u8, Just(56), Just(64),)
+    }
+
+    fn vals_strategy() -> impl Strategy<Value = (u8, Vec<u64>)> {
+        (num_bits_strategy(), 0usize..100usize).prop_flat_map(|(num_bits, len)| {
+            let max_val = if num_bits == 64 {
+                u64::MAX
+            } else {
+                (1u64 << num_bits as u32) - 1
+            };
+            let vals = proptest::collection::vec(0..=max_val, len);
+            vals.prop_map(move |vals| (num_bits, vals))
+        })
+    }
+
+    fn test_bitpacker_aux(num_bits: u8, vals: &[u64]) {
+        let mut buffer: Vec<u8> = Vec::new();
+        let mut bitpacker = BitPacker::new();
+        for &val in vals {
+            bitpacker.write(val, num_bits, &mut buffer).unwrap();
+        }
+        bitpacker.flush(&mut buffer).unwrap();
+        assert_eq!(buffer.len(), (vals.len() * num_bits as usize + 7) / 8);
+        let bitunpacker = BitUnpacker::new(num_bits);
+        let max_val = if num_bits == 64 {
+            u64::MAX
+        } else {
+            (1u64 << num_bits) - 1
+        };
+        for (i, val) in vals.iter().copied().enumerate() {
+            assert!(val <= max_val);
+            assert_eq!(bitunpacker.get(i as u32, &buffer), val);
+        }
+    }
+
+    proptest::proptest! {
+        #[test]
+        fn test_bitpacker_proptest((num_bits, vals) in vals_strategy()) {
+            test_bitpacker_aux(num_bits, &vals);
+        }
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_get_batch_panics_over_32_bits() {
+        let bitunpacker = BitUnpacker::new(33);
+        let mut output: [u32; 1] = [0u32];
+        bitunpacker.get_batch_u32s(0, &[0, 0, 0, 0, 0, 0, 0, 0], &mut output[..]);
+    }
+
+    #[test]
+    fn test_get_batch_limit() {
+        let bitunpacker = BitUnpacker::new(1);
+        let mut output: [u32; 3] = [0u32, 0u32, 0u32];
+        bitunpacker.get_batch_u32s(8 * 4 - 3, &[0u8, 0u8, 0u8, 0u8], &mut output[..]);
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_get_batch_panics_when_off_scope() {
+        let bitunpacker = BitUnpacker::new(1);
+        let mut output: [u32; 3] = [0u32, 0u32, 0u32];
+        // We are missing exactly one bit.
+        bitunpacker.get_batch_u32s(8 * 4 - 2, &[0u8, 0u8, 0u8, 0u8], &mut output[..]);
+    }
+
+    proptest::proptest! {
+        #[test]
+        fn test_get_batch_u32s_proptest(num_bits in 0u8..=32u8) {
+            let mask =
+                if num_bits == 32u8 {
+                    u32::MAX
+                } else {
+                    (1u32 << num_bits) - 1
+                };
+            let mut buffer: Vec<u8> = Vec::new();
+            let mut bitpacker = BitPacker::new();
+            for val in 0..100 {
+                bitpacker.write(val & mask as u64, num_bits, &mut buffer).unwrap();
+            }
+            bitpacker.flush(&mut buffer).unwrap();
+            let bitunpacker = BitUnpacker::new(num_bits);
+            let mut output: Vec<u32> = Vec::new();
+            for len in [0, 1, 2, 32, 33, 34, 64] {
+                for start_idx in 0u32..32u32 {
+                    output.resize(len, 0);
+                    bitunpacker.get_batch_u32s(start_idx, &buffer, &mut output);
+                    for (i, output_byte) in output.iter().enumerate() {
+                        let expected = (start_idx + i as u32) & mask;
+                        assert_eq!(*output_byte, expected);
+                    }
+                }
+            }
+        }
+    }
+}

bitpacker/src/blocked_bitpacker.rs

@@ -0,0 +1,177 @@
+use super::bitpacker::BitPacker;
+use super::compute_num_bits;
+use crate::{BitUnpacker, minmax};
+
+const BLOCK_SIZE: usize = 128;
+
+/// `BlockedBitpacker` compresses data in blocks of
+/// 128 elements, while keeping an index on it
+#[derive(Debug, Clone)]
+pub struct BlockedBitpacker {
+    // bitpacked blocks
+    compressed_blocks: Vec<u8>,
+    // uncompressed data, collected until BLOCK_SIZE
+    buffer: Vec<u64>,
+    offset_and_bits: Vec<BlockedBitpackerEntryMetaData>,
+}
+impl Default for BlockedBitpacker {
+    fn default() -> Self {
+        BlockedBitpacker::new()
+    }
+}
+
+/// `BlockedBitpackerEntryMetaData` encodes the
+/// offset and bit_width into a u64 bit field
+///
+/// This saves some space, since 7byte is more
+/// than enough and also keeps the access fast
+/// because of alignment
+#[derive(Debug, Clone, Default)]
+struct BlockedBitpackerEntryMetaData {
+    encoded: u64,
+    base_value: u64,
+}
+
+impl BlockedBitpackerEntryMetaData {
+    fn new(offset: u64, num_bits: u8, base_value: u64) -> Self {
+        let encoded = offset | (u64::from(num_bits) << (64 - 8));
+        Self {
+            encoded,
+            base_value,
+        }
+    }
+    fn offset(&self) -> u64 {
+        (self.encoded << 8) >> 8
+    }
+    fn num_bits(&self) -> u8 {
+        (self.encoded >> 56) as u8
+    }
+    fn base_value(&self) -> u64 {
+        self.base_value
+    }
+}
+
+#[test]
+fn metadata_test() {
+    let meta = BlockedBitpackerEntryMetaData::new(50000, 6, 40000);
+    assert_eq!(meta.offset(), 50000);
+    assert_eq!(meta.num_bits(), 6);
+}
+
+fn mem_usage<T>(items: &Vec<T>) -> usize {
+    items.capacity() * std::mem::size_of::<T>()
+}
+
+impl BlockedBitpacker {
+    pub fn new() -> Self {
+        Self {
+            compressed_blocks: vec![0; 8],
+            buffer: vec![],
+            offset_and_bits: vec![],
+        }
+    }
+
+    /// The memory used (inclusive childs)
+    pub fn mem_usage(&self) -> usize {
+        std::mem::size_of::<BlockedBitpacker>()
+            + self.compressed_blocks.capacity()
+            + mem_usage(&self.offset_and_bits)
+            + mem_usage(&self.buffer)
+    }
+
+    #[inline]
+    pub fn add(&mut self, val: u64) {
+        self.buffer.push(val);
+        if self.buffer.len() == BLOCK_SIZE {
+            self.flush();
+        }
+    }
+
+    pub fn flush(&mut self) {
+        if let Some((min_value, max_value)) = minmax(self.buffer.iter()) {
+            let mut bit_packer = BitPacker::new();
+            let num_bits_block = compute_num_bits(*max_value - min_value);
+            // todo performance: the padding handling could be done better, e.g. use a slice and
+            // return num_bytes written from bitpacker
+            self.compressed_blocks
+                .resize(self.compressed_blocks.len() - 8, 0); // remove padding for bitpacker
+            let offset = self.compressed_blocks.len() as u64;
+            // todo performance: for some bit_width we
+            // can encode multiple vals into the
+            // mini_buffer before checking to flush
+            // (to be done in BitPacker)
+            for val in self.buffer.iter() {
+                bit_packer
+                    .write(
+                        *val - min_value,
+                        num_bits_block,
+                        &mut self.compressed_blocks,
+                    )
+                    .expect("cannot write bitpacking to output"); // write to in memory can't fail
+            }
+            bit_packer.flush(&mut self.compressed_blocks).unwrap();
+            self.offset_and_bits
+                .push(BlockedBitpackerEntryMetaData::new(
+                    offset,
+                    num_bits_block,
+                    *min_value,
+                ));
+
+            self.buffer.clear();
+            self.compressed_blocks
+                .resize(self.compressed_blocks.len() + 8, 0); // add padding for bitpacker
+        }
+    }
+    #[inline]
+    pub fn get(&self, idx: usize) -> u64 {
+        let metadata_pos = idx / BLOCK_SIZE;
+        let pos_in_block = idx % BLOCK_SIZE;
+        if let Some(metadata) = self.offset_and_bits.get(metadata_pos) {
+            let unpacked = BitUnpacker::new(metadata.num_bits()).get(
+                pos_in_block as u32,
+                &self.compressed_blocks[metadata.offset() as usize..],
+            );
+            unpacked + metadata.base_value()
+        } else {
+            self.buffer[pos_in_block]
+        }
+    }
+
+    pub fn iter(&self) -> impl Iterator<Item = u64> + '_ {
+        // todo performance: we could decompress a whole block and cache it instead
+        let bitpacked_elems = self.offset_and_bits.len() * BLOCK_SIZE;
+        let iter = (0..bitpacked_elems)
+            .map(move |idx| self.get(idx))
+            .chain(self.buffer.iter().cloned());
+        iter
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    #[test]
+    fn blocked_bitpacker_empty() {
+        let blocked_bitpacker = BlockedBitpacker::new();
+        assert_eq!(blocked_bitpacker.iter().collect::<Vec<u64>>(), vec![]);
+    }
+    #[test]
+    fn blocked_bitpacker_one() {
+        let mut blocked_bitpacker = BlockedBitpacker::new();
+        blocked_bitpacker.add(50000);
+        assert_eq!(blocked_bitpacker.get(0), 50000);
+        assert_eq!(blocked_bitpacker.iter().collect::<Vec<u64>>(), vec![50000]);
+    }
+    #[test]
+    fn blocked_bitpacker_test() {
+        let mut blocked_bitpacker = BlockedBitpacker::new();
+        for val in 0..21500 {
+            blocked_bitpacker.add(val);
+        }
+        for val in 0..21500 {
+            assert_eq!(blocked_bitpacker.get(val as usize), val);
+        }
+        assert_eq!(blocked_bitpacker.iter().count(), 21500);
+        assert_eq!(blocked_bitpacker.iter().last().unwrap(), 21499);
+    }
+}

bitpacker/src/filter_vec/avx2.rs

@@ -0,0 +1,365 @@
+//! SIMD filtering of a vector as described in the following blog post.
+//! <https://quickwit.io/blog/filtering%20a%20vector%20with%20simd%20instructions%20avx-2%20and%20avx-512>
+use std::arch::x86_64::{
+    __m256i as DataType, _mm256_add_epi32 as op_add, _mm256_cmpgt_epi32 as op_greater,
+    _mm256_lddqu_si256 as load_unaligned, _mm256_or_si256 as op_or, _mm256_set1_epi32 as set1,
+    _mm256_storeu_si256 as store_unaligned, _mm256_xor_si256 as op_xor, *,
+};
+use std::ops::RangeInclusive;
+
+const NUM_LANES: usize = 8;
+
+const HIGHEST_BIT: u32 = 1 << 31;
+
+#[inline]
+fn u32_to_i32(val: u32) -> i32 {
+    (val ^ HIGHEST_BIT) as i32
+}
+
+#[inline]
+unsafe fn u32_to_i32_avx2(vals_u32x8s: DataType) -> DataType {
+    const HIGHEST_BIT_MASK: DataType = from_u32x8([HIGHEST_BIT; NUM_LANES]);
+    op_xor(vals_u32x8s, HIGHEST_BIT_MASK)
+}
+
+pub fn filter_vec_in_place(range: RangeInclusive<u32>, offset: u32, output: &mut Vec<u32>) {
+    // We use a monotonic mapping from u32 to i32 to make the comparison possible in AVX2.
+    let range_i32: RangeInclusive<i32> = u32_to_i32(*range.start())..=u32_to_i32(*range.end());
+    let num_words = output.len() / NUM_LANES;
+    let mut output_len = unsafe {
+        filter_vec_avx2_aux(
+            output.as_ptr() as *const __m256i,
+            range_i32,
+            output.as_mut_ptr(),
+            offset,
+            num_words,
+        )
+    };
+    let reminder_start = num_words * NUM_LANES;
+    for i in reminder_start..output.len() {
+        let val = output[i];
+        output[output_len] = offset + i as u32;
+        output_len += if range.contains(&val) { 1 } else { 0 };
+    }
+    output.truncate(output_len);
+}
+
+#[target_feature(enable = "avx2")]
+unsafe fn filter_vec_avx2_aux(
+    mut input: *const __m256i,
+    range: RangeInclusive<i32>,
+    output: *mut u32,
+    offset: u32,
+    num_words: usize,
+) -> usize {
+    let mut output_tail = output;
+    let range_simd = set1(*range.start())..=set1(*range.end());
+    let mut ids = from_u32x8([
+        offset,
+        offset + 1,
+        offset + 2,
+        offset + 3,
+        offset + 4,
+        offset + 5,
+        offset + 6,
+        offset + 7,
+    ]);
+    const SHIFT: __m256i = from_u32x8([NUM_LANES as u32; NUM_LANES]);
+    for _ in 0..num_words {
+        let word = load_unaligned(input);
+        let word = u32_to_i32_avx2(word);
+        let keeper_bitset = compute_filter_bitset(word, range_simd.clone());
+        let added_len = keeper_bitset.count_ones();
+        let filtered_doc_ids = compact(ids, keeper_bitset);
+        store_unaligned(output_tail as *mut __m256i, filtered_doc_ids);
+        output_tail = output_tail.offset(added_len as isize);
+        ids = op_add(ids, SHIFT);
+        input = input.offset(1);
+    }
+    output_tail.offset_from(output) as usize
+}
+
+#[inline]
+#[target_feature(enable = "avx2")]
+unsafe fn compact(data: DataType, mask: u8) -> DataType {
+    let vperm_mask = MASK_TO_PERMUTATION[mask as usize];
+    _mm256_permutevar8x32_epi32(data, vperm_mask)
+}
+
+#[inline]
+#[target_feature(enable = "avx2")]
+unsafe fn compute_filter_bitset(val: __m256i, range: std::ops::RangeInclusive<__m256i>) -> u8 {
+    let too_low = op_greater(*range.start(), val);
+    let too_high = op_greater(val, *range.end());
+    let inside = op_or(too_low, too_high);
+    255 - std::arch::x86_64::_mm256_movemask_ps(std::mem::transmute::<DataType, __m256>(inside))
+        as u8
+}
+
+union U8x32 {
+    vector: DataType,
+    vals: [u32; NUM_LANES],
+}
+
+const fn from_u32x8(vals: [u32; NUM_LANES]) -> DataType {
+    unsafe { U8x32 { vals }.vector }
+}
+
+const MASK_TO_PERMUTATION: [DataType; 256] = [
+    from_u32x8([0, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([2, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 0, 0, 0, 0, 0]),
+    from_u32x8([3, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 0, 0, 0, 0, 0]),
+    from_u32x8([2, 3, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 0, 0, 0, 0]),
+    from_u32x8([4, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 4, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 0, 0, 0, 0, 0]),
+    from_u32x8([2, 4, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 4, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 0, 0, 0, 0]),
+    from_u32x8([3, 4, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 4, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 0, 0, 0, 0]),
+    from_u32x8([2, 3, 4, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 4, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 0, 0, 0]),
+    from_u32x8([5, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 5, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 5, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([2, 5, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 5, 0, 0, 0, 0]),
+    from_u32x8([3, 5, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 5, 0, 0, 0, 0]),
+    from_u32x8([2, 3, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 5, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 5, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 5, 0, 0, 0]),
+    from_u32x8([4, 5, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 4, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 5, 0, 0, 0, 0]),
+    from_u32x8([2, 4, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 5, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 4, 5, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 5, 0, 0, 0]),
+    from_u32x8([3, 4, 5, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 5, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 4, 5, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 5, 0, 0, 0]),
+    from_u32x8([2, 3, 4, 5, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 5, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 4, 5, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 5, 0, 0]),
+    from_u32x8([6, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 6, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 6, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([2, 6, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 6, 0, 0, 0, 0]),
+    from_u32x8([3, 6, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 6, 0, 0, 0, 0]),
+    from_u32x8([2, 3, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 6, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 6, 0, 0, 0]),
+    from_u32x8([4, 6, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 4, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 6, 0, 0, 0, 0]),
+    from_u32x8([2, 4, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 6, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 4, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 6, 0, 0, 0]),
+    from_u32x8([3, 4, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 6, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 4, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 6, 0, 0, 0]),
+    from_u32x8([2, 3, 4, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 6, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 4, 6, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 6, 0, 0]),
+    from_u32x8([5, 6, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 5, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 5, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([2, 5, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 5, 6, 0, 0, 0]),
+    from_u32x8([3, 5, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 5, 6, 0, 0, 0]),
+    from_u32x8([2, 3, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 5, 6, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 5, 6, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 5, 6, 0, 0]),
+    from_u32x8([4, 5, 6, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([1, 4, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 5, 6, 0, 0, 0]),
+    from_u32x8([2, 4, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 5, 6, 0, 0, 0]),
+    from_u32x8([1, 2, 4, 5, 6, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 5, 6, 0, 0]),
+    from_u32x8([3, 4, 5, 6, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 5, 6, 0, 0, 0]),
+    from_u32x8([1, 3, 4, 5, 6, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 5, 6, 0, 0]),
+    from_u32x8([2, 3, 4, 5, 6, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 5, 6, 0, 0]),
+    from_u32x8([1, 2, 3, 4, 5, 6, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 5, 6, 0]),
+    from_u32x8([7, 0, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 7, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 7, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([2, 7, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 7, 0, 0, 0, 0]),
+    from_u32x8([3, 7, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 7, 0, 0, 0, 0]),
+    from_u32x8([2, 3, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 7, 0, 0, 0]),
+    from_u32x8([4, 7, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 4, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 7, 0, 0, 0, 0]),
+    from_u32x8([2, 4, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 4, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 7, 0, 0, 0]),
+    from_u32x8([3, 4, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 4, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 7, 0, 0, 0]),
+    from_u32x8([2, 3, 4, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 7, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 4, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 7, 0, 0]),
+    from_u32x8([5, 7, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 5, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 5, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([2, 5, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 5, 7, 0, 0, 0]),
+    from_u32x8([3, 5, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 5, 7, 0, 0, 0]),
+    from_u32x8([2, 3, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 5, 7, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 5, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 5, 7, 0, 0]),
+    from_u32x8([4, 5, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 4, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 5, 7, 0, 0, 0]),
+    from_u32x8([2, 4, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 5, 7, 0, 0, 0]),
+    from_u32x8([1, 2, 4, 5, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 5, 7, 0, 0]),
+    from_u32x8([3, 4, 5, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 5, 7, 0, 0, 0]),
+    from_u32x8([1, 3, 4, 5, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 5, 7, 0, 0]),
+    from_u32x8([2, 3, 4, 5, 7, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 5, 7, 0, 0]),
+    from_u32x8([1, 2, 3, 4, 5, 7, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 5, 7, 0]),
+    from_u32x8([6, 7, 0, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 6, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([1, 6, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([2, 6, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 2, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 6, 7, 0, 0, 0]),
+    from_u32x8([3, 6, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 3, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 6, 7, 0, 0, 0]),
+    from_u32x8([2, 3, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 6, 7, 0, 0, 0]),
+    from_u32x8([1, 2, 3, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 6, 7, 0, 0]),
+    from_u32x8([4, 6, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 4, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 6, 7, 0, 0, 0]),
+    from_u32x8([2, 4, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 6, 7, 0, 0, 0]),
+    from_u32x8([1, 2, 4, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 6, 7, 0, 0]),
+    from_u32x8([3, 4, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 6, 7, 0, 0, 0]),
+    from_u32x8([1, 3, 4, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 6, 7, 0, 0]),
+    from_u32x8([2, 3, 4, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 6, 7, 0, 0]),
+    from_u32x8([1, 2, 3, 4, 6, 7, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 6, 7, 0]),
+    from_u32x8([5, 6, 7, 0, 0, 0, 0, 0]),
+    from_u32x8([0, 5, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([1, 5, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 1, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([2, 5, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 2, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([1, 2, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 2, 5, 6, 7, 0, 0]),
+    from_u32x8([3, 5, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 3, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([1, 3, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 3, 5, 6, 7, 0, 0]),
+    from_u32x8([2, 3, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 2, 3, 5, 6, 7, 0, 0]),
+    from_u32x8([1, 2, 3, 5, 6, 7, 0, 0]),
+    from_u32x8([0, 1, 2, 3, 5, 6, 7, 0]),
+    from_u32x8([4, 5, 6, 7, 0, 0, 0, 0]),
+    from_u32x8([0, 4, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([1, 4, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 1, 4, 5, 6, 7, 0, 0]),
+    from_u32x8([2, 4, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 2, 4, 5, 6, 7, 0, 0]),
+    from_u32x8([1, 2, 4, 5, 6, 7, 0, 0]),
+    from_u32x8([0, 1, 2, 4, 5, 6, 7, 0]),
+    from_u32x8([3, 4, 5, 6, 7, 0, 0, 0]),
+    from_u32x8([0, 3, 4, 5, 6, 7, 0, 0]),
+    from_u32x8([1, 3, 4, 5, 6, 7, 0, 0]),
+    from_u32x8([0, 1, 3, 4, 5, 6, 7, 0]),
+    from_u32x8([2, 3, 4, 5, 6, 7, 0, 0]),
+    from_u32x8([0, 2, 3, 4, 5, 6, 7, 0]),
+    from_u32x8([1, 2, 3, 4, 5, 6, 7, 0]),
+    from_u32x8([0, 1, 2, 3, 4, 5, 6, 7]),
+];

bitpacker/src/filter_vec/mod.rs

@@ -0,0 +1,165 @@
+use std::ops::RangeInclusive;
+
+#[cfg(target_arch = "x86_64")]
+mod avx2;
+
+mod scalar;
+
+#[derive(Clone, Copy, Eq, PartialEq, Debug)]
+#[repr(u8)]
+enum FilterImplPerInstructionSet {
+    #[cfg(target_arch = "x86_64")]
+    AVX2 = 0u8,
+    Scalar = 1u8,
+}
+
+impl FilterImplPerInstructionSet {
+    #[inline]
+    pub fn is_available(&self) -> bool {
+        match *self {
+            #[cfg(target_arch = "x86_64")]
+            FilterImplPerInstructionSet::AVX2 => is_x86_feature_detected!("avx2"),
+            FilterImplPerInstructionSet::Scalar => true,
+        }
+    }
+}
+
+// List of available implementation in preferred order.
+#[cfg(target_arch = "x86_64")]
+const IMPLS: [FilterImplPerInstructionSet; 2] = [
+    FilterImplPerInstructionSet::AVX2,
+    FilterImplPerInstructionSet::Scalar,
+];
+
+#[cfg(not(target_arch = "x86_64"))]
+const IMPLS: [FilterImplPerInstructionSet; 1] = [FilterImplPerInstructionSet::Scalar];
+
+impl FilterImplPerInstructionSet {
+    #[inline]
+    #[allow(unused_variables)] // on non-x86_64, code is unused.
+    fn from(code: u8) -> FilterImplPerInstructionSet {
+        #[cfg(target_arch = "x86_64")]
+        if code == FilterImplPerInstructionSet::AVX2 as u8 {
+            return FilterImplPerInstructionSet::AVX2;
+        }
+        FilterImplPerInstructionSet::Scalar
+    }
+
+    #[inline]
+    fn filter_vec_in_place(self, range: RangeInclusive<u32>, offset: u32, output: &mut Vec<u32>) {
+        match self {
+            #[cfg(target_arch = "x86_64")]
+            FilterImplPerInstructionSet::AVX2 => avx2::filter_vec_in_place(range, offset, output),
+            FilterImplPerInstructionSet::Scalar => {
+                scalar::filter_vec_in_place(range, offset, output)
+            }
+        }
+    }
+}
+
+#[inline]
+fn get_best_available_instruction_set() -> FilterImplPerInstructionSet {
+    use std::sync::atomic::{AtomicU8, Ordering};
+    static INSTRUCTION_SET_BYTE: AtomicU8 = AtomicU8::new(u8::MAX);
+    let instruction_set_byte: u8 = INSTRUCTION_SET_BYTE.load(Ordering::Relaxed);
+    if instruction_set_byte == u8::MAX {
+        // Let's initialize the instruction set and cache it.
+        let instruction_set = IMPLS
+            .into_iter()
+            .find(FilterImplPerInstructionSet::is_available)
+            .unwrap();
+        INSTRUCTION_SET_BYTE.store(instruction_set as u8, Ordering::Relaxed);
+        return instruction_set;
+    }
+    FilterImplPerInstructionSet::from(instruction_set_byte)
+}
+
+pub fn filter_vec_in_place(range: RangeInclusive<u32>, offset: u32, output: &mut Vec<u32>) {
+    get_best_available_instruction_set().filter_vec_in_place(range, offset, output)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_get_best_available_instruction_set() {
+        // This does not test much unfortunately.
+        // We just make sure the function returns without crashing and returns the same result.
+        let instruction_set = get_best_available_instruction_set();
+        assert_eq!(get_best_available_instruction_set(), instruction_set);
+    }
+
+    #[cfg(target_arch = "x86_64")]
+    #[test]
+    fn test_instruction_set_to_code_from_code() {
+        for instruction_set in [
+            FilterImplPerInstructionSet::AVX2,
+            FilterImplPerInstructionSet::Scalar,
+        ] {
+            let code = instruction_set as u8;
+            assert_eq!(instruction_set, FilterImplPerInstructionSet::from(code));
+        }
+    }
+
+    fn test_filter_impl_empty_aux(filter_impl: FilterImplPerInstructionSet) {
+        let mut output = vec![];
+        filter_impl.filter_vec_in_place(0..=u32::MAX, 0, &mut output);
+        assert_eq!(&output, &[]);
+    }
+
+    fn test_filter_impl_simple_aux(filter_impl: FilterImplPerInstructionSet) {
+        let mut output = vec![3, 2, 1, 5, 11, 2, 5, 10, 2];
+        filter_impl.filter_vec_in_place(3..=10, 0, &mut output);
+        assert_eq!(&output, &[0, 3, 6, 7]);
+    }
+
+    fn test_filter_impl_simple_aux_shifted(filter_impl: FilterImplPerInstructionSet) {
+        let mut output = vec![3, 2, 1, 5, 11, 2, 5, 10, 2];
+        filter_impl.filter_vec_in_place(3..=10, 10, &mut output);
+        assert_eq!(&output, &[10, 13, 16, 17]);
+    }
+
+    fn test_filter_impl_simple_outside_i32_range(filter_impl: FilterImplPerInstructionSet) {
+        let mut output = vec![u32::MAX, i32::MAX as u32 + 1, 0, 1, 3, 1, 1, 1, 1];
+        filter_impl.filter_vec_in_place(1..=i32::MAX as u32 + 1u32, 0, &mut output);
+        assert_eq!(&output, &[1, 3, 4, 5, 6, 7, 8]);
+    }
+
+    fn test_filter_impl_test_suite(filter_impl: FilterImplPerInstructionSet) {
+        test_filter_impl_empty_aux(filter_impl);
+        test_filter_impl_simple_aux(filter_impl);
+        test_filter_impl_simple_aux_shifted(filter_impl);
+        test_filter_impl_simple_outside_i32_range(filter_impl);
+    }
+
+    #[test]
+    #[cfg(target_arch = "x86_64")]
+    fn test_filter_implementation_avx2() {
+        if FilterImplPerInstructionSet::AVX2.is_available() {
+            test_filter_impl_test_suite(FilterImplPerInstructionSet::AVX2);
+        }
+    }
+
+    #[test]
+    fn test_filter_implementation_scalar() {
+        test_filter_impl_test_suite(FilterImplPerInstructionSet::Scalar);
+    }
+
+    #[cfg(target_arch = "x86_64")]
+    proptest::proptest! {
+        #[test]
+        fn test_filter_compare_scalar_and_avx2_impl_proptest(
+            start in proptest::prelude::any::<u32>(),
+            end in proptest::prelude::any::<u32>(),
+            offset in 0u32..2u32,
+            mut vals in proptest::collection::vec(0..u32::MAX, 0..30)) {
+            if FilterImplPerInstructionSet::AVX2.is_available() {
+                let mut vals_clone = vals.clone();
+                FilterImplPerInstructionSet::AVX2.filter_vec_in_place(start..=end, offset, &mut vals);
+                FilterImplPerInstructionSet::Scalar.filter_vec_in_place(start..=end, offset, &mut vals_clone);
+                assert_eq!(&vals, &vals_clone);
+            }
+       }
+    }
+}

bitpacker/src/filter_vec/scalar.rs

@@ -0,0 +1,13 @@
+use std::ops::RangeInclusive;
+
+pub fn filter_vec_in_place(range: RangeInclusive<u32>, offset: u32, output: &mut Vec<u32>) {
+    // We restrict the accepted boundary, because unsigned integers & SIMD don't
+    // play well.
+    let mut output_cursor = 0;
+    for i in 0..output.len() {
+        let val = output[i];
+        output[output_cursor] = offset + i as u32;
+        output_cursor += if range.contains(&val) { 1 } else { 0 };
+    }
+    output.truncate(output_cursor);
+}

bitpacker/src/lib.rs

@@ -0,0 +1,139 @@
+mod bitpacker;
+mod blocked_bitpacker;
+mod filter_vec;
+
+use std::cmp::Ordering;
+
+pub use crate::bitpacker::{BitPacker, BitUnpacker};
+pub use crate::blocked_bitpacker::BlockedBitpacker;
+
+/// Computes the number of bits that will be used for bitpacking.
+///
+/// In general the target is the minimum number of bits
+/// required to express the amplitude given in argument.
+///
+/// e.g. If the amplitude is 10, we can store all ints on simply 4bits.
+///
+/// The logic is slightly more convoluted here as for optimization
+/// reasons, we want to ensure that a value spawns over at most 8 bytes
+/// of aligned bytes.
+///
+/// Spanning over 9 bytes is possible for instance, if we do
+/// bitpacking with an amplitude of 63 bits.
+/// In this case, the second int will start on bit
+/// 63 (which belongs to byte 7) and ends at byte 15;
+/// Hence 9 bytes (from byte 7 to byte 15 included).
+///
+/// To avoid this, we force the number of bits to 64bits
+/// when the result is greater than `64-8 = 56 bits`.
+///
+/// Note that this only affects rare use cases spawning over
+/// a very large range of values. Even in this case, it results
+/// in an extra cost of at most 12% compared to the optimal
+/// number of bits.
+pub fn compute_num_bits(n: u64) -> u8 {
+    let amplitude = (64u32 - n.leading_zeros()) as u8;
+    if amplitude <= 64 - 8 { amplitude } else { 64 }
+}
+
+/// Computes the (min, max) of an iterator of `PartialOrd` values.
+///
+/// For values implementing `Ord` (in a way consistent to their `PartialOrd` impl),
+/// this function behaves as expected.
+///
+/// For values with partial ordering, the behavior is non-trivial and may
+/// depends on the order of the values.
+/// For floats however, it simply returns the same results as if NaN were
+/// skipped.
+pub fn minmax<I, T>(mut vals: I) -> Option<(T, T)>
+where
+    I: Iterator<Item = T>,
+    T: Copy + PartialOrd,
+{
+    let first_el = vals.find(|val| {
+        // We use this to make sure we skip all NaN values when
+        // working with a float type.
+        val.partial_cmp(val) == Some(Ordering::Equal)
+    })?;
+    let mut min_so_far: T = first_el;
+    let mut max_so_far: T = first_el;
+    for val in vals {
+        if val.partial_cmp(&min_so_far) == Some(Ordering::Less) {
+            min_so_far = val;
+        }
+        if val.partial_cmp(&max_so_far) == Some(Ordering::Greater) {
+            max_so_far = val;
+        }
+    }
+    Some((min_so_far, max_so_far))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_compute_num_bits() {
+        assert_eq!(compute_num_bits(1), 1u8);
+        assert_eq!(compute_num_bits(0), 0u8);
+        assert_eq!(compute_num_bits(2), 2u8);
+        assert_eq!(compute_num_bits(3), 2u8);
+        assert_eq!(compute_num_bits(4), 3u8);
+        assert_eq!(compute_num_bits(255), 8u8);
+        assert_eq!(compute_num_bits(256), 9u8);
+        assert_eq!(compute_num_bits(5_000_000_000), 33u8);
+    }
+
+    #[test]
+    fn test_minmax_empty() {
+        let vals: Vec<u32> = vec![];
+        assert_eq!(minmax(vals.into_iter()), None);
+    }
+
+    #[test]
+    fn test_minmax_one() {
+        assert_eq!(minmax(vec![1].into_iter()), Some((1, 1)));
+    }
+
+    #[test]
+    fn test_minmax_two() {
+        assert_eq!(minmax(vec![1, 2].into_iter()), Some((1, 2)));
+        assert_eq!(minmax(vec![2, 1].into_iter()), Some((1, 2)));
+    }
+
+    #[test]
+    fn test_minmax_nan() {
+        assert_eq!(
+            minmax(vec![f64::NAN, 1f64, 2f64].into_iter()),
+            Some((1f64, 2f64))
+        );
+        assert_eq!(
+            minmax(vec![2f64, f64::NAN, 1f64].into_iter()),
+            Some((1f64, 2f64))
+        );
+        assert_eq!(
+            minmax(vec![2f64, 1f64, f64::NAN].into_iter()),
+            Some((1f64, 2f64))
+        );
+    }
+
+    #[test]
+    fn test_minmax_inf() {
+        assert_eq!(
+            minmax(vec![f64::INFINITY, 1f64, 2f64].into_iter()),
+            Some((1f64, f64::INFINITY))
+        );
+        assert_eq!(
+            minmax(vec![-f64::INFINITY, 1f64, 2f64].into_iter()),
+            Some((-f64::INFINITY, 2f64))
+        );
+        assert_eq!(
+            minmax(vec![2f64, f64::INFINITY, 1f64].into_iter()),
+            Some((1f64, f64::INFINITY))
+        );
+        assert_eq!(
+            minmax(vec![2f64, 1f64, -f64::INFINITY].into_iter()),
+            Some((-f64::INFINITY, 2f64))
+        );
+    }
+}

ci/before_deploy.ps1

@@ -1,23 +0,0 @@
-# This script takes care of packaging the build artifacts that will go in the
-# release zipfile
-
-$SRC_DIR = $PWD.Path
-$STAGE = [System.Guid]::NewGuid().ToString()
-
-Set-Location $ENV:Temp
-New-Item -Type Directory -Name $STAGE
-Set-Location $STAGE
-
-$ZIP = "$SRC_DIR\$($Env:CRATE_NAME)-$($Env:APPVEYOR_REPO_TAG_NAME)-$($Env:TARGET).zip"
-
-# TODO Update this to package the right artifacts
-Copy-Item "$SRC_DIR\target\$($Env:TARGET)\release\hello.exe" '.\'
-
-7z a "$ZIP" *
-
-Push-AppveyorArtifact "$ZIP"
-
-Remove-Item *.* -Force
-Set-Location ..
-Remove-Item $STAGE
-Set-Location $SRC_DIR

ci/before_deploy.sh

@@ -1,33 +0,0 @@
-# This script takes care of building your crate and packaging it for release
-
-set -ex
-
-main() {
-    local src=$(pwd) \
-          stage=
-
-    case $TRAVIS_OS_NAME in
-        linux)
-            stage=$(mktemp -d)
-            ;;
-        osx)
-            stage=$(mktemp -d -t tmp)
-            ;;
-    esac
-
-    test -f Cargo.lock || cargo generate-lockfile
-
-    # TODO Update this to build the artifacts that matter to you
-    cross rustc --bin hello --target $TARGET --release -- -C lto
-
-    # TODO Update this to package the right artifacts
-    cp target/$TARGET/release/hello $stage/
-
-    cd $stage
-    tar czf $src/$CRATE_NAME-$TRAVIS_TAG-$TARGET.tar.gz *
-    cd $src
-
-    rm -rf $stage
-}
-
-main

ci/install.sh

@@ -1,47 +0,0 @@
-set -ex
-
-main() {
-    local target=
-    if [ $TRAVIS_OS_NAME = linux ]; then
-        target=x86_64-unknown-linux-musl
-        sort=sort
-    else
-        target=x86_64-apple-darwin
-        sort=gsort  # for `sort --sort-version`, from brew's coreutils.
-    fi
-
-    # Builds for iOS are done on OSX, but require the specific target to be
-    # installed.
-    case $TARGET in
-        aarch64-apple-ios)
-            rustup target install aarch64-apple-ios
-            ;;
-        armv7-apple-ios)
-            rustup target install armv7-apple-ios
-            ;;
-        armv7s-apple-ios)
-            rustup target install armv7s-apple-ios
-            ;;
-        i386-apple-ios)
-            rustup target install i386-apple-ios
-            ;;
-        x86_64-apple-ios)
-            rustup target install x86_64-apple-ios
-            ;;
-    esac
-
-    # This fetches latest stable release
-    local tag=$(git ls-remote --tags --refs --exit-code https://github.com/japaric/cross \
-                       | cut -d/ -f3 \
-                       | grep -E '^v[0.1.0-9.]+$' \
-                       | $sort --version-sort \
-                       | tail -n1)
-    curl -LSfs https://japaric.github.io/trust/install.sh | \
-        sh -s -- \
-           --force \
-           --git japaric/cross \
-           --tag $tag \
-           --target $target
-}
-
-main

ci/script.sh

@@ -1,29 +0,0 @@
-#!/usr/bin/env bash
-
-# This script takes care of testing your crate
-
-set -ex
-
-main() {
-    if [ ! -z $CODECOV ]; then
-        echo "Codecov"
-        cargo build --verbose && cargo coverage --verbose && bash <(curl -s https://codecov.io/bash) -s target/kcov
-    else
-        echo "Build"
-        cross build --target $TARGET
-        if [ ! -z $DISABLE_TESTS ]; then
-            return
-        fi
-        echo "Test"
-        cross test --target $TARGET --no-default-features --features mmap -- --test-threads 1
-    fi
-    for example in $(ls examples/*.rs)
-    do
-        cargo run --example  $(basename $example .rs)
-    done
-}
-
-# we don't run the "test phase" when doing deploys
-if [ -z $TRAVIS_TAG ]; then
-    main
-fi

cliff.toml

@@ -0,0 +1,93 @@
+# configuration file for git-cliff{ pattern = "foo", replace = "bar"}
+# see https://github.com/orhun/git-cliff#configuration-file
+
+[remote.github]
+owner = "quickwit-oss"
+repo = "tantivy"
+
+[changelog]
+# changelog header
+header = """
+"""
+# template for the changelog body
+# https://tera.netlify.app/docs/#introduction
+body = """
+## What's Changed
+
+{%- if version %} in {{ version }}{%- endif -%}
+{% for commit in commits %}
+  {% if commit.remote.pr_title -%}
+    {%- set commit_message = commit.remote.pr_title -%}
+  {%- else -%}
+    {%- set commit_message = commit.message -%}
+  {%- endif -%}
+  - {{ commit_message | split(pat="\n") | first | trim }}\
+    {% if commit.remote.pr_number %} \
+      [#{{ commit.remote.pr_number }}]({{ self::remote_url() }}/pull/{{ commit.remote.pr_number }}){% if commit.remote.username %}(@{{ commit.remote.username }}){%- endif -%} \
+    {%- endif %}
+{%- endfor -%}
+
+{% if github.contributors | filter(attribute="is_first_time", value=true) | length != 0 %}
+  {% raw %}\n{% endraw -%}
+  ## New Contributors
+{%- endif %}\
+{% for contributor in github.contributors | filter(attribute="is_first_time", value=true) %}
+  * @{{ contributor.username }} made their first contribution
+    {%- if contributor.pr_number %} in \
+      [#{{ contributor.pr_number }}]({{ self::remote_url() }}/pull/{{ contributor.pr_number }}) \
+    {%- endif %}
+{%- endfor -%}
+
+{% if version %}
+    {% if previous.version %}
+      **Full Changelog**: {{ self::remote_url() }}/compare/{{ previous.version }}...{{ version }}
+    {% endif %}
+{% else -%}
+  {% raw %}\n{% endraw %}
+{% endif %}
+
+{%- macro remote_url() -%}
+  https://github.com/{{ remote.github.owner }}/{{ remote.github.repo }}
+{%- endmacro -%}
+"""
+# remove the leading and trailing whitespace from the template
+trim = true
+# changelog footer
+footer = """
+"""
+
+postprocessors = [
+]
+
+[git]
+# parse the commits based on https://www.conventionalcommits.org
+# This is required or commit.message contains the whole commit message and not just the title
+conventional_commits = false
+# filter out the commits that are not conventional
+filter_unconventional = true
+# process each line of a commit as an individual commit
+split_commits = false
+# regex for preprocessing the commit messages
+commit_preprocessors = [
+    { pattern = '\((\w+\s)?#([0-9]+)\)', replace = ""},
+]
+#link_parsers = [
+    #{ pattern = "#(\\d+)", href = "https://github.com/quickwit-oss/tantivy/pulls/$1"},
+#]
+# regex for parsing and grouping commits
+# protect breaking changes from being skipped due to matching a skipping commit_parser
+protect_breaking_commits = false
+# filter out the commits that are not matched by commit parsers
+filter_commits = false
+# glob pattern for matching git tags
+tag_pattern = "v[0-9]*"
+# regex for skipping tags
+skip_tags = "v0.1.0-beta.1"
+# regex for ignoring tags
+ignore_tags = ""
+# sort the tags topologically
+topo_order = false
+# sort the commits inside sections by oldest/newest order
+sort_commits = "newest"
+# limit the number of commits included in the changelog.
+# limit_commits = 42

columnar/Cargo.toml

@@ -0,0 +1,38 @@
+[package]
+name = "tantivy-columnar"
+version = "0.5.0"
+edition = "2024"
+license = "MIT"
+homepage = "https://github.com/quickwit-oss/tantivy"
+repository = "https://github.com/quickwit-oss/tantivy"
+description = "column oriented storage for tantivy"
+categories = ["database-implementations", "data-structures", "compression"]
+
+[dependencies]
+itertools = "0.14.0"
+fastdivide = "0.4.0"
+
+stacker = { version= "0.5", path = "../stacker", package="tantivy-stacker"}
+sstable = { version= "0.5", path = "../sstable", package = "tantivy-sstable" }
+common = { version= "0.9", path = "../common", package = "tantivy-common" }
+tantivy-bitpacker = { version= "0.8", path = "../bitpacker/" }
+serde = "1.0.152"
+downcast-rs = "2.0.1"
+
+[dev-dependencies]
+proptest = "1"
+more-asserts = "0.3.1"
+rand = "0.8"
+binggan = "0.14.0"
+
+[[bench]]
+name = "bench_merge"
+harness = false
+
+[[bench]]
+name = "bench_access"
+harness = false
+
+[features]
+unstable = []
+zstd-compression = ["sstable/zstd-compression"]

columnar/README.md

@@ -0,0 +1,109 @@
+# Columnar format
+
+This crate describes columnar format used in tantivy.
+
+## Goals
+
+This format is special in the following way.
+- it needs to be compact
+- accessing a specific column does not require to load the entire columnar. It can be done in 2 to 3 random access.
+- columns of several types can be associated with the same column name.
+- it needs to support columns with different types `(str, u64, i64, f64)`
+and different cardinality `(required, optional, multivalued)`.
+- columns, once loaded, offer cheap random access.
+- it is designed to allow range queries.
+
+# Coercion rules
+
+Users can create a columnar by inserting rows to a `ColumnarWriter`,
+and serializing it into a `Write` object.
+Nothing prevents a user from recording values with different type to the same `column_name`.
+
+In that case, `tantivy-columnar`'s behavior is as follows:
+- JsonValues are grouped into 3 types (String, Number, bool).
+Values that corresponds to different groups are mapped to different columns. For instance, String values are treated independently
+from Number or boolean values. `tantivy-columnar` will simply emit several columns associated to a given column_name.
+- Only one column for a given json value type is emitted.  If number values with different number types are recorded (e.g. u64, i64, f64),
+`tantivy-columnar` will pick the first type that can represents the set of appended value, with the following prioriy order (`i64`, `u64`, `f64`).
+`i64` is picked over `u64` as it is likely to  yield less change of types. Most use cases strictly requiring `u64` show the
+restriction on 50% of the values (e.g. a 64-bit hash). On the other hand, a lot of use cases can show rare negative value.
+
+# Columnar format
+
+This columnar format may have more than one column (with different types) associated to the same `column_name` (see [Coercion rules](#coercion-rules) above).
+The `(column_name, column_type)` couple however uniquely identifies a column.
+That couple is serialized as a column `column_key`.  The format of that key is:
+`[column_name][ZERO_BYTE][column_type_header: u8]`
+
+```
+COLUMNAR:=
+    [COLUMNAR_DATA]
+    [COLUMNAR_KEY_TO_DATA_INDEX]
+    [COLUMNAR_FOOTER];
+
+
+# Columns are sorted by their column key.
+COLUMNAR_DATA:=
+    [COLUMN_DATA]+;
+
+COLUMNAR_FOOTER := [RANGE_SSTABLE_BYTES_LEN: 8 bytes little endian]
+
+```
+
+The columnar file starts by the actual column data, concatenated one after the other,
+sorted by column key.
+
+A sstable associates
+`(column name, column_cardinality, column_type) to range of bytes.
+
+Column name may not contain the zero byte `\0`.
+
+Listing all columns associated to `column_name` can therefore
+be done by listing all keys prefixed by
+`[column_name][ZERO_BYTE]`
+
+The associated range of bytes refer to a range of bytes
+
+This crate exposes a columnar format for tantivy.
+This format is described in README.md
+
+
+The crate introduces the following concepts.
+
+`Columnar` is an equivalent of a dataframe.
+It maps `column_key` to `Column`.
+
+A `Column<T>` asssociates a `RowId` (u32) to any
+number of values.
+
+This is made possible by wrapping a `ColumnIndex` and a `ColumnValue` object.
+The `ColumnValue<T>` represents a mapping that associates each `RowId` to
+exactly one single value.
+
+The `ColumnIndex` then maps each RowId to a set of `RowId` in the
+`ColumnValue`.
+
+For optimization, and compression purposes, the `ColumnIndex` has three
+possible representation, each for different cardinalities.
+
+- Full
+
+All RowId have exactly one value. The ColumnIndex is the trivial mapping.
+
+- Optional
+
+All RowIds can have at most one value. The ColumnIndex is the trivial mapping `ColumnRowId -> Option<ColumnValueRowId>`.
+
+- Multivalued
+
+All RowIds can have any number of values.
+The column index is mapping values to a range.
+
+
+All these objects are implemented an unit tested independently
+in their own module:
+
+- columnar
+- column_index
+- column_values
+- column

columnar/benches/bench_access.rs

@@ -0,0 +1,67 @@
+use binggan::{InputGroup, black_box};
+use common::*;
+use tantivy_columnar::Column;
+
+pub mod common;
+
+const NUM_DOCS: u32 = 2_000_000;
+
+pub fn generate_columnar_and_open(card: Card, num_docs: u32) -> Column {
+    let reader = generate_columnar_with_name(card, num_docs, "price");
+    reader.read_columns("price").unwrap()[0]
+        .open_u64_lenient()
+        .unwrap()
+        .unwrap()
+}
+
+fn main() {
+    let mut inputs = Vec::new();
+
+    let mut add_card = |card1: Card| {
+        inputs.push((
+            format!("{card1}"),
+            generate_columnar_and_open(card1, NUM_DOCS),
+        ));
+    };
+
+    add_card(Card::MultiSparse);
+    add_card(Card::Multi);
+    add_card(Card::Sparse);
+    add_card(Card::Dense);
+    add_card(Card::Full);
+
+    bench_group(InputGroup::new_with_inputs(inputs));
+}
+
+fn bench_group(mut runner: InputGroup<Column>) {
+    runner.register("access_values_for_doc", |column| {
+        let mut sum = 0;
+        for i in 0..NUM_DOCS {
+            for value in column.values_for_doc(i) {
+                sum += value;
+            }
+        }
+        black_box(sum);
+    });
+    runner.register("access_first_vals", |column| {
+        let mut sum = 0;
+        const BLOCK_SIZE: usize = 32;
+        let mut docs = vec![0; BLOCK_SIZE];
+        let mut buffer = vec![None; BLOCK_SIZE];
+        for i in (0..NUM_DOCS).step_by(BLOCK_SIZE) {
+            // fill docs
+            for idx in 0..BLOCK_SIZE {
+                docs[idx] = idx as u32 + i;
+            }
+
+            column.first_vals(&docs, &mut buffer);
+            for val in buffer.iter() {
+                let Some(val) = val else { continue };
+                sum += *val;
+            }
+        }
+
+        black_box(sum);
+    });
+    runner.run();
+}

columnar/benches/bench_first_vals.rs

@@ -0,0 +1,155 @@
+#![feature(test)]
+extern crate test;
+
+use std::sync::Arc;
+
+use rand::prelude::*;
+use tantivy_columnar::column_values::{CodecType, serialize_and_load_u64_based_column_values};
+use tantivy_columnar::*;
+use test::{Bencher, black_box};
+
+struct Columns {
+    pub optional: Column,
+    pub full: Column,
+    pub multi: Column,
+}
+
+fn get_test_columns() -> Columns {
+    let data = generate_permutation();
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (idx, val) in data.iter().enumerate() {
+        dataframe_writer.record_numerical(idx as u32, "full_values", NumericalValue::U64(*val));
+        if idx % 2 == 0 {
+            dataframe_writer.record_numerical(
+                idx as u32,
+                "optional_values",
+                NumericalValue::U64(*val),
+            );
+        }
+        dataframe_writer.record_numerical(idx as u32, "multi_values", NumericalValue::U64(*val));
+        dataframe_writer.record_numerical(idx as u32, "multi_values", NumericalValue::U64(*val));
+    }
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(data.len() as u32, &mut buffer)
+        .unwrap();
+    let columnar = ColumnarReader::open(buffer).unwrap();
+
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("optional_values").unwrap();
+    assert_eq!(cols.len(), 1);
+    let optional = cols[0].open_u64_lenient().unwrap().unwrap();
+    assert_eq!(optional.index.get_cardinality(), Cardinality::Optional);
+
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("full_values").unwrap();
+    assert_eq!(cols.len(), 1);
+    let column_full = cols[0].open_u64_lenient().unwrap().unwrap();
+    assert_eq!(column_full.index.get_cardinality(), Cardinality::Full);
+
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("multi_values").unwrap();
+    assert_eq!(cols.len(), 1);
+    let multi = cols[0].open_u64_lenient().unwrap().unwrap();
+    assert_eq!(multi.index.get_cardinality(), Cardinality::Multivalued);
+
+    Columns {
+        optional,
+        full: column_full,
+        multi,
+    }
+}
+
+const NUM_VALUES: u64 = 100_000;
+fn generate_permutation() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..NUM_VALUES).collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+pub fn serialize_and_load(column: &[u64], codec_type: CodecType) -> Arc<dyn ColumnValues<u64>> {
+    serialize_and_load_u64_based_column_values(&column, &[codec_type])
+}
+
+fn run_bench_on_column_full_scan(b: &mut Bencher, column: Column) {
+    let num_iter = black_box(NUM_VALUES);
+    b.iter(|| {
+        let mut sum = 0u64;
+        for i in 0..num_iter as u32 {
+            let val = column.first(i);
+            sum += val.unwrap_or(0);
+        }
+        sum
+    });
+}
+fn run_bench_on_column_block_fetch(b: &mut Bencher, column: Column) {
+    let mut block: Vec<Option<u64>> = vec![None; 64];
+    let fetch_docids = (0..64).collect::<Vec<_>>();
+    b.iter(move || {
+        column.first_vals(&fetch_docids, &mut block);
+        block[0]
+    });
+}
+fn run_bench_on_column_block_single_calls(b: &mut Bencher, column: Column) {
+    let mut block: Vec<Option<u64>> = vec![None; 64];
+    let fetch_docids = (0..64).collect::<Vec<_>>();
+    b.iter(move || {
+        for i in 0..fetch_docids.len() {
+            block[i] = column.first(fetch_docids[i]);
+        }
+        block[0]
+    });
+}
+
+/// Column first method
+#[bench]
+fn bench_get_first_on_full_column_full_scan(b: &mut Bencher) {
+    let column = get_test_columns().full;
+    run_bench_on_column_full_scan(b, column);
+}
+
+#[bench]
+fn bench_get_first_on_optional_column_full_scan(b: &mut Bencher) {
+    let column = get_test_columns().optional;
+    run_bench_on_column_full_scan(b, column);
+}
+
+#[bench]
+fn bench_get_first_on_multi_column_full_scan(b: &mut Bencher) {
+    let column = get_test_columns().multi;
+    run_bench_on_column_full_scan(b, column);
+}
+
+/// Block fetch column accessor
+#[bench]
+fn bench_get_block_first_on_optional_column(b: &mut Bencher) {
+    let column = get_test_columns().optional;
+    run_bench_on_column_block_fetch(b, column);
+}
+
+#[bench]
+fn bench_get_block_first_on_multi_column(b: &mut Bencher) {
+    let column = get_test_columns().multi;
+    run_bench_on_column_block_fetch(b, column);
+}
+
+#[bench]
+fn bench_get_block_first_on_full_column(b: &mut Bencher) {
+    let column = get_test_columns().full;
+    run_bench_on_column_block_fetch(b, column);
+}
+
+#[bench]
+fn bench_get_block_first_on_optional_column_single_calls(b: &mut Bencher) {
+    let column = get_test_columns().optional;
+    run_bench_on_column_block_single_calls(b, column);
+}
+
+#[bench]
+fn bench_get_block_first_on_multi_column_single_calls(b: &mut Bencher) {
+    let column = get_test_columns().multi;
+    run_bench_on_column_block_single_calls(b, column);
+}
+
+#[bench]
+fn bench_get_block_first_on_full_column_single_calls(b: &mut Bencher) {
+    let column = get_test_columns().full;
+    run_bench_on_column_block_single_calls(b, column);
+}

columnar/benches/bench_merge.rs

@@ -0,0 +1,49 @@
+pub mod common;
+
+use binggan::BenchRunner;
+use common::{Card, generate_columnar_with_name};
+use tantivy_columnar::*;
+
+const NUM_DOCS: u32 = 100_000;
+
+fn main() {
+    let mut inputs = Vec::new();
+
+    let mut add_combo = |card1: Card, card2: Card| {
+        inputs.push((
+            format!("merge_{card1}_and_{card2}"),
+            vec![
+                generate_columnar_with_name(card1, NUM_DOCS, "price"),
+                generate_columnar_with_name(card2, NUM_DOCS, "price"),
+            ],
+        ));
+    };
+
+    add_combo(Card::Multi, Card::Multi);
+    add_combo(Card::MultiSparse, Card::MultiSparse);
+    add_combo(Card::Dense, Card::Dense);
+    add_combo(Card::Sparse, Card::Sparse);
+    add_combo(Card::Sparse, Card::Dense);
+    add_combo(Card::MultiSparse, Card::Dense);
+    add_combo(Card::MultiSparse, Card::Sparse);
+    add_combo(Card::Multi, Card::Dense);
+    add_combo(Card::Multi, Card::Sparse);
+
+    let mut runner: BenchRunner = BenchRunner::new();
+    let mut group = runner.new_group();
+    for (input_name, columnar_readers) in inputs.iter() {
+        group.register_with_input(
+            input_name,
+            columnar_readers,
+            move |columnar_readers: &Vec<ColumnarReader>| {
+                let mut out = Vec::new();
+                let columnar_readers = columnar_readers.iter().collect::<Vec<_>>();
+                let merge_row_order = StackMergeOrder::stack(&columnar_readers[..]);
+
+                merge_columnar(&columnar_readers, &[], merge_row_order.into(), &mut out).unwrap();
+                Some(out.len() as u64)
+            },
+        );
+    }
+    group.run();
+}

columnar/benches/bench_values_u128.rs

@@ -0,0 +1,124 @@
+#![feature(test)]
+
+use std::ops::RangeInclusive;
+use std::sync::Arc;
+
+use common::OwnedBytes;
+use rand::rngs::StdRng;
+use rand::seq::SliceRandom;
+use rand::{Rng, SeedableRng, random};
+use tantivy_columnar::ColumnValues;
+use test::Bencher;
+extern crate test;
+
+// TODO does this make sense for IPv6 ?
+fn generate_random() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..100_000u64)
+        .map(|el| el + random::<u16>() as u64)
+        .collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+fn get_u128_column_random() -> Arc<dyn ColumnValues<u128>> {
+    let permutation = generate_random();
+    let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
+    get_u128_column_from_data(&permutation)
+}
+
+fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn ColumnValues<u128>> {
+    let mut out = vec![];
+    tantivy_columnar::column_values::serialize_column_values_u128(&data, &mut out).unwrap();
+    let out = OwnedBytes::new(out);
+    tantivy_columnar::column_values::open_u128_mapped::<u128>(out).unwrap()
+}
+
+const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
+const SINGLE_ITEM: u64 = 90;
+const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
+
+fn get_data_50percent_item() -> Vec<u128> {
+    let mut rng = StdRng::from_seed([1u8; 32]);
+
+    let mut data = vec![];
+    for _ in 0..300_000 {
+        let val = rng.gen_range(1..=100);
+        data.push(val);
+    }
+    data.push(SINGLE_ITEM);
+    data.shuffle(&mut rng);
+    let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
+    data
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let column = get_u128_column_from_data(&data);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(
+            *FIFTY_PERCENT_RANGE.start() as u128..=*FIFTY_PERCENT_RANGE.end() as u128,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let column = get_u128_column_from_data(&data);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(
+            *SINGLE_ITEM_RANGE.start() as u128..=*SINGLE_ITEM_RANGE.end() as u128,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let column = get_u128_column_from_data(&data);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
+        positions
+    });
+}
+// U128 RANGE END
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
+    let column = get_u128_column_random();
+
+    b.iter(|| {
+        let mut a = 0u128;
+        for i in 0u64..column.num_vals() as u64 {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_stride5_u128(b: &mut Bencher) {
+    let column = get_u128_column_random();
+
+    b.iter(|| {
+        let n = column.num_vals();
+        let mut a = 0u128;
+        for i in (0..n / 5).map(|val| val * 5) {
+            a += column.get_val(i);
+        }
+        a
+    });
+}

columnar/benches/bench_values_u64.rs

@@ -0,0 +1,203 @@
+#![feature(test)]
+extern crate test;
+
+use std::ops::RangeInclusive;
+use std::sync::Arc;
+
+use rand::prelude::*;
+use tantivy_columnar::column_values::{CodecType, serialize_and_load_u64_based_column_values};
+use tantivy_columnar::*;
+use test::Bencher;
+
+// Warning: this generates the same permutation at each call
+fn generate_permutation() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+// Warning: this generates the same permutation at each call
+fn generate_permutation_gcd() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+pub fn serialize_and_load(column: &[u64], codec_type: CodecType) -> Arc<dyn ColumnValues<u64>> {
+    serialize_and_load_u64_based_column_values(&column, &[codec_type])
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    b.iter(|| {
+        let mut a = 0u64;
+        for _ in 0..n {
+            a = permutation[a as usize];
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_fflookup_bitpacked(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0u64;
+        for _ in 0..n {
+            a = column.get_val(a as u32);
+        }
+        a
+    });
+}
+
+const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
+const SINGLE_ITEM: u64 = 90;
+const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
+const ONE_PERCENT_ITEM_RANGE: RangeInclusive<u64> = 49..=49;
+fn get_data_50percent_item() -> Vec<u128> {
+    let mut rng = StdRng::from_seed([1u8; 32]);
+
+    let mut data = vec![];
+    for _ in 0..300_000 {
+        let val = rng.gen_range(1..=100);
+        data.push(val);
+    }
+    data.push(SINGLE_ITEM);
+
+    data.shuffle(&mut rng);
+    let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
+    data
+}
+
+// U64 RANGE START
+#[bench]
+fn bench_intfastfield_getrange_u64_50percent_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(
+            FIFTY_PERCENT_RANGE,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u64_1percent_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(
+            ONE_PERCENT_ITEM_RANGE,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u64_single_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(SINGLE_ITEM_RANGE, 0..data.len() as u32, &mut positions);
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u64_hit_all(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
+        positions
+    });
+}
+// U64 RANGE END
+
+#[bench]
+fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in (0..n / 7).map(|val| val * 7) {
+            a += permutation[i as usize];
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0;
+        for i in (0..n / 7).map(|val| val * 7) {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    let column_ref = column.as_ref();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0u32..n as u32 {
+            a += column_ref.get_val(i);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
+    let permutation = generate_permutation_gcd();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0..n {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0..permutation.len() {
+            a += permutation[i as usize] as u64;
+        }
+        a
+    });
+}

columnar/benches/common.rs

@@ -0,0 +1,59 @@
+extern crate tantivy_columnar;
+
+use core::fmt;
+use std::fmt::{Display, Formatter};
+
+use tantivy_columnar::{ColumnarReader, ColumnarWriter};
+
+pub enum Card {
+    MultiSparse,
+    Multi,
+    Sparse,
+    Dense,
+    Full,
+}
+impl Display for Card {
+    fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+        match self {
+            Card::MultiSparse => write!(f, "multi sparse 1/13"),
+            Card::Multi => write!(f, "multi 2x"),
+            Card::Sparse => write!(f, "sparse 1/13"),
+            Card::Dense => write!(f, "dense 1/12"),
+            Card::Full => write!(f, "full"),
+        }
+    }
+}
+pub fn generate_columnar_with_name(card: Card, num_docs: u32, column_name: &str) -> ColumnarReader {
+    let mut columnar_writer = ColumnarWriter::default();
+
+    if let Card::MultiSparse = card {
+        columnar_writer.record_numerical(0, column_name, 10u64);
+        columnar_writer.record_numerical(0, column_name, 10u64);
+    }
+
+    for i in 0..num_docs {
+        match card {
+            Card::MultiSparse | Card::Sparse => {
+                if i % 13 == 0 {
+                    columnar_writer.record_numerical(i, column_name, i as u64);
+                }
+            }
+            Card::Dense => {
+                if i % 12 == 0 {
+                    columnar_writer.record_numerical(i, column_name, i as u64);
+                }
+            }
+            Card::Full => {
+                columnar_writer.record_numerical(i, column_name, i as u64);
+            }
+            Card::Multi => {
+                columnar_writer.record_numerical(i, column_name, i as u64);
+                columnar_writer.record_numerical(i, column_name, i as u64);
+            }
+        }
+    }
+
+    let mut wrt: Vec<u8> = Vec::new();
+    columnar_writer.serialize(num_docs, &mut wrt).unwrap();
+    ColumnarReader::open(wrt).unwrap()
+}

columnar/columnar-cli-inspect/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "tantivy-columnar-inspect"
+version = "0.1.0"
+edition = "2021"
+license = "MIT"
+
+[dependencies]
+tantivy = {path="../..", package="tantivy"}
+columnar = {path="../", package="tantivy-columnar"}
+common = {path="../../common", package="tantivy-common"}
+
+[workspace]
+members = []
+
+[profile.release]
+debug = true
+#debug-assertions = true
+#overflow-checks = true

columnar/columnar-cli-inspect/src/main.rs

@@ -0,0 +1,54 @@
+use columnar::ColumnarReader;
+use common::file_slice::{FileSlice, WrapFile};
+use std::io;
+use std::path::Path;
+use tantivy::directory::footer::Footer;
+
+fn main() -> io::Result<()> {
+    println!("Opens a columnar file written by tantivy and validates it.");
+    let path = std::env::args().nth(1).unwrap();
+
+    let path = Path::new(&path);
+    println!("Reading {:?}", path);
+    let _reader = open_and_validate_columnar(path.to_str().unwrap())?;
+
+    Ok(())
+}
+
+pub fn validate_columnar_reader(reader: &ColumnarReader) {
+    let num_rows = reader.num_rows();
+    println!("num_rows: {}", num_rows);
+    let columns = reader.list_columns().unwrap();
+    println!("num columns: {:?}", columns.len());
+    for (col_name, dynamic_column_handle) in columns {
+        let col = dynamic_column_handle.open().unwrap();
+        match col {
+            columnar::DynamicColumn::Bool(_)
+            | columnar::DynamicColumn::I64(_)
+            | columnar::DynamicColumn::U64(_)
+            | columnar::DynamicColumn::F64(_)
+            | columnar::DynamicColumn::IpAddr(_)
+            | columnar::DynamicColumn::DateTime(_)
+            | columnar::DynamicColumn::Bytes(_) => {}
+            columnar::DynamicColumn::Str(str_column) => {
+                let num_vals = str_column.ords().values.num_vals();
+                let num_terms_dict = str_column.num_terms() as u64;
+                let max_ord = str_column.ords().values.iter().max().unwrap_or_default();
+                println!("{col_name:35}  num_vals {num_vals:10} \t num_terms_dict {num_terms_dict:8} max_ord: {max_ord:8}",);
+                for ord in str_column.ords().values.iter() {
+                    assert!(ord < num_terms_dict);
+                }
+            }
+        }
+    }
+}
+
+/// Opens a columnar file that was written by tantivy and validates it.
+pub fn open_and_validate_columnar(path: &str) -> io::Result<ColumnarReader> {
+    let wrap_file = WrapFile::new(std::fs::File::open(path)?)?;
+    let slice = FileSlice::new(std::sync::Arc::new(wrap_file));
+    let (_footer, slice) = Footer::extract_footer(slice.clone()).unwrap();
+    let reader = ColumnarReader::open(slice).unwrap();
+    validate_columnar_reader(&reader);
+    Ok(reader)
+}

columnar/columnar-cli/Cargo.toml

@@ -0,0 +1,16 @@
+[package]
+name = "tantivy-columnar-cli"
+version = "0.1.0"
+edition = "2021"
+license = "MIT"
+
+[dependencies]
+columnar = {path="../", package="tantivy-columnar"}
+serde_json = "1"
+serde_json_borrow = {git="https://github.com/PSeitz/serde_json_borrow/"}
+
+[workspace]
+members = []
+
+[profile.release]
+debug = true

columnar/columnar-cli/src/main.rs

@@ -0,0 +1,134 @@
+use columnar::ColumnarWriter;
+use columnar::NumericalValue;
+use serde_json_borrow;
+use std::fs::File;
+use std::io;
+use std::io::BufRead;
+use std::io::BufReader;
+use std::time::Instant;
+
+#[derive(Default)]
+struct JsonStack {
+    path: String,
+    stack: Vec<usize>,
+}
+
+impl JsonStack {
+    fn push(&mut self, seg: &str) {
+        let len = self.path.len();
+        self.stack.push(len);
+        self.path.push('.');
+        self.path.push_str(seg);
+    }
+
+    fn pop(&mut self) {
+        if let Some(len) = self.stack.pop() {
+            self.path.truncate(len);
+        }
+    }
+
+    fn path(&self) -> &str {
+        &self.path[1..]
+    }
+}
+
+fn append_json_to_columnar(
+    doc: u32,
+    json_value: &serde_json_borrow::Value,
+    columnar: &mut ColumnarWriter,
+    stack: &mut JsonStack,
+) -> usize {
+    let mut count = 0;
+    match json_value {
+        serde_json_borrow::Value::Null => {}
+        serde_json_borrow::Value::Bool(val) => {
+            columnar.record_numerical(
+                doc,
+                stack.path(),
+                NumericalValue::from(if *val { 1u64 } else { 0u64 }),
+            );
+            count += 1;
+        }
+        serde_json_borrow::Value::Number(num) => {
+            let numerical_value: NumericalValue = if let Some(num_i64) = num.as_i64() {
+                num_i64.into()
+            } else if let Some(num_u64) = num.as_u64() {
+                num_u64.into()
+            } else if let Some(num_f64) = num.as_f64() {
+                num_f64.into()
+            } else {
+                panic!();
+            };
+            count += 1;
+            columnar.record_numerical(
+                doc,
+                stack.path(),
+                numerical_value,
+            );
+        }
+        serde_json_borrow::Value::Str(msg) => {
+            columnar.record_str(
+                doc,
+                stack.path(),
+                msg,
+            );
+            count += 1;
+        },
+        serde_json_borrow::Value::Array(vals) => {
+            for val in vals {
+                count += append_json_to_columnar(doc, val, columnar, stack);
+            }
+        },
+        serde_json_borrow::Value::Object(json_map) => {
+            for (child_key, child_val) in json_map {
+                stack.push(child_key);
+                count += append_json_to_columnar(doc, child_val, columnar, stack);
+                stack.pop();
+            }
+        },
+    }
+    count
+}
+
+fn main() -> io::Result<()> {
+    let file = File::open("gh_small.json")?;
+    let mut reader = BufReader::new(file);
+    let mut line = String::with_capacity(100);
+    let mut columnar = columnar::ColumnarWriter::default();
+    let mut doc = 0;
+    let start = Instant::now();
+    let mut stack = JsonStack::default();
+    let mut total_count = 0;
+
+    let start_build = Instant::now();
+    loop {
+        line.clear();
+        let len = reader.read_line(&mut line)?;
+        if len == 0 {
+            break;
+        }
+        let Ok(json_value) = serde_json::from_str::<serde_json_borrow::Value>(&line) else { continue; };
+        total_count += append_json_to_columnar(doc, &json_value, &mut columnar, &mut stack);
+        doc += 1;
+    }
+    println!("Build in {:?}", start_build.elapsed());
+
+    println!("value count {total_count}");
+
+    let mut buffer = Vec::new();
+    let start_serialize = Instant::now();
+    columnar.serialize(doc, None, &mut buffer)?;
+    println!("Serialized in {:?}", start_serialize.elapsed());
+    println!("num docs: {doc}, {:?}", start.elapsed());
+    println!("buffer len {} MB", buffer.len() / 1_000_000);
+    let columnar = columnar::ColumnarReader::open(buffer)?;
+    for (column_name, dynamic_column) in columnar.list_columns()? {
+        let num_bytes = dynamic_column.num_bytes();
+        let typ = dynamic_column.column_type();
+        if num_bytes > 1_000_000 {
+            println!("{column_name} {typ:?}  {} KB", num_bytes / 1_000);
+        }
+    }
+    println!("{} columns", columnar.num_columns());
+    Ok(())
+}

columnar/src/TODO.md

@@ -0,0 +1,47 @@
+# zero to one
+
+* revisit line codec
+* add columns from schema on merge
+* Plugging JSON
+* replug examples
+* move datetime to quickwit common
+* switch to nanos
+* reintroduce the gcd map.
+
+# Perf and Size
+* remove alloc in `ord_to_term`
++ multivaued range queries restart from the beginning all of the time.
+* re-add ZSTD compression for dictionaries
+no systematic monotonic mapping
+consider removing multilinear
+f32?
+adhoc solution for bool?
+add metrics helper for aggregate. sum(row_id)
+review inline absence/presence
+improv perf of select using PDEP
+compare with roaring bitmap/elias fano etc etc.
+SIMD range? (see blog post)
+Add alignment?
+Consider another codec to bridge the gap between few and 5k elements
+
+# Cleanup and rationalization
+in benchmark, unify percent vs ratio, f32 vs f64.
+investigate if should have better errors? io::Error is overused at the moment.
+rename rank/select in unit tests
+Review the public API via cargo doc
+go through TODOs
+remove all  doc_id occurrences -> row_id
+use the rank & select naming in unit tests branch.
+multi-linear -> blockwise
+linear codec -> simply a multiplication for the index column
+rename columnar to something more explicit, like column_dictionary or columnar_table
+rename fastfield -> column
+document changes
+rationalization FastFieldValue, HasColumnType
+isolate u128_based and uniform naming
+
+# Other
+fix enhance column-cli
+
+# Santa Claus
+autodetect datetime ipaddr, plug customizable tokenizer.

columnar/src/block_accessor.rs

@@ -0,0 +1,158 @@
+use std::cmp::Ordering;
+
+use crate::{Column, DocId, RowId};
+
+#[derive(Debug, Default, Clone)]
+pub struct ColumnBlockAccessor<T> {
+    val_cache: Vec<T>,
+    docid_cache: Vec<DocId>,
+    missing_docids_cache: Vec<DocId>,
+    row_id_cache: Vec<RowId>,
+}
+
+impl<T: PartialOrd + Copy + std::fmt::Debug + Send + Sync + 'static + Default>
+    ColumnBlockAccessor<T>
+{
+    #[inline]
+    pub fn fetch_block<'a>(&'a mut self, docs: &'a [u32], accessor: &Column<T>) {
+        if accessor.index.get_cardinality().is_full() {
+            self.val_cache.resize(docs.len(), T::default());
+            accessor.values.get_vals(docs, &mut self.val_cache);
+        } else {
+            self.docid_cache.clear();
+            self.row_id_cache.clear();
+            accessor.row_ids_for_docs(docs, &mut self.docid_cache, &mut self.row_id_cache);
+            self.val_cache.resize(self.row_id_cache.len(), T::default());
+            accessor
+                .values
+                .get_vals(&self.row_id_cache, &mut self.val_cache);
+        }
+    }
+    #[inline]
+    pub fn fetch_block_with_missing(&mut self, docs: &[u32], accessor: &Column<T>, missing: T) {
+        self.fetch_block(docs, accessor);
+        // no missing values
+        if accessor.index.get_cardinality().is_full() {
+            return;
+        }
+
+        // We can compare docid_cache length with docs to find missing docs
+        // For multi value columns we can't rely on the length and always need to scan
+        if accessor.index.get_cardinality().is_multivalue() || docs.len() != self.docid_cache.len()
+        {
+            self.missing_docids_cache.clear();
+            find_missing_docs(docs, &self.docid_cache, |doc| {
+                self.missing_docids_cache.push(doc);
+                self.val_cache.push(missing);
+            });
+            self.docid_cache
+                .extend_from_slice(&self.missing_docids_cache);
+        }
+    }
+
+    #[inline]
+    pub fn iter_vals(&self) -> impl Iterator<Item = T> + '_ {
+        self.val_cache.iter().cloned()
+    }
+
+    #[inline]
+    /// Returns an iterator over the docids and values
+    /// The passed in `docs` slice needs to be the same slice that was passed to `fetch_block` or
+    /// `fetch_block_with_missing`.
+    ///
+    /// The docs is used if the column is full (each docs has exactly one value), otherwise the
+    /// internal docid vec is used for the iterator, which e.g. may contain duplicate docs.
+    pub fn iter_docid_vals<'a>(
+        &'a self,
+        docs: &'a [u32],
+        accessor: &Column<T>,
+    ) -> impl Iterator<Item = (DocId, T)> + 'a + use<'a, T> {
+        if accessor.index.get_cardinality().is_full() {
+            docs.iter().cloned().zip(self.val_cache.iter().cloned())
+        } else {
+            self.docid_cache
+                .iter()
+                .cloned()
+                .zip(self.val_cache.iter().cloned())
+        }
+    }
+}
+
+/// Given two sorted lists of docids `docs` and `hits`, hits is a subset of `docs`.
+/// Return all docs that are not in `hits`.
+fn find_missing_docs<F>(docs: &[u32], hits: &[u32], mut callback: F)
+where F: FnMut(u32) {
+    let mut docs_iter = docs.iter();
+    let mut hits_iter = hits.iter();
+
+    let mut doc = docs_iter.next();
+    let mut hit = hits_iter.next();
+
+    while let (Some(&current_doc), Some(&current_hit)) = (doc, hit) {
+        match current_doc.cmp(&current_hit) {
+            Ordering::Less => {
+                callback(current_doc);
+                doc = docs_iter.next();
+            }
+            Ordering::Equal => {
+                doc = docs_iter.next();
+                hit = hits_iter.next();
+            }
+            Ordering::Greater => {
+                hit = hits_iter.next();
+            }
+        }
+    }
+
+    while let Some(&current_doc) = doc {
+        callback(current_doc);
+        doc = docs_iter.next();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_find_missing_docs() {
+        let docs: Vec<u32> = vec![1, 2, 3, 4, 5, 6, 7, 8, 9, 10];
+        let hits: Vec<u32> = vec![2, 4, 6, 8, 10];
+
+        let mut missing_docs: Vec<u32> = Vec::new();
+
+        find_missing_docs(&docs, &hits, |missing_doc| {
+            missing_docs.push(missing_doc);
+        });
+
+        assert_eq!(missing_docs, vec![1, 3, 5, 7, 9]);
+    }
+
+    #[test]
+    fn test_find_missing_docs_empty() {
+        let docs: Vec<u32> = Vec::new();
+        let hits: Vec<u32> = vec![2, 4, 6, 8, 10];
+
+        let mut missing_docs: Vec<u32> = Vec::new();
+
+        find_missing_docs(&docs, &hits, |missing_doc| {
+            missing_docs.push(missing_doc);
+        });
+
+        assert_eq!(missing_docs, Vec::<u32>::new());
+    }
+
+    #[test]
+    fn test_find_missing_docs_all_missing() {
+        let docs: Vec<u32> = vec![1, 2, 3, 4, 5];
+        let hits: Vec<u32> = Vec::new();
+
+        let mut missing_docs: Vec<u32> = Vec::new();
+
+        find_missing_docs(&docs, &hits, |missing_doc| {
+            missing_docs.push(missing_doc);
+        });
+
+        assert_eq!(missing_docs, vec![1, 2, 3, 4, 5]);
+    }
+}

columnar/src/column/dictionary_encoded.rs

@@ -0,0 +1,121 @@
+use std::ops::Deref;
+use std::sync::Arc;
+use std::{fmt, io};
+
+use sstable::{Dictionary, VoidSSTable};
+
+use crate::RowId;
+use crate::column::Column;
+
+/// Dictionary encoded column.
+///
+/// The column simply gives access to a regular u64-column that, in
+/// which the values are term-ordinals.
+///
+/// These ordinals are ids uniquely identify the bytes that are stored in
+/// the column. These ordinals are small, and sorted in the same order
+/// as the term_ord_column.
+#[derive(Clone)]
+pub struct BytesColumn {
+    pub(crate) dictionary: Arc<Dictionary<VoidSSTable>>,
+    pub(crate) term_ord_column: Column<u64>,
+}
+
+impl fmt::Debug for BytesColumn {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("BytesColumn")
+            .field("term_ord_column", &self.term_ord_column)
+            .finish()
+    }
+}
+
+impl BytesColumn {
+    pub fn empty(num_docs: u32) -> BytesColumn {
+        BytesColumn {
+            dictionary: Arc::new(Dictionary::empty()),
+            term_ord_column: Column::build_empty_column(num_docs),
+        }
+    }
+
+    /// Fills the given `output` buffer with the term associated to the ordinal `ord`.
+    ///
+    /// Returns `false` if the term does not exist (e.g. `term_ord` is greater or equal to the
+    /// overll number of terms).
+    pub fn ord_to_bytes(&self, ord: u64, output: &mut Vec<u8>) -> io::Result<bool> {
+        self.dictionary.ord_to_term(ord, output)
+    }
+
+    /// Returns the number of rows in the column.
+    pub fn num_rows(&self) -> RowId {
+        self.term_ord_column.num_docs()
+    }
+
+    pub fn term_ords(&self, row_id: RowId) -> impl Iterator<Item = u64> + '_ {
+        self.term_ord_column.values_for_doc(row_id)
+    }
+
+    /// Returns the column of ordinals
+    pub fn ords(&self) -> &Column<u64> {
+        &self.term_ord_column
+    }
+
+    pub fn num_terms(&self) -> usize {
+        self.dictionary.num_terms()
+    }
+
+    pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
+        self.dictionary.as_ref()
+    }
+}
+
+#[derive(Clone)]
+pub struct StrColumn(BytesColumn);
+
+impl fmt::Debug for StrColumn {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "{:?}", self.term_ord_column)
+    }
+}
+
+impl From<StrColumn> for BytesColumn {
+    fn from(str_column: StrColumn) -> BytesColumn {
+        str_column.0
+    }
+}
+
+impl StrColumn {
+    pub fn wrap(bytes_column: BytesColumn) -> StrColumn {
+        StrColumn(bytes_column)
+    }
+
+    pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
+        self.0.dictionary.as_ref()
+    }
+
+    /// Fills the buffer
+    pub fn ord_to_str(&self, term_ord: u64, output: &mut String) -> io::Result<bool> {
+        unsafe {
+            let buf = output.as_mut_vec();
+            if !self.0.dictionary.ord_to_term(term_ord, buf)? {
+                return Ok(false);
+            }
+            // TODO consider remove checks if it hurts performance.
+            if std::str::from_utf8(buf.as_slice()).is_err() {
+                buf.clear();
+                return Err(io::Error::new(
+                    io::ErrorKind::InvalidData,
+                    "Not valid utf-8",
+                ));
+            }
+        }
+        Ok(true)
+    }
+}
+
+impl Deref for StrColumn {
+    type Target = BytesColumn;
+
+    fn deref(&self) -> &Self::Target {
+        &self.0
+    }
+}

columnar/src/column/mod.rs

@@ -0,0 +1,220 @@
+mod dictionary_encoded;
+mod serialize;
+
+use std::fmt::{self, Debug};
+use std::io::Write;
+use std::ops::{Range, RangeInclusive};
+use std::sync::Arc;
+
+use common::BinarySerializable;
+pub use dictionary_encoded::{BytesColumn, StrColumn};
+pub use serialize::{
+    open_column_bytes, open_column_str, open_column_u64, open_column_u128,
+    open_column_u128_as_compact_u64, serialize_column_mappable_to_u64,
+    serialize_column_mappable_to_u128,
+};
+
+use crate::column_index::{ColumnIndex, Set};
+use crate::column_values::monotonic_mapping::StrictlyMonotonicMappingToInternal;
+use crate::column_values::{ColumnValues, monotonic_map_column};
+use crate::{Cardinality, DocId, EmptyColumnValues, MonotonicallyMappableToU64, RowId};
+
+#[derive(Clone)]
+pub struct Column<T = u64> {
+    pub index: ColumnIndex,
+    pub values: Arc<dyn ColumnValues<T>>,
+}
+
+impl<T: Debug + PartialOrd + Send + Sync + Copy + 'static> Debug for Column<T> {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        let num_docs = self.num_docs();
+        let entries = (0..num_docs)
+            .map(|i| (i, self.values_for_doc(i).collect::<Vec<_>>()))
+            .filter(|(_, vals)| !vals.is_empty());
+        f.debug_map().entries(entries).finish()
+    }
+}
+
+impl<T: PartialOrd + Default> Column<T> {
+    pub fn build_empty_column(num_docs: u32) -> Column<T> {
+        Column {
+            index: ColumnIndex::Empty { num_docs },
+            values: Arc::new(EmptyColumnValues),
+        }
+    }
+}
+
+impl<T: MonotonicallyMappableToU64> Column<T> {
+    pub fn to_u64_monotonic(self) -> Column<u64> {
+        let values = Arc::new(monotonic_map_column(
+            self.values,
+            StrictlyMonotonicMappingToInternal::<T>::new(),
+        ));
+        Column {
+            index: self.index,
+            values,
+        }
+    }
+}
+
+impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
+    #[inline]
+    pub fn get_cardinality(&self) -> Cardinality {
+        self.index.get_cardinality()
+    }
+
+    pub fn num_docs(&self) -> RowId {
+        match &self.index {
+            ColumnIndex::Empty { num_docs } => *num_docs,
+            ColumnIndex::Full => self.values.num_vals(),
+            ColumnIndex::Optional(optional_index) => optional_index.num_docs(),
+            ColumnIndex::Multivalued(col_index) => {
+                // The multivalued index contains all value start row_id,
+                // and one extra value at the end with the overall number of rows.
+                col_index.num_docs()
+            }
+        }
+    }
+
+    pub fn min_value(&self) -> T {
+        self.values.min_value()
+    }
+
+    pub fn max_value(&self) -> T {
+        self.values.max_value()
+    }
+
+    #[inline]
+    pub fn first(&self, row_id: RowId) -> Option<T> {
+        self.values_for_doc(row_id).next()
+    }
+
+    /// Load the first value for each docid in the provided slice.
+    #[inline]
+    pub fn first_vals(&self, docids: &[DocId], output: &mut [Option<T>]) {
+        match &self.index {
+            ColumnIndex::Empty { .. } => {}
+            ColumnIndex::Full => self.values.get_vals_opt(docids, output),
+            ColumnIndex::Optional(optional_index) => {
+                for (i, docid) in docids.iter().enumerate() {
+                    output[i] = optional_index
+                        .rank_if_exists(*docid)
+                        .map(|rowid| self.values.get_val(rowid));
+                }
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                for (i, docid) in docids.iter().enumerate() {
+                    let range = multivalued_index.range(*docid);
+                    let is_empty = range.start == range.end;
+                    if !is_empty {
+                        output[i] = Some(self.values.get_val(range.start));
+                    }
+                }
+            }
+        }
+    }
+
+    /// Translates a block of docis to row_ids.
+    ///
+    /// returns the row_ids and the matching docids on the same index
+    /// e.g.
+    /// DocId In:  [0, 5, 6]
+    /// DocId Out: [0, 0, 6, 6]
+    /// RowId Out: [0, 1, 2, 3]
+    #[inline]
+    pub fn row_ids_for_docs(
+        &self,
+        doc_ids: &[DocId],
+        doc_ids_out: &mut Vec<DocId>,
+        row_ids: &mut Vec<RowId>,
+    ) {
+        self.index.docids_to_rowids(doc_ids, doc_ids_out, row_ids)
+    }
+
+    pub fn values_for_doc(&self, doc_id: DocId) -> impl Iterator<Item = T> + '_ {
+        self.index
+            .value_row_ids(doc_id)
+            .map(|value_row_id: RowId| self.values.get_val(value_row_id))
+    }
+
+    /// Get the docids of values which are in the provided value and docid range.
+    #[inline]
+    pub fn get_docids_for_value_range(
+        &self,
+        value_range: RangeInclusive<T>,
+        selected_docid_range: Range<u32>,
+        doc_ids: &mut Vec<u32>,
+    ) {
+        // convert passed docid range to row id range
+        let rowid_range = self
+            .index
+            .docid_range_to_rowids(selected_docid_range.clone());
+
+        // Load rows
+        self.values
+            .get_row_ids_for_value_range(value_range, rowid_range, doc_ids);
+        // Convert rows to docids
+        self.index
+            .select_batch_in_place(selected_docid_range.start, doc_ids);
+    }
+
+    /// Fills the output vector with the (possibly multiple values that are associated_with
+    /// `row_id`.
+    ///
+    /// This method clears the `output` vector.
+    pub fn fill_vals(&self, row_id: RowId, output: &mut Vec<T>) {
+        output.clear();
+        output.extend(self.values_for_doc(row_id));
+    }
+
+    pub fn first_or_default_col(self, default_value: T) -> Arc<dyn ColumnValues<T>> {
+        Arc::new(FirstValueWithDefault {
+            column: self,
+            default_value,
+        })
+    }
+}
+
+impl BinarySerializable for Cardinality {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> std::io::Result<()> {
+        self.to_code().serialize(writer)
+    }
+
+    fn deserialize<R: std::io::Read>(reader: &mut R) -> std::io::Result<Self> {
+        let cardinality_code = u8::deserialize(reader)?;
+        let cardinality = Cardinality::try_from_code(cardinality_code)?;
+        Ok(cardinality)
+    }
+}
+
+// TODO simplify or optimize
+struct FirstValueWithDefault<T: Copy> {
+    column: Column<T>,
+    default_value: T,
+}
+
+impl<T: PartialOrd + Debug + Send + Sync + Copy + 'static> ColumnValues<T>
+    for FirstValueWithDefault<T>
+{
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> T {
+        self.column.first(idx).unwrap_or(self.default_value)
+    }
+
+    fn min_value(&self) -> T {
+        self.column.values.min_value()
+    }
+
+    fn max_value(&self) -> T {
+        self.column.values.max_value()
+    }
+
+    fn num_vals(&self) -> u32 {
+        match &self.column.index {
+            ColumnIndex::Empty { .. } => 0u32,
+            ColumnIndex::Full => self.column.values.num_vals(),
+            ColumnIndex::Optional(optional_idx) => optional_idx.num_docs(),
+            ColumnIndex::Multivalued(multivalue_idx) => multivalue_idx.num_docs(),
+        }
+    }
+}

columnar/src/column/serialize.rs

@@ -0,0 +1,121 @@
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::OwnedBytes;
+use sstable::Dictionary;
+
+use crate::column::{BytesColumn, Column};
+use crate::column_index::{SerializableColumnIndex, serialize_column_index};
+use crate::column_values::{
+    CodecType, MonotonicallyMappableToU64, MonotonicallyMappableToU128,
+    load_u64_based_column_values, serialize_column_values_u128, serialize_u64_based_column_values,
+};
+use crate::iterable::Iterable;
+use crate::{StrColumn, Version};
+
+pub fn serialize_column_mappable_to_u128<T: MonotonicallyMappableToU128>(
+    column_index: SerializableColumnIndex<'_>,
+    iterable: &dyn Iterable<T>,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    let column_index_num_bytes = serialize_column_index(column_index, output)?;
+    serialize_column_values_u128(iterable, output)?;
+    output.write_all(&column_index_num_bytes.to_le_bytes())?;
+    Ok(())
+}
+
+pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64>(
+    column_index: SerializableColumnIndex<'_>,
+    column_values: &impl Iterable<T>,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    let column_index_num_bytes = serialize_column_index(column_index, output)?;
+    serialize_u64_based_column_values(
+        column_values,
+        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
+        output,
+    )?;
+    output.write_all(&column_index_num_bytes.to_le_bytes())?;
+    Ok(())
+}
+
+pub fn open_column_u64<T: MonotonicallyMappableToU64>(
+    bytes: OwnedBytes,
+    format_version: Version,
+) -> io::Result<Column<T>> {
+    let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
+    let column_index_num_bytes = u32::from_le_bytes(
+        column_index_num_bytes_payload
+            .as_slice()
+            .try_into()
+            .unwrap(),
+    );
+    let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
+    let column_index = crate::column_index::open_column_index(column_index_data, format_version)?;
+    let column_values = load_u64_based_column_values(column_values_data)?;
+    Ok(Column {
+        index: column_index,
+        values: column_values,
+    })
+}
+
+pub fn open_column_u128<T: MonotonicallyMappableToU128>(
+    bytes: OwnedBytes,
+    format_version: Version,
+) -> io::Result<Column<T>> {
+    let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
+    let column_index_num_bytes = u32::from_le_bytes(
+        column_index_num_bytes_payload
+            .as_slice()
+            .try_into()
+            .unwrap(),
+    );
+    let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
+    let column_index = crate::column_index::open_column_index(column_index_data, format_version)?;
+    let column_values = crate::column_values::open_u128_mapped(column_values_data)?;
+    Ok(Column {
+        index: column_index,
+        values: column_values,
+    })
+}
+
+/// Open the column as u64.
+///
+/// See [`open_u128_as_compact_u64`] for more details.
+pub fn open_column_u128_as_compact_u64(
+    bytes: OwnedBytes,
+    format_version: Version,
+) -> io::Result<Column<u64>> {
+    let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
+    let column_index_num_bytes = u32::from_le_bytes(
+        column_index_num_bytes_payload
+            .as_slice()
+            .try_into()
+            .unwrap(),
+    );
+    let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
+    let column_index = crate::column_index::open_column_index(column_index_data, format_version)?;
+    let column_values = crate::column_values::open_u128_as_compact_u64(column_values_data)?;
+    Ok(Column {
+        index: column_index,
+        values: column_values,
+    })
+}
+
+pub fn open_column_bytes(data: OwnedBytes, format_version: Version) -> io::Result<BytesColumn> {
+    let (body, dictionary_len_bytes) = data.rsplit(4);
+    let dictionary_len = u32::from_le_bytes(dictionary_len_bytes.as_slice().try_into().unwrap());
+    let (dictionary_bytes, column_bytes) = body.split(dictionary_len as usize);
+    let dictionary = Arc::new(Dictionary::from_bytes(dictionary_bytes)?);
+    let term_ord_column = crate::column::open_column_u64::<u64>(column_bytes, format_version)?;
+    Ok(BytesColumn {
+        dictionary,
+        term_ord_column,
+    })
+}
+
+pub fn open_column_str(data: OwnedBytes, format_version: Version) -> io::Result<StrColumn> {
+    let bytes_column = open_column_bytes(data, format_version)?;
+    Ok(StrColumn::wrap(bytes_column))
+}

columnar/src/column_index/merge/mod.rs

@@ -0,0 +1,223 @@
+mod shuffled;
+mod stacked;
+
+use common::ReadOnlyBitSet;
+use shuffled::merge_column_index_shuffled;
+use stacked::merge_column_index_stacked;
+
+use crate::column_index::SerializableColumnIndex;
+use crate::{Cardinality, ColumnIndex, MergeRowOrder};
+
+fn detect_cardinality_single_column_index(
+    column_index: &ColumnIndex,
+    alive_bitset_opt: &Option<ReadOnlyBitSet>,
+) -> Cardinality {
+    let Some(alive_bitset) = alive_bitset_opt else {
+        return column_index.get_cardinality();
+    };
+    let cardinality_before_deletes = column_index.get_cardinality();
+    if cardinality_before_deletes == Cardinality::Full {
+        // The columnar cardinality can only become more restrictive in the presence of deletes
+        // (where cardinality sorted from the more restrictive to the least restrictive are Full,
+        // Optional, Multivalued)
+        //
+        // If we are already "Full", we are guaranteed to stay "Full" after deletes.
+        return Cardinality::Full;
+    }
+    let mut cardinality_so_far = Cardinality::Full;
+    for doc_id in alive_bitset.iter() {
+        let num_values = column_index.value_row_ids(doc_id).len();
+        let row_cardinality = match num_values {
+            0 => Cardinality::Optional,
+            1 => Cardinality::Full,
+            _ => Cardinality::Multivalued,
+        };
+        cardinality_so_far = cardinality_so_far.max(row_cardinality);
+        if cardinality_so_far >= cardinality_before_deletes {
+            // There won't be any improvement in the cardinality.
+            // We can early exit.
+            return cardinality_before_deletes;
+        }
+    }
+    cardinality_so_far
+}
+
+fn detect_cardinality(
+    column_indexes: &[ColumnIndex],
+    merge_row_order: &MergeRowOrder,
+) -> Cardinality {
+    match merge_row_order {
+        MergeRowOrder::Stack(_) => column_indexes
+            .iter()
+            .map(ColumnIndex::get_cardinality)
+            .max()
+            .unwrap_or(Cardinality::Full),
+        MergeRowOrder::Shuffled(shuffle_merge_order) => {
+            let mut merged_cardinality = Cardinality::Full;
+            for (column_index, alive_bitset_opt) in column_indexes
+                .iter()
+                .zip(shuffle_merge_order.alive_bitsets.iter())
+            {
+                let cardinality: Cardinality =
+                    detect_cardinality_single_column_index(column_index, alive_bitset_opt);
+                if cardinality == Cardinality::Multivalued {
+                    return cardinality;
+                }
+                merged_cardinality = merged_cardinality.max(cardinality);
+            }
+            merged_cardinality
+        }
+    }
+}
+
+pub fn merge_column_index<'a>(
+    columns: &'a [ColumnIndex],
+    merge_row_order: &'a MergeRowOrder,
+) -> SerializableColumnIndex<'a> {
+    // For simplification, we do not try to detect whether the cardinality could be
+    // downgraded thanks to deletes.
+    let cardinality_after_merge = detect_cardinality(columns, merge_row_order);
+    match merge_row_order {
+        MergeRowOrder::Stack(stack_merge_order) => {
+            merge_column_index_stacked(columns, cardinality_after_merge, stack_merge_order)
+        }
+        MergeRowOrder::Shuffled(complex_merge_order) => {
+            merge_column_index_shuffled(columns, cardinality_after_merge, complex_merge_order)
+        }
+    }
+}
+
+// TODO actually, the shuffled code path is a bit too general.
+// In practise, we do not really shuffle everything.
+// The merge order restricted to a specific column keeps the original row order.
+//
+// This may offer some optimization that we have not explored yet.
+
+#[cfg(test)]
+mod tests {
+    use common::OwnedBytes;
+
+    use crate::column_index::merge::detect_cardinality;
+    use crate::column_index::multivalued_index::{
+        MultiValueIndex, open_multivalued_index, serialize_multivalued_index,
+    };
+    use crate::column_index::{OptionalIndex, SerializableColumnIndex, merge_column_index};
+    use crate::{
+        Cardinality, ColumnIndex, MergeRowOrder, RowAddr, RowId, ShuffleMergeOrder, StackMergeOrder,
+    };
+
+    #[test]
+    fn test_detect_cardinality() {
+        assert_eq!(
+            detect_cardinality(&[], &StackMergeOrder::stack_for_test(&[]).into()),
+            Cardinality::Full
+        );
+        let optional_index: ColumnIndex = OptionalIndex::for_test(1, &[]).into();
+        let multivalued_index: ColumnIndex = MultiValueIndex::for_test(&[0, 1]).into();
+        assert_eq!(
+            detect_cardinality(
+                &[optional_index.clone(), ColumnIndex::Empty { num_docs: 0 }],
+                &StackMergeOrder::stack_for_test(&[1, 0]).into()
+            ),
+            Cardinality::Optional
+        );
+        assert_eq!(
+            detect_cardinality(
+                &[optional_index.clone(), ColumnIndex::Full],
+                &StackMergeOrder::stack_for_test(&[1, 1]).into()
+            ),
+            Cardinality::Optional
+        );
+        assert_eq!(
+            detect_cardinality(
+                &[
+                    multivalued_index.clone(),
+                    ColumnIndex::Empty { num_docs: 0 }
+                ],
+                &StackMergeOrder::stack_for_test(&[1, 0]).into()
+            ),
+            Cardinality::Multivalued
+        );
+        assert_eq!(
+            detect_cardinality(
+                &[multivalued_index.clone(), optional_index.clone()],
+                &StackMergeOrder::stack_for_test(&[1, 1]).into()
+            ),
+            Cardinality::Multivalued
+        );
+        assert_eq!(
+            detect_cardinality(
+                &[optional_index, multivalued_index],
+                &StackMergeOrder::stack_for_test(&[1, 1]).into()
+            ),
+            Cardinality::Multivalued
+        );
+    }
+
+    #[test]
+    fn test_merge_index_multivalued_sorted() {
+        let column_indexes: Vec<ColumnIndex> = vec![MultiValueIndex::for_test(&[0, 2, 5]).into()];
+        let merge_row_order: MergeRowOrder = ShuffleMergeOrder::for_test(
+            &[2],
+            vec![
+                RowAddr {
+                    segment_ord: 0u32,
+                    row_id: 1u32,
+                },
+                RowAddr {
+                    segment_ord: 0u32,
+                    row_id: 0u32,
+                },
+            ],
+        )
+        .into();
+        let merged_column_index = merge_column_index(&column_indexes[..], &merge_row_order);
+        let SerializableColumnIndex::Multivalued(start_index_iterable) = merged_column_index else {
+            panic!("Expected a multivalued index")
+        };
+        let mut output = Vec::new();
+        serialize_multivalued_index(&start_index_iterable, &mut output).unwrap();
+        let multivalue =
+            open_multivalued_index(OwnedBytes::new(output), crate::Version::V2).unwrap();
+        let start_indexes: Vec<RowId> = multivalue.get_start_index_column().iter().collect();
+        assert_eq!(&start_indexes, &[0, 3, 5]);
+    }
+
+    #[test]
+    fn test_merge_index_multivalued_sorted_several_segment() {
+        let column_indexes: Vec<ColumnIndex> = vec![
+            MultiValueIndex::for_test(&[0, 2, 5]).into(),
+            ColumnIndex::Empty { num_docs: 0 },
+            MultiValueIndex::for_test(&[0, 1, 4]).into(),
+        ];
+        let merge_row_order: MergeRowOrder = ShuffleMergeOrder::for_test(
+            &[2, 0, 2],
+            vec![
+                RowAddr {
+                    segment_ord: 2u32,
+                    row_id: 1u32,
+                },
+                RowAddr {
+                    segment_ord: 0u32,
+                    row_id: 0u32,
+                },
+                RowAddr {
+                    segment_ord: 2u32,
+                    row_id: 0u32,
+                },
+            ],
+        )
+        .into();
+
+        let merged_column_index = merge_column_index(&column_indexes[..], &merge_row_order);
+        let SerializableColumnIndex::Multivalued(start_index_iterable) = merged_column_index else {
+            panic!("Expected a multivalued index")
+        };
+        let mut output = Vec::new();
+        serialize_multivalued_index(&start_index_iterable, &mut output).unwrap();
+        let multivalue =
+            open_multivalued_index(OwnedBytes::new(output), crate::Version::V2).unwrap();
+        let start_indexes: Vec<RowId> = multivalue.get_start_index_column().iter().collect();
+        assert_eq!(&start_indexes, &[0, 3, 5, 6]);
+    }
+}

columnar/src/column_index/merge/shuffled.rs

@@ -0,0 +1,189 @@
+use std::iter;
+
+use crate::column_index::{
+    SerializableColumnIndex, SerializableMultivalueIndex, SerializableOptionalIndex, Set,
+};
+use crate::iterable::Iterable;
+use crate::{Cardinality, ColumnIndex, RowId, ShuffleMergeOrder};
+
+pub fn merge_column_index_shuffled<'a>(
+    column_indexes: &'a [ColumnIndex],
+    cardinality_after_merge: Cardinality,
+    shuffle_merge_order: &'a ShuffleMergeOrder,
+) -> SerializableColumnIndex<'a> {
+    match cardinality_after_merge {
+        Cardinality::Full => SerializableColumnIndex::Full,
+        Cardinality::Optional => {
+            let non_null_row_ids =
+                merge_column_index_shuffled_optional(column_indexes, shuffle_merge_order);
+            SerializableColumnIndex::Optional(SerializableOptionalIndex {
+                non_null_row_ids,
+                num_rows: shuffle_merge_order.num_rows(),
+            })
+        }
+        Cardinality::Multivalued => {
+            let non_null_row_ids =
+                merge_column_index_shuffled_optional(column_indexes, shuffle_merge_order);
+            SerializableColumnIndex::Multivalued(SerializableMultivalueIndex {
+                doc_ids_with_values: SerializableOptionalIndex {
+                    non_null_row_ids,
+                    num_rows: shuffle_merge_order.num_rows(),
+                },
+                start_offsets: merge_column_index_shuffled_multivalued(
+                    column_indexes,
+                    shuffle_merge_order,
+                ),
+            })
+        }
+    }
+}
+
+/// Merge several column indexes into one, ordering rows according to the merge_order passed as
+/// argument. While it is true that the `merge_order` may imply deletes and hence could in theory a
+/// multivalued index into an optional one, this is not supported today for simplification.
+///
+/// In other words the column_indexes passed as argument may NOT be multivalued.
+fn merge_column_index_shuffled_optional<'a>(
+    column_indexes: &'a [ColumnIndex],
+    merge_order: &'a ShuffleMergeOrder,
+) -> Box<dyn Iterable<RowId> + 'a> {
+    Box::new(ShuffledIndex {
+        column_indexes,
+        merge_order,
+    })
+}
+
+struct ShuffledIndex<'a> {
+    column_indexes: &'a [ColumnIndex],
+    merge_order: &'a ShuffleMergeOrder,
+}
+
+impl Iterable<u32> for ShuffledIndex<'_> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        Box::new(
+            self.merge_order
+                .iter_new_to_old_row_addrs()
+                .enumerate()
+                .filter_map(|(new_row_id, old_row_addr)| {
+                    let column_index = &self.column_indexes[old_row_addr.segment_ord as usize];
+                    let row_id = new_row_id as u32;
+                    if column_index.has_value(old_row_addr.row_id) {
+                        Some(row_id)
+                    } else {
+                        None
+                    }
+                }),
+        )
+    }
+}
+
+fn merge_column_index_shuffled_multivalued<'a>(
+    column_indexes: &'a [ColumnIndex],
+    merge_order: &'a ShuffleMergeOrder,
+) -> Box<dyn Iterable<RowId> + 'a> {
+    Box::new(ShuffledMultivaluedIndex {
+        column_indexes,
+        merge_order,
+    })
+}
+
+struct ShuffledMultivaluedIndex<'a> {
+    column_indexes: &'a [ColumnIndex],
+    merge_order: &'a ShuffleMergeOrder,
+}
+
+fn iter_num_values<'a>(
+    column_indexes: &'a [ColumnIndex],
+    merge_order: &'a ShuffleMergeOrder,
+) -> impl Iterator<Item = u32> + 'a {
+    merge_order.iter_new_to_old_row_addrs().map(|row_addr| {
+        let column_index = &column_indexes[row_addr.segment_ord as usize];
+        match column_index {
+            ColumnIndex::Empty { .. } => 0u32,
+            ColumnIndex::Full => 1,
+            ColumnIndex::Optional(optional_index) => {
+                u32::from(optional_index.contains(row_addr.row_id))
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.range(row_addr.row_id).len() as u32
+            }
+        }
+    })
+}
+
+/// Transforms an iterator containing the number of vals per row (with `num_rows` elements)
+/// into a `start_offset` iterator starting at 0 and (with `num_rows + 1` element)
+///
+/// This will filter values with 0 values as these are covered by the optional index in the
+/// multivalue index.
+fn integrate_num_vals(num_vals: impl Iterator<Item = u32>) -> impl Iterator<Item = RowId> {
+    iter::once(0u32).chain(
+        num_vals
+            .filter(|num_vals| *num_vals != 0)
+            .scan(0, |state, num_vals| {
+                *state += num_vals;
+                Some(*state)
+            }),
+    )
+}
+
+impl Iterable<u32> for ShuffledMultivaluedIndex<'_> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        let num_vals_per_row = iter_num_values(self.column_indexes, self.merge_order);
+        Box::new(integrate_num_vals(num_vals_per_row))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::RowAddr;
+    use crate::column_index::OptionalIndex;
+
+    #[test]
+    fn test_integrate_num_vals_empty() {
+        assert!(integrate_num_vals(iter::empty()).eq(iter::once(0)));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_one_el() {
+        assert!(integrate_num_vals(iter::once(10)).eq([0, 10].into_iter()));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_several() {
+        assert!(integrate_num_vals([3, 0, 10, 20].into_iter()).eq([0, 3, 13, 33].into_iter()));
+    }
+
+    #[test]
+    fn test_merge_column_index_optional_shuffle() {
+        let optional_index: ColumnIndex = OptionalIndex::for_test(2, &[0]).into();
+        let column_indexes = [optional_index, ColumnIndex::Full];
+        let row_addrs = vec![
+            RowAddr {
+                segment_ord: 0u32,
+                row_id: 1u32,
+            },
+            RowAddr {
+                segment_ord: 1u32,
+                row_id: 0u32,
+            },
+        ];
+        let shuffle_merge_order = ShuffleMergeOrder::for_test(&[2, 1], row_addrs);
+        let serializable_index = merge_column_index_shuffled(
+            &column_indexes[..],
+            Cardinality::Optional,
+            &shuffle_merge_order,
+        );
+        let SerializableColumnIndex::Optional(SerializableOptionalIndex {
+            non_null_row_ids,
+            num_rows,
+        }) = serializable_index
+        else {
+            panic!()
+        };
+        assert_eq!(num_rows, 2);
+        let non_null_rows: Vec<RowId> = non_null_row_ids.boxed_iter().collect();
+        assert_eq!(&non_null_rows, &[1]);
+    }
+}

columnar/src/column_index/merge/stacked.rs

@@ -0,0 +1,192 @@
+use std::ops::Range;
+
+use crate::column_index::SerializableColumnIndex;
+use crate::column_index::multivalued_index::{MultiValueIndex, SerializableMultivalueIndex};
+use crate::column_index::serialize::SerializableOptionalIndex;
+use crate::iterable::Iterable;
+use crate::{Cardinality, ColumnIndex, RowId, StackMergeOrder};
+
+/// Simple case:
+/// The new mapping just consists in stacking the different column indexes.
+///
+/// There are no sort nor deletes involved.
+pub fn merge_column_index_stacked<'a>(
+    columns: &'a [ColumnIndex],
+    cardinality_after_merge: Cardinality,
+    stack_merge_order: &'a StackMergeOrder,
+) -> SerializableColumnIndex<'a> {
+    match cardinality_after_merge {
+        Cardinality::Full => SerializableColumnIndex::Full,
+        Cardinality::Optional => SerializableColumnIndex::Optional(SerializableOptionalIndex {
+            non_null_row_ids: Box::new(StackedOptionalIndex {
+                columns,
+                stack_merge_order,
+            }),
+            num_rows: stack_merge_order.num_rows(),
+        }),
+        Cardinality::Multivalued => {
+            let serializable_multivalue_index =
+                make_serializable_multivalued_index(columns, stack_merge_order);
+            SerializableColumnIndex::Multivalued(serializable_multivalue_index)
+        }
+    }
+}
+
+struct StackedDocIdsWithValues<'a> {
+    column_indexes: &'a [ColumnIndex],
+    stack_merge_order: &'a StackMergeOrder,
+}
+
+impl Iterable<u32> for StackedDocIdsWithValues<'_> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        Box::new((0..self.column_indexes.len()).flat_map(|i| {
+            let column_index = &self.column_indexes[i];
+            let doc_range = self.stack_merge_order.columnar_range(i);
+            get_doc_ids_with_values(column_index, doc_range)
+        }))
+    }
+}
+
+fn get_doc_ids_with_values<'a>(
+    column_index: &'a ColumnIndex,
+    doc_range: Range<u32>,
+) -> Box<dyn Iterator<Item = u32> + 'a> {
+    match column_index {
+        ColumnIndex::Empty { .. } => Box::new(0..0),
+        ColumnIndex::Full => Box::new(doc_range),
+        ColumnIndex::Optional(optional_index) => Box::new(
+            optional_index
+                .iter_docs()
+                .map(move |row| row + doc_range.start),
+        ),
+        ColumnIndex::Multivalued(multivalued_index) => match multivalued_index {
+            MultiValueIndex::MultiValueIndexV1(multivalued_index) => {
+                Box::new((0..multivalued_index.num_docs()).filter_map(move |docid| {
+                    let range = multivalued_index.range(docid);
+                    if range.is_empty() {
+                        None
+                    } else {
+                        Some(docid + doc_range.start)
+                    }
+                }))
+            }
+            MultiValueIndex::MultiValueIndexV2(multivalued_index) => Box::new(
+                multivalued_index
+                    .optional_index
+                    .iter_docs()
+                    .map(move |row| row + doc_range.start),
+            ),
+        },
+    }
+}
+
+fn stack_doc_ids_with_values<'a>(
+    column_indexes: &'a [ColumnIndex],
+    stack_merge_order: &'a StackMergeOrder,
+) -> SerializableOptionalIndex<'a> {
+    let num_rows = stack_merge_order.num_rows();
+    SerializableOptionalIndex {
+        non_null_row_ids: Box::new(StackedDocIdsWithValues {
+            column_indexes,
+            stack_merge_order,
+        }),
+        num_rows,
+    }
+}
+
+struct StackedStartOffsets<'a> {
+    column_indexes: &'a [ColumnIndex],
+    stack_merge_order: &'a StackMergeOrder,
+}
+
+fn get_num_values_iterator<'a>(
+    column_index: &'a ColumnIndex,
+    num_docs: u32,
+) -> Box<dyn Iterator<Item = u32> + 'a> {
+    match column_index {
+        ColumnIndex::Empty { .. } => Box::new(std::iter::empty()),
+        ColumnIndex::Full => Box::new(std::iter::repeat(1u32).take(num_docs as usize)),
+        ColumnIndex::Optional(optional_index) => {
+            Box::new(std::iter::repeat(1u32).take(optional_index.num_non_nulls() as usize))
+        }
+        ColumnIndex::Multivalued(multivalued_index) => Box::new(
+            multivalued_index
+                .get_start_index_column()
+                .iter()
+                .scan(0u32, |previous_start_offset, current_start_offset| {
+                    let num_vals = current_start_offset - *previous_start_offset;
+                    *previous_start_offset = current_start_offset;
+                    Some(num_vals)
+                })
+                .skip(1),
+        ),
+    }
+}
+
+impl Iterable<u32> for StackedStartOffsets<'_> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        let num_values_it = (0..self.column_indexes.len()).flat_map(|columnar_id| {
+            let num_docs = self.stack_merge_order.columnar_range(columnar_id).len() as u32;
+            let column_index = &self.column_indexes[columnar_id];
+            get_num_values_iterator(column_index, num_docs)
+        });
+        Box::new(std::iter::once(0u32).chain(num_values_it.into_iter().scan(
+            0u32,
+            |cumulated, el| {
+                *cumulated += el;
+                Some(*cumulated)
+            },
+        )))
+    }
+}
+
+fn stack_start_offsets<'a>(
+    column_indexes: &'a [ColumnIndex],
+    stack_merge_order: &'a StackMergeOrder,
+) -> Box<dyn Iterable<u32> + 'a> {
+    Box::new(StackedStartOffsets {
+        column_indexes,
+        stack_merge_order,
+    })
+}
+
+fn make_serializable_multivalued_index<'a>(
+    columns: &'a [ColumnIndex],
+    stack_merge_order: &'a StackMergeOrder,
+) -> SerializableMultivalueIndex<'a> {
+    SerializableMultivalueIndex {
+        doc_ids_with_values: stack_doc_ids_with_values(columns, stack_merge_order),
+        start_offsets: stack_start_offsets(columns, stack_merge_order),
+    }
+}
+
+struct StackedOptionalIndex<'a> {
+    columns: &'a [ColumnIndex],
+    stack_merge_order: &'a StackMergeOrder,
+}
+
+impl<'a> Iterable<RowId> for StackedOptionalIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(
+            self.columns
+                .iter()
+                .enumerate()
+                .flat_map(|(columnar_id, column_index_opt)| {
+                    let columnar_row_range = self.stack_merge_order.columnar_range(columnar_id);
+                    let rows_it: Box<dyn Iterator<Item = RowId>> = match column_index_opt {
+                        ColumnIndex::Full => Box::new(columnar_row_range),
+                        ColumnIndex::Optional(optional_index) => Box::new(
+                            optional_index
+                                .iter_docs()
+                                .map(move |row_id: RowId| columnar_row_range.start + row_id),
+                        ),
+                        ColumnIndex::Multivalued(_) => {
+                            panic!("No multivalued index is allowed when stacking column index");
+                        }
+                        ColumnIndex::Empty { .. } => Box::new(std::iter::empty()),
+                    };
+                    rows_it
+                }),
+        )
+    }
+}

columnar/src/column_index/mod.rs

@@ -0,0 +1,210 @@
+//! # `column_index`
+//!
+//! `column_index` provides rank and select operations to associate positions when not all
+//! documents have exactly one element.
+
+mod merge;
+mod multivalued_index;
+mod optional_index;
+mod serialize;
+
+use std::ops::Range;
+
+pub use merge::merge_column_index;
+pub(crate) use multivalued_index::SerializableMultivalueIndex;
+pub use optional_index::{OptionalIndex, Set};
+pub use serialize::{
+    SerializableColumnIndex, SerializableOptionalIndex, open_column_index, serialize_column_index,
+};
+
+use crate::column_index::multivalued_index::MultiValueIndex;
+use crate::{Cardinality, DocId, RowId};
+
+#[derive(Clone, Debug)]
+pub enum ColumnIndex {
+    Empty {
+        num_docs: u32,
+    },
+    Full,
+    Optional(OptionalIndex),
+    /// In addition, at index num_rows, an extra value is added
+    /// containing the overall number of values.
+    Multivalued(MultiValueIndex),
+}
+
+impl From<OptionalIndex> for ColumnIndex {
+    fn from(optional_index: OptionalIndex) -> ColumnIndex {
+        ColumnIndex::Optional(optional_index)
+    }
+}
+
+impl From<MultiValueIndex> for ColumnIndex {
+    fn from(multi_value_index: MultiValueIndex) -> ColumnIndex {
+        ColumnIndex::Multivalued(multi_value_index)
+    }
+}
+
+impl ColumnIndex {
+    /// Returns the cardinality of the column index.
+    ///
+    /// By convention, if the column contains no docs, we consider that it is
+    /// full.
+    #[inline]
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            ColumnIndex::Empty { num_docs: 0 } | ColumnIndex::Full => Cardinality::Full,
+            ColumnIndex::Empty { .. } => Cardinality::Optional,
+            ColumnIndex::Optional(_) => Cardinality::Optional,
+            ColumnIndex::Multivalued(_) => Cardinality::Multivalued,
+        }
+    }
+
+    /// Returns true if and only if there are at least one value associated to the row.
+    pub fn has_value(&self, doc_id: DocId) -> bool {
+        match self {
+            ColumnIndex::Empty { .. } => false,
+            ColumnIndex::Full => true,
+            ColumnIndex::Optional(optional_index) => optional_index.contains(doc_id),
+            ColumnIndex::Multivalued(multivalued_index) => {
+                !multivalued_index.range(doc_id).is_empty()
+            }
+        }
+    }
+
+    pub fn value_row_ids(&self, doc_id: DocId) -> Range<RowId> {
+        match self {
+            ColumnIndex::Empty { .. } => 0..0,
+            ColumnIndex::Full => doc_id..doc_id + 1,
+            ColumnIndex::Optional(optional_index) => {
+                if let Some(val) = optional_index.rank_if_exists(doc_id) {
+                    val..val + 1
+                } else {
+                    0..0
+                }
+            }
+            ColumnIndex::Multivalued(multivalued_index) => multivalued_index.range(doc_id),
+        }
+    }
+
+    /// Translates a block of docis to row_ids.
+    ///
+    /// returns the row_ids and the matching docids on the same index
+    /// e.g.
+    /// DocId In:  [0, 5, 6]
+    /// DocId Out: [0, 0, 6, 6]
+    /// RowId Out: [0, 1, 2, 3]
+    #[inline]
+    pub fn docids_to_rowids(
+        &self,
+        doc_ids: &[DocId],
+        doc_ids_out: &mut Vec<DocId>,
+        row_ids: &mut Vec<RowId>,
+    ) {
+        match self {
+            ColumnIndex::Empty { .. } => {}
+            ColumnIndex::Full => {
+                doc_ids_out.extend_from_slice(doc_ids);
+                row_ids.extend_from_slice(doc_ids);
+            }
+            ColumnIndex::Optional(optional_index) => {
+                for doc_id in doc_ids {
+                    if let Some(row_id) = optional_index.rank_if_exists(*doc_id) {
+                        doc_ids_out.push(*doc_id);
+                        row_ids.push(row_id);
+                    }
+                }
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                for doc_id in doc_ids {
+                    for row_id in multivalued_index.range(*doc_id) {
+                        doc_ids_out.push(*doc_id);
+                        row_ids.push(row_id);
+                    }
+                }
+            }
+        }
+    }
+
+    pub fn docid_range_to_rowids(&self, doc_id_range: Range<DocId>) -> Range<RowId> {
+        match self {
+            ColumnIndex::Empty { .. } => 0..0,
+            ColumnIndex::Full => doc_id_range,
+            ColumnIndex::Optional(optional_index) => {
+                let row_start = optional_index.rank(doc_id_range.start);
+                let row_end = optional_index.rank(doc_id_range.end);
+                row_start..row_end
+            }
+            ColumnIndex::Multivalued(multivalued_index) => match multivalued_index {
+                MultiValueIndex::MultiValueIndexV1(index) => {
+                    let row_start = index.start_index_column.get_val(doc_id_range.start);
+                    let row_end = index.start_index_column.get_val(doc_id_range.end);
+                    row_start..row_end
+                }
+                MultiValueIndex::MultiValueIndexV2(index) => {
+                    // In this case we will use the optional_index select the next values
+                    // that are valid. There are different cases to consider:
+                    // Not exists below means does not exist in the optional
+                    // index, because it has no values.
+                    // * doc_id_range may cover a range of docids which are non existent
+                    // => rank
+                    //   will give us the next document outside the range with a value. They both
+                    //   get the same rank and therefore return a zero range
+                    //
+                    // * doc_id_range.start and doc_id_range.end may not exist, but docids in
+                    // between may have values
+                    // => rank will give us the next document outside the range with a value.
+                    //
+                    // * doc_id_range.start may be not existent but doc_id_range.end may exist
+                    // * doc_id_range.start may exist but doc_id_range.end may not exist
+                    // * doc_id_range.start and doc_id_range.end may exist
+                    // => rank on doc_id_range.end will give use the next value, which matches
+                    // how the `start_index_column` works, so we get the value start of the next
+                    // docid which we use to create the exclusive range.
+                    //
+                    let rank_start = index.optional_index.rank(doc_id_range.start);
+                    let row_start = index.start_index_column.get_val(rank_start);
+                    let rank_end = index.optional_index.rank(doc_id_range.end);
+                    let row_end = index.start_index_column.get_val(rank_end);
+
+                    row_start..row_end
+                }
+            },
+        }
+    }
+
+    pub fn select_batch_in_place(&self, doc_id_start: DocId, rank_ids: &mut Vec<RowId>) {
+        match self {
+            ColumnIndex::Empty { .. } => {
+                rank_ids.clear();
+            }
+            ColumnIndex::Full => {
+                // No need to do anything:
+                // value_idx and row_idx are the same.
+            }
+            ColumnIndex::Optional(optional_index) => {
+                optional_index.select_batch(&mut rank_ids[..]);
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.select_batch_in_place(doc_id_start, rank_ids)
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::{Cardinality, ColumnIndex};
+
+    #[test]
+    fn test_column_index_get_cardinality() {
+        assert_eq!(
+            ColumnIndex::Empty { num_docs: 0 }.get_cardinality(),
+            Cardinality::Full
+        );
+        assert_eq!(ColumnIndex::Full.get_cardinality(), Cardinality::Full);
+        assert_eq!(
+            ColumnIndex::Empty { num_docs: 1 }.get_cardinality(),
+            Cardinality::Optional
+        );
+    }
+}

columnar/src/column_index/multivalued_index.rs

@@ -0,0 +1,400 @@
+use std::io;
+use std::io::Write;
+use std::ops::Range;
+use std::sync::Arc;
+
+use common::{CountingWriter, OwnedBytes};
+
+use super::optional_index::{open_optional_index, serialize_optional_index};
+use super::{OptionalIndex, SerializableOptionalIndex, Set};
+use crate::column_values::{
+    CodecType, ColumnValues, load_u64_based_column_values, serialize_u64_based_column_values,
+};
+use crate::iterable::Iterable;
+use crate::{DocId, RowId, Version};
+
+pub struct SerializableMultivalueIndex<'a> {
+    pub doc_ids_with_values: SerializableOptionalIndex<'a>,
+    pub start_offsets: Box<dyn Iterable<u32> + 'a>,
+}
+
+pub fn serialize_multivalued_index(
+    multivalued_index: &SerializableMultivalueIndex,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    let SerializableMultivalueIndex {
+        doc_ids_with_values,
+        start_offsets,
+    } = multivalued_index;
+    let mut count_writer = CountingWriter::wrap(output);
+    let SerializableOptionalIndex {
+        non_null_row_ids,
+        num_rows,
+    } = doc_ids_with_values;
+    serialize_optional_index(&**non_null_row_ids, *num_rows, &mut count_writer)?;
+    let optional_len = count_writer.written_bytes() as u32;
+    let output = count_writer.finish();
+    serialize_u64_based_column_values(
+        &**start_offsets,
+        &[CodecType::Bitpacked, CodecType::Linear],
+        output,
+    )?;
+    output.write_all(&optional_len.to_le_bytes())?;
+    Ok(())
+}
+
+pub fn open_multivalued_index(
+    bytes: OwnedBytes,
+    format_version: Version,
+) -> io::Result<MultiValueIndex> {
+    match format_version {
+        Version::V1 => {
+            let start_index_column: Arc<dyn ColumnValues<RowId>> =
+                load_u64_based_column_values(bytes)?;
+            Ok(MultiValueIndex::MultiValueIndexV1(MultiValueIndexV1 {
+                start_index_column,
+            }))
+        }
+        Version::V2 => {
+            let (body_bytes, optional_index_len) = bytes.rsplit(4);
+            let optional_index_len =
+                u32::from_le_bytes(optional_index_len.as_slice().try_into().unwrap());
+            let (optional_index_bytes, start_index_bytes) =
+                body_bytes.split(optional_index_len as usize);
+            let optional_index = open_optional_index(optional_index_bytes)?;
+            let start_index_column: Arc<dyn ColumnValues<RowId>> =
+                load_u64_based_column_values(start_index_bytes)?;
+            Ok(MultiValueIndex::MultiValueIndexV2(MultiValueIndexV2 {
+                optional_index,
+                start_index_column,
+            }))
+        }
+    }
+}
+
+#[derive(Clone)]
+/// Index to resolve value range for given doc_id.
+/// Starts at 0.
+pub enum MultiValueIndex {
+    MultiValueIndexV1(MultiValueIndexV1),
+    MultiValueIndexV2(MultiValueIndexV2),
+}
+
+#[derive(Clone)]
+/// Index to resolve value range for given doc_id.
+/// Starts at 0.
+pub struct MultiValueIndexV1 {
+    pub start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
+}
+
+impl MultiValueIndexV1 {
+    /// Returns `[start, end)`, such that the values associated with
+    /// the given document are `start..end`.
+    #[inline]
+    pub(crate) fn range(&self, doc_id: DocId) -> Range<RowId> {
+        if doc_id >= self.num_docs() {
+            return 0..0;
+        }
+        let start = self.start_index_column.get_val(doc_id);
+        let end = self.start_index_column.get_val(doc_id + 1);
+        start..end
+    }
+
+    /// Returns the number of documents in the index.
+    #[inline]
+    pub fn num_docs(&self) -> u32 {
+        self.start_index_column.num_vals() - 1
+    }
+
+    /// Converts a list of ranks (row ids of values) in a 1:n index to the corresponding list of
+    /// docids. Positions are converted inplace to docids.
+    ///
+    /// Since there is no index for value pos -> docid, but docid -> value pos range, we scan the
+    /// index.
+    ///
+    /// Correctness: positions needs to be sorted. idx_reader needs to contain monotonically
+    /// increasing positions.
+    ///
+    /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
+    /// match a docid to its value position.
+    pub(crate) fn select_batch_in_place(&self, docid_start: DocId, ranks: &mut Vec<u32>) {
+        if ranks.is_empty() {
+            return;
+        }
+        let mut cur_doc = docid_start;
+        let mut last_doc = None;
+
+        assert!(self.start_index_column.get_val(docid_start) <= ranks[0]);
+
+        let mut write_doc_pos = 0;
+        for i in 0..ranks.len() {
+            let pos = ranks[i];
+            loop {
+                let end = self.start_index_column.get_val(cur_doc + 1);
+                if end > pos {
+                    ranks[write_doc_pos] = cur_doc;
+                    write_doc_pos += if last_doc == Some(cur_doc) { 0 } else { 1 };
+                    last_doc = Some(cur_doc);
+                    break;
+                }
+                cur_doc += 1;
+            }
+        }
+        ranks.truncate(write_doc_pos);
+    }
+}
+
+#[derive(Clone)]
+/// Index to resolve value range for given doc_id.
+/// Starts at 0.
+pub struct MultiValueIndexV2 {
+    pub optional_index: OptionalIndex,
+    pub start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
+}
+
+impl std::fmt::Debug for MultiValueIndex {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        let index = match self {
+            MultiValueIndex::MultiValueIndexV1(idx) => &idx.start_index_column,
+            MultiValueIndex::MultiValueIndexV2(idx) => &idx.start_index_column,
+        };
+        f.debug_struct("MultiValuedIndex")
+            .field("num_rows", &index.num_vals())
+            .finish_non_exhaustive()
+    }
+}
+
+impl MultiValueIndex {
+    pub fn for_test(start_offsets: &[RowId]) -> MultiValueIndex {
+        assert!(!start_offsets.is_empty());
+        assert_eq!(start_offsets[0], 0);
+        let mut doc_with_values = Vec::new();
+        let mut compact_start_offsets: Vec<u32> = vec![0];
+        for doc in 0..start_offsets.len() - 1 {
+            if start_offsets[doc] < start_offsets[doc + 1] {
+                doc_with_values.push(doc as RowId);
+                compact_start_offsets.push(start_offsets[doc + 1]);
+            }
+        }
+        let serializable_multivalued_index = SerializableMultivalueIndex {
+            doc_ids_with_values: SerializableOptionalIndex {
+                non_null_row_ids: Box::new(&doc_with_values[..]),
+                num_rows: start_offsets.len() as u32 - 1,
+            },
+            start_offsets: Box::new(&compact_start_offsets[..]),
+        };
+        let mut buffer = Vec::new();
+        serialize_multivalued_index(&serializable_multivalued_index, &mut buffer).unwrap();
+        let bytes = OwnedBytes::new(buffer);
+        open_multivalued_index(bytes, Version::V2).unwrap()
+    }
+
+    pub fn get_start_index_column(&self) -> &Arc<dyn crate::ColumnValues<RowId>> {
+        match self {
+            MultiValueIndex::MultiValueIndexV1(idx) => &idx.start_index_column,
+            MultiValueIndex::MultiValueIndexV2(idx) => &idx.start_index_column,
+        }
+    }
+
+    /// Returns `[start, end)` values range, such that the values associated with
+    /// the given document are `start..end`.
+    #[inline]
+    pub(crate) fn range(&self, doc_id: DocId) -> Range<RowId> {
+        match self {
+            MultiValueIndex::MultiValueIndexV1(idx) => idx.range(doc_id),
+            MultiValueIndex::MultiValueIndexV2(idx) => idx.range(doc_id),
+        }
+    }
+
+    /// Returns the number of documents in the index.
+    #[inline]
+    pub fn num_docs(&self) -> u32 {
+        match self {
+            MultiValueIndex::MultiValueIndexV1(idx) => idx.start_index_column.num_vals() - 1,
+            MultiValueIndex::MultiValueIndexV2(idx) => idx.optional_index.num_docs(),
+        }
+    }
+
+    /// Converts a list of ranks (row ids of values) in a 1:n index to the corresponding list of
+    /// docids. Positions are converted inplace to docids.
+    ///
+    /// Since there is no index for value pos -> docid, but docid -> value pos range, we scan the
+    /// index.
+    ///
+    /// Correctness: positions needs to be sorted. idx_reader needs to contain monotonically
+    /// increasing positions.
+    ///
+    /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
+    /// match a docid to its value position.
+    pub(crate) fn select_batch_in_place(&self, docid_start: DocId, ranks: &mut Vec<u32>) {
+        match self {
+            MultiValueIndex::MultiValueIndexV1(idx) => {
+                idx.select_batch_in_place(docid_start, ranks)
+            }
+            MultiValueIndex::MultiValueIndexV2(idx) => {
+                idx.select_batch_in_place(docid_start, ranks)
+            }
+        }
+    }
+}
+impl MultiValueIndexV2 {
+    /// Returns `[start, end)`, such that the values associated with
+    /// the given document are `start..end`.
+    #[inline]
+    pub(crate) fn range(&self, doc_id: DocId) -> Range<RowId> {
+        let Some(rank) = self.optional_index.rank_if_exists(doc_id) else {
+            return 0..0;
+        };
+        let start = self.start_index_column.get_val(rank);
+        let end = self.start_index_column.get_val(rank + 1);
+        start..end
+    }
+
+    /// Returns the number of documents in the index.
+    #[inline]
+    pub fn num_docs(&self) -> u32 {
+        self.optional_index.num_docs()
+    }
+
+    /// Converts a list of ranks (row ids of values) in a 1:n index to the corresponding list of
+    /// docids. Positions are converted inplace to docids.
+    ///
+    /// Since there is no index for value pos -> docid, but docid -> value pos range, we scan the
+    /// index.
+    ///
+    /// Correctness: positions needs to be sorted. idx_reader needs to contain monotonically
+    /// increasing positions.
+    ///
+    /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
+    /// match a docid to its value position.
+    pub(crate) fn select_batch_in_place(&self, docid_start: DocId, ranks: &mut Vec<u32>) {
+        if ranks.is_empty() {
+            return;
+        }
+        let mut cur_pos_in_idx = self.optional_index.rank(docid_start);
+        let mut last_doc = None;
+
+        assert!(cur_pos_in_idx <= ranks[0]);
+
+        let mut write_doc_pos = 0;
+        for i in 0..ranks.len() {
+            let pos = ranks[i];
+            loop {
+                let end = self.start_index_column.get_val(cur_pos_in_idx + 1);
+                if end > pos {
+                    ranks[write_doc_pos] = cur_pos_in_idx;
+                    write_doc_pos += if last_doc == Some(cur_pos_in_idx) {
+                        0
+                    } else {
+                        1
+                    };
+                    last_doc = Some(cur_pos_in_idx);
+                    break;
+                }
+                cur_pos_in_idx += 1;
+            }
+        }
+        ranks.truncate(write_doc_pos);
+
+        for rank in ranks.iter_mut() {
+            *rank = self.optional_index.select(*rank);
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::ops::Range;
+
+    use super::MultiValueIndex;
+    use crate::{ColumnarReader, DynamicColumn};
+
+    fn index_to_pos_helper(
+        index: &MultiValueIndex,
+        doc_id_range: Range<u32>,
+        positions: &[u32],
+    ) -> Vec<u32> {
+        let mut positions = positions.to_vec();
+        index.select_batch_in_place(doc_id_range.start, &mut positions);
+        positions
+    }
+
+    #[test]
+    fn test_positions_to_docid() {
+        let index = MultiValueIndex::for_test(&[0, 10, 12, 15, 22, 23]);
+        assert_eq!(index.num_docs(), 5);
+        let positions = &[10u32, 11, 15, 20, 21, 22];
+        assert_eq!(index_to_pos_helper(&index, 0..5, positions), vec![1, 3, 4]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, positions), vec![1, 3, 4]);
+
+        assert_eq!(index_to_pos_helper(&index, 0..5, &[9]), vec![0]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, &[10]), vec![1]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, &[11]), vec![1]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12]), vec![2]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14]), vec![2]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14, 15]), vec![2, 3]);
+    }
+
+    #[test]
+    fn test_range_to_rowids() {
+        use crate::ColumnarWriter;
+
+        let mut columnar_writer = ColumnarWriter::default();
+
+        // This column gets coerced to u64
+        columnar_writer.record_numerical(1, "full", u64::MAX);
+        columnar_writer.record_numerical(1, "full", u64::MAX);
+
+        columnar_writer.record_numerical(5, "full", u64::MAX);
+        columnar_writer.record_numerical(5, "full", u64::MAX);
+
+        let mut wrt: Vec<u8> = Vec::new();
+        columnar_writer.serialize(7, &mut wrt).unwrap();
+
+        let reader = ColumnarReader::open(wrt).unwrap();
+        // Open the column as u64
+        let column = reader.read_columns("full").unwrap()[0]
+            .open()
+            .unwrap()
+            .coerce_numerical(crate::NumericalType::U64)
+            .unwrap();
+        let DynamicColumn::U64(column) = column else {
+            panic!();
+        };
+
+        let row_id_range = column.index.docid_range_to_rowids(1..2);
+        assert_eq!(row_id_range, 0..2);
+
+        let row_id_range = column.index.docid_range_to_rowids(0..2);
+        assert_eq!(row_id_range, 0..2);
+
+        let row_id_range = column.index.docid_range_to_rowids(0..4);
+        assert_eq!(row_id_range, 0..2);
+
+        let row_id_range = column.index.docid_range_to_rowids(3..4);
+        assert_eq!(row_id_range, 2..2);
+
+        let row_id_range = column.index.docid_range_to_rowids(1..6);
+        assert_eq!(row_id_range, 0..4);
+
+        let row_id_range = column.index.docid_range_to_rowids(3..6);
+        assert_eq!(row_id_range, 2..4);
+
+        let row_id_range = column.index.docid_range_to_rowids(0..6);
+        assert_eq!(row_id_range, 0..4);
+
+        let row_id_range = column.index.docid_range_to_rowids(0..6);
+        assert_eq!(row_id_range, 0..4);
+
+        let check = |range, expected| {
+            let full_range = 0..=u64::MAX;
+            let mut docids = Vec::new();
+            column.get_docids_for_value_range(full_range, range, &mut docids);
+            assert_eq!(docids, expected);
+        };
+
+        // check(0..1, vec![]);
+        // check(0..2, vec![1]);
+        check(1..2, vec![1]);
+    }
+}

columnar/src/column_index/optional_index/mod.rs

@@ -0,0 +1,540 @@
+use std::io::{self, Write};
+use std::sync::Arc;
+
+mod set;
+mod set_block;
+
+use common::{BinarySerializable, OwnedBytes, VInt};
+pub use set::{SelectCursor, Set, SetCodec};
+use set_block::{
+    DENSE_BLOCK_NUM_BYTES, DenseBlock, DenseBlockCodec, SparseBlock, SparseBlockCodec,
+};
+
+use crate::iterable::Iterable;
+use crate::{DocId, InvalidData, RowId};
+
+/// The threshold for for number of elements after which we switch to dense block encoding.
+///
+/// We simply pick the value that minimize the size of the blocks.
+const DENSE_BLOCK_THRESHOLD: u32 =
+    set_block::DENSE_BLOCK_NUM_BYTES / std::mem::size_of::<u16>() as u32; //< 5_120
+
+const ELEMENTS_PER_BLOCK: u32 = u16::MAX as u32 + 1;
+
+#[derive(Copy, Clone, Debug)]
+struct BlockMeta {
+    non_null_rows_before_block: u32,
+    start_byte_offset: u32,
+    block_variant: BlockVariant,
+}
+
+#[derive(Clone, Copy, Debug)]
+enum BlockVariant {
+    Dense,
+    Sparse { num_vals: u16 },
+}
+
+impl BlockVariant {
+    pub fn empty() -> Self {
+        Self::Sparse { num_vals: 0 }
+    }
+    pub fn num_bytes_in_block(&self) -> u32 {
+        match *self {
+            BlockVariant::Dense => set_block::DENSE_BLOCK_NUM_BYTES,
+            BlockVariant::Sparse { num_vals } => num_vals as u32 * 2,
+        }
+    }
+}
+
+/// This codec is inspired by roaring bitmaps.
+/// In the dense blocks, however, in order to accelerate `select`
+/// we interleave an offset over two bytes. (more on this lower)
+///
+/// The lower 16 bits of doc ids are stored as u16 while the upper 16 bits are given by the block
+/// id. Each block contains 1<<16 docids.
+///
+/// # Serialized Data Layout
+/// The data starts with the block data. Each block is either dense or sparse encoded, depending on
+/// the number of values in the block. A block is sparse when it contains less than
+/// DENSE_BLOCK_THRESHOLD (6144) values.
+/// [Sparse data block | dense data block, .. #repeat*; Desc: Either a sparse or dense encoded
+/// block]
+/// ### Sparse block data
+/// [u16 LE, .. #repeat*; Desc: Positions with values in a block]
+/// ### Dense block data
+/// [Dense codec for the whole block; Desc: Similar to a bitvec(0..ELEMENTS_PER_BLOCK) + Metadata
+/// for faster lookups. See dense.rs]
+///
+/// The data is followed by block metadata, to know which area of the raw block data belongs to
+/// which block. Only metadata for blocks with elements is recorded to
+/// keep the overhead low for scenarios with many very sparse columns. The block metadata consists
+/// of the block index and the number of values in the block. Since we don't store empty blocks
+/// num_vals is incremented by 1, e.g. 0 means 1 value.
+///
+/// The last u16 is storing the number of metadata blocks.
+/// [u16 LE, .. #repeat*; Desc: Positions with values in a block][(u16 LE, u16 LE), .. #repeat*;
+/// Desc: (Block Id u16, Num Elements u16)][u16 LE; Desc: num blocks with values u16]
+///
+/// # Opening
+/// When opening the data layout, the data is expanded to `Vec<SparseCodecBlockVariant>`, where the
+/// index is the block index. For each block `byte_start` and `offset` is computed.
+#[derive(Clone)]
+pub struct OptionalIndex {
+    num_docs: RowId,
+    num_non_null_docs: RowId,
+    block_data: OwnedBytes,
+    block_metas: Arc<[BlockMeta]>,
+}
+
+impl Iterable<u32> for &OptionalIndex {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        Box::new(self.iter_docs())
+    }
+}
+
+impl std::fmt::Debug for OptionalIndex {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        f.debug_struct("OptionalIndex")
+            .field("num_docs", &self.num_docs)
+            .field("num_non_null_docs", &self.num_non_null_docs)
+            .finish_non_exhaustive()
+    }
+}
+
+/// Splits a value address into lower and upper 16bits.
+/// The lower 16 bits are the value in the block
+/// The upper 16 bits are the block index
+#[derive(Copy, Debug, Clone)]
+struct RowAddr {
+    block_id: u16,
+    in_block_row_id: u16,
+}
+
+#[inline(always)]
+fn row_addr_from_row_id(row_id: RowId) -> RowAddr {
+    RowAddr {
+        block_id: (row_id / ELEMENTS_PER_BLOCK) as u16,
+        in_block_row_id: (row_id % ELEMENTS_PER_BLOCK) as u16,
+    }
+}
+
+enum BlockSelectCursor<'a> {
+    Dense(<DenseBlock<'a> as Set<u16>>::SelectCursor<'a>),
+    Sparse(<SparseBlock<'a> as Set<u16>>::SelectCursor<'a>),
+}
+
+impl BlockSelectCursor<'_> {
+    fn select(&mut self, rank: u16) -> u16 {
+        match self {
+            BlockSelectCursor::Dense(dense_select_cursor) => dense_select_cursor.select(rank),
+            BlockSelectCursor::Sparse(sparse_select_cursor) => sparse_select_cursor.select(rank),
+        }
+    }
+}
+pub struct OptionalIndexSelectCursor<'a> {
+    current_block_cursor: BlockSelectCursor<'a>,
+    current_block_id: u16,
+    // The current block is guaranteed to contain ranks < end_rank.
+    current_block_end_rank: RowId,
+    optional_index: &'a OptionalIndex,
+    block_doc_idx_start: RowId,
+    num_null_rows_before_block: RowId,
+}
+
+impl OptionalIndexSelectCursor<'_> {
+    fn search_and_load_block(&mut self, rank: RowId) {
+        if rank < self.current_block_end_rank {
+            // we are already in the right block
+            return;
+        }
+        self.current_block_id = self.optional_index.find_block(rank, self.current_block_id);
+        self.current_block_end_rank = self
+            .optional_index
+            .block_metas
+            .get(self.current_block_id as usize + 1)
+            .map(|block_meta| block_meta.non_null_rows_before_block)
+            .unwrap_or(u32::MAX);
+        self.block_doc_idx_start = (self.current_block_id as u32) * ELEMENTS_PER_BLOCK;
+        let block_meta = self.optional_index.block_metas[self.current_block_id as usize];
+        self.num_null_rows_before_block = block_meta.non_null_rows_before_block;
+        let block: Block<'_> = self.optional_index.block(block_meta);
+        self.current_block_cursor = match block {
+            Block::Dense(dense_block) => BlockSelectCursor::Dense(dense_block.select_cursor()),
+            Block::Sparse(sparse_block) => BlockSelectCursor::Sparse(sparse_block.select_cursor()),
+        };
+    }
+}
+
+impl SelectCursor<RowId> for OptionalIndexSelectCursor<'_> {
+    fn select(&mut self, rank: RowId) -> RowId {
+        self.search_and_load_block(rank);
+        let index_in_block = (rank - self.num_null_rows_before_block) as u16;
+        self.current_block_cursor.select(index_in_block) as RowId + self.block_doc_idx_start
+    }
+}
+
+impl Set<RowId> for OptionalIndex {
+    type SelectCursor<'b>
+        = OptionalIndexSelectCursor<'b>
+    where Self: 'b;
+    // Check if value at position is not null.
+    #[inline]
+    fn contains(&self, row_id: RowId) -> bool {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(row_id);
+        let block_meta = self.block_metas[block_id as usize];
+        match self.block(block_meta) {
+            Block::Dense(dense_block) => dense_block.contains(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.contains(in_block_row_id),
+        }
+    }
+
+    /// Any value doc_id is allowed.
+    /// In particular, doc_id = num_rows.
+    #[inline]
+    fn rank(&self, doc_id: DocId) -> RowId {
+        if doc_id >= self.num_docs() {
+            return self.num_non_nulls();
+        }
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(doc_id);
+        let block_meta = self.block_metas[block_id as usize];
+        let block = self.block(block_meta);
+
+        let block_offset_row_id = match block {
+            Block::Dense(dense_block) => dense_block.rank(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.rank(in_block_row_id),
+        } as u32;
+        block_meta.non_null_rows_before_block + block_offset_row_id
+    }
+
+    /// Any value doc_id is allowed.
+    /// In particular, doc_id = num_rows.
+    #[inline]
+    fn rank_if_exists(&self, doc_id: DocId) -> Option<RowId> {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(doc_id);
+        let block_meta = *self.block_metas.get(block_id as usize)?;
+        let block = self.block(block_meta);
+        let block_offset_row_id = match block {
+            Block::Dense(dense_block) => dense_block.rank_if_exists(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.rank_if_exists(in_block_row_id),
+        }? as u32;
+        Some(block_meta.non_null_rows_before_block + block_offset_row_id)
+    }
+
+    #[inline]
+    fn select(&self, rank: RowId) -> RowId {
+        let block_pos = self.find_block(rank, 0);
+        let block_doc_idx_start = (block_pos as u32) * ELEMENTS_PER_BLOCK;
+        let block_meta = self.block_metas[block_pos as usize];
+        let block: Block<'_> = self.block(block_meta);
+        let index_in_block = (rank - block_meta.non_null_rows_before_block) as u16;
+        let in_block_rank = match block {
+            Block::Dense(dense_block) => dense_block.select(index_in_block),
+            Block::Sparse(sparse_block) => sparse_block.select(index_in_block),
+        };
+        block_doc_idx_start + in_block_rank as u32
+    }
+
+    fn select_cursor(&self) -> OptionalIndexSelectCursor<'_> {
+        OptionalIndexSelectCursor {
+            current_block_cursor: BlockSelectCursor::Sparse(
+                SparseBlockCodec::open(b"").select_cursor(),
+            ),
+            current_block_id: 0u16,
+            current_block_end_rank: 0u32, //< this is sufficient to force the first load
+            optional_index: self,
+            block_doc_idx_start: 0u32,
+            num_null_rows_before_block: 0u32,
+        }
+    }
+}
+
+impl OptionalIndex {
+    pub fn for_test(num_rows: RowId, row_ids: &[RowId]) -> OptionalIndex {
+        assert!(
+            row_ids
+                .last()
+                .copied()
+                .map(|last_row_id| last_row_id < num_rows)
+                .unwrap_or(true)
+        );
+        let mut buffer = Vec::new();
+        serialize_optional_index(&row_ids, num_rows, &mut buffer).unwrap();
+        let bytes = OwnedBytes::new(buffer);
+        open_optional_index(bytes).unwrap()
+    }
+
+    pub fn num_docs(&self) -> RowId {
+        self.num_docs
+    }
+
+    pub fn num_non_nulls(&self) -> RowId {
+        self.num_non_null_docs
+    }
+
+    pub fn iter_docs(&self) -> impl Iterator<Item = RowId> + '_ {
+        // TODO optimize
+        let mut select_batch = self.select_cursor();
+        (0..self.num_non_null_docs).map(move |rank| select_batch.select(rank))
+    }
+    pub fn select_batch(&self, ranks: &mut [RowId]) {
+        let mut select_cursor = self.select_cursor();
+        for rank in ranks.iter_mut() {
+            *rank = select_cursor.select(*rank);
+        }
+    }
+
+    #[inline]
+    fn block(&self, block_meta: BlockMeta) -> Block<'_> {
+        let BlockMeta {
+            start_byte_offset,
+            block_variant,
+            ..
+        } = block_meta;
+        let start_byte_offset = start_byte_offset as usize;
+        let bytes = self.block_data.as_slice();
+        match block_variant {
+            BlockVariant::Dense => Block::Dense(DenseBlockCodec::open(
+                &bytes[start_byte_offset..start_byte_offset + DENSE_BLOCK_NUM_BYTES as usize],
+            )),
+            BlockVariant::Sparse { num_vals } => {
+                let end_byte_offset = start_byte_offset + num_vals as usize * 2;
+                let sparse_bytes = &bytes[start_byte_offset..end_byte_offset];
+                Block::Sparse(SparseBlockCodec::open(sparse_bytes))
+            }
+        }
+    }
+
+    #[inline]
+    fn find_block(&self, dense_idx: u32, start_block_pos: u16) -> u16 {
+        for block_pos in start_block_pos..self.block_metas.len() as u16 {
+            let offset = self.block_metas[block_pos as usize].non_null_rows_before_block;
+            if offset > dense_idx {
+                return block_pos - 1u16;
+            }
+        }
+        self.block_metas.len() as u16 - 1u16
+    }
+
+    // TODO Add a good API for the codec_idx to original_idx translation.
+    // The Iterator API is a probably a bad idea
+}
+
+#[derive(Copy, Clone)]
+enum Block<'a> {
+    Dense(DenseBlock<'a>),
+    Sparse(SparseBlock<'a>),
+}
+
+#[derive(Debug, Copy, Clone)]
+enum OptionalIndexCodec {
+    Dense = 0,
+    Sparse = 1,
+}
+
+impl OptionalIndexCodec {
+    fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        match code {
+            0 => Ok(Self::Dense),
+            1 => Ok(Self::Sparse),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+impl BinarySerializable for OptionalIndexCodec {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&[self.to_code()])
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let optional_codec_code = u8::deserialize(reader)?;
+        let optional_codec = Self::try_from_code(optional_codec_code)?;
+        Ok(optional_codec)
+    }
+}
+
+fn serialize_optional_index_block(block_els: &[u16], out: &mut impl io::Write) -> io::Result<()> {
+    let is_sparse = is_sparse(block_els.len() as u32);
+    if is_sparse {
+        SparseBlockCodec::serialize(block_els.iter().copied(), out)?;
+    } else {
+        DenseBlockCodec::serialize(block_els.iter().copied(), out)?;
+    }
+    Ok(())
+}
+
+pub fn serialize_optional_index<W: io::Write>(
+    non_null_rows: &dyn Iterable<RowId>,
+    num_rows: RowId,
+    output: &mut W,
+) -> io::Result<()> {
+    VInt(num_rows as u64).serialize(output)?;
+
+    let mut rows_it = non_null_rows.boxed_iter();
+    let mut block_metadata: Vec<SerializedBlockMeta> = Vec::new();
+    let mut current_block = Vec::new();
+
+    // This if-statement for the first element ensures that
+    // `block_metadata` is not empty in the loop below.
+    let Some(idx) = rows_it.next() else {
+        output.write_all(&0u16.to_le_bytes())?;
+        return Ok(());
+    };
+
+    let row_addr = row_addr_from_row_id(idx);
+
+    let mut current_block_id = row_addr.block_id;
+    current_block.push(row_addr.in_block_row_id);
+
+    for idx in rows_it {
+        let value_addr = row_addr_from_row_id(idx);
+        if current_block_id != value_addr.block_id {
+            serialize_optional_index_block(&current_block[..], output)?;
+            block_metadata.push(SerializedBlockMeta {
+                block_id: current_block_id,
+                num_non_null_rows: current_block.len() as u32,
+            });
+            current_block.clear();
+            current_block_id = value_addr.block_id;
+        }
+        current_block.push(value_addr.in_block_row_id);
+    }
+
+    // handle last block
+    serialize_optional_index_block(&current_block[..], output)?;
+
+    block_metadata.push(SerializedBlockMeta {
+        block_id: current_block_id,
+        num_non_null_rows: current_block.len() as u32,
+    });
+
+    for block in &block_metadata {
+        output.write_all(&block.to_bytes())?;
+    }
+
+    output.write_all((block_metadata.len() as u16).to_le_bytes().as_ref())?;
+
+    Ok(())
+}
+
+const SERIALIZED_BLOCK_META_NUM_BYTES: usize = 4;
+
+#[derive(Clone, Copy, Debug)]
+struct SerializedBlockMeta {
+    block_id: u16,
+    num_non_null_rows: u32, //< takes values in 1..=u16::MAX
+}
+
+// TODO unit tests
+impl SerializedBlockMeta {
+    #[inline]
+    fn from_bytes(bytes: [u8; SERIALIZED_BLOCK_META_NUM_BYTES]) -> SerializedBlockMeta {
+        let block_id = u16::from_le_bytes(bytes[0..2].try_into().unwrap());
+        let num_non_null_rows: u32 =
+            u16::from_le_bytes(bytes[2..4].try_into().unwrap()) as u32 + 1u32;
+        SerializedBlockMeta {
+            block_id,
+            num_non_null_rows,
+        }
+    }
+
+    #[inline]
+    fn to_bytes(self) -> [u8; SERIALIZED_BLOCK_META_NUM_BYTES] {
+        assert!(self.num_non_null_rows > 0);
+        let mut bytes = [0u8; SERIALIZED_BLOCK_META_NUM_BYTES];
+        bytes[0..2].copy_from_slice(&self.block_id.to_le_bytes());
+        // We don't store empty blocks, therefore we can subtract 1.
+        // This way we will be able to use u16 when the number of elements is 1 << 16 or u16::MAX+1
+        bytes[2..4].copy_from_slice(&((self.num_non_null_rows - 1u32) as u16).to_le_bytes());
+        bytes
+    }
+}
+
+#[inline]
+fn is_sparse(num_rows_in_block: u32) -> bool {
+    num_rows_in_block < DENSE_BLOCK_THRESHOLD
+}
+
+fn deserialize_optional_index_block_metadatas(
+    data: &[u8],
+    num_rows: u32,
+) -> (Box<[BlockMeta]>, u32) {
+    let num_blocks = data.len() / SERIALIZED_BLOCK_META_NUM_BYTES;
+    let mut block_metas = Vec::with_capacity(num_blocks + 1);
+    let mut start_byte_offset = 0;
+    let mut non_null_rows_before_block = 0;
+    for block_meta_bytes in data.chunks_exact(SERIALIZED_BLOCK_META_NUM_BYTES) {
+        let block_meta_bytes: [u8; SERIALIZED_BLOCK_META_NUM_BYTES] =
+            block_meta_bytes.try_into().unwrap();
+        let SerializedBlockMeta {
+            block_id,
+            num_non_null_rows,
+        } = SerializedBlockMeta::from_bytes(block_meta_bytes);
+        block_metas.resize(
+            block_id as usize,
+            BlockMeta {
+                non_null_rows_before_block,
+                start_byte_offset,
+                block_variant: BlockVariant::empty(),
+            },
+        );
+        let block_variant = if is_sparse(num_non_null_rows) {
+            BlockVariant::Sparse {
+                num_vals: num_non_null_rows as u16,
+            }
+        } else {
+            BlockVariant::Dense
+        };
+        block_metas.push(BlockMeta {
+            non_null_rows_before_block,
+            start_byte_offset,
+            block_variant,
+        });
+        start_byte_offset += block_variant.num_bytes_in_block();
+        non_null_rows_before_block += num_non_null_rows;
+    }
+    block_metas.resize(
+        num_rows.div_ceil(ELEMENTS_PER_BLOCK) as usize,
+        BlockMeta {
+            non_null_rows_before_block,
+            start_byte_offset,
+            block_variant: BlockVariant::empty(),
+        },
+    );
+    (block_metas.into_boxed_slice(), non_null_rows_before_block)
+}
+
+pub fn open_optional_index(bytes: OwnedBytes) -> io::Result<OptionalIndex> {
+    let (mut bytes, num_non_empty_blocks_bytes) = bytes.rsplit(2);
+    let num_non_empty_block_bytes =
+        u16::from_le_bytes(num_non_empty_blocks_bytes.as_slice().try_into().unwrap());
+    let num_docs = VInt::deserialize_u64(&mut bytes)? as u32;
+    let block_metas_num_bytes =
+        num_non_empty_block_bytes as usize * SERIALIZED_BLOCK_META_NUM_BYTES;
+    let (block_data, block_metas) = bytes.rsplit(block_metas_num_bytes);
+    let (block_metas, num_non_null_docs) =
+        deserialize_optional_index_block_metadatas(block_metas.as_slice(), num_docs);
+    let optional_index = OptionalIndex {
+        num_docs,
+        num_non_null_docs,
+        block_data,
+        block_metas: block_metas.into(),
+    };
+    Ok(optional_index)
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/column_index/optional_index/set.rs

@@ -0,0 +1,49 @@
+use std::io;
+
+/// A codec makes it possible to serialize a set of
+/// elements, and open the resulting Set representation.
+pub trait SetCodec {
+    type Item: Copy + TryFrom<usize> + Eq + std::hash::Hash + std::fmt::Debug;
+    type Reader<'a>: Set<Self::Item>;
+
+    /// Serializes a set of unique sorted u16 elements.
+    ///
+    /// May panic if the elements are not sorted.
+    fn serialize(els: impl Iterator<Item = Self::Item>, wrt: impl io::Write) -> io::Result<()>;
+    fn open(data: &[u8]) -> Self::Reader<'_>;
+}
+
+/// Stateful object that makes it possible to compute several select in a row,
+/// provided the rank passed as argument are increasing.
+pub trait SelectCursor<T> {
+    // May panic if rank is greater than the number of elements in the Set,
+    // or if rank is < than value provided in the previous call.
+    fn select(&mut self, rank: T) -> T;
+}
+
+pub trait Set<T> {
+    type SelectCursor<'b>: SelectCursor<T>
+    where Self: 'b;
+
+    /// Returns true if the elements is contained in the Set
+    fn contains(&self, el: T) -> bool;
+
+    /// Returns the element's rank (its position in the set).
+    /// If the set does not contain the element, it will return the next existing elements rank.
+    fn rank(&self, el: T) -> T;
+
+    /// If the set contains `el`, returns the element's rank (its position in the set).
+    /// If the set does not contain the element, it returns `None`.
+    fn rank_if_exists(&self, el: T) -> Option<T>;
+
+    /// Return the rank-th value stored in this bitmap.
+    ///
+    /// # Panics
+    ///
+    /// May panic if rank is greater or equal to the number of
+    /// elements in the Set.
+    fn select(&self, rank: T) -> T;
+
+    /// Creates a brand new select cursor.
+    fn select_cursor(&self) -> Self::SelectCursor<'_>;
+}

columnar/src/column_index/optional_index/set_block/dense.rs

@@ -0,0 +1,278 @@
+use std::io::{self, Write};
+
+use common::BinarySerializable;
+
+use crate::column_index::optional_index::{ELEMENTS_PER_BLOCK, SelectCursor, Set, SetCodec};
+
+#[inline(always)]
+fn get_bit_at(input: u64, n: u16) -> bool {
+    input & (1 << n) != 0
+}
+
+#[inline]
+fn set_bit_at(input: &mut u64, n: u16) {
+    *input |= 1 << n;
+}
+
+/// For the `DenseCodec`, `data` which contains the encoded blocks.
+/// Each block consists of [u8; 12]. The first 8 bytes is a bitvec for 64 elements.
+/// The last 4 bytes are the offset, the number of set bits so far.
+///
+/// When translating the original index to a dense index, the correct block can be computed
+/// directly `orig_idx/64`. Inside the block the position is `orig_idx%64`.
+///
+/// When translating a dense index to the original index, we can use the offset to find the correct
+/// block. Direct computation is not possible, but we can employ a linear or binary search.
+const ELEMENTS_PER_MINI_BLOCK: u16 = 64;
+const MINI_BLOCK_BITVEC_NUM_BYTES: usize = 8;
+const MINI_BLOCK_OFFSET_NUM_BYTES: usize = 2;
+pub const MINI_BLOCK_NUM_BYTES: usize = MINI_BLOCK_BITVEC_NUM_BYTES + MINI_BLOCK_OFFSET_NUM_BYTES;
+
+/// Number of bytes in a dense block.
+pub const DENSE_BLOCK_NUM_BYTES: u32 =
+    (ELEMENTS_PER_BLOCK / ELEMENTS_PER_MINI_BLOCK as u32) * MINI_BLOCK_NUM_BYTES as u32;
+
+pub struct DenseBlockCodec;
+
+impl SetCodec for DenseBlockCodec {
+    type Item = u16;
+    type Reader<'a> = DenseBlock<'a>;
+
+    fn serialize(els: impl Iterator<Item = u16>, wrt: impl io::Write) -> io::Result<()> {
+        serialize_dense_codec(els, wrt)
+    }
+
+    #[inline]
+    fn open(data: &[u8]) -> Self::Reader<'_> {
+        assert_eq!(data.len(), DENSE_BLOCK_NUM_BYTES as usize);
+        DenseBlock(data)
+    }
+}
+
+/// Interpreting the bitvec as a set of integer within 0..=63
+/// and given an element, returns the number of elements in the
+/// set lesser than the element.
+///
+/// # Panics
+///
+/// May panic or return a wrong result if el <= 64.
+#[inline(always)]
+fn rank_u64(bitvec: u64, el: u16) -> u16 {
+    debug_assert!(el < 64);
+    let mask = (1u64 << el) - 1;
+    let masked_bitvec = bitvec & mask;
+    masked_bitvec.count_ones() as u16
+}
+
+#[inline(always)]
+fn select_u64(mut bitvec: u64, rank: u16) -> u16 {
+    for _ in 0..rank {
+        bitvec &= bitvec - 1;
+    }
+    bitvec.trailing_zeros() as u16
+}
+
+// TODO test the following solution on Intel... on Ryzen Zen <3 it is a catastrophy.
+// #[target_feature(enable = "bmi2")]
+// unsafe fn select_bitvec_unsafe(bitvec: u64, rank: u16) -> u16 {
+//     let pdep = _pdep_u64(1u64 << rank, bitvec);
+//     pdep.trailing_zeros() as u16
+// }
+
+#[derive(Clone, Copy, Debug)]
+struct DenseMiniBlock {
+    bitvec: u64,
+    rank: u16,
+}
+
+impl DenseMiniBlock {
+    fn from_bytes(data: [u8; MINI_BLOCK_NUM_BYTES]) -> Self {
+        let bitvec = u64::from_le_bytes(data[..MINI_BLOCK_BITVEC_NUM_BYTES].try_into().unwrap());
+        let rank = u16::from_le_bytes(data[MINI_BLOCK_BITVEC_NUM_BYTES..].try_into().unwrap());
+        Self { bitvec, rank }
+    }
+
+    fn to_bytes(self) -> [u8; MINI_BLOCK_NUM_BYTES] {
+        let mut bytes = [0u8; MINI_BLOCK_NUM_BYTES];
+        bytes[..MINI_BLOCK_BITVEC_NUM_BYTES].copy_from_slice(&self.bitvec.to_le_bytes());
+        bytes[MINI_BLOCK_BITVEC_NUM_BYTES..].copy_from_slice(&self.rank.to_le_bytes());
+        bytes
+    }
+}
+
+#[derive(Copy, Clone)]
+pub struct DenseBlock<'a>(&'a [u8]);
+
+pub struct DenseBlockSelectCursor<'a> {
+    block_id: u16,
+    dense_block: DenseBlock<'a>,
+}
+
+impl SelectCursor<u16> for DenseBlockSelectCursor<'_> {
+    #[inline]
+    fn select(&mut self, rank: u16) -> u16 {
+        self.block_id = self
+            .dense_block
+            .find_miniblock_containing_rank(rank, self.block_id)
+            .unwrap();
+        let index_block = self.dense_block.mini_block(self.block_id);
+        let in_block_rank = rank - index_block.rank;
+        self.block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
+    }
+}
+
+impl<'a> Set<u16> for DenseBlock<'a> {
+    type SelectCursor<'b>
+        = DenseBlockSelectCursor<'a>
+    where Self: 'b;
+
+    #[inline(always)]
+    fn contains(&self, el: u16) -> bool {
+        let mini_block_id = el / ELEMENTS_PER_MINI_BLOCK;
+        let bitvec = self.mini_block(mini_block_id).bitvec;
+        let pos_in_bitvec = el % ELEMENTS_PER_MINI_BLOCK;
+        get_bit_at(bitvec, pos_in_bitvec)
+    }
+
+    #[inline(always)]
+    fn rank_if_exists(&self, el: u16) -> Option<u16> {
+        let block_pos = el / ELEMENTS_PER_MINI_BLOCK;
+        let index_block = self.mini_block(block_pos);
+        let pos_in_block_bit_vec = el % ELEMENTS_PER_MINI_BLOCK;
+        let ones_in_block = rank_u64(index_block.bitvec, pos_in_block_bit_vec);
+        let rank = index_block.rank + ones_in_block;
+        if get_bit_at(index_block.bitvec, pos_in_block_bit_vec) {
+            Some(rank)
+        } else {
+            None
+        }
+    }
+
+    #[inline(always)]
+    fn rank(&self, el: u16) -> u16 {
+        let block_pos = el / ELEMENTS_PER_MINI_BLOCK;
+        let index_block = self.mini_block(block_pos);
+        let pos_in_block_bit_vec = el % ELEMENTS_PER_MINI_BLOCK;
+        let ones_in_block = rank_u64(index_block.bitvec, pos_in_block_bit_vec);
+        index_block.rank + ones_in_block
+    }
+
+    #[inline(always)]
+    fn select(&self, rank: u16) -> u16 {
+        let block_id = self.find_miniblock_containing_rank(rank, 0).unwrap();
+        let index_block = self.mini_block(block_id);
+        let in_block_rank = rank - index_block.rank;
+        block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
+    }
+
+    #[inline(always)]
+    fn select_cursor(&self) -> Self::SelectCursor<'_> {
+        DenseBlockSelectCursor {
+            block_id: 0,
+            dense_block: *self,
+        }
+    }
+}
+
+impl DenseBlock<'_> {
+    #[inline]
+    fn mini_block(&self, mini_block_id: u16) -> DenseMiniBlock {
+        let data_start_pos = mini_block_id as usize * MINI_BLOCK_NUM_BYTES;
+        DenseMiniBlock::from_bytes(
+            self.0[data_start_pos..data_start_pos + MINI_BLOCK_NUM_BYTES]
+                .try_into()
+                .unwrap(),
+        )
+    }
+
+    #[inline]
+    fn iter_miniblocks(
+        &self,
+        from_block_id: u16,
+    ) -> impl Iterator<Item = (u16, DenseMiniBlock)> + '_ {
+        self.0
+            .chunks_exact(MINI_BLOCK_NUM_BYTES)
+            .enumerate()
+            .skip(from_block_id as usize)
+            .map(|(block_id, bytes)| {
+                let mini_block = DenseMiniBlock::from_bytes(bytes.try_into().unwrap());
+                (block_id as u16, mini_block)
+            })
+    }
+
+    /// Finds the block position containing the dense_idx.
+    ///
+    /// # Correctness
+    /// dense_idx needs to be smaller than the number of values in the index
+    ///
+    /// The last offset number is equal to the number of values in the index.
+    #[inline]
+    fn find_miniblock_containing_rank(&self, rank: u16, from_block_id: u16) -> Option<u16> {
+        self.iter_miniblocks(from_block_id)
+            .take_while(|(_, block)| block.rank <= rank)
+            .map(|(block_id, _)| block_id)
+            .last()
+    }
+}
+
+/// Iterator over all values, true if set, otherwise false
+pub fn serialize_dense_codec(
+    els: impl Iterator<Item = u16>,
+    mut output: impl Write,
+) -> io::Result<()> {
+    let mut non_null_rows_before: u16 = 0u16;
+    let mut block = 0u64;
+    let mut current_block_id = 0u16;
+    for el in els {
+        let block_id = el / ELEMENTS_PER_MINI_BLOCK;
+        let in_offset = el % ELEMENTS_PER_MINI_BLOCK;
+        while block_id > current_block_id {
+            let dense_mini_block = DenseMiniBlock {
+                bitvec: block,
+                rank: non_null_rows_before,
+            };
+            output.write_all(&dense_mini_block.to_bytes())?;
+            non_null_rows_before += block.count_ones() as u16;
+            block = 0u64;
+            current_block_id += 1u16;
+        }
+        set_bit_at(&mut block, in_offset);
+    }
+    while current_block_id <= u16::MAX / ELEMENTS_PER_MINI_BLOCK {
+        block.serialize(&mut output)?;
+        non_null_rows_before.serialize(&mut output)?;
+        // This will overflow to 0 exactly if all bits are set.
+        // This is however not problem as we won't use this last value.
+        non_null_rows_before = non_null_rows_before.wrapping_add(block.count_ones() as u16);
+        block = 0u64;
+        current_block_id += 1u16;
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_select_bitvec() {
+        assert_eq!(select_u64(1u64, 0), 0);
+        assert_eq!(select_u64(2u64, 0), 1);
+        assert_eq!(select_u64(4u64, 0), 2);
+        assert_eq!(select_u64(8u64, 0), 3);
+        assert_eq!(select_u64(1 | 8u64, 0), 0);
+        assert_eq!(select_u64(1 | 8u64, 1), 3);
+    }
+
+    #[test]
+    fn test_count_ones() {
+        for i in 0..=63 {
+            assert_eq!(rank_u64(u64::MAX, i), i);
+        }
+    }
+
+    #[test]
+    fn test_dense() {
+        assert_eq!(DENSE_BLOCK_NUM_BYTES, 10_240);
+    }
+}

columnar/src/column_index/optional_index/set_block/mod.rs

@@ -0,0 +1,8 @@
+mod dense;
+mod sparse;
+
+pub use dense::{DENSE_BLOCK_NUM_BYTES, DenseBlock, DenseBlockCodec};
+pub use sparse::{SparseBlock, SparseBlockCodec};
+
+#[cfg(test)]
+mod tests;

columnar/src/column_index/optional_index/set_block/sparse.rs

@@ -0,0 +1,113 @@
+use crate::column_index::optional_index::{SelectCursor, Set, SetCodec};
+
+pub struct SparseBlockCodec;
+
+impl SetCodec for SparseBlockCodec {
+    type Item = u16;
+    type Reader<'a> = SparseBlock<'a>;
+
+    fn serialize(
+        els: impl Iterator<Item = u16>,
+        mut wrt: impl std::io::Write,
+    ) -> std::io::Result<()> {
+        for el in els {
+            wrt.write_all(&el.to_le_bytes())?;
+        }
+        Ok(())
+    }
+
+    fn open(data: &[u8]) -> Self::Reader<'_> {
+        SparseBlock(data)
+    }
+}
+
+#[derive(Copy, Clone)]
+pub struct SparseBlock<'a>(&'a [u8]);
+
+impl<'a> SelectCursor<u16> for SparseBlock<'a> {
+    #[inline]
+    fn select(&mut self, rank: u16) -> u16 {
+        <SparseBlock<'a> as Set<u16>>::select(self, rank)
+    }
+}
+
+impl Set<u16> for SparseBlock<'_> {
+    type SelectCursor<'b>
+        = Self
+    where Self: 'b;
+
+    #[inline(always)]
+    fn contains(&self, el: u16) -> bool {
+        self.binary_search(el).is_ok()
+    }
+
+    #[inline(always)]
+    fn rank_if_exists(&self, el: u16) -> Option<u16> {
+        self.binary_search(el).ok()
+    }
+
+    #[inline(always)]
+    fn rank(&self, el: u16) -> u16 {
+        self.binary_search(el).unwrap_or_else(|el| el)
+    }
+
+    #[inline(always)]
+    fn select(&self, rank: u16) -> u16 {
+        let offset = rank as usize * 2;
+        u16::from_le_bytes(self.0[offset..offset + 2].try_into().unwrap())
+    }
+
+    #[inline(always)]
+    fn select_cursor(&self) -> Self::SelectCursor<'_> {
+        *self
+    }
+}
+
+#[inline(always)]
+fn get_u16(data: &[u8], byte_position: usize) -> u16 {
+    let bytes: [u8; 2] = data[byte_position..byte_position + 2].try_into().unwrap();
+    u16::from_le_bytes(bytes)
+}
+
+impl SparseBlock<'_> {
+    #[inline(always)]
+    fn value_at_idx(&self, data: &[u8], idx: u16) -> u16 {
+        let start_offset: usize = idx as usize * 2;
+        get_u16(data, start_offset)
+    }
+
+    #[inline]
+    fn num_vals(&self) -> u16 {
+        (self.0.len() / 2) as u16
+    }
+
+    #[inline]
+    #[expect(clippy::comparison_chain)]
+    // Looks for the element in the block. Returns the positions if found.
+    fn binary_search(&self, target: u16) -> Result<u16, u16> {
+        let data = &self.0;
+        let mut size = self.num_vals();
+        let mut left = 0;
+        let mut right = size;
+        // TODO try different implem.
+        //  e.g. exponential search into binary search
+        while left < right {
+            let mid = left + size / 2;
+
+            // TODO do boundary check only once, and then use an
+            // unsafe `value_at_idx`
+            let mid_val = self.value_at_idx(data, mid);
+
+            if target > mid_val {
+                left = mid + 1;
+            } else if target < mid_val {
+                right = mid;
+            } else {
+                return Ok(mid);
+            }
+
+            size = right - left;
+        }
+        Err(left)
+    }
+}

columnar/src/column_index/optional_index/set_block/tests.rs

@@ -0,0 +1,147 @@
+use std::collections::HashMap;
+
+use crate::column_index::optional_index::set_block::dense::DENSE_BLOCK_NUM_BYTES;
+use crate::column_index::optional_index::set_block::{DenseBlockCodec, SparseBlockCodec};
+use crate::column_index::optional_index::{SelectCursor, Set, SetCodec};
+
+fn test_set_helper<C: SetCodec<Item = u16>>(vals: &[u16]) -> usize {
+    let mut buffer = Vec::new();
+    C::serialize(vals.iter().copied(), &mut buffer).unwrap();
+    let tested_set = C::open(buffer.as_slice());
+    let hash_set: HashMap<C::Item, C::Item> = vals
+        .iter()
+        .copied()
+        .enumerate()
+        .map(|(ord, val)| (val, C::Item::try_from(ord).ok().unwrap()))
+        .collect();
+    for val in 0u16..=u16::MAX {
+        assert_eq!(tested_set.contains(val), hash_set.contains_key(&val));
+        assert_eq!(tested_set.rank_if_exists(val), hash_set.get(&val).copied());
+        assert_eq!(
+            tested_set.rank(val),
+            vals.iter().cloned().take_while(|v| *v < val).count() as u16
+        );
+    }
+    for (rank, val) in vals.iter().enumerate() {
+        assert_eq!(tested_set.select(rank as u16), *val);
+    }
+    buffer.len()
+}
+
+#[test]
+fn test_dense_block_set_u16_empty() {
+    let buffer_len = test_set_helper::<DenseBlockCodec>(&[]);
+    assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
+}
+
+#[test]
+fn test_dense_block_set_u16_max() {
+    let buffer_len = test_set_helper::<DenseBlockCodec>(&[u16::MAX]);
+    assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
+}
+
+#[test]
+fn test_sparse_block_set_u16_empty() {
+    let buffer_len = test_set_helper::<SparseBlockCodec>(&[]);
+    assert_eq!(buffer_len, 0);
+}
+
+#[test]
+fn test_sparse_block_set_u16_max() {
+    let buffer_len = test_set_helper::<SparseBlockCodec>(&[u16::MAX]);
+    assert_eq!(buffer_len, 2);
+}
+
+use proptest::prelude::*;
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(1))]
+    #[test]
+    fn test_prop_test_dense(els in proptest::collection::btree_set(0..=u16::MAX, 0..=u16::MAX as usize)) {
+        let vals: Vec<u16> = els.into_iter().collect();
+        let buffer_len = test_set_helper::<DenseBlockCodec>(&vals);
+        assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
+    }
+
+    #[test]
+    fn test_prop_test_sparse(els in proptest::collection::btree_set(0..=u16::MAX, 0..=u16::MAX as usize)) {
+        let vals: Vec<u16> = els.into_iter().collect();
+        let buffer_len = test_set_helper::<SparseBlockCodec>(&vals);
+        assert_eq!(buffer_len, vals.len() * 2);
+    }
+}
+
+#[test]
+fn test_simple_translate_codec_codec_idx_to_original_idx_dense() {
+    let mut buffer = Vec::new();
+    DenseBlockCodec::serialize([1, 3, 17, 32, 30_000, 30_001].iter().copied(), &mut buffer)
+        .unwrap();
+    let tested_set = DenseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    let mut select_cursor = tested_set.select_cursor();
+    assert_eq!(select_cursor.select(0), 1);
+    assert_eq!(select_cursor.select(1), 3);
+    assert_eq!(select_cursor.select(2), 17);
+}
+
+#[test]
+fn test_simple_translate_codec_idx_to_original_idx_sparse() {
+    let mut buffer = Vec::new();
+    SparseBlockCodec::serialize([1, 3, 17].iter().copied(), &mut buffer).unwrap();
+    let tested_set = SparseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    let mut select_cursor = tested_set.select_cursor();
+    assert_eq!(SelectCursor::select(&mut select_cursor, 0), 1);
+    assert_eq!(SelectCursor::select(&mut select_cursor, 1), 3);
+    assert_eq!(SelectCursor::select(&mut select_cursor, 2), 17);
+}
+
+#[test]
+fn test_simple_translate_codec_idx_to_original_idx_dense() {
+    let mut buffer = Vec::new();
+    DenseBlockCodec::serialize(0u16..150u16, &mut buffer).unwrap();
+    let tested_set = DenseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    let mut select_cursor = tested_set.select_cursor();
+    for i in 0..150 {
+        assert_eq!(i, select_cursor.select(i));
+    }
+}
+
+#[test]
+fn test_simple_translate_idx_to_value_idx_dense() {
+    let mut buffer = Vec::new();
+    DenseBlockCodec::serialize([1, 10].iter().copied(), &mut buffer).unwrap();
+    let tested_set = DenseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    assert!(!tested_set.contains(2));
+    assert_eq!(tested_set.rank(0), 0);
+    assert_eq!(tested_set.rank(1), 0);
+    for rank in 2..10 {
+        // ranks that don't exist select the next highest one
+        assert_eq!(tested_set.rank_if_exists(rank), None);
+        assert_eq!(tested_set.rank(rank), 1);
+    }
+    assert_eq!(tested_set.rank(10), 1);
+}
+
+#[test]
+fn test_simple_translate_idx_to_value_idx_sparse() {
+    let mut buffer = Vec::new();
+    SparseBlockCodec::serialize([1, 10].iter().copied(), &mut buffer).unwrap();
+    let tested_set = SparseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    assert!(!tested_set.contains(2));
+    assert_eq!(tested_set.rank(0), 0);
+    assert_eq!(tested_set.select(tested_set.rank(0)), 1);
+    assert_eq!(tested_set.rank(1), 0);
+    assert_eq!(tested_set.select(tested_set.rank(1)), 1);
+    for rank in 2..10 {
+        // ranks that don't exist select the next highest one
+        assert_eq!(tested_set.rank_if_exists(rank), None);
+        assert_eq!(tested_set.rank(rank), 1);
+        assert_eq!(tested_set.select(tested_set.rank(rank)), 10);
+    }
+    assert_eq!(tested_set.rank(10), 1);
+    assert_eq!(tested_set.select(tested_set.rank(10)), 10);
+}

columnar/src/column_index/optional_index/tests.rs

@@ -0,0 +1,388 @@
+use proptest::prelude::*;
+use proptest::{prop_oneof, proptest};
+
+use super::*;
+use crate::{ColumnarReader, ColumnarWriter, DynamicColumnHandle};
+
+#[test]
+fn test_optional_index_bug_2293() {
+    // tests for panic in docid_range_to_rowids for docid == num_docs
+    test_optional_index_with_num_docs(ELEMENTS_PER_BLOCK - 1);
+    test_optional_index_with_num_docs(ELEMENTS_PER_BLOCK);
+    test_optional_index_with_num_docs(ELEMENTS_PER_BLOCK + 1);
+}
+fn test_optional_index_with_num_docs(num_docs: u32) {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_numerical(100, "score", 80i64);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(num_docs, &mut buffer).unwrap();
+    let columnar = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar.num_columns(), 1);
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("score").unwrap();
+    assert_eq!(cols.len(), 1);
+
+    let col = cols[0].open().unwrap();
+    col.column_index().docid_range_to_rowids(0..num_docs);
+}
+
+#[test]
+fn test_dense_block_threshold() {
+    assert_eq!(super::DENSE_BLOCK_THRESHOLD, 5_120);
+}
+
+fn random_bitvec() -> BoxedStrategy<Vec<bool>> {
+    prop_oneof![
+        1 => prop::collection::vec(proptest::bool::weighted(1.0), 0..100),
+        1 => prop::collection::vec(proptest::bool::weighted(0.00), 0..(ELEMENTS_PER_BLOCK as usize * 3)), // empty blocks
+        1 => prop::collection::vec(proptest::bool::weighted(1.00), 0..(ELEMENTS_PER_BLOCK as usize + 10)), // full block
+        1 => prop::collection::vec(proptest::bool::weighted(0.01), 0..100),
+        1 => prop::collection::vec(proptest::bool::weighted(0.01), 0..u16::MAX as usize),
+        8 => vec![any::<bool>()],
+    ]
+    .boxed()
+}
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(50))]
+    #[test]
+    fn test_with_random_bitvecs(bitvec1 in random_bitvec(), bitvec2 in random_bitvec(), bitvec3 in random_bitvec()) {
+        let mut bitvec = Vec::new();
+        bitvec.extend_from_slice(&bitvec1);
+        bitvec.extend_from_slice(&bitvec2);
+        bitvec.extend_from_slice(&bitvec3);
+        test_null_index(&bitvec[..]);
+    }
+}
+
+#[test]
+fn test_with_random_sets_simple() {
+    let vals = 10..ELEMENTS_PER_BLOCK * 2;
+    let mut out: Vec<u8> = Vec::new();
+    serialize_optional_index(&vals, 100, &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    let ranks: Vec<u32> = (65_472u32..65_473u32).collect();
+    let els: Vec<u32> = ranks.iter().copied().map(|rank| rank + 10).collect();
+    let mut select_cursor = null_index.select_cursor();
+    for (rank, el) in ranks.iter().copied().zip(els.iter().copied()) {
+        assert_eq!(select_cursor.select(rank), el);
+    }
+}
+
+#[test]
+fn test_optional_index_trailing_empty_blocks() {
+    test_null_index(&[false]);
+}
+
+#[test]
+fn test_optional_index_one_block_false() {
+    let mut iter = vec![false; ELEMENTS_PER_BLOCK as usize];
+    iter.push(true);
+    test_null_index(&iter[..]);
+}
+
+#[test]
+fn test_optional_index_one_block_true() {
+    let mut iter = vec![true; ELEMENTS_PER_BLOCK as usize];
+    iter.push(true);
+    test_null_index(&iter[..]);
+}
+
+impl<'a> Iterable<RowId> for &'a [bool] {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(
+            self.iter()
+                .cloned()
+                .enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+        )
+    }
+}
+
+fn test_null_index(data: &[bool]) {
+    let mut out: Vec<u8> = Vec::new();
+    serialize_optional_index(&data, data.len() as RowId, &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    let orig_idx_with_value: Vec<u32> = data
+        .iter()
+        .enumerate()
+        .filter(|(_pos, val)| **val)
+        .map(|(pos, _val)| pos as u32)
+        .collect();
+    let mut select_iter = null_index.select_cursor();
+    for (i, expected) in orig_idx_with_value.iter().enumerate() {
+        assert_eq!(select_iter.select(i as u32), *expected);
+    }
+
+    let step_size = (orig_idx_with_value.len() / 100).max(1);
+    for (dense_idx, orig_idx) in orig_idx_with_value.iter().enumerate().step_by(step_size) {
+        assert_eq!(null_index.rank_if_exists(*orig_idx), Some(dense_idx as u32));
+    }
+
+    // 100 samples
+    let step_size = (data.len() / 100).max(1);
+    for (pos, value) in data.iter().enumerate().step_by(step_size) {
+        assert_eq!(null_index.contains(pos as u32), *value);
+    }
+}
+
+#[test]
+fn test_optional_index_test_translation() {
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    let mut select_cursor = optional_index.select_cursor();
+    assert_eq!(select_cursor.select(0), 0);
+    assert_eq!(select_cursor.select(1), 2);
+}
+
+#[test]
+fn test_optional_index_translate() {
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    assert_eq!(optional_index.rank_if_exists(0), Some(0));
+    assert_eq!(optional_index.rank_if_exists(2), Some(1));
+}
+
+#[test]
+fn test_optional_index_small() {
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    assert!(optional_index.contains(0));
+    assert!(!optional_index.contains(1));
+    assert!(optional_index.contains(2));
+    assert!(!optional_index.contains(3));
+}
+
+#[test]
+fn test_optional_index_large() {
+    let row_ids = &[ELEMENTS_PER_BLOCK, ELEMENTS_PER_BLOCK + 1];
+    let optional_index = OptionalIndex::for_test(ELEMENTS_PER_BLOCK + 2, row_ids);
+    assert!(!optional_index.contains(0));
+    assert!(!optional_index.contains(100));
+    assert!(!optional_index.contains(ELEMENTS_PER_BLOCK - 1));
+    assert!(optional_index.contains(ELEMENTS_PER_BLOCK));
+    assert!(optional_index.contains(ELEMENTS_PER_BLOCK + 1));
+}
+
+fn test_optional_index_iter_aux(row_ids: &[RowId], num_rows: RowId) {
+    let optional_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(optional_index.num_docs(), num_rows);
+    assert!(optional_index.iter_docs().eq(row_ids.iter().copied()));
+}
+
+#[test]
+fn test_optional_index_iter_empty() {
+    test_optional_index_iter_aux(&[], 0u32);
+}
+
+fn test_optional_index_rank_aux(row_ids: &[RowId]) {
+    let num_rows = row_ids.last().copied().unwrap_or(0u32) + 1;
+    let null_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(null_index.num_docs(), num_rows);
+    for (row_id, row_val) in row_ids.iter().copied().enumerate() {
+        assert_eq!(null_index.rank(row_val), row_id as u32);
+        assert_eq!(null_index.rank_if_exists(row_val), Some(row_id as u32));
+        if row_val > 0 && !null_index.contains(&row_val - 1) {
+            assert_eq!(null_index.rank(row_val - 1), row_id as u32);
+        }
+        assert_eq!(null_index.rank(row_val + 1), row_id as u32 + 1);
+    }
+}
+
+#[test]
+fn test_optional_index_rank() {
+    test_optional_index_rank_aux(&[1u32]);
+    test_optional_index_rank_aux(&[0u32, 1u32]);
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..ELEMENTS_PER_BLOCK).map(|i| i + ELEMENTS_PER_BLOCK + 1));
+    test_optional_index_rank_aux(&block);
+}
+
+#[test]
+fn test_optional_index_iter_empty_one() {
+    test_optional_index_iter_aux(&[1], 2u32);
+    test_optional_index_iter_aux(&[100_000], 200_000u32);
+}
+
+#[test]
+fn test_optional_index_iter_dense_block() {
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..ELEMENTS_PER_BLOCK).map(|i| i + ELEMENTS_PER_BLOCK + 1));
+    test_optional_index_iter_aux(&block, 3 * ELEMENTS_PER_BLOCK);
+}
+
+#[test]
+fn test_optional_index_for_tests() {
+    let optional_index = OptionalIndex::for_test(4, &[1, 2]);
+    assert!(!optional_index.contains(0));
+    assert!(optional_index.contains(1));
+    assert!(optional_index.contains(2));
+    assert!(!optional_index.contains(3));
+    assert_eq!(optional_index.num_docs(), 4);
+}
+
+#[cfg(all(test, feature = "unstable"))]
+mod bench {
+
+    use rand::rngs::StdRng;
+    use rand::{Rng, SeedableRng};
+    use test::Bencher;
+
+    use super::*;
+
+    const TOTAL_NUM_VALUES: u32 = 1_000_000;
+    fn gen_bools(fill_ratio: f64) -> OptionalIndex {
+        let mut out = Vec::new();
+        let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
+        let vals: Vec<RowId> = (0..TOTAL_NUM_VALUES)
+            .map(|_| rng.gen_bool(fill_ratio))
+            .enumerate()
+            .filter(|(_pos, val)| *val)
+            .map(|(pos, _)| pos as RowId)
+            .collect();
+        serialize_optional_index(&&vals[..], TOTAL_NUM_VALUES, &mut out).unwrap();
+
+        open_optional_index(OwnedBytes::new(out)).unwrap()
+    }
+
+    fn random_range_iterator(
+        start: u32,
+        end: u32,
+        avg_step_size: u32,
+        avg_deviation: u32,
+    ) -> impl Iterator<Item = u32> {
+        let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
+        let mut current = start;
+        std::iter::from_fn(move || {
+            current += rng.gen_range(avg_step_size - avg_deviation..=avg_step_size + avg_deviation);
+            if current >= end { None } else { Some(current) }
+        })
+    }
+
+    fn n_percent_step_iterator(percent: f32, num_values: u32) -> impl Iterator<Item = u32> {
+        let ratio = percent / 100.0;
+        let step_size = (1f32 / ratio) as u32;
+        let deviation = step_size - 1;
+        random_range_iterator(0, num_values, step_size, deviation)
+    }
+
+    fn walk_over_data(codec: &OptionalIndex, avg_step_size: u32) -> Option<u32> {
+        walk_over_data_from_positions(
+            codec,
+            random_range_iterator(0, TOTAL_NUM_VALUES, avg_step_size, 0),
+        )
+    }
+
+    fn walk_over_data_from_positions(
+        codec: &OptionalIndex,
+        positions: impl Iterator<Item = u32>,
+    ) -> Option<u32> {
+        let mut dense_idx: Option<u32> = None;
+        for idx in positions {
+            dense_idx = dense_idx.or(codec.rank_if_exists(idx));
+        }
+        dense_idx
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_1percent_filled_10percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_5percent_filled_10percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_5percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 1000));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_1percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_10percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_90percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_10percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_50percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.5f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_90percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_1percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.01f64, 0.005f32, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_10percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.1f64, 0.005f32, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_1percent_filled_10percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.01f64, 10f32, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_1percent_filled_full_scan(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.01f64, 100f32, bench);
+    }
+
+    fn bench_translate_codec_to_orig_util(
+        percent_filled: f64,
+        percent_hit: f32,
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(percent_filled);
+        let num_non_nulls = codec.num_non_nulls();
+        let idxs: Vec<u32> = if percent_hit == 100.0f32 {
+            (0..num_non_nulls).collect()
+        } else {
+            n_percent_step_iterator(percent_hit, num_non_nulls).collect()
+        };
+        let mut output = vec![0u32; idxs.len()];
+        bench.iter(|| {
+            output.copy_from_slice(&idxs[..]);
+            codec.select_batch(&mut output);
+        });
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_90percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.9f64, 0.005, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.9f64, 100.0f32, bench);
+    }
+}

columnar/src/column_index/serialize.rs

@@ -0,0 +1,94 @@
+use std::io;
+use std::io::Write;
+
+use common::{CountingWriter, OwnedBytes};
+
+use super::OptionalIndex;
+use super::multivalued_index::SerializableMultivalueIndex;
+use crate::column_index::ColumnIndex;
+use crate::column_index::multivalued_index::serialize_multivalued_index;
+use crate::column_index::optional_index::serialize_optional_index;
+use crate::iterable::Iterable;
+use crate::{Cardinality, RowId, Version};
+
+pub struct SerializableOptionalIndex<'a> {
+    pub non_null_row_ids: Box<dyn Iterable<RowId> + 'a>,
+    pub num_rows: RowId,
+}
+
+impl<'a> From<&'a OptionalIndex> for SerializableOptionalIndex<'a> {
+    fn from(optional_index: &'a OptionalIndex) -> Self {
+        SerializableOptionalIndex {
+            non_null_row_ids: Box::new(optional_index),
+            num_rows: optional_index.num_docs(),
+        }
+    }
+}
+
+pub enum SerializableColumnIndex<'a> {
+    Full,
+    Optional(SerializableOptionalIndex<'a>),
+    Multivalued(SerializableMultivalueIndex<'a>),
+}
+
+impl SerializableColumnIndex<'_> {
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            SerializableColumnIndex::Full => Cardinality::Full,
+            SerializableColumnIndex::Optional(_) => Cardinality::Optional,
+            SerializableColumnIndex::Multivalued(_) => Cardinality::Multivalued,
+        }
+    }
+}
+
+/// Serialize a column index.
+pub fn serialize_column_index(
+    column_index: SerializableColumnIndex,
+    output: &mut impl Write,
+) -> io::Result<u32> {
+    let mut output = CountingWriter::wrap(output);
+    let cardinality = column_index.get_cardinality().to_code();
+    output.write_all(&[cardinality])?;
+    match column_index {
+        SerializableColumnIndex::Full => {}
+        SerializableColumnIndex::Optional(SerializableOptionalIndex {
+            non_null_row_ids,
+            num_rows,
+        }) => serialize_optional_index(non_null_row_ids.as_ref(), num_rows, &mut output)?,
+        SerializableColumnIndex::Multivalued(multivalued_index) => {
+            serialize_multivalued_index(&multivalued_index, &mut output)?
+        }
+    }
+    let column_index_num_bytes = output.written_bytes() as u32;
+    Ok(column_index_num_bytes)
+}
+
+/// Open a serialized column index.
+pub fn open_column_index(
+    mut bytes: OwnedBytes,
+    format_version: Version,
+) -> io::Result<ColumnIndex> {
+    if bytes.is_empty() {
+        return Err(io::Error::new(
+            io::ErrorKind::UnexpectedEof,
+            "Failed to deserialize column index. Empty buffer.",
+        ));
+    }
+    let cardinality_code = bytes[0];
+    let cardinality = Cardinality::try_from_code(cardinality_code)?;
+    bytes.advance(1);
+    match cardinality {
+        Cardinality::Full => Ok(ColumnIndex::Full),
+        Cardinality::Optional => {
+            let optional_index = super::optional_index::open_optional_index(bytes)?;
+            Ok(ColumnIndex::Optional(optional_index))
+        }
+        Cardinality::Multivalued => {
+            let multivalue_index =
+                super::multivalued_index::open_multivalued_index(bytes, format_version)?;
+            Ok(ColumnIndex::Multivalued(multivalue_index))
+        }
+    }
+}
+
+// TODO unit tests

columnar/src/column_values/bench.rs

@@ -0,0 +1,139 @@
+use std::sync::Arc;
+
+use common::OwnedBytes;
+use rand::rngs::StdRng;
+use rand::{Rng, SeedableRng};
+use test::{self, Bencher};
+
+use super::*;
+use crate::column_values::u64_based::*;
+
+fn get_data() -> Vec<u64> {
+    let mut rng = StdRng::seed_from_u64(2u64);
+    let mut data: Vec<_> = (100..55000_u64)
+        .map(|num| num + rng.r#gen::<u8>() as u64)
+        .collect();
+    data.push(99_000);
+    data.insert(1000, 2000);
+    data.insert(2000, 100);
+    data.insert(3000, 4100);
+    data.insert(4000, 100);
+    data.insert(5000, 800);
+    data
+}
+
+fn compute_stats(vals: impl Iterator<Item = u64>) -> ColumnStats {
+    let mut stats_collector = StatsCollector::default();
+    for val in vals {
+        stats_collector.collect(val);
+    }
+    stats_collector.stats()
+}
+
+#[inline(never)]
+fn value_iter() -> impl Iterator<Item = u64> {
+    0..20_000
+}
+
+fn get_reader_for_bench<Codec: ColumnCodec>(data: &[u64]) -> Codec::ColumnValues {
+    let mut bytes = Vec::new();
+    let stats = compute_stats(data.iter().cloned());
+    let mut codec_serializer = Codec::estimator();
+    for val in data {
+        codec_serializer.collect(*val);
+    }
+    codec_serializer
+        .serialize(&stats, Box::new(data.iter().copied()).as_mut(), &mut bytes)
+        .unwrap();
+
+    Codec::load(OwnedBytes::new(bytes)).unwrap()
+}
+
+fn bench_get<Codec: ColumnCodec>(b: &mut Bencher, data: &[u64]) {
+    let col = get_reader_for_bench::<Codec>(data);
+    b.iter(|| {
+        let mut sum = 0u64;
+        for pos in value_iter() {
+            let val = col.get_val(pos as u32);
+            sum = sum.wrapping_add(val);
+        }
+        sum
+    });
+}
+
+#[inline(never)]
+fn bench_get_dynamic_helper(b: &mut Bencher, col: Arc<dyn ColumnValues>) {
+    b.iter(|| {
+        let mut sum = 0u64;
+        for pos in value_iter() {
+            let val = col.get_val(pos as u32);
+            sum = sum.wrapping_add(val);
+        }
+        sum
+    });
+}
+
+fn bench_get_dynamic<Codec: ColumnCodec>(b: &mut Bencher, data: &[u64]) {
+    let col = Arc::new(get_reader_for_bench::<Codec>(data));
+    bench_get_dynamic_helper(b, col);
+}
+fn bench_create<Codec: ColumnCodec>(b: &mut Bencher, data: &[u64]) {
+    let stats = compute_stats(data.iter().cloned());
+
+    let mut bytes = Vec::new();
+    b.iter(|| {
+        bytes.clear();
+        let mut codec_serializer = Codec::estimator();
+        for val in data.iter().take(1024) {
+            codec_serializer.collect(*val);
+        }
+
+        codec_serializer.serialize(&stats, Box::new(data.iter().copied()).as_mut(), &mut bytes)
+    });
+}
+
+#[bench]
+fn bench_fastfield_bitpack_create(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_create::<BitpackedCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_linearinterpol_create(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_create::<LinearCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_multilinearinterpol_create(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_create::<BlockwiseLinearCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_bitpack_get(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_get::<BitpackedCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_bitpack_get_dynamic(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_get_dynamic::<BitpackedCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_linearinterpol_get(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_get::<LinearCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_linearinterpol_get_dynamic(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_get_dynamic::<LinearCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_multilinearinterpol_get(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_get::<BlockwiseLinearCodec>(b, &data);
+}
+#[bench]
+fn bench_fastfield_multilinearinterpol_get_dynamic(b: &mut Bencher) {
+    let data: Vec<_> = get_data();
+    bench_get_dynamic::<BlockwiseLinearCodec>(b, &data);
+}

columnar/src/column_values/merge.rs

@@ -0,0 +1,40 @@
+use std::fmt::Debug;
+use std::sync::Arc;
+
+use crate::iterable::Iterable;
+use crate::{ColumnIndex, ColumnValues, MergeRowOrder};
+
+pub(crate) struct MergedColumnValues<'a, T> {
+    pub(crate) column_indexes: &'a [ColumnIndex],
+    pub(crate) column_values: &'a [Option<Arc<dyn ColumnValues<T>>>],
+    pub(crate) merge_row_order: &'a MergeRowOrder,
+}
+
+impl<T: Copy + PartialOrd + Debug + 'static> Iterable<T> for MergedColumnValues<'_, T> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
+        match self.merge_row_order {
+            MergeRowOrder::Stack(_) => Box::new(
+                self.column_values
+                    .iter()
+                    .flatten()
+                    .flat_map(|column_value| column_value.iter()),
+            ),
+            MergeRowOrder::Shuffled(shuffle_merge_order) => Box::new(
+                shuffle_merge_order
+                    .iter_new_to_old_row_addrs()
+                    .flat_map(|row_addr| {
+                        let column_index = &self.column_indexes[row_addr.segment_ord as usize];
+                        let column_values =
+                            self.column_values[row_addr.segment_ord as usize].as_ref()?;
+                        let value_range = column_index.value_row_ids(row_addr.row_id);
+                        Some((value_range, column_values))
+                    })
+                    .flat_map(|(value_range, column_values)| {
+                        value_range
+                            .into_iter()
+                            .map(|val| column_values.get_val(val))
+                    }),
+            ),
+        }
+    }
+}

columnar/src/column_values/mod.rs

@@ -0,0 +1,247 @@
+#![warn(missing_docs)]
+
+//! # `fastfield_codecs`
+//!
+//! - Columnar storage of data for tantivy [`crate::Column`].
+//! - Encode data in different codecs.
+//! - Monotonically map values to u64/u128
+
+use std::fmt::Debug;
+use std::ops::{Range, RangeInclusive};
+use std::sync::Arc;
+
+use downcast_rs::DowncastSync;
+pub use monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
+pub use monotonic_mapping_u128::MonotonicallyMappableToU128;
+
+mod merge;
+pub(crate) mod monotonic_mapping;
+pub(crate) mod monotonic_mapping_u128;
+mod stats;
+mod u128_based;
+mod u64_based;
+mod vec_column;
+
+mod monotonic_column;
+
+pub(crate) use merge::MergedColumnValues;
+pub use stats::ColumnStats;
+pub use u64_based::{
+    ALL_U64_CODEC_TYPES, CodecType, load_u64_based_column_values,
+    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,
+    serialize_column_values_u128,
+};
+pub use vec_column::VecColumn;
+
+pub use self::monotonic_column::monotonic_map_column;
+use crate::RowId;
+
+/// `ColumnValues` provides access to a dense field column.
+///
+/// `Column` are just a wrapper over `ColumnValues` and a `ColumnIndex`.
+///
+/// Any methods with a default and specialized implementation need to be called in the
+/// wrappers that implement the trait: Arc and MonotonicMappingColumn
+pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync + DowncastSync {
+    /// Return the value associated with the given idx.
+    ///
+    /// This accessor should return as fast as possible.
+    ///
+    /// # Panics
+    ///
+    /// May panic if `idx` is greater than the column length.
+    fn get_val(&self, idx: u32) -> T;
+
+    /// Allows to push down multiple fetch calls, to avoid dynamic dispatch overhead.
+    ///
+    /// idx and output should have the same length
+    ///
+    /// # Panics
+    ///
+    /// May panic if `idx` is greater than the column length.
+    fn get_vals(&self, indexes: &[u32], output: &mut [T]) {
+        assert!(indexes.len() == output.len());
+        let out_and_idx_chunks = output.chunks_exact_mut(4).zip(indexes.chunks_exact(4));
+        for (out_x4, idx_x4) in out_and_idx_chunks {
+            out_x4[0] = self.get_val(idx_x4[0]);
+            out_x4[1] = self.get_val(idx_x4[1]);
+            out_x4[2] = self.get_val(idx_x4[2]);
+            out_x4[3] = self.get_val(idx_x4[3]);
+        }
+
+        let out_and_idx_chunks = output
+            .chunks_exact_mut(4)
+            .into_remainder()
+            .iter_mut()
+            .zip(indexes.chunks_exact(4).remainder());
+        for (out, idx) in out_and_idx_chunks {
+            *out = self.get_val(*idx);
+        }
+    }
+
+    /// Allows to push down multiple fetch calls, to avoid dynamic dispatch overhead.
+    /// The slightly weird `Option<T>` in output allows pushdown to full columns.
+    ///
+    /// idx and output should have the same length
+    ///
+    /// # Panics
+    ///
+    /// May panic if `idx` is greater than the column length.
+    fn get_vals_opt(&self, indexes: &[u32], output: &mut [Option<T>]) {
+        assert!(indexes.len() == output.len());
+        let out_and_idx_chunks = output.chunks_exact_mut(4).zip(indexes.chunks_exact(4));
+        for (out_x4, idx_x4) in out_and_idx_chunks {
+            out_x4[0] = Some(self.get_val(idx_x4[0]));
+            out_x4[1] = Some(self.get_val(idx_x4[1]));
+            out_x4[2] = Some(self.get_val(idx_x4[2]));
+            out_x4[3] = Some(self.get_val(idx_x4[3]));
+        }
+        let out_and_idx_chunks = output
+            .chunks_exact_mut(4)
+            .into_remainder()
+            .iter_mut()
+            .zip(indexes.chunks_exact(4).remainder());
+        for (out, idx) in out_and_idx_chunks {
+            *out = Some(self.get_val(*idx));
+        }
+    }
+
+    /// Fills an output buffer with the fast field values
+    /// associated with the `DocId` going from
+    /// `start` to `start + output.len()`.
+    ///
+    /// # Panics
+    ///
+    /// Must panic if `start + output.len()` is greater than
+    /// the segment's `maxdoc`.
+    #[inline(always)]
+    fn get_range(&self, start: u64, output: &mut [T]) {
+        for (out, idx) in output.iter_mut().zip(start..) {
+            *out = self.get_val(idx as u32);
+        }
+    }
+
+    /// Get the row ids of values which are in the provided value range.
+    ///
+    /// Note that position == docid for single value fast fields
+    fn get_row_ids_for_value_range(
+        &self,
+        value_range: RangeInclusive<T>,
+        row_id_range: Range<RowId>,
+        row_id_hits: &mut Vec<RowId>,
+    ) {
+        let row_id_range = row_id_range.start..row_id_range.end.min(self.num_vals());
+        for idx in row_id_range {
+            let val = self.get_val(idx);
+            if value_range.contains(&val) {
+                row_id_hits.push(idx);
+            }
+        }
+    }
+
+    /// Returns a lower bound for this column of values.
+    ///
+    /// All values are guaranteed to be higher than `.min_value()`
+    /// but this value is not necessary the best boundary value.
+    ///
+    /// We have
+    /// ∀i < self.num_vals(), self.get_val(i) >= self.min_value()
+    /// But we don't have necessarily
+    /// ∃i < self.num_vals(), self.get_val(i) == self.min_value()
+    fn min_value(&self) -> T;
+
+    /// Returns an upper bound for this column of values.
+    ///
+    /// All values are guaranteed to be lower than `.max_value()`
+    /// but this value is not necessary the best boundary value.
+    ///
+    /// We have
+    /// ∀i < self.num_vals(), self.get_val(i) <= self.max_value()
+    /// But we don't have necessarily
+    /// ∃i < self.num_vals(), self.get_val(i) == self.max_value()
+    fn max_value(&self) -> T;
+
+    /// The number of values in the column.
+    fn num_vals(&self) -> u32;
+
+    /// Returns a iterator over the data
+    fn iter<'a>(&'a self) -> Box<dyn Iterator<Item = T> + 'a> {
+        Box::new((0..self.num_vals()).map(|idx| self.get_val(idx)))
+    }
+}
+downcast_rs::impl_downcast!(sync ColumnValues<T> where T: PartialOrd);
+
+/// Empty column of values.
+pub struct EmptyColumnValues;
+
+impl<T: PartialOrd + Default> ColumnValues<T> for EmptyColumnValues {
+    fn get_val(&self, _idx: u32) -> T {
+        panic!("Internal Error: Called get_val of empty column.")
+    }
+
+    fn min_value(&self) -> T {
+        T::default()
+    }
+
+    fn max_value(&self) -> T {
+        T::default()
+    }
+
+    fn num_vals(&self) -> u32 {
+        0
+    }
+}
+
+impl<T: Copy + PartialOrd + Debug + 'static> ColumnValues<T> for Arc<dyn ColumnValues<T>> {
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> T {
+        self.as_ref().get_val(idx)
+    }
+
+    #[inline(always)]
+    fn get_vals_opt(&self, indexes: &[u32], output: &mut [Option<T>]) {
+        self.as_ref().get_vals_opt(indexes, output)
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> T {
+        self.as_ref().min_value()
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> T {
+        self.as_ref().max_value()
+    }
+
+    #[inline(always)]
+    fn num_vals(&self) -> u32 {
+        self.as_ref().num_vals()
+    }
+
+    #[inline(always)]
+    fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
+        self.as_ref().iter()
+    }
+
+    #[inline(always)]
+    fn get_range(&self, start: u64, output: &mut [T]) {
+        self.as_ref().get_range(start, output)
+    }
+
+    #[inline(always)]
+    fn get_row_ids_for_value_range(
+        &self,
+        range: RangeInclusive<T>,
+        doc_id_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        self.as_ref()
+            .get_row_ids_for_value_range(range, doc_id_range, positions)
+    }
+}
+
+#[cfg(all(test, feature = "unstable"))]
+mod bench;

columnar/src/column_values/monotonic_column.rs

@@ -0,0 +1,120 @@
+use std::fmt::Debug;
+use std::marker::PhantomData;
+use std::ops::{Range, RangeInclusive};
+
+use crate::ColumnValues;
+use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
+
+struct MonotonicMappingColumn<C, T, Input> {
+    from_column: C,
+    monotonic_mapping: T,
+    _phantom: PhantomData<Input>,
+}
+
+/// Creates a view of a column transformed by a strictly monotonic mapping. See
+/// [`StrictlyMonotonicFn`].
+///
+/// E.g. apply a gcd monotonic_mapping([100, 200, 300]) == [1, 2, 3]
+/// monotonic_mapping.mapping() is expected to be injective, and we should always have
+/// monotonic_mapping.inverse(monotonic_mapping.mapping(el)) == el
+///
+/// The inverse of the mapping is required for:
+/// `fn get_positions_for_value_range(&self, range: RangeInclusive<T>) -> Vec<u64> `
+/// The user provides the original value range and we need to monotonic map them in the same way the
+/// serialization does before calling the underlying column.
+///
+/// Note that when opening a codec, the monotonic_mapping should be the inverse of the mapping
+/// during serialization. And therefore the monotonic_mapping_inv when opening is the same as
+/// monotonic_mapping during serialization.
+pub fn monotonic_map_column<C, T, Input, Output>(
+    from_column: C,
+    monotonic_mapping: T,
+) -> impl ColumnValues<Output>
+where
+    C: ColumnValues<Input> + 'static,
+    T: StrictlyMonotonicFn<Input, Output> + Send + Sync + 'static,
+    Input: PartialOrd + Debug + Send + Sync + Clone + 'static,
+    Output: PartialOrd + Debug + Send + Sync + Clone + 'static,
+{
+    MonotonicMappingColumn {
+        from_column,
+        monotonic_mapping,
+        _phantom: PhantomData,
+    }
+}
+
+impl<C, T, Input, Output> ColumnValues<Output> for MonotonicMappingColumn<C, T, Input>
+where
+    C: ColumnValues<Input> + 'static,
+    T: StrictlyMonotonicFn<Input, Output> + Send + Sync + 'static,
+    Input: PartialOrd + Send + Debug + Sync + Clone + 'static,
+    Output: PartialOrd + Send + Debug + Sync + Clone + 'static,
+{
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> Output {
+        let from_val = self.from_column.get_val(idx);
+        self.monotonic_mapping.mapping(from_val)
+    }
+
+    fn min_value(&self) -> Output {
+        let from_min_value = self.from_column.min_value();
+        self.monotonic_mapping.mapping(from_min_value)
+    }
+
+    fn max_value(&self) -> Output {
+        let from_max_value = self.from_column.max_value();
+        self.monotonic_mapping.mapping(from_max_value)
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.from_column.num_vals()
+    }
+
+    fn iter(&self) -> Box<dyn Iterator<Item = Output> + '_> {
+        Box::new(
+            self.from_column
+                .iter()
+                .map(|el| self.monotonic_mapping.mapping(el)),
+        )
+    }
+
+    fn get_row_ids_for_value_range(
+        &self,
+        range: RangeInclusive<Output>,
+        doc_id_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        self.from_column.get_row_ids_for_value_range(
+            self.monotonic_mapping.inverse(range.start().clone())
+                ..=self.monotonic_mapping.inverse(range.end().clone()),
+            doc_id_range,
+            positions,
+        )
+    }
+
+    // We voluntarily do not implement get_range as it yields a regression,
+    // and we do not have any specialized implementation anyway.
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::VecColumn;
+    use crate::column_values::monotonic_mapping::{
+        StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+    };
+
+    #[test]
+    fn test_monotonic_mapping_iter() {
+        let vals: Vec<u64> = (0..100u64).map(|el| el * 10).collect();
+        let col = VecColumn::from(vals);
+        let mapped = monotonic_map_column(
+            col,
+            StrictlyMonotonicMappingInverter::from(StrictlyMonotonicMappingToInternal::<i64>::new()),
+        );
+        let val_i64s: Vec<u64> = mapped.iter().collect();
+        for i in 0..100 {
+            assert_eq!(val_i64s[i as usize], mapped.get_val(i));
+        }
+    }
+}

columnar/src/column_values/monotonic_mapping.rs

@@ -0,0 +1,211 @@
+use std::fmt::Debug;
+use std::marker::PhantomData;
+
+use common::DateTime;
+
+use super::MonotonicallyMappableToU128;
+use crate::RowId;
+
+/// Monotonic maps a value to u64 value space.
+/// Monotonic mapping enables `PartialOrd` on u64 space without conversion to original space.
+pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Debug + Copy + Send + Sync {
+    /// Converts a value to u64.
+    ///
+    /// Internally all fast field values are encoded as u64.
+    fn to_u64(self) -> u64;
+
+    /// Converts a value from u64
+    ///
+    /// Internally all fast field values are encoded as u64.
+    /// **Note: To be used for converting encoded Term, Posting values.**
+    fn from_u64(val: u64) -> Self;
+}
+
+/// Values need to be strictly monotonic mapped to a `Internal` value (u64 or u128) that can be
+/// used in fast field codecs.
+///
+/// The monotonic mapping is required so that `PartialOrd` can be used on `Internal` without
+/// converting to `External`.
+///
+/// All strictly monotonic functions are invertible because they are guaranteed to have a one-to-one
+/// mapping from their range to their domain. The `inverse` method is required when opening a codec,
+/// so a value can be converted back to its original domain (e.g. ip address or f64) from its
+/// internal representation.
+pub trait StrictlyMonotonicFn<External, Internal> {
+    /// Strictly monotonically maps the value from External to Internal.
+    fn mapping(&self, inp: External) -> Internal;
+    /// Inverse of `mapping`. Maps the value from Internal to External.
+    fn inverse(&self, out: Internal) -> External;
+}
+
+/// Inverts a strictly monotonic mapping from `StrictlyMonotonicFn<A, B>` to
+/// `StrictlyMonotonicFn<B, A>`.
+///
+/// # Warning
+///
+/// This type comes with a footgun. A type being strictly monotonic does not impose that the inverse
+/// mapping is strictly monotonic over the entire space External. e.g. a -> a * 2. Use at your own
+/// risks.
+pub(crate) struct StrictlyMonotonicMappingInverter<T> {
+    orig_mapping: T,
+}
+impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
+    fn from(orig_mapping: T) -> Self {
+        Self { orig_mapping }
+    }
+}
+
+impl<From, To, T> StrictlyMonotonicFn<To, From> for StrictlyMonotonicMappingInverter<T>
+where T: StrictlyMonotonicFn<From, To>
+{
+    #[inline(always)]
+    fn mapping(&self, val: To) -> From {
+        self.orig_mapping.inverse(val)
+    }
+
+    #[inline(always)]
+    fn inverse(&self, val: From) -> To {
+        self.orig_mapping.mapping(val)
+    }
+}
+
+/// Applies the strictly monotonic mapping from `T` without any additional changes.
+pub(crate) struct StrictlyMonotonicMappingToInternal<T> {
+    _phantom: PhantomData<T>,
+}
+
+impl<T> StrictlyMonotonicMappingToInternal<T> {
+    pub(crate) fn new() -> StrictlyMonotonicMappingToInternal<T> {
+        Self {
+            _phantom: PhantomData,
+        }
+    }
+}
+
+impl<External: MonotonicallyMappableToU128, T: MonotonicallyMappableToU128>
+    StrictlyMonotonicFn<External, u128> for StrictlyMonotonicMappingToInternal<T>
+where T: MonotonicallyMappableToU128
+{
+    #[inline(always)]
+    fn mapping(&self, inp: External) -> u128 {
+        External::to_u128(inp)
+    }
+
+    #[inline(always)]
+    fn inverse(&self, out: u128) -> External {
+        External::from_u128(out)
+    }
+}
+
+impl<External: MonotonicallyMappableToU64, T: MonotonicallyMappableToU64>
+    StrictlyMonotonicFn<External, u64> for StrictlyMonotonicMappingToInternal<T>
+where T: MonotonicallyMappableToU64
+{
+    #[inline(always)]
+    fn mapping(&self, inp: External) -> u64 {
+        External::to_u64(inp)
+    }
+
+    #[inline(always)]
+    fn inverse(&self, out: u64) -> External {
+        External::from_u64(out)
+    }
+}
+
+impl MonotonicallyMappableToU64 for u64 {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        self
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        val
+    }
+}
+
+impl MonotonicallyMappableToU64 for i64 {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        common::i64_to_u64(self)
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        common::u64_to_i64(val)
+    }
+}
+
+impl MonotonicallyMappableToU64 for DateTime {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        common::i64_to_u64(self.into_timestamp_nanos())
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        DateTime::from_timestamp_nanos(common::u64_to_i64(val))
+    }
+}
+
+impl MonotonicallyMappableToU64 for bool {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        u64::from(self)
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        val > 0
+    }
+}
+
+impl MonotonicallyMappableToU64 for RowId {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        u64::from(self)
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> RowId {
+        val as RowId
+    }
+}
+
+// TODO remove me.
+// Tantivy should refuse NaN values and work with NotNaN internally.
+impl MonotonicallyMappableToU64 for f64 {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        common::f64_to_u64(self)
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        common::u64_to_f64(val)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+
+    use super::*;
+
+    #[test]
+    fn strictly_monotonic_test() {
+        // identity mapping
+        test_round_trip(&StrictlyMonotonicMappingToInternal::<u64>::new(), 100u64);
+        // round trip to i64
+        test_round_trip(&StrictlyMonotonicMappingToInternal::<i64>::new(), 100u64);
+        // TODO
+        // identity mapping
+        // test_round_trip(&StrictlyMonotonicMappingToInternal::<u128>::new(), 100u128);
+    }
+
+    fn test_round_trip<T: StrictlyMonotonicFn<K, L>, K: std::fmt::Debug + Eq + Copy, L>(
+        mapping: &T,
+        test_val: K,
+    ) {
+        assert_eq!(mapping.inverse(mapping.mapping(test_val)), test_val);
+    }
+}

columnar/src/column_values/monotonic_mapping_u128.rs

@@ -0,0 +1,41 @@
+use std::fmt::Debug;
+use std::net::Ipv6Addr;
+
+/// Montonic maps a value to u128 value space
+/// Monotonic mapping enables `PartialOrd` on u128 space without conversion to original space.
+pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Debug + Send + Sync {
+    /// Converts a value to u128.
+    ///
+    /// Internally all fast field values are encoded as u64.
+    fn to_u128(self) -> u128;
+
+    /// Converts a value from u128
+    ///
+    /// Internally all fast field values are encoded as u64.
+    /// **Note: To be used for converting encoded Term, Posting values.**
+    fn from_u128(val: u128) -> Self;
+}
+
+impl MonotonicallyMappableToU128 for u128 {
+    fn to_u128(self) -> u128 {
+        self
+    }
+
+    fn from_u128(val: u128) -> Self {
+        val
+    }
+}
+
+impl MonotonicallyMappableToU128 for Ipv6Addr {
+    fn to_u128(self) -> u128 {
+        ip_to_u128(self)
+    }
+
+    fn from_u128(val: u128) -> Self {
+        Ipv6Addr::from(val.to_be_bytes())
+    }
+}
+
+fn ip_to_u128(ip_addr: Ipv6Addr) -> u128 {
+    u128::from_be_bytes(ip_addr.octets())
+}

columnar/src/column_values/stats.rs

@@ -0,0 +1,103 @@
+use std::io;
+use std::io::Write;
+use std::num::NonZeroU64;
+
+use common::{BinarySerializable, VInt};
+
+use crate::RowId;
+
+/// Column statistics.
+#[derive(Debug, Clone, Eq, PartialEq)]
+pub struct ColumnStats {
+    /// GCD of the elements `el - min(column)`.
+    pub gcd: NonZeroU64,
+    /// Minimum value of the column.
+    pub min_value: u64,
+    /// Maximum value of the column.
+    pub max_value: u64,
+    /// Number of rows in the column.
+    pub num_rows: RowId,
+}
+
+impl ColumnStats {
+    /// Amplitude of value.
+    /// Difference between the maximum and the minimum value.
+    pub fn amplitude(&self) -> u64 {
+        self.max_value - self.min_value
+    }
+}
+
+impl BinarySerializable for ColumnStats {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.min_value).serialize(writer)?;
+        VInt(self.gcd.get()).serialize(writer)?;
+        VInt(self.amplitude() / self.gcd).serialize(writer)?;
+        VInt(self.num_rows as u64).serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let min_value = VInt::deserialize(reader)?.0;
+        let gcd = VInt::deserialize(reader)?.0;
+        let gcd = NonZeroU64::new(gcd)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "GCD of 0 is forbidden"))?;
+        let amplitude = VInt::deserialize(reader)?.0 * gcd.get();
+        let max_value = min_value + amplitude;
+        let num_rows = VInt::deserialize(reader)?.0 as RowId;
+        Ok(ColumnStats {
+            min_value,
+            max_value,
+            num_rows,
+            gcd,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use common::BinarySerializable;
+
+    use crate::column_values::ColumnStats;
+
+    #[track_caller]
+    fn test_stats_ser_deser_aux(stats: &ColumnStats, num_bytes: usize) {
+        let mut buffer: Vec<u8> = Vec::new();
+        stats.serialize(&mut buffer).unwrap();
+        assert_eq!(buffer.len(), num_bytes);
+        let deser_stats = ColumnStats::deserialize(&mut &buffer[..]).unwrap();
+        assert_eq!(stats, &deser_stats);
+    }
+
+    #[test]
+    fn test_stats_serialization() {
+        test_stats_ser_deser_aux(
+            &(ColumnStats {
+                gcd: NonZeroU64::new(3).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(ColumnStats {
+                gcd: NonZeroU64::new(1_000).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 0,
+                num_rows: 0,
+            }),
+            4,
+        );
+    }
+}

columnar/src/column_values/u128_based/compact_space/blank_range.rs

@@ -0,0 +1,43 @@
+use std::ops::RangeInclusive;
+
+/// The range of a blank in value space.
+///
+/// A blank is an unoccupied space in the data.
+/// Use try_into() to construct.
+/// A range has to have at least length of 3. Invalid ranges will be rejected.
+///
+/// Ordered by range length.
+#[derive(Debug, Eq, PartialEq, Clone)]
+pub(crate) struct BlankRange {
+    blank_range: RangeInclusive<u128>,
+}
+impl TryFrom<RangeInclusive<u128>> for BlankRange {
+    type Error = &'static str;
+    fn try_from(range: RangeInclusive<u128>) -> Result<Self, Self::Error> {
+        let blank_size = range.end().saturating_sub(*range.start());
+        if blank_size < 2 {
+            Err("invalid range")
+        } else {
+            Ok(BlankRange { blank_range: range })
+        }
+    }
+}
+impl BlankRange {
+    pub(crate) fn blank_size(&self) -> u128 {
+        self.blank_range.end() - self.blank_range.start() + 1
+    }
+    pub(crate) fn blank_range(&self) -> RangeInclusive<u128> {
+        self.blank_range.clone()
+    }
+}
+
+impl Ord for BlankRange {
+    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
+        self.blank_size().cmp(&other.blank_size())
+    }
+}
+impl PartialOrd for BlankRange {
+    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
+        Some(self.cmp(other))
+    }
+}

columnar/src/column_values/u128_based/compact_space/build_compact_space.rs

@@ -0,0 +1,252 @@
+use std::collections::{BTreeSet, BinaryHeap};
+use std::iter;
+use std::ops::RangeInclusive;
+
+use itertools::Itertools;
+
+use super::blank_range::BlankRange;
+use super::{CompactSpace, RangeMapping};
+
+/// Put the blanks for the sorted values into a binary heap
+fn get_blanks(values_sorted: &BTreeSet<u128>) -> BinaryHeap<BlankRange> {
+    let mut blanks: BinaryHeap<BlankRange> = BinaryHeap::new();
+    for (first, second) in values_sorted.iter().copied().tuple_windows() {
+        // Correctness Overflow: the values are deduped and sorted (BTreeSet property), that means
+        // there's always space between two values.
+        let blank_range = first + 1..=second - 1;
+        let blank_range: Result<BlankRange, _> = blank_range.try_into();
+        if let Ok(blank_range) = blank_range {
+            blanks.push(blank_range);
+        }
+    }
+
+    blanks
+}
+
+struct BlankCollector {
+    blanks: Vec<BlankRange>,
+    staged_blanks_sum: u128,
+}
+impl BlankCollector {
+    fn new() -> Self {
+        Self {
+            blanks: vec![],
+            staged_blanks_sum: 0,
+        }
+    }
+    fn stage_blank(&mut self, blank: BlankRange) {
+        self.staged_blanks_sum += blank.blank_size();
+        self.blanks.push(blank);
+    }
+    fn drain(&mut self) -> impl Iterator<Item = BlankRange> + '_ {
+        self.staged_blanks_sum = 0;
+        self.blanks.drain(..)
+    }
+    fn staged_blanks_sum(&self) -> u128 {
+        self.staged_blanks_sum
+    }
+    fn num_staged_blanks(&self) -> usize {
+        self.blanks.len()
+    }
+}
+fn num_bits(val: u128) -> u8 {
+    (128u32 - val.leading_zeros()) as u8
+}
+
+/// Will collect blanks and add them to compact space if more bits are saved than cost from
+/// metadata.
+pub fn get_compact_space(
+    values_deduped_sorted: &BTreeSet<u128>,
+    total_num_values: u32,
+    cost_per_blank: usize,
+) -> CompactSpace {
+    let mut compact_space_builder = CompactSpaceBuilder::new();
+    if values_deduped_sorted.is_empty() {
+        return compact_space_builder.finish();
+    }
+
+    // We start by space that's limited to min_value..=max_value
+    // Replace after stabilization of https://github.com/rust-lang/rust/issues/62924
+    let min_value = values_deduped_sorted.iter().next().copied().unwrap_or(0);
+    let max_value = values_deduped_sorted.iter().last().copied().unwrap_or(0);
+
+    let mut blanks: BinaryHeap<BlankRange> = get_blanks(values_deduped_sorted);
+
+    // +1 for null, in case min and max covers the whole space, we are off by one.
+    let mut amplitude_compact_space = (max_value - min_value).saturating_add(1);
+    if min_value != 0 {
+        compact_space_builder.add_blanks(iter::once(0..=min_value - 1));
+    }
+    if max_value != u128::MAX {
+        compact_space_builder.add_blanks(iter::once(max_value + 1..=u128::MAX));
+    }
+
+    let mut amplitude_bits: u8 = num_bits(amplitude_compact_space);
+
+    let mut blank_collector = BlankCollector::new();
+
+    // We will stage blanks until they reduce the compact space by at least 1 bit and then flush
+    // them if the metadata cost is lower than the total number of saved bits.
+    // Binary heap to process the gaps by their size
+    while let Some(blank_range) = blanks.pop() {
+        blank_collector.stage_blank(blank_range);
+
+        let staged_spaces_sum: u128 = blank_collector.staged_blanks_sum();
+        let amplitude_new_compact_space = amplitude_compact_space - staged_spaces_sum;
+        let amplitude_new_bits = num_bits(amplitude_new_compact_space);
+
+        if amplitude_bits == amplitude_new_bits {
+            continue;
+        }
+        let saved_bits = (amplitude_bits - amplitude_new_bits) as usize * total_num_values as usize;
+        // TODO: Maybe calculate exact cost of blanks and run this more expensive computation only,
+        // when amplitude_new_bits changes
+        let cost = blank_collector.num_staged_blanks() * cost_per_blank;
+
+        // We want to end up with a compact space that fits into 32 bits.
+        // In order to deal with pathological cases, we force the algorithm to keep
+        // refining the compact space the amplitude bits is lower than 32.
+        //
+        // The worst case scenario happens for a large number of u128s regularly
+        // spread over the full u128 space.
+        //
+        // This change will force the algorithm to degenerate into dictionary encoding.
+        if amplitude_bits <= 32 && cost >= saved_bits {
+            // Continue here, since although we walk over the blanks by size,
+            // we can potentially save a lot at the last bits, which are smaller blanks
+            //
+            // E.g. if the first range reduces the compact space by 1000 from 2000 to 1000, which
+            // saves 11-10=1 bit and the next range reduces the compact space by 950 to
+            // 50, which saves 10-6=4 bit
+            continue;
+        }
+
+        amplitude_compact_space = amplitude_new_compact_space;
+        amplitude_bits = amplitude_new_bits;
+        compact_space_builder.add_blanks(blank_collector.drain().map(|blank| blank.blank_range()));
+    }
+
+    assert!(amplitude_bits <= 32);
+
+    // special case, when we don't collected any blanks because:
+    // * the data is empty (early exit)
+    // * the algorithm did decide it's not worth the cost, which can be the case for single values
+    //
+    // We drain one collected blank unconditionally, so the empty case is reserved for empty
+    // data, and therefore empty compact_space means the data is empty and no data is covered
+    // (conversely to all data) and we can assign null to it.
+    if compact_space_builder.is_empty() {
+        compact_space_builder.add_blanks(
+            blank_collector
+                .drain()
+                .map(|blank| blank.blank_range())
+                .take(1),
+        );
+    }
+
+    let compact_space = compact_space_builder.finish();
+    if max_value - min_value != u128::MAX {
+        debug_assert_eq!(
+            compact_space.amplitude_compact_space(),
+            amplitude_compact_space
+        );
+    }
+    compact_space
+}
+
+#[derive(Debug, Clone, Eq, PartialEq)]
+struct CompactSpaceBuilder {
+    blanks: Vec<RangeInclusive<u128>>,
+}
+
+impl CompactSpaceBuilder {
+    /// Creates a new compact space builder which will initially cover the whole space.
+    fn new() -> Self {
+        Self { blanks: Vec::new() }
+    }
+
+    /// Assumes that repeated add_blank calls don't overlap and are not adjacent,
+    /// e.g. [3..=5, 5..=10] is not allowed
+    ///
+    /// Both of those assumptions are true when blanks are produced from sorted values.
+    fn add_blanks(&mut self, blank: impl Iterator<Item = RangeInclusive<u128>>) {
+        self.blanks.extend(blank);
+    }
+
+    fn is_empty(&self) -> bool {
+        self.blanks.is_empty()
+    }
+
+    /// Convert blanks to covered space and assign null value
+    fn finish(mut self) -> CompactSpace {
+        // sort by start. ranges are not allowed to overlap
+        self.blanks.sort_unstable_by_key(|blank| *blank.start());
+
+        let mut covered_space = Vec::with_capacity(self.blanks.len());
+
+        // beginning of the blanks
+        if let Some(first_blank_start) = self.blanks.first().map(RangeInclusive::start) {
+            if *first_blank_start != 0 {
+                covered_space.push(0..=first_blank_start - 1);
+            }
+        }
+
+        // Between the blanks
+        let between_blanks = self.blanks.iter().tuple_windows().map(|(left, right)| {
+            assert!(
+                left.end() < right.start(),
+                "overlapping or adjacent ranges detected"
+            );
+            *left.end() + 1..=*right.start() - 1
+        });
+        covered_space.extend(between_blanks);
+
+        // end of the blanks
+        if let Some(last_blank_end) = self.blanks.last().map(RangeInclusive::end) {
+            if *last_blank_end != u128::MAX {
+                covered_space.push(last_blank_end + 1..=u128::MAX);
+            }
+        }
+
+        if covered_space.is_empty() {
+            covered_space.push(0..=0); // empty data case
+        };
+
+        let mut compact_start: u32 = 1; // 0 is reserved for `null`
+        let mut ranges_mapping: Vec<RangeMapping> = Vec::with_capacity(covered_space.len());
+        for cov in covered_space {
+            let range_mapping = super::RangeMapping {
+                value_range: cov,
+                compact_start,
+            };
+            let covered_range_len = range_mapping.range_length();
+            ranges_mapping.push(range_mapping);
+            compact_start += covered_range_len;
+        }
+        // println!("num ranges {}", ranges_mapping.len());
+        CompactSpace { ranges_mapping }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u128_based::compact_space::COST_PER_BLANK_IN_BITS;
+
+    #[test]
+    fn test_binary_heap_pop_order() {
+        let mut blanks: BinaryHeap<BlankRange> = BinaryHeap::new();
+        blanks.push((0..=10).try_into().unwrap());
+        blanks.push((100..=200).try_into().unwrap());
+        blanks.push((100..=110).try_into().unwrap());
+        assert_eq!(blanks.pop().unwrap().blank_size(), 101);
+        assert_eq!(blanks.pop().unwrap().blank_size(), 11);
+    }
+
+    #[test]
+    fn test_worst_case_scenario() {
+        let vals: BTreeSet<u128> = (0..8).map(|i| i * ((1u128 << 34) / 8)).collect();
+        let compact_space = get_compact_space(&vals, vals.len() as u32, COST_PER_BLANK_IN_BITS);
+        assert!(compact_space.amplitude_compact_space() < u32::MAX as u128);
+    }
+}

columnar/src/column_values/u128_based/compact_space/mod.rs

@@ -0,0 +1,846 @@
+/// This codec takes a large number space (u128) and reduces it to a compact number space.
+///
+/// It will find spaces in the number range. For example:
+///
+/// 100, 101, 102, 103, 104, 50000, 50001
+/// could be mapped to
+/// 100..104 -> 0..4
+/// 50000..50001 -> 5..6
+///
+/// Compact space 0..=6 requires much less bits than 100..=50001
+///
+/// The codec is created to compress ip addresses, but may be employed in other use cases.
+use std::{
+    cmp::Ordering,
+    collections::BTreeSet,
+    io::{self, Write},
+    ops::{Range, RangeInclusive},
+};
+
+mod blank_range;
+mod build_compact_space;
+
+use build_compact_space::get_compact_space;
+use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt, VIntU128};
+use tantivy_bitpacker::{BitPacker, BitUnpacker};
+
+use crate::RowId;
+use crate::column_values::ColumnValues;
+
+/// The cost per blank is quite hard actually, since blanks are delta encoded, the actual cost of
+/// blanks depends on the number of blanks.
+///
+/// The number is taken by looking at a real dataset. It is optimized for larger datasets.
+const COST_PER_BLANK_IN_BITS: usize = 36;
+
+#[derive(Debug, Clone, Eq, PartialEq)]
+pub struct CompactSpace {
+    ranges_mapping: Vec<RangeMapping>,
+}
+
+/// Maps the range from the original space to compact_start + range.len()
+#[derive(Debug, Clone, Eq, PartialEq)]
+struct RangeMapping {
+    value_range: RangeInclusive<u128>,
+    compact_start: u32,
+}
+impl RangeMapping {
+    fn range_length(&self) -> u32 {
+        (self.value_range.end() - self.value_range.start()) as u32 + 1
+    }
+
+    // The last value of the compact space in this range
+    fn compact_end(&self) -> u32 {
+        self.compact_start + self.range_length() - 1
+    }
+}
+
+impl BinarySerializable for CompactSpace {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
+
+        let mut prev_value = 0;
+        for value_range in self
+            .ranges_mapping
+            .iter()
+            .map(|range_mapping| &range_mapping.value_range)
+        {
+            let blank_delta_start = value_range.start() - prev_value;
+            VIntU128(blank_delta_start).serialize(writer)?;
+            prev_value = *value_range.start();
+
+            let blank_delta_end = value_range.end() - prev_value;
+            VIntU128(blank_delta_end).serialize(writer)?;
+            prev_value = *value_range.end();
+        }
+
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let num_ranges = VInt::deserialize(reader)?.0;
+        let mut ranges_mapping: Vec<RangeMapping> = vec![];
+        let mut value = 0u128;
+        let mut compact_start = 1u32; // 0 is reserved for `null`
+        for _ in 0..num_ranges {
+            let blank_delta_start = VIntU128::deserialize(reader)?.0;
+            value += blank_delta_start;
+            let blank_start = value;
+
+            let blank_delta_end = VIntU128::deserialize(reader)?.0;
+            value += blank_delta_end;
+            let blank_end = value;
+
+            let range_mapping = RangeMapping {
+                value_range: blank_start..=blank_end,
+                compact_start,
+            };
+            let range_length = range_mapping.range_length();
+            ranges_mapping.push(range_mapping);
+            compact_start += range_length;
+        }
+
+        Ok(Self { ranges_mapping })
+    }
+}
+
+impl CompactSpace {
+    /// Amplitude is the value range of the compact space including the sentinel value used to
+    /// identify null values. The compact space is 0..=amplitude .
+    ///
+    /// It's only used to verify we don't exceed u64 number space, which would indicate a bug.
+    fn amplitude_compact_space(&self) -> u128 {
+        self.ranges_mapping
+            .last()
+            .map(|last_range| last_range.compact_end() as u128)
+            .unwrap_or(1) // compact space starts at 1, 0 == null
+    }
+
+    fn get_range_mapping(&self, pos: usize) -> &RangeMapping {
+        &self.ranges_mapping[pos]
+    }
+
+    /// Returns either Ok(the value in the compact space) or if it is outside the compact space the
+    /// Err(position where it would be inserted)
+    fn u128_to_compact(&self, value: u128) -> Result<u32, usize> {
+        self.ranges_mapping
+            .binary_search_by(|probe| {
+                let value_range: &RangeInclusive<u128> = &probe.value_range;
+                if value < *value_range.start() {
+                    Ordering::Greater
+                } else if value > *value_range.end() {
+                    Ordering::Less
+                } else {
+                    Ordering::Equal
+                }
+            })
+            .map(|pos| {
+                let range_mapping = &self.ranges_mapping[pos];
+                let pos_in_range: u32 = (value - range_mapping.value_range.start()) as u32;
+                range_mapping.compact_start + pos_in_range
+            })
+    }
+
+    /// Unpacks a value from compact space u32 to u128 space
+    fn compact_to_u128(&self, compact: u32) -> u128 {
+        let pos = self
+            .ranges_mapping
+            .binary_search_by_key(&compact, |range_mapping| range_mapping.compact_start)
+            // Correctness: Overflow. The first range starts at compact space 0, the error from
+            // binary search can never be 0
+            .unwrap_or_else(|e| e - 1);
+
+        let range_mapping = &self.ranges_mapping[pos];
+        let diff = compact - range_mapping.compact_start;
+        range_mapping.value_range.start() + diff as u128
+    }
+}
+
+pub struct CompactSpaceCompressor {
+    params: IPCodecParams,
+}
+
+#[derive(Debug, Clone)]
+pub struct IPCodecParams {
+    compact_space: CompactSpace,
+    bit_unpacker: BitUnpacker,
+    min_value: u128,
+    max_value: u128,
+    num_vals: RowId,
+    num_bits: u8,
+}
+
+impl CompactSpaceCompressor {
+    pub fn num_vals(&self) -> RowId {
+        self.params.num_vals
+    }
+
+    /// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
+    pub fn train_from(iter: impl Iterator<Item = u128>) -> Self {
+        let mut values_sorted = BTreeSet::new();
+        // Total number of values, with their redundancy.
+        let mut total_num_values = 0u32;
+        for val in iter {
+            total_num_values += 1u32;
+            values_sorted.insert(val);
+        }
+        let min_value = *values_sorted.iter().next().unwrap_or(&0);
+        let max_value = *values_sorted.iter().last().unwrap_or(&0);
+
+        let compact_space =
+            get_compact_space(&values_sorted, total_num_values, COST_PER_BLANK_IN_BITS);
+        let amplitude_compact_space = compact_space.amplitude_compact_space();
+
+        assert!(
+            amplitude_compact_space <= u64::MAX as u128,
+            "case unsupported."
+        );
+
+        let num_bits = tantivy_bitpacker::compute_num_bits(amplitude_compact_space as u64);
+
+        assert_eq!(
+            compact_space
+                .u128_to_compact(max_value)
+                .expect("could not convert max value to compact space"),
+            amplitude_compact_space as u32
+        );
+        CompactSpaceCompressor {
+            params: IPCodecParams {
+                compact_space,
+                bit_unpacker: BitUnpacker::new(num_bits),
+                min_value,
+                max_value,
+                num_vals: total_num_values,
+                num_bits,
+            },
+        }
+    }
+
+    fn write_footer(self, writer: &mut impl Write) -> io::Result<()> {
+        let writer = &mut CountingWriter::wrap(writer);
+        self.params.serialize(writer)?;
+
+        let footer_len = writer.written_bytes() as u32;
+        footer_len.serialize(writer)?;
+
+        Ok(())
+    }
+
+    pub fn compress_into(
+        self,
+        vals: impl Iterator<Item = u128>,
+        write: &mut impl Write,
+    ) -> io::Result<()> {
+        let mut bitpacker = BitPacker::default();
+        for val in vals {
+            let compact = self
+                .params
+                .compact_space
+                .u128_to_compact(val)
+                .map_err(|_| {
+                    io::Error::new(
+                        io::ErrorKind::InvalidData,
+                        "Could not convert value to compact_space. This is a bug.",
+                    )
+                })?;
+            bitpacker.write(compact as u64, self.params.num_bits, write)?;
+        }
+        bitpacker.close(write)?;
+        self.write_footer(write)?;
+        Ok(())
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct CompactSpaceDecompressor {
+    data: OwnedBytes,
+    params: IPCodecParams,
+}
+
+impl BinarySerializable for IPCodecParams {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        // header flags for future optional dictionary encoding
+        let footer_flags = 0u64;
+        footer_flags.serialize(writer)?;
+
+        VIntU128(self.min_value).serialize(writer)?;
+        VIntU128(self.max_value).serialize(writer)?;
+        VIntU128(self.num_vals as u128).serialize(writer)?;
+        self.num_bits.serialize(writer)?;
+
+        self.compact_space.serialize(writer)?;
+
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let _header_flags = u64::deserialize(reader)?;
+        let min_value = VIntU128::deserialize(reader)?.0;
+        let max_value = VIntU128::deserialize(reader)?.0;
+        let num_vals = VIntU128::deserialize(reader)?.0 as u32;
+        let num_bits = u8::deserialize(reader)?;
+        let compact_space = CompactSpace::deserialize(reader)?;
+
+        Ok(Self {
+            compact_space,
+            bit_unpacker: BitUnpacker::new(num_bits),
+            min_value,
+            max_value,
+            num_vals,
+            num_bits,
+        })
+    }
+}
+
+/// Exposes the compact space compressed values as u64.
+///
+/// This allows faster access to the values, as u64 is faster to work with than u128.
+/// It also allows to handle u128 values like u64, via the `open_u64_lenient` as a uniform
+/// access interface.
+///
+/// When converting from the internal u64 to u128 `compact_to_u128` can be used.
+pub struct CompactSpaceU64Accessor(CompactSpaceDecompressor);
+impl CompactSpaceU64Accessor {
+    pub(crate) fn open(data: OwnedBytes) -> io::Result<CompactSpaceU64Accessor> {
+        let decompressor = CompactSpaceU64Accessor(CompactSpaceDecompressor::open(data)?);
+        Ok(decompressor)
+    }
+    /// Convert a compact space value to u128
+    pub fn compact_to_u128(&self, compact: u32) -> u128 {
+        self.0.compact_to_u128(compact)
+    }
+}
+
+impl ColumnValues<u64> for CompactSpaceU64Accessor {
+    #[inline]
+    fn get_val(&self, doc: u32) -> u64 {
+        let compact = self.0.get_compact(doc);
+        compact as u64
+    }
+
+    fn min_value(&self) -> u64 {
+        self.0.u128_to_compact(self.0.min_value()).unwrap() as u64
+    }
+
+    fn max_value(&self) -> u64 {
+        self.0.u128_to_compact(self.0.max_value()).unwrap() as u64
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.0.params.num_vals
+    }
+
+    #[inline]
+    fn iter(&self) -> Box<dyn Iterator<Item = u64> + '_> {
+        Box::new(self.0.iter_compact().map(|el| el as u64))
+    }
+
+    #[inline]
+    fn get_row_ids_for_value_range(
+        &self,
+        value_range: RangeInclusive<u64>,
+        position_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        let value_range = self.0.compact_to_u128(*value_range.start() as u32)
+            ..=self.0.compact_to_u128(*value_range.end() as u32);
+        self.0
+            .get_row_ids_for_value_range(value_range, position_range, positions)
+    }
+}
+
+impl ColumnValues<u128> for CompactSpaceDecompressor {
+    #[inline]
+    fn get_val(&self, doc: u32) -> u128 {
+        self.get(doc)
+    }
+
+    fn min_value(&self) -> u128 {
+        self.min_value()
+    }
+
+    fn max_value(&self) -> u128 {
+        self.max_value()
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.params.num_vals
+    }
+
+    #[inline]
+    fn iter(&self) -> Box<dyn Iterator<Item = u128> + '_> {
+        Box::new(self.iter())
+    }
+
+    #[inline]
+    fn get_row_ids_for_value_range(
+        &self,
+        value_range: RangeInclusive<u128>,
+        position_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        if value_range.start() > value_range.end() {
+            return;
+        }
+        let position_range = position_range.start..position_range.end.min(self.num_vals());
+        let from_value = *value_range.start();
+        let to_value = *value_range.end();
+        assert!(to_value >= from_value);
+        let compact_from = self.u128_to_compact(from_value);
+        let compact_to = self.u128_to_compact(to_value);
+
+        // Quick return, if both ranges fall into the same non-mapped space, the range can't cover
+        // any values, so we can early exit
+        match (compact_to, compact_from) {
+            (Err(pos1), Err(pos2)) if pos1 == pos2 => return,
+            _ => {}
+        }
+
+        let compact_from = compact_from.unwrap_or_else(|pos| {
+            // Correctness: Out of bounds, if this value is Err(last_index + 1), we early exit,
+            // since the to_value also mapps into the same non-mapped space
+            let range_mapping = self.params.compact_space.get_range_mapping(pos);
+            range_mapping.compact_start
+        });
+        // If there is no compact space, we go to the closest upperbound compact space
+        let compact_to = compact_to.unwrap_or_else(|pos| {
+            // Correctness: Overflow, if this value is Err(0), we early exit,
+            // since the from_value also mapps into the same non-mapped space
+
+            // Get end of previous range
+            let pos = pos - 1;
+            let range_mapping = self.params.compact_space.get_range_mapping(pos);
+            range_mapping.compact_end()
+        });
+
+        let value_range = compact_from..=compact_to;
+        self.get_positions_for_compact_value_range(value_range, position_range, positions);
+    }
+}
+
+impl CompactSpaceDecompressor {
+    pub fn open(data: OwnedBytes) -> io::Result<CompactSpaceDecompressor> {
+        let (data_slice, footer_len_bytes) = data.split_at(data.len() - 4);
+        let footer_len = u32::deserialize(&mut &footer_len_bytes[..])?;
+
+        let data_footer = &data_slice[data_slice.len() - footer_len as usize..];
+        let params = IPCodecParams::deserialize(&mut &data_footer[..])?;
+        let decompressor = CompactSpaceDecompressor { data, params };
+
+        Ok(decompressor)
+    }
+
+    /// Converting to compact space for the decompressor is more complex, since we may get values
+    /// which are outside the compact space. e.g. if we map
+    /// 1000 => 5
+    /// 2000 => 6
+    ///
+    /// and we want a mapping for 1005, there is no equivalent compact space. We instead return an
+    /// error with the index of the next range.
+    fn u128_to_compact(&self, value: u128) -> Result<u32, usize> {
+        self.params.compact_space.u128_to_compact(value)
+    }
+
+    fn compact_to_u128(&self, compact: u32) -> u128 {
+        self.params.compact_space.compact_to_u128(compact)
+    }
+
+    #[inline]
+    fn iter_compact(&self) -> impl Iterator<Item = u32> + '_ {
+        (0..self.params.num_vals)
+            .map(move |idx| self.params.bit_unpacker.get(idx, &self.data) as u32)
+    }
+
+    #[inline]
+    fn iter(&self) -> impl Iterator<Item = u128> + '_ {
+        // TODO: Performance. It would be better to iterate on the ranges and check existence via
+        // the bit_unpacker.
+        self.iter_compact()
+            .map(|compact| self.compact_to_u128(compact))
+    }
+
+    #[inline]
+    pub fn get_compact(&self, idx: u32) -> u32 {
+        self.params.bit_unpacker.get(idx, &self.data) as u32
+    }
+
+    #[inline]
+    pub fn get(&self, idx: u32) -> u128 {
+        let compact = self.get_compact(idx);
+        self.compact_to_u128(compact)
+    }
+
+    pub fn min_value(&self) -> u128 {
+        self.params.min_value
+    }
+
+    pub fn max_value(&self) -> u128 {
+        self.params.max_value
+    }
+
+    fn get_positions_for_compact_value_range(
+        &self,
+        value_range: RangeInclusive<u32>,
+        position_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        self.params.bit_unpacker.get_ids_for_value_range(
+            *value_range.start() as u64..=*value_range.end() as u64,
+            position_range,
+            &self.data,
+            positions,
+        );
+    }
+}
+
+#[cfg(test)]
+mod tests {
+
+    use itertools::Itertools;
+
+    use super::*;
+    use crate::column_values::u128_based::U128Header;
+    use crate::column_values::{open_u128_mapped, serialize_column_values_u128};
+
+    #[test]
+    fn compact_space_test() {
+        let ips: BTreeSet<u128> = [
+            2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
+        ]
+        .into_iter()
+        .collect();
+        let compact_space = get_compact_space(&ips, ips.len() as u32, 11);
+        let amplitude = compact_space.amplitude_compact_space();
+        assert_eq!(amplitude, 17);
+        assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
+        assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
+        assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
+
+        for (num1, num2) in (0..3).tuple_windows() {
+            assert_eq!(
+                compact_space.get_range_mapping(num1).compact_end() + 1,
+                compact_space.get_range_mapping(num2).compact_start
+            );
+        }
+
+        let mut output: Vec<u8> = Vec::new();
+        compact_space.serialize(&mut output).unwrap();
+
+        assert_eq!(
+            compact_space,
+            CompactSpace::deserialize(&mut &output[..]).unwrap()
+        );
+
+        for ip in ips {
+            let compact = compact_space.u128_to_compact(ip).unwrap();
+            assert_eq!(compact_space.compact_to_u128(compact), ip);
+        }
+    }
+
+    #[test]
+    fn compact_space_amplitude_test() {
+        let ips = &[100000u128, 1000000].into_iter().collect();
+        let compact_space = get_compact_space(ips, ips.len() as u32, 1);
+        let amplitude = compact_space.amplitude_compact_space();
+        assert_eq!(amplitude, 2);
+    }
+
+    fn test_all(mut data: OwnedBytes, expected: &[u128]) {
+        let _header = U128Header::deserialize(&mut data);
+        let decompressor = CompactSpaceDecompressor::open(data).unwrap();
+        for (idx, expected_val) in expected.iter().cloned().enumerate() {
+            let val = decompressor.get(idx as u32);
+            assert_eq!(val, expected_val);
+
+            let test_range = |range: RangeInclusive<u128>| {
+                let expected_positions = expected
+                    .iter()
+                    .positions(|val| range.contains(val))
+                    .map(|pos| pos as u32)
+                    .collect::<Vec<_>>();
+                let mut positions = Vec::new();
+                decompressor.get_row_ids_for_value_range(
+                    range,
+                    0..decompressor.num_vals(),
+                    &mut positions,
+                );
+                assert_eq!(positions, expected_positions);
+            };
+
+            test_range(expected_val.saturating_sub(1)..=expected_val);
+            test_range(expected_val..=expected_val);
+            test_range(expected_val..=expected_val.saturating_add(1));
+            test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
+        }
+    }
+
+    fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
+        let mut out = Vec::new();
+        serialize_column_values_u128(&u128_vals, &mut out).unwrap();
+        let data = OwnedBytes::new(out);
+        test_all(data.clone(), u128_vals);
+        data
+    }
+
+    #[test]
+    fn test_range_1() {
+        let vals = &[
+            1u128,
+            100u128,
+            3u128,
+            99999u128,
+            100000u128,
+            100001u128,
+            4_000_211_221u128,
+            4_000_211_222u128,
+            333u128,
+        ];
+        let mut data = test_aux_vals(vals);
+
+        let _header = U128Header::deserialize(&mut data);
+        let decomp = CompactSpaceDecompressor::open(data).unwrap();
+        let complete_range = 0..vals.len() as u32;
+        for (pos, val) in vals.iter().enumerate() {
+            let val = *val;
+            let pos = pos as u32;
+            let mut positions = Vec::new();
+            decomp.get_row_ids_for_value_range(val..=val, pos..pos + 1, &mut positions);
+            assert_eq!(positions, vec![pos]);
+        }
+
+        // handle docid range out of bounds
+        let positions: Vec<u32> = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
+        assert!(positions.is_empty());
+
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
+        assert_eq!(positions, vec![0]);
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
+        assert_eq!(positions, vec![0]);
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
+        assert_eq!(positions, vec![0, 2]);
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99999u128..=99999u128,
+                complete_range.clone()
+            ),
+            vec![3]
+        );
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99999u128..=100000u128,
+                complete_range.clone()
+            ),
+            vec![3, 4]
+        );
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99998u128..=100000u128,
+                complete_range.clone()
+            ),
+            vec![3, 4]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                99998u128..=99999u128,
+                complete_range.clone()
+            ),
+            &[3]
+        );
+        assert!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99998u128..=99998u128,
+                complete_range.clone()
+            )
+            .is_empty()
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                333u128..=333u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                332u128..=333u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                332u128..=334u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                333u128..=334u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                4_000_211_221u128..=5_000_000_000u128,
+                complete_range
+            ),
+            &[6, 7]
+        );
+    }
+
+    #[test]
+    fn test_empty() {
+        let vals = &[];
+        let data = test_aux_vals(vals);
+        let _decomp = CompactSpaceDecompressor::open(data).unwrap();
+    }
+
+    #[test]
+    fn test_range_2() {
+        let vals = &[
+            100u128,
+            99999u128,
+            100000u128,
+            100001u128,
+            4_000_211_221u128,
+            4_000_211_222u128,
+            333u128,
+        ];
+        let mut data = test_aux_vals(vals);
+        let _header = U128Header::deserialize(&mut data);
+        let decomp = CompactSpaceDecompressor::open(data).unwrap();
+        let complete_range = 0..vals.len() as u32;
+        assert!(
+            &get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone())
+                .is_empty(),
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
+            &[0]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
+            &[0]
+        );
+    }
+
+    fn get_positions_for_value_range_helper<C: ColumnValues<T> + ?Sized, T: PartialOrd>(
+        column: &C,
+        value_range: RangeInclusive<T>,
+        doc_id_range: Range<u32>,
+    ) -> Vec<u32> {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(value_range, doc_id_range, &mut positions);
+        positions
+    }
+
+    #[test]
+    fn test_range_3() {
+        let vals = &[
+            200u128,
+            201,
+            202,
+            203,
+            204,
+            204,
+            206,
+            207,
+            208,
+            209,
+            210,
+            1_000_000,
+            5_000_000_000,
+        ];
+        let mut out = Vec::new();
+        serialize_column_values_u128(&&vals[..], &mut out).unwrap();
+        let decomp = open_u128_mapped(OwnedBytes::new(out)).unwrap();
+        let complete_range = 0..vals.len() as u32;
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
+            vec![0]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
+            vec![0, 1]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
+            vec![0]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
+            vec![11]
+        );
+    }
+
+    #[test]
+    fn test_bug1() {
+        let vals = &[9223372036854775806];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug2() {
+        let vals = &[340282366920938463463374607431768211455u128];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug3() {
+        let vals = &[340282366920938463463374607431768211454];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug4() {
+        let vals = &[340282366920938463463374607431768211455, 0];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_first_large_gaps() {
+        let vals = &[1_000_000_000u128; 100];
+        let _data = test_aux_vals(vals);
+    }
+
+    use proptest::prelude::*;
+
+    fn num_strategy() -> impl Strategy<Value = u128> {
+        prop_oneof![
+            1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
+            20 => prop::num::u128::ANY,
+        ]
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(10))]
+
+        #[test]
+        fn compress_decompress_random(vals in proptest::collection::vec(num_strategy() , 1..1000)) {
+            let _data = test_aux_vals(&vals);
+        }
+    }
+}

columnar/src/column_values/u128_based/mod.rs

@@ -0,0 +1,197 @@
+use std::fmt::Debug;
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+mod compact_space;
+
+use common::{BinarySerializable, OwnedBytes, VInt};
+pub use compact_space::{
+    CompactSpaceCompressor, CompactSpaceDecompressor, CompactSpaceU64Accessor,
+};
+
+use crate::column_values::monotonic_map_column;
+use crate::column_values::monotonic_mapping::{
+    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+};
+use crate::iterable::Iterable;
+use crate::{ColumnValues, MonotonicallyMappableToU128};
+
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+pub(crate) struct U128Header {
+    pub num_vals: u32,
+    pub codec_type: U128FastFieldCodecType,
+}
+
+impl BinarySerializable for U128Header {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.num_vals as u64).serialize(writer)?;
+        self.codec_type.serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let num_vals = VInt::deserialize(reader)?.0 as u32;
+        let codec_type = U128FastFieldCodecType::deserialize(reader)?;
+        Ok(U128Header {
+            num_vals,
+            codec_type,
+        })
+    }
+}
+
+/// Serializes u128 values with the compact space codec.
+pub fn serialize_column_values_u128<T: MonotonicallyMappableToU128>(
+    iterable: &dyn Iterable<T>,
+    output: &mut impl io::Write,
+) -> io::Result<()> {
+    let compressor = CompactSpaceCompressor::train_from(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+    );
+    let header = U128Header {
+        num_vals: compressor.num_vals(),
+        codec_type: U128FastFieldCodecType::CompactSpace,
+    };
+    header.serialize(output)?;
+    compressor.compress_into(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+        output,
+    )?;
+    Ok(())
+}
+
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+/// Available codecs to use to encode the u128 (via [`MonotonicallyMappableToU128`]) converted data.
+pub(crate) enum U128FastFieldCodecType {
+    /// This codec takes a large number space (u128) and reduces it to a compact number space, by
+    /// removing the holes.
+    CompactSpace = 1,
+}
+
+impl BinarySerializable for U128FastFieldCodecType {
+    fn serialize<W: Write + ?Sized>(&self, wrt: &mut W) -> io::Result<()> {
+        self.to_code().serialize(wrt)
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let code = u8::deserialize(reader)?;
+        let codec_type: Self = Self::from_code(code)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
+        Ok(codec_type)
+    }
+}
+
+impl U128FastFieldCodecType {
+    pub(crate) fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub(crate) fn from_code(code: u8) -> Option<Self> {
+        match code {
+            1 => Some(Self::CompactSpace),
+            _ => None,
+        }
+    }
+}
+
+/// Returns the correct codec reader wrapped in the `Arc` for the data.
+pub fn open_u128_mapped<T: MonotonicallyMappableToU128 + Debug>(
+    mut bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let header = U128Header::deserialize(&mut bytes)?;
+    assert_eq!(header.codec_type, U128FastFieldCodecType::CompactSpace);
+    let reader = CompactSpaceDecompressor::open(bytes)?;
+    let inverted: StrictlyMonotonicMappingInverter<StrictlyMonotonicMappingToInternal<T>> =
+        StrictlyMonotonicMappingToInternal::<T>::new().into();
+    Ok(Arc::new(monotonic_map_column(reader, inverted)))
+}
+
+/// Returns the u64 representation of the u128 data.
+/// The internal representation of the data as u64 is useful for faster processing.
+///
+/// In order to convert to u128 back cast to `CompactSpaceU64Accessor` and call
+/// `compact_to_u128`.
+///
+/// # Notice
+/// In case there are new codecs added, check for usages of `CompactSpaceDecompressorU64` and
+/// also handle the new codecs.
+pub fn open_u128_as_compact_u64(mut bytes: OwnedBytes) -> io::Result<Arc<dyn ColumnValues<u64>>> {
+    let header = U128Header::deserialize(&mut bytes)?;
+    assert_eq!(header.codec_type, U128FastFieldCodecType::CompactSpace);
+    let reader = CompactSpaceU64Accessor::open(bytes)?;
+    Ok(Arc::new(reader))
+}
+
+#[cfg(test)]
+pub(crate) mod tests {
+    use super::*;
+    use crate::column_values::CodecType;
+    use crate::column_values::u64_based::{
+        ALL_U64_CODEC_TYPES, serialize_and_load_u64_based_column_values,
+        serialize_u64_based_column_values,
+    };
+
+    #[test]
+    fn test_serialize_deserialize_u128_header() {
+        let original = U128Header {
+            num_vals: 11,
+            codec_type: U128FastFieldCodecType::CompactSpace,
+        };
+        let mut out = Vec::new();
+        original.serialize(&mut out).unwrap();
+        let restored = U128Header::deserialize(&mut &out[..]).unwrap();
+        assert_eq!(restored, original);
+    }
+
+    #[test]
+    fn test_serialize_deserialize() {
+        let original = [1u64, 5u64, 10u64];
+        let restored: Vec<u64> =
+            serialize_and_load_u64_based_column_values(&&original[..], &ALL_U64_CODEC_TYPES)
+                .iter()
+                .collect();
+        assert_eq!(&restored, &original[..]);
+    }
+
+    #[test]
+    fn test_fastfield_bool_size_bitwidth_1() {
+        let mut buffer = Vec::new();
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_CODEC_TYPES,
+            &mut buffer,
+        )
+        .unwrap();
+        // TODO put the header as a footer so that it serves as a padding.
+        // 5 bytes of header, 1 byte of value, 7 bytes of padding.
+        assert_eq!(buffer.len(), 5 + 1);
+    }
+
+    #[test]
+    fn test_fastfield_bool_bit_size_bitwidth_0() {
+        let mut buffer = Vec::new();
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_CODEC_TYPES,
+            &mut buffer,
+        )
+        .unwrap();
+        // 6 bytes of header, 0 bytes of value, 7 bytes of padding.
+        assert_eq!(buffer.len(), 6);
+    }
+
+    #[test]
+    fn test_fastfield_gcd() {
+        let mut buffer = Vec::new();
+        let vals: Vec<u64> = (0..80).map(|val| (val % 7) * 1_000u64).collect();
+        serialize_u64_based_column_values(&&vals[..], &[CodecType::Bitpacked], &mut buffer)
+            .unwrap();
+        // Values are stored over 3 bits.
+        assert_eq!(buffer.len(), 6 + (3 * 80 / 8));
+    }
+}

columnar/src/column_values/u64_based/bitpacked.rs

@@ -0,0 +1,184 @@
+use std::io::{self, Write};
+use std::num::NonZeroU64;
+use std::ops::{Range, RangeInclusive};
+
+use common::{BinarySerializable, OwnedBytes};
+use fastdivide::DividerU64;
+use tantivy_bitpacker::{BitPacker, BitUnpacker, compute_num_bits};
+
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
+use crate::{ColumnValues, RowId};
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct BitpackedReader {
+    data: OwnedBytes,
+    bit_unpacker: BitUnpacker,
+    stats: ColumnStats,
+}
+
+#[inline(always)]
+const fn div_ceil(n: u64, q: NonZeroU64) -> u64 {
+    // copied from unstable rust standard library.
+    let d = n / q.get();
+    let r = n % q.get();
+    if r > 0 { d + 1 } else { d }
+}
+
+// The bitpacked codec applies a linear transformation `f` over data that are bitpacked.
+// f is defined by:
+// f: bitpacked -> stats.min_value + stats.gcd * bitpacked
+//
+// In order to run range queries, we invert the transformation.
+// `transform_range_before_linear_transformation` returns the range of values
+// [min_bipacked_value..max_bitpacked_value] such that
+// f(bitpacked) ∈ [min_value, max_value] <=> bitpacked ∈ [min_bitpacked_value, max_bitpacked_value]
+fn transform_range_before_linear_transformation(
+    stats: &ColumnStats,
+    range: RangeInclusive<u64>,
+) -> Option<RangeInclusive<u64>> {
+    if range.is_empty() {
+        return None;
+    }
+    if stats.min_value > *range.end() {
+        return None;
+    }
+    if stats.max_value < *range.start() {
+        return None;
+    }
+    let shifted_range =
+        range.start().saturating_sub(stats.min_value)..=range.end().saturating_sub(stats.min_value);
+    let start_before_gcd_multiplication: u64 = div_ceil(*shifted_range.start(), stats.gcd);
+    let end_before_gcd_multiplication: u64 = *shifted_range.end() / stats.gcd;
+    Some(start_before_gcd_multiplication..=end_before_gcd_multiplication)
+}
+
+impl ColumnValues for BitpackedReader {
+    #[inline(always)]
+    fn get_val(&self, doc: u32) -> u64 {
+        self.stats.min_value + self.stats.gcd.get() * self.bit_unpacker.get(doc, &self.data)
+    }
+    #[inline]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+    #[inline]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+    #[inline]
+    fn num_vals(&self) -> RowId {
+        self.stats.num_rows
+    }
+
+    fn get_row_ids_for_value_range(
+        &self,
+        range: RangeInclusive<u64>,
+        doc_id_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        let Some(transformed_range) =
+            transform_range_before_linear_transformation(&self.stats, range)
+        else {
+            positions.clear();
+            return;
+        };
+        self.bit_unpacker.get_ids_for_value_range(
+            transformed_range,
+            doc_id_range,
+            &self.data,
+            positions,
+        );
+    }
+}
+
+fn num_bits(stats: &ColumnStats) -> u8 {
+    compute_num_bits(stats.amplitude() / stats.gcd)
+}
+
+#[derive(Default)]
+pub struct BitpackedCodecEstimator;
+
+impl ColumnCodecEstimator for BitpackedCodecEstimator {
+    fn collect(&mut self, _value: u64) {}
+
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64> {
+        let num_bits_per_value = num_bits(stats);
+        Some(stats.num_bytes() + (stats.num_rows as u64 * (num_bits_per_value as u64) + 7) / 8)
+    }
+
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let num_bits = num_bits(stats);
+        let mut bit_packer = BitPacker::new();
+        let divider = DividerU64::divide_by(stats.gcd.get());
+        for val in vals {
+            bit_packer.write(divider.divide(val - stats.min_value), num_bits, wrt)?;
+        }
+        bit_packer.close(wrt)?;
+        Ok(())
+    }
+}
+
+pub struct BitpackedCodec;
+
+impl ColumnCodec for BitpackedCodec {
+    type ColumnValues = BitpackedReader;
+    type Estimator = BitpackedCodecEstimator;
+
+    /// Opens a fast field given a file.
+    fn load(mut data: OwnedBytes) -> io::Result<Self::ColumnValues> {
+        let stats = ColumnStats::deserialize(&mut data)?;
+        let num_bits = num_bits(&stats);
+        let bit_unpacker = BitUnpacker::new(num_bits);
+        Ok(BitpackedReader {
+            data,
+            bit_unpacker,
+            stats,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u64_based::tests::create_and_validate;
+
+    #[test]
+    fn test_with_codec_data_sets_simple() {
+        create_and_validate::<BitpackedCodec>(&[4, 3, 12], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets_simple_gcd() {
+        create_and_validate::<BitpackedCodec>(&[1000, 2000, 3000], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::u64_based::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<BitpackedCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, name);
+        }
+    }
+
+    #[test]
+    fn bitpacked_fast_field_rand() {
+        for _ in 0..500 {
+            let mut data = (0..1 + rand::random::<u8>() as usize)
+                .map(|_| rand::random::<i64>() as u64 / 2)
+                .collect::<Vec<_>>();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/u64_based/blockwise_linear.rs

@@ -0,0 +1,284 @@
+use std::io::Write;
+use std::sync::Arc;
+use std::{io, iter};
+
+use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
+use fastdivide::DividerU64;
+use tantivy_bitpacker::{BitPacker, BitUnpacker, compute_num_bits};
+
+use crate::MonotonicallyMappableToU64;
+use crate::column_values::u64_based::line::Line;
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
+use crate::column_values::{ColumnValues, VecColumn};
+
+const BLOCK_SIZE: u32 = 512u32;
+
+#[derive(Debug, Default)]
+struct Block {
+    line: Line,
+    bit_unpacker: BitUnpacker,
+    data_start_offset: usize,
+}
+
+impl BinarySerializable for Block {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        self.line.serialize(writer)?;
+        self.bit_unpacker.bit_width().serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let line = Line::deserialize(reader)?;
+        let bit_width = u8::deserialize(reader)?;
+        Ok(Block {
+            line,
+            bit_unpacker: BitUnpacker::new(bit_width),
+            data_start_offset: 0,
+        })
+    }
+}
+
+fn compute_num_blocks(num_vals: u32) -> u32 {
+    num_vals.div_ceil(BLOCK_SIZE)
+}
+
+pub struct BlockwiseLinearEstimator {
+    block: Vec<u64>,
+    values_num_bytes: u64,
+    meta_num_bytes: u64,
+}
+
+impl Default for BlockwiseLinearEstimator {
+    fn default() -> Self {
+        Self {
+            block: Vec::with_capacity(BLOCK_SIZE as usize),
+            values_num_bytes: 0u64,
+            meta_num_bytes: 0u64,
+        }
+    }
+}
+
+impl BlockwiseLinearEstimator {
+    fn flush_block_estimate(&mut self) {
+        if self.block.is_empty() {
+            return;
+        }
+        let column = VecColumn::from(std::mem::take(&mut self.block));
+        let line = Line::train(&column);
+        self.block = column.into();
+
+        let mut max_value = 0u64;
+        for (i, buffer_val) in self.block.iter().enumerate() {
+            let interpolated_val = line.eval(i as u32);
+            let val = buffer_val.wrapping_sub(interpolated_val);
+            max_value = val.max(max_value);
+        }
+        let bit_width = compute_num_bits(max_value) as usize;
+        self.values_num_bytes += (bit_width * self.block.len() + 7) as u64 / 8;
+        self.meta_num_bytes += 1 + line.num_bytes();
+    }
+}
+
+impl ColumnCodecEstimator for BlockwiseLinearEstimator {
+    fn collect(&mut self, value: u64) {
+        self.block.push(value);
+        if self.block.len() == BLOCK_SIZE as usize {
+            self.flush_block_estimate();
+            self.block.clear();
+        }
+    }
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64> {
+        let mut estimate = 4 + stats.num_bytes() + self.meta_num_bytes + self.values_num_bytes;
+        if stats.gcd.get() > 1 {
+            let estimate_gain_from_gcd =
+                (stats.gcd.get() as f32).log2().floor() * stats.num_rows as f32 / 8.0f32;
+            estimate = estimate.saturating_sub(estimate_gain_from_gcd as u64);
+        }
+        Some(estimate)
+    }
+
+    fn finalize(&mut self) {
+        self.flush_block_estimate();
+    }
+
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        mut vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let mut buffer = Vec::with_capacity(BLOCK_SIZE as usize);
+        let num_blocks = compute_num_blocks(stats.num_rows) as usize;
+        let mut blocks = Vec::with_capacity(num_blocks);
+
+        let mut bit_packer = BitPacker::new();
+
+        let gcd_divider = DividerU64::divide_by(stats.gcd.get());
+
+        for _ in 0..num_blocks {
+            buffer.clear();
+            buffer.extend(
+                (&mut vals)
+                    .map(MonotonicallyMappableToU64::to_u64)
+                    .take(BLOCK_SIZE as usize),
+            );
+
+            for buffer_val in buffer.iter_mut() {
+                *buffer_val = gcd_divider.divide(*buffer_val - stats.min_value);
+            }
+
+            let line = Line::train(&VecColumn::from(buffer.to_vec()));
+
+            assert!(!buffer.is_empty());
+
+            for (i, buffer_val) in buffer.iter_mut().enumerate() {
+                let interpolated_val = line.eval(i as u32);
+                *buffer_val = buffer_val.wrapping_sub(interpolated_val);
+            }
+
+            let bit_width = buffer.iter().copied().map(compute_num_bits).max().unwrap();
+
+            for &buffer_val in &buffer {
+                bit_packer.write(buffer_val, bit_width, wrt)?;
+            }
+
+            blocks.push(Block {
+                line,
+                bit_unpacker: BitUnpacker::new(bit_width),
+                data_start_offset: 0,
+            });
+        }
+
+        bit_packer.close(wrt)?;
+
+        assert_eq!(blocks.len(), num_blocks);
+
+        let mut counting_wrt = CountingWriter::wrap(wrt);
+        for block in &blocks {
+            block.serialize(&mut counting_wrt)?;
+        }
+        let footer_len = counting_wrt.written_bytes();
+        (footer_len as u32).serialize(&mut counting_wrt)?;
+
+        Ok(())
+    }
+}
+
+pub struct BlockwiseLinearCodec;
+
+impl ColumnCodec<u64> for BlockwiseLinearCodec {
+    type ColumnValues = BlockwiseLinearReader;
+
+    type Estimator = BlockwiseLinearEstimator;
+
+    fn load(mut bytes: OwnedBytes) -> io::Result<Self::ColumnValues> {
+        let stats = ColumnStats::deserialize(&mut bytes)?;
+        let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
+        let footer_offset = bytes.len() - 4 - footer_len as usize;
+        let (data, mut footer) = bytes.split(footer_offset);
+        let num_blocks = compute_num_blocks(stats.num_rows);
+        let mut blocks: Vec<Block> = iter::repeat_with(|| Block::deserialize(&mut footer))
+            .take(num_blocks as usize)
+            .collect::<io::Result<_>>()?;
+        let mut start_offset = 0;
+        for block in &mut blocks {
+            block.data_start_offset = start_offset;
+            start_offset += (block.bit_unpacker.bit_width() as usize) * BLOCK_SIZE as usize / 8;
+        }
+        Ok(BlockwiseLinearReader {
+            blocks: blocks.into_boxed_slice().into(),
+            data,
+            stats,
+        })
+    }
+}
+
+#[derive(Clone)]
+pub struct BlockwiseLinearReader {
+    blocks: Arc<[Block]>,
+    data: OwnedBytes,
+    stats: ColumnStats,
+}
+
+impl ColumnValues for BlockwiseLinearReader {
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> u64 {
+        let block_id = (idx / BLOCK_SIZE) as usize;
+        let idx_within_block = idx % BLOCK_SIZE;
+        let block = &self.blocks[block_id];
+        let interpoled_val: u64 = block.line.eval(idx_within_block);
+        let block_bytes = &self.data[block.data_start_offset..];
+        let bitpacked_diff = block.bit_unpacker.get(idx_within_block, block_bytes);
+        // TODO optimize me! the line parameters could be tweaked to include the multiplication and
+        // remove the dependency.
+        self.stats.min_value
+            + self
+                .stats
+                .gcd
+                .get()
+                .wrapping_mul(interpoled_val.wrapping_add(bitpacked_diff))
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+
+    #[inline(always)]
+    fn num_vals(&self) -> u32 {
+        self.stats.num_rows
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u64_based::tests::create_and_validate;
+
+    #[test]
+    fn test_with_codec_data_sets_simple() {
+        create_and_validate::<BlockwiseLinearCodec>(
+            &[11, 20, 40, 20, 10, 10, 10, 10, 10, 10],
+            "simple test",
+        )
+        .unwrap();
+    }
+
+    #[test]
+    fn test_with_codec_data_sets_simple_gcd() {
+        let (_, actual_compression_rate) = create_and_validate::<BlockwiseLinearCodec>(
+            &[10, 20, 40, 20, 10, 10, 10, 10, 10, 10],
+            "name",
+        )
+        .unwrap();
+        assert_eq!(actual_compression_rate, 0.175);
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::u64_based::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<BlockwiseLinearCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<BlockwiseLinearCodec>(&data, name);
+        }
+    }
+
+    #[test]
+    fn test_blockwise_linear_fast_field_rand() {
+        for _ in 0..500 {
+            let mut data = (0..1 + rand::random::<u8>() as usize)
+                .map(|_| rand::random::<i64>() as u64 / 2)
+                .collect::<Vec<_>>();
+            create_and_validate::<BlockwiseLinearCodec>(&data, "rand");
+            data.reverse();
+            create_and_validate::<BlockwiseLinearCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/u64_based/line.rs

@@ -0,0 +1,209 @@
+use std::io;
+use std::num::NonZeroU32;
+
+use common::{BinarySerializable, VInt};
+
+use crate::column_values::ColumnValues;
+
+const MID_POINT: u64 = (1u64 << 32) - 1u64;
+
+/// `Line` describes a line function `y: ax + b` using integer
+/// arithmetics.
+///
+/// The slope is in fact a decimal split into a 32 bit integer value,
+/// and a 32-bit decimal value.
+///
+/// The multiplication then becomes.
+/// `y = m * x >> 32 + b`
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Line {
+    pub(crate) slope: u64,
+    pub(crate) intercept: u64,
+}
+
+/// Compute the line slope.
+///
+/// This function has the nice property of being
+/// invariant by translation.
+/// `
+///   compute_slope(y0, y1)
+/// = compute_slope(y0 + X % 2^64, y1 + X % 2^64)
+/// `
+fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU32) -> u64 {
+    let dy = y1.wrapping_sub(y0);
+    let sign = dy <= (1 << 63);
+    let abs_dy = if sign {
+        y1.wrapping_sub(y0)
+    } else {
+        y0.wrapping_sub(y1)
+    };
+    if abs_dy >= 1 << 32 {
+        // This is outside of realm we handle.
+        // Let's just bail.
+        return 0u64;
+    }
+
+    let abs_slope = (abs_dy << 32) / num_vals.get() as u64;
+    if sign {
+        abs_slope
+    } else {
+        // The complement does indeed create the
+        // opposite decreasing slope...
+        //
+        // Intuitively (without the bitshifts and % u64::MAX)
+        // ```
+        //    (x + shift)*(u64::MAX - abs_slope)
+        // -  (x * (u64::MAX - abs_slope))
+        // = - shift * abs_slope
+        // ```
+        u64::MAX - abs_slope
+    }
+}
+
+impl Line {
+    #[inline(always)]
+    pub fn eval(&self, x: u32) -> u64 {
+        let linear_part = ((x as u64).wrapping_mul(self.slope) >> 32) as i32 as u64;
+        self.intercept.wrapping_add(linear_part)
+    }
+
+    // Intercept is only computed from provided positions
+    pub fn train_from(
+        first_val: u64,
+        last_val: u64,
+        num_vals: u32,
+        positions_and_values: impl Iterator<Item = (u64, u64)>,
+    ) -> Self {
+        // TODO replace with let else
+        let idx_last_val = if let Some(idx_last_val) = NonZeroU32::new(num_vals - 1) {
+            idx_last_val
+        } else {
+            return Line::default();
+        };
+
+        let y0 = first_val;
+        let y1 = last_val;
+
+        // We first independently pick our slope.
+        let slope = compute_slope(y0, y1, idx_last_val);
+
+        // We picked our slope. Note that it does not have to be perfect.
+        // Now we need to compute the best intercept.
+        //
+        // Intuitively, the best intercept is such that line passes through one of the
+        // `(i, ys[])`.
+        //
+        // The best intercept therefore has the form
+        // `y[i] - line.eval(i)` (using wrapping arithmetics).
+        // In other words, the best intercept is one of the `y - Line::eval(ys[i])`
+        // and our task is just to pick the one that minimizes our error.
+        //
+        // Without sorting our values, this is a difficult problem.
+        // We however rely on the following trick...
+        //
+        // We only focus on the case where the interpolation is half decent.
+        // If the line interpolation is doing its job on a dataset suited for it,
+        // we can hope that the maximum error won't be larger than `u64::MAX / 2`.
+        //
+        // In other words, even without the intercept the values `y - Line::eval(ys[i])` will all be
+        // within an interval that takes less than half of the modulo space of `u64`.
+        //
+        // Our task is therefore to identify this interval.
+        // Here we simply translate all of our values by `y0 - 2^63` and pick the min.
+        let mut line = Line {
+            slope,
+            intercept: 0,
+        };
+        let heuristic_shift = y0.wrapping_sub(MID_POINT);
+        line.intercept = positions_and_values
+            .map(|(pos, y)| y.wrapping_sub(line.eval(pos as u32)))
+            .min_by_key(|&val| val.wrapping_sub(heuristic_shift))
+            .unwrap_or(0u64); //< Never happens.
+        line
+    }
+
+    /// Returns a line that attempts to approximate a function
+    /// f: i in 0..[ys.num_vals()) -> ys[i].
+    ///
+    /// - The approximation is always lower than the actual value. Or more rigorously, formally
+    ///   `f(i).wrapping_sub(ys[i])` is small for any i in [0..ys.len()).
+    /// - It computes without panicking for any value of it.
+    ///
+    /// This function is only invariable by translation if all of the
+    /// `ys` are packaged into half of the space. (See heuristic below)
+    /// TODO USE array
+    pub fn train(ys: &dyn ColumnValues) -> Self {
+        let first_val = ys.iter().next().unwrap();
+        let last_val = ys.iter().nth(ys.num_vals() as usize - 1).unwrap();
+        Self::train_from(
+            first_val,
+            last_val,
+            ys.num_vals(),
+            ys.iter().enumerate().map(|(pos, val)| (pos as u64, val)),
+        )
+    }
+}
+
+impl BinarySerializable for Line {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.slope).serialize(writer)?;
+        VInt(self.intercept).serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let slope = VInt::deserialize(reader)?.0;
+        let intercept = VInt::deserialize(reader)?.0;
+        Ok(Line { slope, intercept })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::VecColumn;
+
+    /// Test training a line and ensuring that the maximum difference between
+    /// the data points and the line is `expected`.
+    ///
+    /// This function operates translation over the data for better coverage.
+    #[track_caller]
+    fn test_line_interpol_with_translation(ys: &[u64], expected: Option<u64>) {
+        let mut translations = vec![0, 100, u64::MAX / 2, u64::MAX, u64::MAX - 1];
+        translations.extend_from_slice(ys);
+        for translation in translations {
+            let translated_ys: Vec<u64> = ys
+                .iter()
+                .copied()
+                .map(|y| y.wrapping_add(translation))
+                .collect();
+            let largest_err = test_eval_max_err(&translated_ys);
+            assert_eq!(largest_err, expected);
+        }
+    }
+
+    fn test_eval_max_err(ys: &[u64]) -> Option<u64> {
+        let line = Line::train(&VecColumn::from(ys.to_vec()));
+        ys.iter()
+            .enumerate()
+            .map(|(x, y)| y.wrapping_sub(line.eval(x as u32)))
+            .max()
+    }
+
+    #[test]
+    fn test_train() {
+        test_line_interpol_with_translation(&[11, 11, 11, 12, 12, 13], Some(1));
+        test_line_interpol_with_translation(&[13, 12, 12, 11, 11, 11], Some(1));
+        test_line_interpol_with_translation(&[13, 13, 12, 11, 11, 11], Some(1));
+        test_line_interpol_with_translation(&[13, 13, 12, 11, 11, 11], Some(1));
+        test_line_interpol_with_translation(&[u64::MAX - 1, 0, 0, 1], Some(1));
+        test_line_interpol_with_translation(&[u64::MAX - 1, u64::MAX, 0, 1], Some(0));
+        test_line_interpol_with_translation(&[0, 1, 2, 3, 5], Some(0));
+        test_line_interpol_with_translation(&[1, 2, 3, 4], Some(0));
+
+        let data: Vec<u64> = (0..255).collect();
+        test_line_interpol_with_translation(&data, Some(0));
+        let data: Vec<u64> = (0..255).map(|el| el * 2).collect();
+        test_line_interpol_with_translation(&data, Some(0));
+    }
+}

columnar/src/column_values/u64_based/linear.rs

@@ -0,0 +1,279 @@
+use std::io;
+
+use common::{BinarySerializable, OwnedBytes};
+use tantivy_bitpacker::{BitPacker, BitUnpacker, compute_num_bits};
+
+use super::ColumnValues;
+use super::line::Line;
+use crate::RowId;
+use crate::column_values::VecColumn;
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
+
+const HALF_SPACE: u64 = u64::MAX / 2;
+const LINE_ESTIMATION_BLOCK_LEN: usize = 512;
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct LinearReader {
+    data: OwnedBytes,
+    linear_params: LinearParams,
+    stats: ColumnStats,
+}
+
+impl ColumnValues for LinearReader {
+    #[inline]
+    fn get_val(&self, doc: u32) -> u64 {
+        let interpoled_val: u64 = self.linear_params.line.eval(doc);
+        let bitpacked_diff = self.linear_params.bit_unpacker.get(doc, &self.data);
+        interpoled_val.wrapping_add(bitpacked_diff)
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+
+    #[inline]
+    fn num_vals(&self) -> u32 {
+        self.stats.num_rows
+    }
+}
+
+/// Fastfield serializer, which tries to guess values by linear interpolation
+/// and stores the difference bitpacked.
+pub struct LinearCodec;
+
+#[derive(Debug, Clone)]
+struct LinearParams {
+    line: Line,
+    bit_unpacker: BitUnpacker,
+}
+
+impl BinarySerializable for LinearParams {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        self.line.serialize(writer)?;
+        self.bit_unpacker.bit_width().serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let line = Line::deserialize(reader)?;
+        let bit_width = u8::deserialize(reader)?;
+        Ok(Self {
+            line,
+            bit_unpacker: BitUnpacker::new(bit_width),
+        })
+    }
+}
+
+pub struct LinearCodecEstimator {
+    block: Vec<u64>,
+    line: Option<Line>,
+    row_id: RowId,
+    min_deviation: u64,
+    max_deviation: u64,
+    first_val: u64,
+    last_val: u64,
+}
+
+impl Default for LinearCodecEstimator {
+    fn default() -> LinearCodecEstimator {
+        LinearCodecEstimator {
+            block: Vec::with_capacity(LINE_ESTIMATION_BLOCK_LEN),
+            line: None,
+            row_id: 0,
+            min_deviation: u64::MAX,
+            max_deviation: u64::MIN,
+            first_val: 0u64,
+            last_val: 0u64,
+        }
+    }
+}
+
+impl ColumnCodecEstimator for LinearCodecEstimator {
+    fn finalize(&mut self) {
+        if let Some(line) = self.line.as_mut() {
+            line.intercept = line
+                .intercept
+                .wrapping_add(self.min_deviation)
+                .wrapping_sub(HALF_SPACE);
+        }
+    }
+
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64> {
+        let line = self.line?;
+        let amplitude = self.max_deviation - self.min_deviation;
+        let num_bits = compute_num_bits(amplitude);
+        let linear_params = LinearParams {
+            line,
+            bit_unpacker: BitUnpacker::new(num_bits),
+        };
+        Some(
+            stats.num_bytes()
+                + linear_params.num_bytes()
+                + (num_bits as u64 * stats.num_rows as u64 + 7) / 8,
+        )
+    }
+
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let line = self.line.unwrap();
+        let amplitude = self.max_deviation - self.min_deviation;
+        let num_bits = compute_num_bits(amplitude);
+        let linear_params = LinearParams {
+            line,
+            bit_unpacker: BitUnpacker::new(num_bits),
+        };
+        linear_params.serialize(wrt)?;
+        let mut bit_packer = BitPacker::new();
+        for (pos, value) in vals.enumerate() {
+            let calculated_value = line.eval(pos as u32);
+            let offset = value.wrapping_sub(calculated_value);
+            bit_packer.write(offset, num_bits, wrt)?;
+        }
+        bit_packer.close(wrt)?;
+        Ok(())
+    }
+
+    fn collect(&mut self, value: u64) {
+        if let Some(line) = self.line {
+            self.collect_after_line_estimation(&line, value);
+        } else {
+            self.collect_before_line_estimation(value);
+        }
+    }
+}
+
+impl LinearCodecEstimator {
+    #[inline]
+    fn collect_after_line_estimation(&mut self, line: &Line, value: u64) {
+        let interpoled_val: u64 = line.eval(self.row_id);
+        let deviation = value.wrapping_add(HALF_SPACE).wrapping_sub(interpoled_val);
+        self.min_deviation = self.min_deviation.min(deviation);
+        self.max_deviation = self.max_deviation.max(deviation);
+        if self.row_id == 0 {
+            self.first_val = value;
+        }
+        self.last_val = value;
+        self.row_id += 1u32;
+    }
+
+    #[inline]
+    fn collect_before_line_estimation(&mut self, value: u64) {
+        self.block.push(value);
+        if self.block.len() == LINE_ESTIMATION_BLOCK_LEN {
+            let column = VecColumn::from(std::mem::take(&mut self.block));
+            let line = Line::train(&column);
+            self.block = column.into();
+            let block = std::mem::take(&mut self.block);
+            for val in block {
+                self.collect_after_line_estimation(&line, val);
+            }
+            self.line = Some(line);
+        }
+    }
+}
+
+impl ColumnCodec for LinearCodec {
+    type ColumnValues = LinearReader;
+
+    type Estimator = LinearCodecEstimator;
+
+    fn load(mut data: OwnedBytes) -> io::Result<Self::ColumnValues> {
+        let stats = ColumnStats::deserialize(&mut data)?;
+        let linear_params = LinearParams::deserialize(&mut data)?;
+        Ok(LinearReader {
+            stats,
+            linear_params,
+            data,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use rand::RngCore;
+
+    use super::*;
+    use crate::column_values::u64_based::tests::{create_and_validate, get_codec_test_datasets};
+
+    #[test]
+    fn test_compression_simple() {
+        let vals = (100u64..)
+            .take(super::LINE_ESTIMATION_BLOCK_LEN)
+            .collect::<Vec<_>>();
+        create_and_validate::<LinearCodec>(&vals, "simple monotonically large").unwrap();
+    }
+
+    #[test]
+    fn test_compression() {
+        let data = (10..=6_000_u64).collect::<Vec<_>>();
+        let (estimate, actual_compression) =
+            create_and_validate::<LinearCodec>(&data, "simple monotonically large").unwrap();
+        assert_le!(actual_compression, 0.001);
+        assert_le!(estimate, 0.02);
+    }
+
+    #[test]
+    fn test_with_codec_datasets() {
+        let data_sets = get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<LinearCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<LinearCodec>(&data, name);
+        }
+    }
+    #[test]
+    fn linear_interpol_fast_field_test_large_amplitude() {
+        let data = vec![
+            i64::MAX as u64 / 2,
+            i64::MAX as u64 / 3,
+            i64::MAX as u64 / 2,
+        ];
+        create_and_validate::<LinearCodec>(&data, "large amplitude");
+    }
+
+    #[test]
+    fn overflow_error_test() {
+        let data = vec![1572656989877777, 1170935903116329, 720575940379279, 0];
+        create_and_validate::<LinearCodec>(&data, "overflow test");
+    }
+
+    #[test]
+    fn linear_interpol_fast_concave_data() {
+        let data = vec![0, 1, 2, 5, 8, 10, 20, 50];
+        create_and_validate::<LinearCodec>(&data, "concave data");
+    }
+    #[test]
+    fn linear_interpol_fast_convex_data() {
+        let data = vec![0, 40, 60, 70, 75, 77];
+        create_and_validate::<LinearCodec>(&data, "convex data");
+    }
+    #[test]
+    fn linear_interpol_fast_field_test_simple() {
+        let data = (10..=20_u64).collect::<Vec<_>>();
+        create_and_validate::<LinearCodec>(&data, "simple monotonically");
+    }
+
+    #[test]
+    fn linear_interpol_fast_field_rand() {
+        let mut rng = rand::thread_rng();
+        for _ in 0..50 {
+            let mut data = (0..10_000).map(|_| rng.next_u64()).collect::<Vec<_>>();
+            create_and_validate::<LinearCodec>(&data, "random");
+            data.reverse();
+            create_and_validate::<LinearCodec>(&data, "random");
+        }
+    }
+}

columnar/src/column_values/u64_based/mod.rs

@@ -0,0 +1,214 @@
+mod bitpacked;
+mod blockwise_linear;
+mod line;
+mod linear;
+mod stats_collector;
+
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::{BinarySerializable, OwnedBytes};
+
+use crate::column_values::monotonic_mapping::{
+    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+};
+pub use crate::column_values::u64_based::bitpacked::BitpackedCodec;
+pub use crate::column_values::u64_based::blockwise_linear::BlockwiseLinearCodec;
+pub use crate::column_values::u64_based::linear::LinearCodec;
+pub use crate::column_values::u64_based::stats_collector::StatsCollector;
+use crate::column_values::{ColumnStats, monotonic_map_column};
+use crate::iterable::Iterable;
+use crate::{ColumnValues, MonotonicallyMappableToU64};
+
+/// A `ColumnCodecEstimator` is in charge of gathering all
+/// data required to serialize a column.
+///
+/// This happens during a first pass on data of the column elements.
+/// During that pass, all column estimators receive a call to their
+/// `.collect(el)`.
+///
+/// After this first pass, finalize is called.
+/// `.estimate(..)` then should return an accurate estimation of the
+/// size of the serialized column (were we to pick this codec.).
+/// `.serialize(..)` then serializes the column using this codec.
+pub trait ColumnCodecEstimator<T = u64>: 'static {
+    /// Records a new value for estimation.
+    /// This method will be called for each element of the column during
+    /// `estimation`.
+    fn collect(&mut self, value: u64);
+    /// Finalizes the first pass phase.
+    fn finalize(&mut self) {}
+    /// Returns an accurate estimation of the number of bytes that will
+    /// be used to represent this column.
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64>;
+    /// Serializes the column using the given codec.
+    /// This constitutes a second pass over the columns values.
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        vals: &mut dyn Iterator<Item = T>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()>;
+}
+
+/// A column codec describes a colunm serialization format.
+pub trait ColumnCodec<T: PartialOrd = u64> {
+    /// Specialized `ColumnValues` type.
+    type ColumnValues: ColumnValues<T> + 'static;
+    /// `Estimator` for the given codec.
+    type Estimator: ColumnCodecEstimator + Default;
+
+    /// Loads a column that has been serialized using this codec.
+    fn load(bytes: OwnedBytes) -> io::Result<Self::ColumnValues>;
+
+    /// Returns an estimator.
+    fn estimator() -> Self::Estimator {
+        Self::Estimator::default()
+    }
+
+    /// Returns a boxed estimator.
+    fn boxed_estimator() -> Box<dyn ColumnCodecEstimator> {
+        Box::new(Self::estimator())
+    }
+}
+
+/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+pub enum CodecType {
+    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
+    /// `column.max_value() - column.min_value()`
+    Bitpacked = 0u8,
+    /// Linear interpolation puts a line between the first and last value and then bitpacks the
+    /// values by the offset from the line. The number of bits is defined by the max deviation from
+    /// the line.
+    Linear = 1u8,
+    /// Same as [`CodecType::Linear`], but encodes in blocks of 512 elements.
+    BlockwiseLinear = 2u8,
+}
+
+/// List of all available u64-base codecs.
+pub const ALL_U64_CODEC_TYPES: [CodecType; 3] = [
+    CodecType::Bitpacked,
+    CodecType::Linear,
+    CodecType::BlockwiseLinear,
+];
+
+impl CodecType {
+    fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    fn try_from_code(code: u8) -> Option<CodecType> {
+        match code {
+            0u8 => Some(CodecType::Bitpacked),
+            1u8 => Some(CodecType::Linear),
+            2u8 => Some(CodecType::BlockwiseLinear),
+            _ => None,
+        }
+    }
+
+    fn load<T: MonotonicallyMappableToU64>(
+        &self,
+        bytes: OwnedBytes,
+    ) -> io::Result<Arc<dyn ColumnValues<T>>> {
+        match self {
+            CodecType::Bitpacked => load_specific_codec::<BitpackedCodec, T>(bytes),
+            CodecType::Linear => load_specific_codec::<LinearCodec, T>(bytes),
+            CodecType::BlockwiseLinear => load_specific_codec::<BlockwiseLinearCodec, T>(bytes),
+        }
+    }
+}
+
+fn load_specific_codec<C: ColumnCodec, T: MonotonicallyMappableToU64>(
+    bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let reader = C::load(bytes)?;
+    let reader_typed = monotonic_map_column(
+        reader,
+        StrictlyMonotonicMappingInverter::from(StrictlyMonotonicMappingToInternal::<T>::new()),
+    );
+    Ok(Arc::new(reader_typed))
+}
+
+impl CodecType {
+    /// Returns a boxed codec estimator associated to a given `CodecType`.
+    pub fn estimator(&self) -> Box<dyn ColumnCodecEstimator> {
+        match self {
+            CodecType::Bitpacked => BitpackedCodec::boxed_estimator(),
+            CodecType::Linear => LinearCodec::boxed_estimator(),
+            CodecType::BlockwiseLinear => BlockwiseLinearCodec::boxed_estimator(),
+        }
+    }
+}
+
+/// Serializes a given column of u64-mapped values.
+pub fn serialize_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable<T>,
+    codec_types: &[CodecType],
+    wrt: &mut dyn Write,
+) -> io::Result<()> {
+    let mut stats_collector = StatsCollector::default();
+    let mut estimators: Vec<(CodecType, Box<dyn ColumnCodecEstimator>)> =
+        Vec::with_capacity(codec_types.len());
+    for &codec_type in codec_types {
+        estimators.push((codec_type, codec_type.estimator()));
+    }
+    for val in vals.boxed_iter() {
+        let val_u64 = val.to_u64();
+        stats_collector.collect(val_u64);
+        for (_, estimator) in &mut estimators {
+            estimator.collect(val_u64);
+        }
+    }
+    for (_, estimator) in &mut estimators {
+        estimator.finalize();
+    }
+    let stats = stats_collector.stats();
+    let (_, best_codec, best_codec_estimator) = estimators
+        .into_iter()
+        .flat_map(|(codec_type, estimator)| {
+            let num_bytes = estimator.estimate(&stats)?;
+            Some((num_bytes, codec_type, estimator))
+        })
+        .min_by_key(|(num_bytes, _, _)| *num_bytes)
+        .ok_or_else(|| {
+            io::Error::new(io::ErrorKind::InvalidData, "No available applicable codec.")
+        })?;
+    best_codec.to_code().serialize(wrt)?;
+    best_codec_estimator.serialize(
+        &stats,
+        &mut vals.boxed_iter().map(MonotonicallyMappableToU64::to_u64),
+        wrt,
+    )?;
+    Ok(())
+}
+
+/// Load u64-based column values.
+///
+/// This method first identifies the codec off the first byte.
+pub fn load_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    mut bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let codec_type: CodecType = bytes
+        .first()
+        .copied()
+        .and_then(CodecType::try_from_code)
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Failed to read codec type"))?;
+    bytes.advance(1);
+    codec_type.load(bytes)
+}
+
+/// Helper function to serialize a column (autodetect from all codecs) and then open it
+pub fn serialize_and_load_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable,
+    codec_types: &[CodecType],
+) -> Arc<dyn ColumnValues<T>> {
+    let mut buffer = Vec::new();
+    serialize_u64_based_column_values(vals, codec_types, &mut buffer).unwrap();
+    load_u64_based_column_values::<T>(OwnedBytes::new(buffer)).unwrap()
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/column_values/u64_based/stats_collector.rs

@@ -0,0 +1,200 @@
+use std::num::NonZeroU64;
+
+use fastdivide::DividerU64;
+
+use crate::RowId;
+use crate::column_values::ColumnStats;
+
+/// Compute the gcd of two non null numbers.
+///
+/// It is recommended, but not required, to feed values such that `large >= small`.
+fn compute_gcd(mut large: NonZeroU64, mut small: NonZeroU64) -> NonZeroU64 {
+    loop {
+        let rem: u64 = large.get() % small;
+        if let Some(new_small) = NonZeroU64::new(rem) {
+            (large, small) = (small, new_small);
+        } else {
+            return small;
+        }
+    }
+}
+
+#[derive(Default)]
+pub struct StatsCollector {
+    min_max_opt: Option<(u64, u64)>,
+    num_rows: RowId,
+    // We measure the GCD of the difference between the values and the minimal value.
+    // This is the same as computing the difference between the values and the first value.
+    //
+    // This way, we can compress i64-converted-to-u64 (e.g. timestamp that were supplied in
+    // seconds, only to be converted in nanoseconds).
+    increment_gcd_opt: Option<(NonZeroU64, DividerU64)>,
+    first_value_opt: Option<u64>,
+}
+
+impl StatsCollector {
+    pub fn stats(&self) -> ColumnStats {
+        let (min_value, max_value) = self.min_max_opt.unwrap_or((0u64, 0u64));
+        let increment_gcd = if let Some((increment_gcd, _)) = self.increment_gcd_opt {
+            increment_gcd
+        } else {
+            NonZeroU64::new(1u64).unwrap()
+        };
+        ColumnStats {
+            min_value,
+            max_value,
+            num_rows: self.num_rows,
+            gcd: increment_gcd,
+        }
+    }
+
+    #[inline]
+    fn update_increment_gcd(&mut self, value: u64) {
+        let Some(first_value) = self.first_value_opt else {
+            // We set the first value and just quit.
+            self.first_value_opt = Some(value);
+            return;
+        };
+        let Some(non_zero_value) = NonZeroU64::new(value.abs_diff(first_value)) else {
+            // We can simply skip 0 values.
+            return;
+        };
+        let Some((gcd, gcd_divider)) = self.increment_gcd_opt else {
+            self.set_increment_gcd(non_zero_value);
+            return;
+        };
+        if gcd.get() == 1 {
+            // It won't see any update now.
+            return;
+        }
+        let remainder =
+            non_zero_value.get() - (gcd_divider.divide(non_zero_value.get())) * gcd.get();
+        if remainder == 0 {
+            return;
+        }
+        let new_gcd = compute_gcd(non_zero_value, gcd);
+        self.set_increment_gcd(new_gcd);
+    }
+
+    fn set_increment_gcd(&mut self, gcd: NonZeroU64) {
+        let new_divider = DividerU64::divide_by(gcd.get());
+        self.increment_gcd_opt = Some((gcd, new_divider));
+    }
+
+    pub fn collect(&mut self, value: u64) {
+        self.min_max_opt = Some(if let Some((min, max)) = self.min_max_opt {
+            (min.min(value), max.max(value))
+        } else {
+            (value, value)
+        });
+        self.num_rows += 1;
+        self.update_increment_gcd(value);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use crate::column_values::u64_based::ColumnStats;
+    use crate::column_values::u64_based::stats_collector::{StatsCollector, compute_gcd};
+
+    fn compute_stats(vals: impl Iterator<Item = u64>) -> ColumnStats {
+        let mut stats_collector = StatsCollector::default();
+        for val in vals {
+            stats_collector.collect(val);
+        }
+        stats_collector.stats()
+    }
+
+    fn find_gcd(vals: impl Iterator<Item = u64>) -> u64 {
+        compute_stats(vals).gcd.get()
+    }
+
+    #[test]
+    fn test_compute_gcd() {
+        let test_compute_gcd_aux = |large, small, expected| {
+            let large = NonZeroU64::new(large).unwrap();
+            let small = NonZeroU64::new(small).unwrap();
+            let expected = NonZeroU64::new(expected).unwrap();
+            assert_eq!(compute_gcd(small, large), expected);
+            assert_eq!(compute_gcd(large, small), expected);
+        };
+        test_compute_gcd_aux(1, 4, 1);
+        test_compute_gcd_aux(2, 4, 2);
+        test_compute_gcd_aux(10, 25, 5);
+        test_compute_gcd_aux(25, 25, 25);
+    }
+
+    #[test]
+    fn test_gcd() {
+        assert_eq!(find_gcd([0].into_iter()), 1);
+        assert_eq!(find_gcd([0, 10].into_iter()), 10);
+        assert_eq!(find_gcd([10, 0].into_iter()), 10);
+        assert_eq!(find_gcd([].into_iter()), 1);
+        assert_eq!(find_gcd([15, 30, 5, 10].into_iter()), 5);
+        assert_eq!(find_gcd([15, 16, 10].into_iter()), 1);
+        assert_eq!(find_gcd([0, 5, 5, 5].into_iter()), 5);
+        assert_eq!(find_gcd([0, 0].into_iter()), 1);
+        assert_eq!(find_gcd([1, 10, 4, 1, 7, 10].into_iter()), 3);
+        assert_eq!(find_gcd([1, 10, 0, 4, 1, 7, 10].into_iter()), 1);
+    }
+
+    #[test]
+    fn test_stats() {
+        assert_eq!(
+            compute_stats([].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 0,
+                num_rows: 0
+            }
+        );
+        assert_eq!(
+            compute_stats([0, 1].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 1,
+                num_rows: 2
+            }
+        );
+        assert_eq!(
+            compute_stats([0, 1].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 1,
+                num_rows: 2
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 20, 30].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(10).unwrap(),
+                min_value: 10,
+                max_value: 30,
+                num_rows: 3
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 50, 10, 30].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(20).unwrap(),
+                min_value: 10,
+                max_value: 50,
+                num_rows: 4
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 0, 30].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(10).unwrap(),
+                min_value: 0,
+                max_value: 30,
+                num_rows: 3
+            }
+        );
+    }
+}

columnar/src/column_values/u64_based/tests.rs

@@ -0,0 +1,415 @@
+use proptest::prelude::*;
+use proptest::{prop_oneof, proptest};
+use rand::Rng;
+
+#[test]
+fn test_serialize_and_load_simple() {
+    let mut buffer = Vec::new();
+    let vals = &[1u64, 2u64, 5u64];
+    serialize_u64_based_column_values(
+        &&vals[..],
+        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
+        &mut buffer,
+    )
+    .unwrap();
+    assert_eq!(buffer.len(), 7);
+    let col = load_u64_based_column_values::<u64>(OwnedBytes::new(buffer)).unwrap();
+    assert_eq!(col.num_vals(), 3);
+    assert_eq!(col.get_val(0), 1);
+    assert_eq!(col.get_val(1), 2);
+    assert_eq!(col.get_val(2), 5);
+}
+
+#[test]
+fn test_empty_column_i64() {
+    let vals: [i64; 0] = [];
+    let mut num_acceptable_codecs = 0;
+    for codec in ALL_U64_CODEC_TYPES {
+        let mut buffer = Vec::new();
+        if serialize_u64_based_column_values(&&vals[..], &[codec], &mut buffer).is_err() {
+            continue;
+        }
+        num_acceptable_codecs += 1;
+        let col = load_u64_based_column_values::<i64>(OwnedBytes::new(buffer)).unwrap();
+        assert_eq!(col.num_vals(), 0);
+        assert_eq!(col.min_value(), i64::MIN);
+        assert_eq!(col.max_value(), i64::MIN);
+    }
+    assert!(num_acceptable_codecs > 0);
+}
+
+#[test]
+fn test_empty_column_u64() {
+    let vals: [u64; 0] = [];
+    let mut num_acceptable_codecs = 0;
+    for codec in ALL_U64_CODEC_TYPES {
+        let mut buffer = Vec::new();
+        if serialize_u64_based_column_values(&&vals[..], &[codec], &mut buffer).is_err() {
+            continue;
+        }
+        num_acceptable_codecs += 1;
+        let col = load_u64_based_column_values::<u64>(OwnedBytes::new(buffer)).unwrap();
+        assert_eq!(col.num_vals(), 0);
+        assert_eq!(col.min_value(), u64::MIN);
+        assert_eq!(col.max_value(), u64::MIN);
+    }
+    assert!(num_acceptable_codecs > 0);
+}
+
+#[test]
+fn test_empty_column_f64() {
+    let vals: [f64; 0] = [];
+    let mut num_acceptable_codecs = 0;
+    for codec in ALL_U64_CODEC_TYPES {
+        let mut buffer = Vec::new();
+        if serialize_u64_based_column_values(&&vals[..], &[codec], &mut buffer).is_err() {
+            continue;
+        }
+        num_acceptable_codecs += 1;
+        let col = load_u64_based_column_values::<f64>(OwnedBytes::new(buffer)).unwrap();
+        assert_eq!(col.num_vals(), 0);
+        // FIXME. f64::MIN would be better!
+        assert!(col.min_value().is_nan());
+        assert!(col.max_value().is_nan());
+    }
+    assert!(num_acceptable_codecs > 0);
+}
+
+pub(crate) fn create_and_validate<TColumnCodec: ColumnCodec>(
+    vals: &[u64],
+    name: &str,
+) -> Option<(f32, f32)> {
+    let mut stats_collector = StatsCollector::default();
+    let mut codec_estimator: TColumnCodec::Estimator = Default::default();
+
+    for val in vals.boxed_iter() {
+        stats_collector.collect(val);
+        codec_estimator.collect(val);
+    }
+    codec_estimator.finalize();
+    let stats = stats_collector.stats();
+    let estimation = codec_estimator.estimate(&stats)?;
+
+    let mut buffer = Vec::new();
+    codec_estimator
+        .serialize(&stats, vals.boxed_iter().as_mut(), &mut buffer)
+        .unwrap();
+
+    let actual_compression = buffer.len() as u64;
+
+    let reader = TColumnCodec::load(OwnedBytes::new(buffer)).unwrap();
+    assert_eq!(reader.num_vals(), vals.len() as u32);
+    let mut buffer = Vec::new();
+    for (doc, orig_val) in vals.iter().copied().enumerate() {
+        let val = reader.get_val(doc as u32);
+        assert_eq!(
+            val, orig_val,
+            "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data `{vals:?}`",
+        );
+
+        buffer.resize(1, 0);
+        reader.get_vals(&[doc as u32], &mut buffer);
+        let val = buffer[0];
+        assert_eq!(
+            val, orig_val,
+            "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data `{vals:?}`",
+        );
+    }
+
+    let all_docs: Vec<u32> = (0..vals.len() as u32).collect();
+    buffer.resize(all_docs.len(), 0);
+    reader.get_vals(&all_docs, &mut buffer);
+    assert_eq!(vals, buffer);
+
+    if !vals.is_empty() {
+        let test_rand_idx = rand::thread_rng().gen_range(0..=vals.len() - 1);
+        let expected_positions: Vec<u32> = vals
+            .iter()
+            .enumerate()
+            .filter(|(_, el)| **el == vals[test_rand_idx])
+            .map(|(pos, _)| pos as u32)
+            .collect();
+        let mut positions = Vec::new();
+        reader.get_row_ids_for_value_range(
+            vals[test_rand_idx]..=vals[test_rand_idx],
+            0..vals.len() as u32,
+            &mut positions,
+        );
+        assert_eq!(expected_positions, positions);
+    }
+    if actual_compression > 1000 {
+        assert!(relative_difference(estimation, actual_compression) < 0.10f32);
+    }
+    Some((
+        compression_rate(estimation, stats.num_rows),
+        compression_rate(actual_compression, stats.num_rows),
+    ))
+}
+
+fn compression_rate(num_bytes: u64, num_values: u32) -> f32 {
+    num_bytes as f32 / (num_values as f32 * 8.0)
+}
+
+fn relative_difference(left: u64, right: u64) -> f32 {
+    let left = left as f32;
+    let right = right as f32;
+    2.0f32 * (left - right).abs() / (left + right)
+}
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(100))]
+
+    #[test]
+    fn test_proptest_small_bitpacked(data in proptest::collection::vec(num_strategy(), 1..10)) {
+        create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
+    }
+
+    #[test]
+    fn test_proptest_small_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
+        create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
+    }
+
+
+    #[test]
+    fn test_proptest_small_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
+        create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
+    }
+}
+
+#[test]
+fn test_small_blockwise_linear_example() {
+    create_and_validate::<BlockwiseLinearCodec>(
+        &[9223372036854775808, 9223370937344622593],
+        "proptest multilinearinterpol",
+    );
+}
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(10))]
+
+    #[test]
+    fn test_proptest_large_bitpacked(data in proptest::collection::vec(num_strategy(), 1..6000)) {
+        create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
+    }
+
+    #[test]
+    fn test_proptest_large_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
+        create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
+    }
+
+    #[test]
+    fn test_proptest_large_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
+        create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
+    }
+}
+
+fn num_strategy() -> impl Strategy<Value = u64> {
+    prop_oneof![
+        1 => prop::num::u64::ANY.prop_map(|num| u64::MAX - (num % 10) ),
+        1 => prop::num::u64::ANY.prop_map(|num| num % 10 ),
+        20 => prop::num::u64::ANY,
+    ]
+}
+
+pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
+    let mut data_and_names = vec![];
+
+    let data = (10..=10_000_u64).collect::<Vec<_>>();
+    data_and_names.push((data, "simple monotonically increasing"));
+
+    data_and_names.push((
+        vec![5, 6, 7, 8, 9, 10, 99, 100],
+        "offset in linear interpol",
+    ));
+    data_and_names.push((vec![5, 50, 3, 13, 1, 1000, 35], "rand small"));
+    data_and_names.push((vec![10], "single value"));
+
+    data_and_names.push((
+        vec![1572656989877777, 1170935903116329, 720575940379279, 0],
+        "overflow error",
+    ));
+
+    data_and_names
+}
+
+fn test_codec<C: ColumnCodec>() {
+    let codec_name = std::any::type_name::<C>();
+    for (data, dataset_name) in get_codec_test_datasets() {
+        let estimate_actual_opt: Option<(f32, f32)> =
+            tests::create_and_validate::<C>(&data, dataset_name);
+        let result = if let Some((estimate, actual)) = estimate_actual_opt {
+            format!("Estimate `{estimate}` Actual `{actual}`")
+        } else {
+            "Disabled".to_string()
+        };
+        println!("Codec {codec_name}, DataSet {dataset_name}, {result}");
+    }
+}
+#[test]
+fn test_codec_bitpacking() {
+    test_codec::<BitpackedCodec>();
+}
+#[test]
+fn test_codec_interpolation() {
+    test_codec::<LinearCodec>();
+}
+#[test]
+fn test_codec_multi_interpolation() {
+    test_codec::<BlockwiseLinearCodec>();
+}
+
+use super::*;
+
+fn estimate<C: ColumnCodec>(vals: &[u64]) -> Option<f32> {
+    let mut stats_collector = StatsCollector::default();
+    let mut estimator = C::Estimator::default();
+    for &val in vals {
+        stats_collector.collect(val);
+        estimator.collect(val);
+    }
+    estimator.finalize();
+    let stats = stats_collector.stats();
+    let num_bytes = estimator.estimate(&stats)?;
+    if stats.num_rows == 0 {
+        return None;
+    }
+    Some(num_bytes as f32 / (8.0 * stats.num_rows as f32))
+}
+
+#[test]
+fn estimation_good_interpolation_case() {
+    let data = (10..=20000_u64).collect::<Vec<_>>();
+
+    let linear_interpol_estimation = estimate::<LinearCodec>(&data).unwrap();
+    assert_le!(linear_interpol_estimation, 0.01);
+
+    let multi_linear_interpol_estimation = estimate::<BlockwiseLinearCodec>(&data).unwrap();
+    assert_le!(multi_linear_interpol_estimation, 0.2);
+    assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
+
+    let bitpacked_estimation = estimate::<BitpackedCodec>(&data).unwrap();
+    assert_lt!(linear_interpol_estimation, bitpacked_estimation);
+}
+
+#[test]
+fn estimation_test_bad_interpolation_case_monotonically_increasing() {
+    let mut data: Vec<u64> = (201..=20000_u64).collect();
+    data.push(1_000_000);
+
+    // in this case the linear interpolation can't in fact not be worse than bitpacking,
+    // but the estimator adds some threshold, which leads to estimated worse behavior
+    let linear_interpol_estimation = estimate::<LinearCodec>(&data[..]).unwrap();
+    assert_le!(linear_interpol_estimation, 0.35);
+
+    let bitpacked_estimation = estimate::<BitpackedCodec>(&data).unwrap();
+    assert_le!(bitpacked_estimation, 0.32);
+    assert_le!(bitpacked_estimation, linear_interpol_estimation);
+}
+
+#[test]
+fn test_fast_field_codec_type_to_code() {
+    let mut count_codec = 0;
+    for code in 0..=255 {
+        if let Some(codec_type) = CodecType::try_from_code(code) {
+            assert_eq!(codec_type.to_code(), code);
+            count_codec += 1;
+        }
+    }
+    assert_eq!(count_codec, 3);
+}
+
+fn test_fastfield_gcd_i64_with_codec(codec_type: CodecType, num_vals: usize) -> io::Result<()> {
+    let mut vals: Vec<i64> = (-4..=(num_vals as i64) - 5).map(|val| val * 1000).collect();
+    let mut buffer: Vec<u8> = Vec::new();
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
+        &[codec_type],
+        &mut buffer,
+    )?;
+    let buffer = OwnedBytes::new(buffer);
+    let column = crate::column_values::load_u64_based_column_values::<i64>(buffer.clone())?;
+    assert_eq!(column.get_val(0), -4000i64);
+    assert_eq!(column.get_val(1), -3000i64);
+    assert_eq!(column.get_val(2), -2000i64);
+    assert_eq!(column.max_value(), (num_vals as i64 - 5) * 1000);
+    assert_eq!(column.min_value(), -4000i64);
+
+    // Can't apply gcd
+    let mut buffer_without_gcd = Vec::new();
+    vals.pop();
+    vals.push(1001i64);
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
+        &[codec_type],
+        &mut buffer_without_gcd,
+    )?;
+    let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
+    assert!(buffer_without_gcd.len() > buffer.len());
+
+    Ok(())
+}
+
+#[test]
+fn test_fastfield_gcd_i64() -> io::Result<()> {
+    for &codec_type in &[
+        CodecType::Bitpacked,
+        CodecType::BlockwiseLinear,
+        CodecType::Linear,
+    ] {
+        test_fastfield_gcd_i64_with_codec(codec_type, 5500)?;
+    }
+    Ok(())
+}
+
+fn test_fastfield_gcd_u64_with_codec(codec_type: CodecType, num_vals: usize) -> io::Result<()> {
+    let mut vals: Vec<u64> = (1..=num_vals).map(|i| i as u64 * 1000u64).collect();
+    let mut buffer: Vec<u8> = Vec::new();
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
+        &[codec_type],
+        &mut buffer,
+    )?;
+    let buffer = OwnedBytes::new(buffer);
+    let column = crate::column_values::load_u64_based_column_values::<u64>(buffer.clone())?;
+    assert_eq!(column.get_val(0), 1000u64);
+    assert_eq!(column.get_val(1), 2000u64);
+    assert_eq!(column.get_val(2), 3000u64);
+    assert_eq!(column.max_value(), num_vals as u64 * 1000);
+    assert_eq!(column.min_value(), 1000u64);
+
+    // Can't apply gcd
+    let mut buffer_without_gcd = Vec::new();
+    vals.pop();
+    vals.push(1001u64);
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
+        &[codec_type],
+        &mut buffer_without_gcd,
+    )?;
+    let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
+    assert!(buffer_without_gcd.len() > buffer.len());
+    Ok(())
+}
+
+#[test]
+fn test_fastfield_gcd_u64() -> io::Result<()> {
+    for &codec_type in &[
+        CodecType::Bitpacked,
+        CodecType::BlockwiseLinear,
+        CodecType::Linear,
+    ] {
+        test_fastfield_gcd_u64_with_codec(codec_type, 5500)?;
+    }
+    Ok(())
+}
+
+#[test]
+pub fn test_fastfield2() {
+    let test_fastfield = crate::column_values::serialize_and_load_u64_based_column_values::<u64>(
+        &&[100u64, 200u64, 300u64][..],
+        &ALL_U64_CODEC_TYPES,
+    );
+    assert_eq!(test_fastfield.get_val(0), 100);
+    assert_eq!(test_fastfield.get_val(1), 200);
+    assert_eq!(test_fastfield.get_val(2), 300);
+}

columnar/src/column_values/vec_column.rs

@@ -0,0 +1,54 @@
+use std::fmt::Debug;
+
+use tantivy_bitpacker::minmax;
+
+use crate::ColumnValues;
+
+/// VecColumn provides `Column` over a `Vec<T>`.
+pub struct VecColumn<T = u64> {
+    pub(crate) values: Vec<T>,
+    pub(crate) min_value: T,
+    pub(crate) max_value: T,
+}
+
+impl<T: Copy + PartialOrd + Send + Sync + Debug + 'static> ColumnValues<T> for VecColumn<T> {
+    fn get_val(&self, position: u32) -> T {
+        self.values[position as usize]
+    }
+
+    fn iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
+        Box::new(self.values.iter().copied())
+    }
+
+    fn min_value(&self) -> T {
+        self.min_value
+    }
+
+    fn max_value(&self) -> T {
+        self.max_value
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.values.len() as u32
+    }
+
+    fn get_range(&self, start: u64, output: &mut [T]) {
+        output.copy_from_slice(&self.values[start as usize..][..output.len()])
+    }
+}
+
+impl<T: Copy + PartialOrd + Default> From<Vec<T>> for VecColumn<T> {
+    fn from(values: Vec<T>) -> Self {
+        let (min_value, max_value) = minmax(values.iter().copied()).unwrap_or_default();
+        Self {
+            values,
+            min_value,
+            max_value,
+        }
+    }
+}
+impl From<VecColumn> for Vec<u64> {
+    fn from(column: VecColumn) -> Self {
+        column.values
+    }
+}

columnar/src/columnar/column_type.rs

@@ -0,0 +1,183 @@
+use std::fmt;
+use std::fmt::Debug;
+use std::net::Ipv6Addr;
+
+use serde::{Deserialize, Serialize};
+
+use crate::InvalidData;
+use crate::value::NumericalType;
+
+/// The column type represents the column type.
+/// Any changes need to be propagated to `COLUMN_TYPES`.
+#[derive(Hash, Eq, PartialEq, Debug, Clone, Copy, Ord, PartialOrd, Serialize, Deserialize)]
+#[repr(u8)]
+pub enum ColumnType {
+    I64 = 0u8,
+    U64 = 1u8,
+    F64 = 2u8,
+    Bytes = 3u8,
+    Str = 4u8,
+    Bool = 5u8,
+    IpAddr = 6u8,
+    DateTime = 7u8,
+}
+
+impl fmt::Display for ColumnType {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        let short_str = match self {
+            ColumnType::I64 => "i64",
+            ColumnType::U64 => "u64",
+            ColumnType::F64 => "f64",
+            ColumnType::Bytes => "bytes",
+            ColumnType::Str => "str",
+            ColumnType::Bool => "bool",
+            ColumnType::IpAddr => "ip",
+            ColumnType::DateTime => "datetime",
+        };
+        write!(f, "{short_str}")
+    }
+}
+
+// The order needs to match _exactly_ the order in the enum
+const COLUMN_TYPES: [ColumnType; 8] = [
+    ColumnType::I64,
+    ColumnType::U64,
+    ColumnType::F64,
+    ColumnType::Bytes,
+    ColumnType::Str,
+    ColumnType::Bool,
+    ColumnType::IpAddr,
+    ColumnType::DateTime,
+];
+
+impl ColumnType {
+    pub fn to_code(self) -> u8 {
+        self as u8
+    }
+    pub fn is_date_time(&self) -> bool {
+        self == &ColumnType::DateTime
+    }
+
+    pub(crate) fn try_from_code(code: u8) -> Result<ColumnType, InvalidData> {
+        COLUMN_TYPES.get(code as usize).copied().ok_or(InvalidData)
+    }
+}
+
+impl From<NumericalType> for ColumnType {
+    fn from(numerical_type: NumericalType) -> Self {
+        match numerical_type {
+            NumericalType::I64 => ColumnType::I64,
+            NumericalType::U64 => ColumnType::U64,
+            NumericalType::F64 => ColumnType::F64,
+        }
+    }
+}
+
+impl ColumnType {
+    pub fn numerical_type(&self) -> Option<NumericalType> {
+        match self {
+            ColumnType::I64 => Some(NumericalType::I64),
+            ColumnType::U64 => Some(NumericalType::U64),
+            ColumnType::F64 => Some(NumericalType::F64),
+            ColumnType::Bytes
+            | ColumnType::Str
+            | ColumnType::Bool
+            | ColumnType::IpAddr
+            | ColumnType::DateTime => None,
+        }
+    }
+}
+
+// TODO remove if possible
+pub trait HasAssociatedColumnType: 'static + Debug + Send + Sync + Copy + PartialOrd {
+    fn column_type() -> ColumnType;
+    fn default_value() -> Self;
+}
+
+impl HasAssociatedColumnType for u64 {
+    fn column_type() -> ColumnType {
+        ColumnType::U64
+    }
+
+    fn default_value() -> Self {
+        0u64
+    }
+}
+
+impl HasAssociatedColumnType for i64 {
+    fn column_type() -> ColumnType {
+        ColumnType::I64
+    }
+
+    fn default_value() -> Self {
+        0i64
+    }
+}
+
+impl HasAssociatedColumnType for f64 {
+    fn column_type() -> ColumnType {
+        ColumnType::F64
+    }
+
+    fn default_value() -> Self {
+        Default::default()
+    }
+}
+
+impl HasAssociatedColumnType for bool {
+    fn column_type() -> ColumnType {
+        ColumnType::Bool
+    }
+    fn default_value() -> Self {
+        Default::default()
+    }
+}
+
+impl HasAssociatedColumnType for common::DateTime {
+    fn column_type() -> ColumnType {
+        ColumnType::DateTime
+    }
+    fn default_value() -> Self {
+        Default::default()
+    }
+}
+
+impl HasAssociatedColumnType for Ipv6Addr {
+    fn column_type() -> ColumnType {
+        ColumnType::IpAddr
+    }
+
+    fn default_value() -> Self {
+        Ipv6Addr::from([0u8; 16])
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::Cardinality;
+
+    #[test]
+    fn test_column_type_to_code() {
+        for (code, expected_column_type) in super::COLUMN_TYPES.iter().copied().enumerate() {
+            if let Ok(column_type) = ColumnType::try_from_code(code as u8) {
+                assert_eq!(column_type, expected_column_type);
+            }
+        }
+        for code in COLUMN_TYPES.len() as u8..=u8::MAX {
+            assert!(ColumnType::try_from_code(code).is_err());
+        }
+    }
+
+    #[test]
+    fn test_cardinality_to_code() {
+        let mut num_cardinality = 0;
+        for code in u8::MIN..=u8::MAX {
+            if let Ok(cardinality) = Cardinality::try_from_code(code) {
+                assert_eq!(cardinality.to_code(), code);
+                num_cardinality += 1;
+            }
+        }
+        assert_eq!(num_cardinality, 3);
+    }
+}

columnar/src/columnar/format_version.rs

@@ -0,0 +1,88 @@
+use core::fmt;
+use std::fmt::{Display, Formatter};
+
+use crate::InvalidData;
+
+pub const VERSION_FOOTER_NUM_BYTES: usize = MAGIC_BYTES.len() + std::mem::size_of::<u32>();
+
+/// We end the file by these 4 bytes just to somewhat identify that
+/// this is indeed a columnar file.
+const MAGIC_BYTES: [u8; 4] = [2, 113, 119, 66];
+
+pub fn footer() -> [u8; VERSION_FOOTER_NUM_BYTES] {
+    let mut footer_bytes = [0u8; VERSION_FOOTER_NUM_BYTES];
+    footer_bytes[0..4].copy_from_slice(&CURRENT_VERSION.to_bytes());
+    footer_bytes[4..8].copy_from_slice(&MAGIC_BYTES[..]);
+    footer_bytes
+}
+
+pub fn parse_footer(footer_bytes: [u8; VERSION_FOOTER_NUM_BYTES]) -> Result<Version, InvalidData> {
+    if footer_bytes[4..8] != MAGIC_BYTES {
+        return Err(InvalidData);
+    }
+    Version::try_from_bytes(footer_bytes[0..4].try_into().unwrap())
+}
+
+pub const CURRENT_VERSION: Version = Version::V2;
+
+#[derive(Debug, Copy, Clone, Eq, PartialEq)]
+#[repr(u32)]
+pub enum Version {
+    V1 = 1u32,
+    V2 = 2u32,
+}
+
+impl Display for Version {
+    fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+        match self {
+            Version::V1 => write!(f, "v1"),
+            Version::V2 => write!(f, "v2"),
+        }
+    }
+}
+
+impl Version {
+    fn to_bytes(self) -> [u8; 4] {
+        (self as u32).to_le_bytes()
+    }
+
+    fn try_from_bytes(bytes: [u8; 4]) -> Result<Version, InvalidData> {
+        let code = u32::from_le_bytes(bytes);
+        match code {
+            1u32 => Ok(Version::V1),
+            2u32 => Ok(Version::V2),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashSet;
+
+    use super::*;
+
+    #[test]
+    fn test_footer_deserialization() {
+        let parsed_version: Version = parse_footer(footer()).unwrap();
+        assert_eq!(Version::V2, parsed_version);
+    }
+
+    #[test]
+    fn test_version_serialization() {
+        let version_to_tests: Vec<u32> = [0, 1 << 8, 1 << 16, 1 << 24]
+            .iter()
+            .copied()
+            .flat_map(|offset| (0..255).map(move |el| el + offset))
+            .collect();
+        let mut valid_versions: HashSet<u32> = HashSet::default();
+        for &i in &version_to_tests {
+            let version_res = Version::try_from_bytes(i.to_le_bytes());
+            if let Ok(version) = version_res {
+                assert_eq!(version.to_bytes(), i.to_le_bytes());
+                valid_versions.insert(i);
+            }
+        }
+        assert_eq!(valid_versions.len(), 2);
+    }
+}

columnar/src/columnar/merge/merge_dict_column.rs

@@ -0,0 +1,214 @@
+use std::io::{self, Write};
+
+use common::{BitSet, CountingWriter, ReadOnlyBitSet};
+use sstable::{SSTable, Streamer, TermOrdinal, VoidSSTable};
+
+use super::term_merger::{TermMerger, TermsWithSegmentOrd};
+use crate::column::serialize_column_mappable_to_u64;
+use crate::column_index::SerializableColumnIndex;
+use crate::iterable::Iterable;
+use crate::{BytesColumn, MergeRowOrder, ShuffleMergeOrder};
+
+// Serialize [Dictionary, Column, dictionary num bytes U32::LE]
+// Column: [Column Index, Column Values, column index num bytes U32::LE]
+pub fn merge_bytes_or_str_column(
+    column_index: SerializableColumnIndex<'_>,
+    bytes_columns: &[Option<BytesColumn>],
+    merge_row_order: &MergeRowOrder,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    // Serialize dict and generate mapping for values
+    let mut output = CountingWriter::wrap(output);
+    // TODO !!! Remove useless terms.
+    let term_ord_mapping = serialize_merged_dict(bytes_columns, merge_row_order, &mut output)?;
+    let dictionary_num_bytes: u32 = output.written_bytes() as u32;
+    let output = output.finish();
+    let remapped_term_ordinals_values = RemappedTermOrdinalsValues {
+        bytes_columns,
+        term_ord_mapping: &term_ord_mapping,
+        merge_row_order,
+    };
+    serialize_column_mappable_to_u64(column_index, &remapped_term_ordinals_values, output)?;
+    output.write_all(&dictionary_num_bytes.to_le_bytes())?;
+    Ok(())
+}
+
+struct RemappedTermOrdinalsValues<'a> {
+    bytes_columns: &'a [Option<BytesColumn>],
+    term_ord_mapping: &'a TermOrdinalMapping,
+    merge_row_order: &'a MergeRowOrder,
+}
+
+impl Iterable for RemappedTermOrdinalsValues<'_> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u64> + '_> {
+        match self.merge_row_order {
+            MergeRowOrder::Stack(_) => self.boxed_iter_stacked(),
+            MergeRowOrder::Shuffled(shuffle_merge_order) => {
+                self.boxed_iter_shuffled(shuffle_merge_order)
+            }
+        }
+    }
+}
+
+impl RemappedTermOrdinalsValues<'_> {
+    fn boxed_iter_stacked(&self) -> Box<dyn Iterator<Item = u64> + '_> {
+        let iter = self
+            .bytes_columns
+            .iter()
+            .enumerate()
+            .flat_map(|(seg_ord, bytes_column_opt)| {
+                let bytes_column = bytes_column_opt.as_ref()?;
+                Some((seg_ord, bytes_column))
+            })
+            .flat_map(move |(seg_ord, bytes_column)| {
+                let term_ord_after_merge_mapping =
+                    self.term_ord_mapping.get_segment(seg_ord as u32);
+                bytes_column
+                    .ords()
+                    .values
+                    .iter()
+                    .map(move |term_ord| term_ord_after_merge_mapping[term_ord as usize])
+            });
+        Box::new(iter)
+    }
+
+    fn boxed_iter_shuffled<'b>(
+        &'b self,
+        shuffle_merge_order: &'b ShuffleMergeOrder,
+    ) -> Box<dyn Iterator<Item = u64> + 'b> {
+        Box::new(
+            shuffle_merge_order
+                .iter_new_to_old_row_addrs()
+                .flat_map(move |old_addr| {
+                    let segment_ord = self.term_ord_mapping.get_segment(old_addr.segment_ord);
+                    self.bytes_columns[old_addr.segment_ord as usize]
+                        .as_ref()
+                        .into_iter()
+                        .flat_map(move |bytes_column| {
+                            bytes_column
+                                .term_ords(old_addr.row_id)
+                                .map(|old_term_ord: u64| segment_ord[old_term_ord as usize])
+                        })
+                }),
+        )
+    }
+}
+
+fn compute_term_bitset(column: &BytesColumn, row_bitset: &ReadOnlyBitSet) -> BitSet {
+    let num_terms = column.dictionary().num_terms();
+    let mut term_bitset = BitSet::with_max_value(num_terms as u32);
+    for row_id in row_bitset.iter() {
+        for term_ord in column.term_ord_column.values_for_doc(row_id) {
+            term_bitset.insert(term_ord as u32);
+        }
+    }
+    term_bitset
+}
+
+fn is_term_present(bitsets: &[Option<BitSet>], term_merger: &TermMerger) -> bool {
+    for (segment_ord, from_term_ord) in term_merger.matching_segments() {
+        if let Some(bitset) = bitsets[segment_ord].as_ref() {
+            if bitset.contains(from_term_ord as u32) {
+                return true;
+            }
+        } else {
+            return true;
+        }
+    }
+    false
+}
+
+fn serialize_merged_dict(
+    bytes_columns: &[Option<BytesColumn>],
+    merge_row_order: &MergeRowOrder,
+    output: &mut impl Write,
+) -> io::Result<TermOrdinalMapping> {
+    let mut term_ord_mapping = TermOrdinalMapping::default();
+
+    let mut field_term_streams = Vec::new();
+    for (segment_ord, column_opt) in bytes_columns.iter().enumerate() {
+        if let Some(column) = column_opt {
+            term_ord_mapping.add_segment(column.dictionary.num_terms());
+            let terms: Streamer<VoidSSTable> = column.dictionary.stream()?;
+            field_term_streams.push(TermsWithSegmentOrd { terms, segment_ord });
+        } else {
+            term_ord_mapping.add_segment(0);
+            field_term_streams.push(TermsWithSegmentOrd {
+                terms: Streamer::empty(),
+                segment_ord,
+            });
+        }
+    }
+
+    let mut merged_terms = TermMerger::new(field_term_streams);
+    let mut sstable_builder = sstable::VoidSSTable::writer(output);
+
+    match merge_row_order {
+        MergeRowOrder::Stack(_) => {
+            let mut current_term_ord = 0;
+            while merged_terms.advance() {
+                let term_bytes: &[u8] = merged_terms.key();
+                sstable_builder.insert(term_bytes, &())?;
+                for (segment_ord, from_term_ord) in merged_terms.matching_segments() {
+                    term_ord_mapping.register_from_to(segment_ord, from_term_ord, current_term_ord);
+                }
+                current_term_ord += 1;
+            }
+            sstable_builder.finish()?;
+        }
+        MergeRowOrder::Shuffled(shuffle_merge_order) => {
+            assert_eq!(shuffle_merge_order.alive_bitsets.len(), bytes_columns.len());
+            let mut term_bitsets: Vec<Option<BitSet>> = Vec::with_capacity(bytes_columns.len());
+            for (alive_bitset_opt, bytes_column_opt) in shuffle_merge_order
+                .alive_bitsets
+                .iter()
+                .zip(bytes_columns.iter())
+            {
+                match (alive_bitset_opt, bytes_column_opt) {
+                    (Some(alive_bitset), Some(bytes_column)) => {
+                        let term_bitset = compute_term_bitset(bytes_column, alive_bitset);
+                        term_bitsets.push(Some(term_bitset));
+                    }
+                    _ => {
+                        term_bitsets.push(None);
+                    }
+                }
+            }
+            let mut current_term_ord = 0;
+            while merged_terms.advance() {
+                let term_bytes: &[u8] = merged_terms.key();
+                if !is_term_present(&term_bitsets[..], &merged_terms) {
+                    continue;
+                }
+                sstable_builder.insert(term_bytes, &())?;
+                for (segment_ord, from_term_ord) in merged_terms.matching_segments() {
+                    term_ord_mapping.register_from_to(segment_ord, from_term_ord, current_term_ord);
+                }
+                current_term_ord += 1;
+            }
+            sstable_builder.finish()?;
+        }
+    }
+    Ok(term_ord_mapping)
+}
+
+#[derive(Default, Debug)]
+struct TermOrdinalMapping {
+    /// Contains the new term ordinals for each segment.
+    per_segment_new_term_ordinals: Vec<Vec<TermOrdinal>>,
+}
+
+impl TermOrdinalMapping {
+    fn add_segment(&mut self, max_term_ord: usize) {
+        self.per_segment_new_term_ordinals
+            .push(vec![TermOrdinal::default(); max_term_ord]);
+    }
+
+    fn register_from_to(&mut self, segment_ord: usize, from_ord: TermOrdinal, to_ord: TermOrdinal) {
+        self.per_segment_new_term_ordinals[segment_ord][from_ord as usize] = to_ord;
+    }
+
+    fn get_segment(&self, segment_ord: u32) -> &[TermOrdinal] {
+        &self.per_segment_new_term_ordinals[segment_ord as usize]
+    }
+}

columnar/src/columnar/merge/merge_mapping.rs

@@ -0,0 +1,129 @@
+use std::ops::Range;
+
+use common::{BitSet, OwnedBytes, ReadOnlyBitSet};
+
+use crate::{ColumnarReader, RowAddr, RowId};
+
+pub struct StackMergeOrder {
+    // This does not start at 0. The first row is the number of
+    // rows in the first columnar.
+    cumulated_row_ids: Vec<RowId>,
+}
+
+impl StackMergeOrder {
+    #[cfg(test)]
+    pub fn stack_for_test(num_rows_per_columnar: &[u32]) -> StackMergeOrder {
+        let mut cumulated_row_ids: Vec<RowId> = Vec::with_capacity(num_rows_per_columnar.len());
+        let mut cumulated_row_id = 0;
+        for &num_rows in num_rows_per_columnar {
+            cumulated_row_id += num_rows;
+            cumulated_row_ids.push(cumulated_row_id);
+        }
+        StackMergeOrder { cumulated_row_ids }
+    }
+
+    pub fn stack(columnars: &[&ColumnarReader]) -> StackMergeOrder {
+        let mut cumulated_row_ids: Vec<RowId> = Vec::with_capacity(columnars.len());
+        let mut cumulated_row_id = 0;
+        for columnar in columnars {
+            cumulated_row_id += columnar.num_docs();
+            cumulated_row_ids.push(cumulated_row_id);
+        }
+        StackMergeOrder { cumulated_row_ids }
+    }
+
+    pub fn num_rows(&self) -> RowId {
+        self.cumulated_row_ids.last().copied().unwrap_or(0)
+    }
+
+    pub fn offset(&self, columnar_id: usize) -> RowId {
+        if columnar_id == 0 {
+            return 0;
+        }
+        self.cumulated_row_ids[columnar_id - 1]
+    }
+
+    pub fn columnar_range(&self, columnar_id: usize) -> Range<RowId> {
+        self.offset(columnar_id)..self.offset(columnar_id + 1)
+    }
+}
+
+pub enum MergeRowOrder {
+    /// Columnar tables are simply stacked one above the other.
+    /// If the i-th columnar_readers has n_rows_i rows, then
+    /// in the resulting columnar,
+    /// rows [r0..n_row_0) contains the row of `columnar_readers[0]`, in ordder
+    /// rows [n_row_0..n_row_0 + n_row_1 contains the row of `columnar_readers[1]`, in order.
+    /// ..
+    /// No documents is deleted.
+    Stack(StackMergeOrder),
+    /// Some more complex mapping, that may interleaves rows from the different readers and
+    /// drop rows, or do both.
+    Shuffled(ShuffleMergeOrder),
+}
+
+impl From<StackMergeOrder> for MergeRowOrder {
+    fn from(stack_merge_order: StackMergeOrder) -> MergeRowOrder {
+        MergeRowOrder::Stack(stack_merge_order)
+    }
+}
+
+impl From<ShuffleMergeOrder> for MergeRowOrder {
+    fn from(shuffle_merge_order: ShuffleMergeOrder) -> MergeRowOrder {
+        MergeRowOrder::Shuffled(shuffle_merge_order)
+    }
+}
+
+impl MergeRowOrder {
+    pub fn num_rows(&self) -> RowId {
+        match self {
+            MergeRowOrder::Stack(stack_row_order) => stack_row_order.num_rows(),
+            MergeRowOrder::Shuffled(complex_mapping) => complex_mapping.num_rows(),
+        }
+    }
+}
+
+pub struct ShuffleMergeOrder {
+    pub new_row_id_to_old_row_id: Vec<RowAddr>,
+    pub alive_bitsets: Vec<Option<ReadOnlyBitSet>>,
+}
+
+impl ShuffleMergeOrder {
+    pub fn for_test(
+        segment_num_rows: &[RowId],
+        new_row_id_to_old_row_id: Vec<RowAddr>,
+    ) -> ShuffleMergeOrder {
+        let mut alive_bitsets: Vec<BitSet> = segment_num_rows
+            .iter()
+            .map(|&num_rows| BitSet::with_max_value(num_rows))
+            .collect();
+        for &RowAddr {
+            segment_ord,
+            row_id,
+        } in &new_row_id_to_old_row_id
+        {
+            alive_bitsets[segment_ord as usize].insert(row_id);
+        }
+        let alive_bitsets: Vec<Option<ReadOnlyBitSet>> = alive_bitsets
+            .into_iter()
+            .map(|alive_bitset| {
+                let mut buffer = Vec::new();
+                alive_bitset.serialize(&mut buffer).unwrap();
+                let data = OwnedBytes::new(buffer);
+                Some(ReadOnlyBitSet::open(data))
+            })
+            .collect();
+        ShuffleMergeOrder {
+            new_row_id_to_old_row_id,
+            alive_bitsets,
+        }
+    }
+
+    pub fn num_rows(&self) -> RowId {
+        self.new_row_id_to_old_row_id.len() as RowId
+    }
+
+    pub fn iter_new_to_old_row_addrs(&self) -> impl Iterator<Item = RowAddr> + '_ {
+        self.new_row_id_to_old_row_id.iter().copied()
+    }
+}

columnar/src/columnar/merge/mod.rs

@@ -0,0 +1,477 @@
+mod merge_dict_column;
+mod merge_mapping;
+mod term_merger;
+
+use std::collections::{BTreeMap, HashSet};
+use std::io;
+use std::net::Ipv6Addr;
+use std::sync::Arc;
+
+pub use merge_mapping::{MergeRowOrder, ShuffleMergeOrder, StackMergeOrder};
+
+use super::writer::ColumnarSerializer;
+use crate::column::{serialize_column_mappable_to_u64, serialize_column_mappable_to_u128};
+use crate::column_values::MergedColumnValues;
+use crate::columnar::ColumnarReader;
+use crate::columnar::merge::merge_dict_column::merge_bytes_or_str_column;
+use crate::columnar::writer::CompatibleNumericalTypes;
+use crate::dynamic_column::DynamicColumn;
+use crate::{
+    BytesColumn, Column, ColumnIndex, ColumnType, ColumnValues, DynamicColumnHandle, NumericalType,
+    NumericalValue,
+};
+
+/// Column types are grouped into different categories.
+/// After merge, all columns belonging to the same category are coerced to
+/// the same column type.
+///
+/// In practise, today, only Numerical columns are coerced into one type today.
+///
+/// See also [README.md].
+///
+/// The ordering has to match the ordering of the variants in [ColumnType].
+#[derive(Copy, Clone, Eq, PartialOrd, Ord, PartialEq, Hash, Debug)]
+pub(crate) enum ColumnTypeCategory {
+    Numerical,
+    Bytes,
+    Str,
+    Bool,
+    IpAddr,
+    DateTime,
+}
+
+impl From<ColumnType> for ColumnTypeCategory {
+    fn from(column_type: ColumnType) -> Self {
+        match column_type {
+            ColumnType::I64 => ColumnTypeCategory::Numerical,
+            ColumnType::U64 => ColumnTypeCategory::Numerical,
+            ColumnType::F64 => ColumnTypeCategory::Numerical,
+            ColumnType::Bytes => ColumnTypeCategory::Bytes,
+            ColumnType::Str => ColumnTypeCategory::Str,
+            ColumnType::Bool => ColumnTypeCategory::Bool,
+            ColumnType::IpAddr => ColumnTypeCategory::IpAddr,
+            ColumnType::DateTime => ColumnTypeCategory::DateTime,
+        }
+    }
+}
+
+/// Merge several columnar table together.
+///
+/// If several columns with the same name are conflicting with the numerical types in the
+/// input columnars, the first type compatible out of i64, u64, f64 in that order will be used.
+///
+/// `require_columns` makes it possible to ensure that some columns will be present in the
+/// resulting columnar. When a required column is a numerical column type, one of two things can
+/// happen:
+/// - If the required column type is compatible with all of the input columnar, the resulting merged
+///   columnar will simply coerce the input column and use the required column type.
+/// - If the required column type is incompatible with one of the input columnar, the merged will
+///   fail with an InvalidData error.
+///
+/// `merge_row_order` makes it possible to remove or reorder row in the resulting
+/// `Columnar` table.
+///
+/// Reminder: a string and a numerical column may bare the same column name. This is not
+/// considered a conflict.
+pub fn merge_columnar(
+    columnar_readers: &[&ColumnarReader],
+    required_columns: &[(String, ColumnType)],
+    merge_row_order: MergeRowOrder,
+    output: &mut impl io::Write,
+) -> io::Result<()> {
+    let mut serializer = ColumnarSerializer::new(output);
+    let num_docs_per_columnar = columnar_readers
+        .iter()
+        .map(|reader| reader.num_docs())
+        .collect::<Vec<u32>>();
+
+    let columns_to_merge = group_columns_for_merge(columnar_readers, required_columns)?;
+    for res in columns_to_merge {
+        let ((column_name, _column_type_category), grouped_columns) = res;
+        let grouped_columns = grouped_columns.open(&merge_row_order)?;
+        if grouped_columns.is_empty() {
+            continue;
+        }
+
+        let column_type_after_merge = grouped_columns.column_type_after_merge();
+        let mut columns = grouped_columns.columns;
+        // Make sure the number of columns is the same as the number of columnar readers.
+        // Or num_docs_per_columnar would be incorrect.
+        assert_eq!(columns.len(), columnar_readers.len());
+        coerce_columns(column_type_after_merge, &mut columns)?;
+
+        let mut column_serializer =
+            serializer.start_serialize_column(column_name.as_bytes(), column_type_after_merge);
+        merge_column(
+            column_type_after_merge,
+            &num_docs_per_columnar,
+            columns,
+            &merge_row_order,
+            &mut column_serializer,
+        )?;
+        column_serializer.finalize()?;
+    }
+
+    serializer.finalize(merge_row_order.num_rows())?;
+    Ok(())
+}
+
+fn dynamic_column_to_u64_monotonic(dynamic_column: DynamicColumn) -> Option<Column<u64>> {
+    match dynamic_column {
+        DynamicColumn::Bool(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::I64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::U64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::F64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::DateTime(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::IpAddr(_) | DynamicColumn::Bytes(_) | DynamicColumn::Str(_) => None,
+    }
+}
+
+fn merge_column(
+    column_type: ColumnType,
+    num_docs_per_column: &[u32],
+    columns_to_merge: Vec<Option<DynamicColumn>>,
+    merge_row_order: &MergeRowOrder,
+    wrt: &mut impl io::Write,
+) -> io::Result<()> {
+    match column_type {
+        ColumnType::I64
+        | ColumnType::U64
+        | ColumnType::F64
+        | ColumnType::DateTime
+        | ColumnType::Bool => {
+            let mut column_indexes: Vec<ColumnIndex> = Vec::with_capacity(columns_to_merge.len());
+            let mut column_values: Vec<Option<Arc<dyn ColumnValues>>> =
+                Vec::with_capacity(columns_to_merge.len());
+            for (i, dynamic_column_opt) in columns_to_merge.into_iter().enumerate() {
+                match dynamic_column_opt.and_then(dynamic_column_to_u64_monotonic) {
+                    Some(Column { index: idx, values }) => {
+                        column_indexes.push(idx);
+                        column_values.push(Some(values));
+                    }
+                    None => {
+                        column_indexes.push(ColumnIndex::Empty {
+                            num_docs: num_docs_per_column[i],
+                        });
+                        column_values.push(None);
+                    }
+                }
+            }
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            let merge_column_values = MergedColumnValues {
+                column_indexes: &column_indexes[..],
+                column_values: &column_values[..],
+                merge_row_order,
+            };
+            serialize_column_mappable_to_u64(merged_column_index, &merge_column_values, wrt)?;
+        }
+        ColumnType::IpAddr => {
+            let mut column_indexes: Vec<ColumnIndex> = Vec::with_capacity(columns_to_merge.len());
+            let mut column_values: Vec<Option<Arc<dyn ColumnValues<Ipv6Addr>>>> =
+                Vec::with_capacity(columns_to_merge.len());
+            for (i, dynamic_column_opt) in columns_to_merge.into_iter().enumerate() {
+                if let Some(DynamicColumn::IpAddr(Column { index: idx, values })) =
+                    dynamic_column_opt
+                {
+                    column_indexes.push(idx);
+                    column_values.push(Some(values));
+                } else {
+                    column_indexes.push(ColumnIndex::Empty {
+                        num_docs: num_docs_per_column[i],
+                    });
+                    column_values.push(None);
+                }
+            }
+
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            let merge_column_values = MergedColumnValues {
+                column_indexes: &column_indexes[..],
+                column_values: &column_values,
+                merge_row_order,
+            };
+
+            serialize_column_mappable_to_u128(merged_column_index, &merge_column_values, wrt)?;
+        }
+        ColumnType::Bytes | ColumnType::Str => {
+            let mut column_indexes: Vec<ColumnIndex> = Vec::with_capacity(columns_to_merge.len());
+            let mut bytes_columns: Vec<Option<BytesColumn>> =
+                Vec::with_capacity(columns_to_merge.len());
+            for (i, dynamic_column_opt) in columns_to_merge.into_iter().enumerate() {
+                match dynamic_column_opt {
+                    Some(DynamicColumn::Str(str_column)) => {
+                        column_indexes.push(str_column.term_ord_column.index.clone());
+                        bytes_columns.push(Some(str_column.into()));
+                    }
+                    Some(DynamicColumn::Bytes(bytes_column)) => {
+                        column_indexes.push(bytes_column.term_ord_column.index.clone());
+                        bytes_columns.push(Some(bytes_column));
+                    }
+                    _ => {
+                        column_indexes.push(ColumnIndex::Empty {
+                            num_docs: num_docs_per_column[i],
+                        });
+                        bytes_columns.push(None);
+                    }
+                }
+            }
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            merge_bytes_or_str_column(merged_column_index, &bytes_columns, merge_row_order, wrt)?;
+        }
+    }
+    Ok(())
+}
+
+struct GroupedColumns {
+    required_column_type: Option<ColumnType>,
+    columns: Vec<Option<DynamicColumn>>,
+}
+
+impl GroupedColumns {
+    /// Check is column group can be skipped during serialization.
+    fn is_empty(&self) -> bool {
+        self.required_column_type.is_none() && self.columns.iter().all(Option::is_none)
+    }
+
+    /// Returns the column type after merge.
+    ///
+    /// This method does not check if the column types can actually be coerced to
+    /// this type.
+    fn column_type_after_merge(&self) -> ColumnType {
+        if let Some(required_type) = self.required_column_type {
+            return required_type;
+        }
+        let column_type: HashSet<ColumnType> = self
+            .columns
+            .iter()
+            .flatten()
+            .map(|column| column.column_type())
+            .collect();
+        if column_type.len() == 1 {
+            return column_type.into_iter().next().unwrap();
+        }
+        // At the moment, only the numerical column type category has more than one possible
+        // column type.
+        assert!(
+            self.columns
+                .iter()
+                .flatten()
+                .all(|el| ColumnTypeCategory::from(el.column_type())
+                    == ColumnTypeCategory::Numerical)
+        );
+        merged_numerical_columns_type(self.columns.iter().flatten()).into()
+    }
+}
+
+struct GroupedColumnsHandle {
+    required_column_type: Option<ColumnType>,
+    columns: Vec<Option<DynamicColumnHandle>>,
+}
+
+impl GroupedColumnsHandle {
+    fn new(num_columnars: usize) -> Self {
+        GroupedColumnsHandle {
+            required_column_type: None,
+            columns: vec![None; num_columnars],
+        }
+    }
+    fn open(self, merge_row_order: &MergeRowOrder) -> io::Result<GroupedColumns> {
+        let mut columns: Vec<Option<DynamicColumn>> = Vec::new();
+        for (columnar_id, column) in self.columns.iter().enumerate() {
+            if let Some(column) = column {
+                let column = column.open()?;
+                // We skip columns that end up with 0 documents.
+                // That way, we make sure they don't end up influencing the merge type or
+                // creating empty columns.
+
+                if is_empty_after_merge(merge_row_order, &column, columnar_id) {
+                    columns.push(None);
+                } else {
+                    columns.push(Some(column));
+                }
+            } else {
+                columns.push(None);
+            }
+        }
+        Ok(GroupedColumns {
+            required_column_type: self.required_column_type,
+            columns,
+        })
+    }
+
+    /// Set the dynamic column for a given columnar.
+    fn set_column(&mut self, columnar_id: usize, column: DynamicColumnHandle) {
+        self.columns[columnar_id] = Some(column);
+    }
+
+    /// Force the existence of a column, as well as its type.
+    fn require_type(&mut self, required_type: ColumnType) -> io::Result<()> {
+        if let Some(existing_required_type) = self.required_column_type {
+            if existing_required_type == required_type {
+                // This was just a duplicate in the `required_columns`.
+                // Nothing to do.
+                return Ok(());
+            } else {
+                return Err(io::Error::new(
+                    io::ErrorKind::InvalidInput,
+                    "Required column conflicts with another required column of the same type \
+                     category.",
+                ));
+            }
+        }
+        self.required_column_type = Some(required_type);
+        Ok(())
+    }
+}
+
+/// Returns the type of the merged numerical column.
+///
+/// This function picks the first numerical type out of i64, u64, f64 (order matters
+/// here), that is compatible with all the `columns`.
+///
+/// # Panics
+/// Panics if one of the column is not numerical.
+fn merged_numerical_columns_type<'a>(
+    columns: impl Iterator<Item = &'a DynamicColumn>,
+) -> NumericalType {
+    let mut compatible_numerical_types = CompatibleNumericalTypes::default();
+    for column in columns {
+        let (min_value, max_value) =
+            min_max_if_numerical(column).expect("All columns re required to be numerical");
+        compatible_numerical_types.accept_value(min_value);
+        compatible_numerical_types.accept_value(max_value);
+    }
+    compatible_numerical_types.to_numerical_type()
+}
+
+fn is_empty_after_merge(
+    merge_row_order: &MergeRowOrder,
+    column: &DynamicColumn,
+    columnar_ord: usize,
+) -> bool {
+    if column.num_values() == 0u32 {
+        // It was empty before the merge.
+        return true;
+    }
+    match merge_row_order {
+        MergeRowOrder::Stack(_) => {
+            // If we are stacking the columnar, no rows are being deleted.
+            false
+        }
+        MergeRowOrder::Shuffled(shuffled) => {
+            if let Some(alive_bitset) = &shuffled.alive_bitsets[columnar_ord] {
+                let column_index = column.column_index();
+                match column_index {
+                    ColumnIndex::Empty { .. } => true,
+                    ColumnIndex::Full => alive_bitset.len() == 0,
+                    ColumnIndex::Optional(optional_index) => {
+                        for doc in optional_index.iter_docs() {
+                            if alive_bitset.contains(doc) {
+                                return false;
+                            }
+                        }
+                        true
+                    }
+                    ColumnIndex::Multivalued(multivalued_index) => {
+                        for alive_docid in alive_bitset.iter() {
+                            if !multivalued_index.range(alive_docid).is_empty() {
+                                return false;
+                            }
+                        }
+                        true
+                    }
+                }
+            } else {
+                // No document is being deleted.
+                // The shuffle is applying a permutation.
+                false
+            }
+        }
+    }
+}
+
+/// Iterates over the columns of the columnar readers, grouped by column name.
+/// Key functionality is that `open` of the Columns is done lazy per group.
+fn group_columns_for_merge<'a>(
+    columnar_readers: &'a [&'a ColumnarReader],
+    required_columns: &'a [(String, ColumnType)],
+) -> io::Result<BTreeMap<(String, ColumnTypeCategory), GroupedColumnsHandle>> {
+    let mut columns: BTreeMap<(String, ColumnTypeCategory), GroupedColumnsHandle> = BTreeMap::new();
+
+    for &(ref column_name, column_type) in required_columns {
+        columns
+            .entry((column_name.clone(), column_type.into()))
+            .or_insert_with(|| GroupedColumnsHandle::new(columnar_readers.len()))
+            .require_type(column_type)?;
+    }
+
+    for (columnar_id, columnar_reader) in columnar_readers.iter().enumerate() {
+        let column_name_and_handle = columnar_reader.iter_columns()?;
+
+        for (column_name, handle) in column_name_and_handle {
+            let column_category: ColumnTypeCategory = handle.column_type().into();
+            columns
+                .entry((column_name, column_category))
+                .or_insert_with(|| GroupedColumnsHandle::new(columnar_readers.len()))
+                .set_column(columnar_id, handle);
+        }
+    }
+    Ok(columns)
+}
+
+fn coerce_columns(
+    column_type: ColumnType,
+    columns: &mut [Option<DynamicColumn>],
+) -> io::Result<()> {
+    for column_opt in columns.iter_mut() {
+        if let Some(column) = column_opt.take() {
+            *column_opt = Some(coerce_column(column_type, column)?);
+        }
+    }
+    Ok(())
+}
+
+fn coerce_column(column_type: ColumnType, column: DynamicColumn) -> io::Result<DynamicColumn> {
+    if let Some(numerical_type) = column_type.numerical_type() {
+        column
+            .coerce_numerical(numerical_type)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, ""))
+    } else {
+        if column.column_type() != column_type {
+            return Err(io::Error::new(
+                io::ErrorKind::InvalidInput,
+                format!(
+                    "Cannot coerce column of type `{:?}` to `{column_type:?}`",
+                    column.column_type()
+                ),
+            ));
+        }
+        Ok(column)
+    }
+}
+
+/// Returns the (min, max) of a column provided it is numerical (i64, u64. f64).
+///
+/// The min and the max are simply the numerical value as defined by `ColumnValue::min_value()`,
+/// and `ColumnValue::max_value()`.
+///
+/// It is important to note that these values are only guaranteed to be lower/upper bound
+/// (as opposed to min/max value).
+/// If a column is empty, the min and max values are currently set to 0.
+fn min_max_if_numerical(column: &DynamicColumn) -> Option<(NumericalValue, NumericalValue)> {
+    match column {
+        DynamicColumn::I64(column) => Some((column.min_value().into(), column.max_value().into())),
+        DynamicColumn::U64(column) => Some((column.min_value().into(), column.max_value().into())),
+        DynamicColumn::F64(column) => Some((column.min_value().into(), column.max_value().into())),
+        DynamicColumn::Bool(_)
+        | DynamicColumn::IpAddr(_)
+        | DynamicColumn::DateTime(_)
+        | DynamicColumn::Bytes(_)
+        | DynamicColumn::Str(_) => None,
+    }
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/columnar/merge/term_merger.rs

@@ -0,0 +1,101 @@
+use std::cmp::Ordering;
+use std::collections::BinaryHeap;
+
+use sstable::TermOrdinal;
+
+use crate::Streamer;
+
+/// The terms of a column with the ordinal of the segment.
+pub struct TermsWithSegmentOrd<'a> {
+    pub terms: Streamer<'a>,
+    pub segment_ord: usize,
+}
+
+impl PartialEq for TermsWithSegmentOrd<'_> {
+    fn eq(&self, other: &Self) -> bool {
+        self.segment_ord == other.segment_ord
+    }
+}
+
+impl Eq for TermsWithSegmentOrd<'_> {}
+
+impl<'a> PartialOrd for TermsWithSegmentOrd<'a> {
+    fn partial_cmp(&self, other: &TermsWithSegmentOrd<'a>) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl<'a> Ord for TermsWithSegmentOrd<'a> {
+    fn cmp(&self, other: &TermsWithSegmentOrd<'a>) -> Ordering {
+        (&other.terms.key(), &other.segment_ord).cmp(&(&self.terms.key(), &self.segment_ord))
+    }
+}
+
+/// Given a list of sorted term streams,
+/// returns an iterator over sorted unique terms.
+///
+/// The item yield is actually a pair with
+/// - the term
+/// - a slice with the ordinal of the segments containing the terms.
+pub struct TermMerger<'a> {
+    heap: BinaryHeap<TermsWithSegmentOrd<'a>>,
+    term_streams_with_segment: Vec<TermsWithSegmentOrd<'a>>,
+}
+
+impl<'a> TermMerger<'a> {
+    /// Stream of merged term dictionary
+    pub fn new(term_streams_with_segment: Vec<TermsWithSegmentOrd<'a>>) -> TermMerger<'a> {
+        TermMerger {
+            heap: BinaryHeap::new(),
+            term_streams_with_segment,
+        }
+    }
+
+    pub(crate) fn matching_segments<'b: 'a>(
+        &'b self,
+    ) -> impl 'b + Iterator<Item = (usize, TermOrdinal)> {
+        self.term_streams_with_segment
+            .iter()
+            .map(|heap_item| (heap_item.segment_ord, heap_item.terms.term_ord()))
+    }
+
+    fn advance_segments(&mut self) {
+        let streamers = &mut self.term_streams_with_segment;
+        let heap = &mut self.heap;
+        for mut heap_item in streamers.drain(..) {
+            if heap_item.terms.advance() {
+                heap.push(heap_item);
+            }
+        }
+    }
+
+    /// Advance the term iterator to the next term.
+    /// Returns true if there is indeed another term
+    /// False if there is none.
+    pub fn advance(&mut self) -> bool {
+        self.advance_segments();
+        match self.heap.pop() {
+            Some(head) => {
+                self.term_streams_with_segment.push(head);
+                while let Some(next_streamer) = self.heap.peek() {
+                    if self.term_streams_with_segment[0].terms.key() != next_streamer.terms.key() {
+                        break;
+                    }
+                    let next_heap_it = self.heap.pop().unwrap(); // safe : we peeked beforehand
+                    self.term_streams_with_segment.push(next_heap_it);
+                }
+                true
+            }
+            _ => false,
+        }
+    }
+
+    /// Returns the current term.
+    ///
+    /// This method may be called
+    /// if and only if advance() has been called before
+    /// and "true" was returned.
+    pub fn key(&self) -> &[u8] {
+        self.term_streams_with_segment[0].terms.key()
+    }
+}

columnar/src/columnar/merge/tests.rs

@@ -0,0 +1,588 @@
+use itertools::Itertools;
+use proptest::collection::vec;
+use proptest::prelude::*;
+
+use super::*;
+use crate::columnar::{ColumnarReader, MergeRowOrder, StackMergeOrder, merge_columnar};
+use crate::{Cardinality, ColumnarWriter, DynamicColumn, HasAssociatedColumnType, RowId};
+
+fn make_columnar<T: Into<NumericalValue> + HasAssociatedColumnType + Copy>(
+    column_name: &str,
+    vals: &[T],
+) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_column_type(column_name, T::column_type(), false);
+    for (row_id, val) in vals.iter().copied().enumerate() {
+        dataframe_writer.record_numerical(row_id as RowId, column_name, val.into());
+    }
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(vals.len() as RowId, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+#[test]
+fn test_column_coercion_to_u64() {
+    // i64 type
+    let columnar1 = make_columnar("numbers", &[1i64]);
+    // u64 type
+    let columnar2 = make_columnar("numbers", &[u64::MAX]);
+    let columnars = &[&columnar1, &columnar2];
+    let column_map: BTreeMap<(String, ColumnTypeCategory), GroupedColumnsHandle> =
+        group_columns_for_merge(columnars, &[]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnTypeCategory::Numerical)));
+}
+
+#[test]
+fn test_column_coercion_to_i64() {
+    let columnar1 = make_columnar("numbers", &[-1i64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let columnars = &[&columnar1, &columnar2];
+    let column_map: BTreeMap<(String, ColumnTypeCategory), GroupedColumnsHandle> =
+        group_columns_for_merge(columnars, &[]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnTypeCategory::Numerical)));
+}
+
+//#[test]
+// fn test_impossible_coercion_returns_an_error() {
+// let columnar1 = make_columnar("numbers", &[u64::MAX]);
+// let merge_order = StackMergeOrder::stack(&[&columnar1]).into();
+// let group_error = group_columns_for_merge_iter(
+//&[&columnar1],
+//&[("numbers".to_string(), ColumnType::I64)],
+//&merge_order,
+//)
+//.unwrap_err();
+// assert_eq!(group_error.kind(), io::ErrorKind::InvalidInput);
+//}
+
+#[test]
+fn test_group_columns_with_required_column() {
+    let columnar1 = make_columnar("numbers", &[1i64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let columnars = &[&columnar1, &columnar2];
+    let column_map: BTreeMap<(String, ColumnTypeCategory), GroupedColumnsHandle> =
+        group_columns_for_merge(columnars, &[("numbers".to_string(), ColumnType::U64)]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnTypeCategory::Numerical)));
+}
+
+#[test]
+fn test_group_columns_required_column_with_no_existing_columns() {
+    let columnar1 = make_columnar("numbers", &[2u64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let columnars = &[&columnar1, &columnar2];
+    let column_map: BTreeMap<_, _> =
+        group_columns_for_merge(columnars, &[("required_col".to_string(), ColumnType::Str)])
+            .unwrap();
+    assert_eq!(column_map.len(), 2);
+    let columns = &column_map
+        .get(&("required_col".to_string(), ColumnTypeCategory::Str))
+        .unwrap()
+        .columns;
+    assert_eq!(columns.len(), 2);
+    assert!(columns[0].is_none());
+    assert!(columns[1].is_none());
+}
+
+#[test]
+fn test_group_columns_required_column_is_above_all_columns_have_the_same_type_rule() {
+    let columnar1 = make_columnar("numbers", &[2i64]);
+    let columnar2 = make_columnar("numbers", &[2i64]);
+    let columnars = &[&columnar1, &columnar2];
+    let column_map: BTreeMap<(String, ColumnTypeCategory), GroupedColumnsHandle> =
+        group_columns_for_merge(columnars, &[("numbers".to_string(), ColumnType::U64)]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnTypeCategory::Numerical)));
+}
+
+#[test]
+fn test_missing_column() {
+    let columnar1 = make_columnar("numbers", &[-1i64]);
+    let columnar2 = make_columnar("numbers2", &[2u64]);
+    let columnars = &[&columnar1, &columnar2];
+    let column_map: BTreeMap<(String, ColumnTypeCategory), GroupedColumnsHandle> =
+        group_columns_for_merge(columnars, &[]).unwrap();
+    assert_eq!(column_map.len(), 2);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnTypeCategory::Numerical)));
+    {
+        let columns = &column_map
+            .get(&("numbers".to_string(), ColumnTypeCategory::Numerical))
+            .unwrap()
+            .columns;
+        assert!(columns[0].is_some());
+        assert!(columns[1].is_none());
+    }
+    {
+        let columns = &column_map
+            .get(&("numbers2".to_string(), ColumnTypeCategory::Numerical))
+            .unwrap()
+            .columns;
+        assert!(columns[0].is_none());
+        assert!(columns[1].is_some());
+    }
+}
+
+fn make_numerical_columnar_multiple_columns(
+    columns: &[(&str, &[&[NumericalValue]])],
+) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_numerical(row_id as u32, column_name, *val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(num_rows, &mut buffer).unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+#[track_caller]
+fn make_byte_columnar_multiple_columns(
+    columns: &[(&str, &[&[&[u8]]])],
+    num_rows: u32,
+) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        assert_eq!(
+            column_values.len(),
+            num_rows as usize,
+            "All columns must have `{num_rows}` rows"
+        );
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_bytes(row_id as u32, column_name, val);
+            }
+        }
+    }
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(num_rows, &mut buffer).unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+fn make_text_columnar_multiple_columns(columns: &[(&str, &[&[&str]])]) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_str(row_id as u32, column_name, val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(num_rows, &mut buffer).unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+#[test]
+fn test_merge_columnar_numbers() {
+    let columnar1 =
+        make_numerical_columnar_multiple_columns(&[("numbers", &[&[NumericalValue::from(-1f64)]])]);
+    let columnar2 = make_numerical_columnar_multiple_columns(&[(
+        "numbers",
+        &[&[], &[NumericalValue::from(-3f64)]],
+    )]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_docs(), 3);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("numbers").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::F64(vals) = dynamic_column else {
+        panic!()
+    };
+    assert_eq!(vals.get_cardinality(), Cardinality::Optional);
+    assert_eq!(vals.first(0u32), Some(-1f64));
+    assert_eq!(vals.first(1u32), None);
+    assert_eq!(vals.first(2u32), Some(-3f64));
+}
+
+#[test]
+fn test_merge_columnar_texts() {
+    let columnar1 = make_text_columnar_multiple_columns(&[("texts", &[&["a"]])]);
+    let columnar2 = make_text_columnar_multiple_columns(&[("texts", &[&[], &["b"]])]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_docs(), 3);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("texts").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::Str(vals) = dynamic_column else {
+        panic!()
+    };
+    assert_eq!(vals.ords().get_cardinality(), Cardinality::Optional);
+
+    let get_str_for_ord = |ord| {
+        let mut out = String::new();
+        vals.ord_to_str(ord, &mut out).unwrap();
+        out
+    };
+
+    assert_eq!(vals.dictionary.num_terms(), 2);
+    assert_eq!(get_str_for_ord(0), "a");
+    assert_eq!(get_str_for_ord(1), "b");
+
+    let get_str_for_row = |row_id| {
+        let term_ords: Vec<u64> = vals.term_ords(row_id).collect();
+        assert!(term_ords.len() <= 1);
+        let mut out = String::new();
+        if term_ords.len() == 1 {
+            vals.ord_to_str(term_ords[0], &mut out).unwrap();
+        }
+        out
+    };
+
+    assert_eq!(get_str_for_row(0), "a");
+    assert_eq!(get_str_for_row(1), "");
+    assert_eq!(get_str_for_row(2), "b");
+}
+
+#[test]
+fn test_merge_columnar_byte() {
+    let columnar1 = make_byte_columnar_multiple_columns(&[("bytes", &[&[b"bbbb"], &[b"baaa"]])], 2);
+    let columnar2 = make_byte_columnar_multiple_columns(&[("bytes", &[&[], &[b"a"]])], 2);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_docs(), 4);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("bytes").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::Bytes(vals) = dynamic_column else {
+        panic!()
+    };
+    let get_bytes_for_ord = |ord| {
+        let mut out = Vec::new();
+        vals.ord_to_bytes(ord, &mut out).unwrap();
+        out
+    };
+
+    assert_eq!(vals.dictionary.num_terms(), 3);
+    assert_eq!(get_bytes_for_ord(0), b"a");
+    assert_eq!(get_bytes_for_ord(1), b"baaa");
+    assert_eq!(get_bytes_for_ord(2), b"bbbb");
+
+    let get_bytes_for_row = |row_id| {
+        let term_ords: Vec<u64> = vals.term_ords(row_id).collect();
+        assert!(term_ords.len() <= 1);
+        let mut out = Vec::new();
+        if term_ords.len() == 1 {
+            vals.ord_to_bytes(term_ords[0], &mut out).unwrap();
+        }
+        out
+    };
+
+    assert_eq!(get_bytes_for_row(0), b"bbbb");
+    assert_eq!(get_bytes_for_row(1), b"baaa");
+    assert_eq!(get_bytes_for_row(2), b"");
+    assert_eq!(get_bytes_for_row(3), b"a");
+}
+
+#[test]
+fn test_merge_columnar_byte_with_missing() {
+    let columnar1 = make_byte_columnar_multiple_columns(&[], 3);
+    let columnar2 = make_byte_columnar_multiple_columns(&[("col", &[&[b"b"], &[]])], 2);
+    let columnar3 = make_byte_columnar_multiple_columns(
+        &[
+            ("col", &[&[], &[b"b"], &[b"a", b"b"]]),
+            ("col2", &[&[b"hello"], &[], &[b"a", b"b"]]),
+        ],
+        3,
+    );
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2, &columnar3];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_docs(), 3 + 2 + 3);
+    assert_eq!(columnar_reader.num_columns(), 2);
+    let cols = columnar_reader.read_columns("col").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::Bytes(vals) = dynamic_column else {
+        panic!()
+    };
+    let get_bytes_for_ord = |ord| {
+        let mut out = Vec::new();
+        vals.ord_to_bytes(ord, &mut out).unwrap();
+        out
+    };
+    assert_eq!(vals.dictionary.num_terms(), 2);
+    assert_eq!(get_bytes_for_ord(0), b"a");
+    assert_eq!(get_bytes_for_ord(1), b"b");
+    let get_bytes_for_row = |row_id| {
+        let terms: Vec<Vec<u8>> = vals
+            .term_ords(row_id)
+            .map(|term_ord| {
+                let mut out = Vec::new();
+                vals.ord_to_bytes(term_ord, &mut out).unwrap();
+                out
+            })
+            .collect();
+        terms
+    };
+    assert!(get_bytes_for_row(0).is_empty());
+    assert!(get_bytes_for_row(1).is_empty());
+    assert!(get_bytes_for_row(2).is_empty());
+    assert_eq!(get_bytes_for_row(3), vec![b"b".to_vec()]);
+    assert!(get_bytes_for_row(4).is_empty());
+    assert!(get_bytes_for_row(5).is_empty());
+    assert_eq!(get_bytes_for_row(6), vec![b"b".to_vec()]);
+    assert_eq!(get_bytes_for_row(7), vec![b"a".to_vec(), b"b".to_vec()]);
+}
+
+#[test]
+fn test_merge_columnar_different_types() {
+    let columnar1 = make_text_columnar_multiple_columns(&[("mixed", &[&["a"]])]);
+    let columnar2 = make_text_columnar_multiple_columns(&[("mixed", &[&[], &["b"]])]);
+    let columnar3 = make_columnar("mixed", &[1i64]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2, &columnar3];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_docs(), 4);
+    assert_eq!(columnar_reader.num_columns(), 2);
+    let cols = columnar_reader.read_columns("mixed").unwrap();
+
+    // numeric column
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::I64(vals) = dynamic_column else {
+        panic!()
+    };
+    assert_eq!(vals.get_cardinality(), Cardinality::Optional);
+    assert_eq!(vals.values_for_doc(0).collect_vec(), Vec::<i64>::new());
+    assert_eq!(vals.values_for_doc(1).collect_vec(), Vec::<i64>::new());
+    assert_eq!(vals.values_for_doc(2).collect_vec(), Vec::<i64>::new());
+    assert_eq!(vals.values_for_doc(3).collect_vec(), vec![1]);
+    assert_eq!(vals.values_for_doc(4).collect_vec(), Vec::<i64>::new());
+
+    // text column
+    let dynamic_column = cols[1].open().unwrap();
+    let DynamicColumn::Str(vals) = dynamic_column else {
+        panic!()
+    };
+    assert_eq!(vals.ords().get_cardinality(), Cardinality::Optional);
+    let get_str_for_ord = |ord| {
+        let mut out = String::new();
+        vals.ord_to_str(ord, &mut out).unwrap();
+        out
+    };
+
+    assert_eq!(vals.dictionary.num_terms(), 2);
+    assert_eq!(get_str_for_ord(0), "a");
+    assert_eq!(get_str_for_ord(1), "b");
+
+    let get_str_for_row = |row_id| {
+        let term_ords: Vec<String> = vals
+            .term_ords(row_id)
+            .map(|el| {
+                let mut out = String::new();
+                vals.ord_to_str(el, &mut out).unwrap();
+                out
+            })
+            .collect();
+        term_ords
+    };
+
+    assert_eq!(get_str_for_row(0), vec!["a".to_string()]);
+    assert_eq!(get_str_for_row(1), Vec::<String>::new());
+    assert_eq!(get_str_for_row(2), vec!["b".to_string()]);
+    assert_eq!(get_str_for_row(3), Vec::<String>::new());
+}
+
+#[test]
+fn test_merge_columnar_different_empty_cardinality() {
+    let columnar1 = make_text_columnar_multiple_columns(&[("mixed", &[&["a"]])]);
+    let columnar2 = make_columnar("mixed", &[1i64]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_docs(), 2);
+    assert_eq!(columnar_reader.num_columns(), 2);
+    let cols = columnar_reader.read_columns("mixed").unwrap();
+
+    // numeric column
+    let dynamic_column = cols[0].open().unwrap();
+    assert_eq!(dynamic_column.get_cardinality(), Cardinality::Optional);
+
+    // text column
+    let dynamic_column = cols[1].open().unwrap();
+    assert_eq!(dynamic_column.get_cardinality(), Cardinality::Optional);
+}
+
+#[derive(Debug, Clone)]
+struct ColumnSpec {
+    column_name: String,
+    /// (row_id, term)
+    terms: Vec<(RowId, Vec<u8>)>,
+}
+
+#[derive(Clone, Debug)]
+struct ColumnarSpec {
+    columns: Vec<ColumnSpec>,
+}
+
+/// Generate a random (row_id, term) pair:
+///  - row_id in [0..10]
+///  - term is either from POSSIBLE_TERMS or random bytes
+fn rowid_and_term_strategy() -> impl Strategy<Value = (RowId, Vec<u8>)> {
+    const POSSIBLE_TERMS: &[&[u8]] = &[b"a", b"b", b"allo"];
+
+    let term_strat = prop_oneof![
+        // pick from the fixed list
+        (0..POSSIBLE_TERMS.len()).prop_map(|i| POSSIBLE_TERMS[i].to_vec()),
+        // or random bytes (length 0..10)
+        prop::collection::vec(any::<u8>(), 0..10),
+    ];
+
+    (0u32..11, term_strat)
+}
+
+/// Generate one ColumnSpec, with a random name and a random list of (row_id, term).
+/// We sort it by row_id so that data is in ascending order.
+fn column_spec_strategy() -> impl Strategy<Value = ColumnSpec> {
+    let column_name = prop_oneof![
+        Just("col".to_string()),
+        Just("col2".to_string()),
+        "col.*".prop_map(|s| s),
+    ];
+
+    // We'll produce 0..8 (rowid,term) entries for this column
+    let data_strat = vec(rowid_and_term_strategy(), 0..8).prop_map(|mut pairs| {
+        // Sort by row_id
+        pairs.sort_by_key(|(row_id, _)| *row_id);
+        pairs
+    });
+
+    (column_name, data_strat).prop_map(|(name, data)| ColumnSpec {
+        column_name: name,
+        terms: data,
+    })
+}
+
+/// Strategy to generate an ColumnarSpec
+fn columnar_strategy() -> impl Strategy<Value = ColumnarSpec> {
+    vec(column_spec_strategy(), 0..3).prop_map(|columns| ColumnarSpec { columns })
+}
+
+/// Strategy to generate multiple ColumnarSpecs, each of which we will treat
+/// as one "columnar" to be merged together.
+fn columnars_strategy() -> impl Strategy<Value = Vec<ColumnarSpec>> {
+    vec(columnar_strategy(), 1..4)
+}
+
+/// Build a `ColumnarReader` from a `ColumnarSpec`
+fn build_columnar(spec: &ColumnarSpec) -> ColumnarReader {
+    let mut writer = ColumnarWriter::default();
+    let mut max_row_id = 0;
+    for col in &spec.columns {
+        for &(row_id, ref term) in &col.terms {
+            writer.record_bytes(row_id, &col.column_name, term);
+            max_row_id = max_row_id.max(row_id);
+        }
+    }
+
+    let mut buffer = Vec::new();
+    writer.serialize(max_row_id + 1, &mut buffer).unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+proptest! {
+    // We just test that the merge_columnar function doesn't crash.
+    #![proptest_config(ProptestConfig::with_cases(256))]
+    #[test]
+    fn test_merge_columnar_bytes_no_crash(columnars in columnars_strategy(), second_merge_columnars in columnars_strategy()) {
+        let columnars: Vec<ColumnarReader> = columnars.iter()
+            .map(build_columnar)
+            .collect();
+
+        let mut out = Vec::new();
+        let columnar_refs: Vec<&ColumnarReader> = columnars.iter().collect();
+        let stack_merge_order = StackMergeOrder::stack(&columnar_refs);
+        merge_columnar(
+            &columnar_refs,
+            &[],
+            MergeRowOrder::Stack(stack_merge_order),
+            &mut out,
+        ).unwrap();
+
+        let merged_reader = ColumnarReader::open(out).unwrap();
+
+        // Merge the second set of columnars with the result of the first merge
+        let mut columnars: Vec<ColumnarReader> = second_merge_columnars.iter()
+            .map(build_columnar)
+            .collect();
+        columnars.push(merged_reader);
+        let mut out = Vec::new();
+        let columnar_refs: Vec<&ColumnarReader> = columnars.iter().collect();
+        let stack_merge_order = StackMergeOrder::stack(&columnar_refs);
+        merge_columnar(
+            &columnar_refs,
+            &[],
+            MergeRowOrder::Stack(stack_merge_order),
+            &mut out,
+        ).unwrap();
+
+    }
+}

columnar/src/columnar/mod.rs

@@ -0,0 +1,13 @@
+mod column_type;
+mod format_version;
+mod merge;
+mod reader;
+mod writer;
+
+pub use column_type::{ColumnType, HasAssociatedColumnType};
+pub use format_version::{CURRENT_VERSION, Version};
+#[cfg(test)]
+pub(crate) use merge::ColumnTypeCategory;
+pub use merge::{MergeRowOrder, ShuffleMergeOrder, StackMergeOrder, merge_columnar};
+pub use reader::ColumnarReader;
+pub use writer::ColumnarWriter;

columnar/src/columnar/reader/mod.rs

@@ -0,0 +1,318 @@
+use std::{fmt, io, mem};
+
+use common::BinarySerializable;
+use common::file_slice::FileSlice;
+use common::json_path_writer::JSON_PATH_SEGMENT_SEP;
+use sstable::{Dictionary, RangeSSTable};
+
+use crate::columnar::{ColumnType, format_version};
+use crate::dynamic_column::DynamicColumnHandle;
+use crate::{RowId, Version};
+
+fn io_invalid_data(msg: String) -> io::Error {
+    io::Error::new(io::ErrorKind::InvalidData, msg)
+}
+
+/// The ColumnarReader makes it possible to access a set of columns
+/// associated to field names.
+#[derive(Clone)]
+pub struct ColumnarReader {
+    column_dictionary: Dictionary<RangeSSTable>,
+    column_data: FileSlice,
+    num_docs: RowId,
+    format_version: Version,
+}
+
+impl fmt::Debug for ColumnarReader {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        let num_rows = self.num_docs();
+        let columns = self.list_columns().unwrap();
+        let num_cols = columns.len();
+        let mut debug_struct = f.debug_struct("Columnar");
+        debug_struct
+            .field("num_rows", &num_rows)
+            .field("num_cols", &num_cols);
+        for (col_name, dynamic_column_handle) in columns.into_iter().take(5) {
+            let col = dynamic_column_handle.open().unwrap();
+            if col.num_values() > 10 {
+                debug_struct.field(&col_name, &"..");
+            } else {
+                debug_struct.field(&col_name, &col);
+            }
+        }
+        if num_cols > 5 {
+            debug_struct.finish_non_exhaustive()?;
+        } else {
+            debug_struct.finish()?;
+        }
+        Ok(())
+    }
+}
+
+/// Functions by both the async/sync code listing columns.
+/// It takes a stream from the column sstable and return the list of
+/// `DynamicColumn` available in it.
+fn read_all_columns_in_stream(
+    mut stream: sstable::Streamer<'_, RangeSSTable>,
+    column_data: &FileSlice,
+    format_version: Version,
+) -> io::Result<Vec<DynamicColumnHandle>> {
+    let mut results = Vec::new();
+    while stream.advance() {
+        let key_bytes: &[u8] = stream.key();
+        let Some(column_code) = key_bytes.last().copied() else {
+            return Err(io_invalid_data("Empty column name.".to_string()));
+        };
+        let column_type = ColumnType::try_from_code(column_code)
+            .map_err(|_| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
+        let range = stream.value();
+        let file_slice = column_data.slice(range.start as usize..range.end as usize);
+        let dynamic_column_handle = DynamicColumnHandle {
+            file_slice,
+            column_type,
+            format_version,
+        };
+        results.push(dynamic_column_handle);
+    }
+    Ok(results)
+}
+
+fn column_dictionary_prefix_for_column_name(column_name: &str) -> String {
+    // Each column is a associated to a given `column_key`,
+    // that starts by `column_name\0column_header`.
+    //
+    // Listing the columns associated to the given column name is therefore equivalent to
+    // listing `column_key` with the prefix `column_name\0`.
+    format!("{}{}", column_name, '\0')
+}
+
+fn column_dictionary_prefix_for_subpath(root_path: &str) -> String {
+    format!("{}{}", root_path, JSON_PATH_SEGMENT_SEP as char)
+}
+
+impl ColumnarReader {
+    /// Opens a new Columnar file.
+    pub fn open<F>(file_slice: F) -> io::Result<ColumnarReader>
+    where FileSlice: From<F> {
+        Self::open_inner(file_slice.into())
+    }
+
+    fn open_inner(file_slice: FileSlice) -> io::Result<ColumnarReader> {
+        let (file_slice_without_sstable_len, footer_slice) = file_slice
+            .split_from_end(mem::size_of::<u64>() + 4 + format_version::VERSION_FOOTER_NUM_BYTES);
+        let footer_bytes = footer_slice.read_bytes()?;
+        let sstable_len = u64::deserialize(&mut &footer_bytes[0..8])?;
+        let num_rows = u32::deserialize(&mut &footer_bytes[8..12])?;
+        let version_footer_bytes: [u8; format_version::VERSION_FOOTER_NUM_BYTES] =
+            footer_bytes[12..].try_into().unwrap();
+        let format_version = format_version::parse_footer(version_footer_bytes)?;
+        let (column_data, sstable) =
+            file_slice_without_sstable_len.split_from_end(sstable_len as usize);
+        let column_dictionary = Dictionary::open(sstable)?;
+        Ok(ColumnarReader {
+            column_dictionary,
+            column_data,
+            num_docs: num_rows,
+            format_version,
+        })
+    }
+
+    pub fn num_docs(&self) -> RowId {
+        self.num_docs
+    }
+    // Iterate over the columns in a sorted way
+    pub fn iter_columns(
+        &self,
+    ) -> io::Result<impl Iterator<Item = (String, DynamicColumnHandle)> + '_> {
+        let mut stream = self.column_dictionary.stream()?;
+        Ok(std::iter::from_fn(move || {
+            if stream.advance() {
+                let key_bytes: &[u8] = stream.key();
+                let column_code: u8 = key_bytes.last().cloned().unwrap();
+                // TODO Error Handling. The API gets quite ugly when returning the error here, so
+                // instead we could just check the first N columns upfront.
+                let column_type: ColumnType = ColumnType::try_from_code(column_code)
+                    .map_err(|_| io_invalid_data(format!("Unknown column code `{column_code}`")))
+                    .unwrap();
+                let range = stream.value().clone();
+                let column_name =
+                // The last two bytes are respectively the 0u8 separator and the column_type.
+                String::from_utf8_lossy(&key_bytes[..key_bytes.len() - 2]).to_string();
+                let file_slice = self
+                    .column_data
+                    .slice(range.start as usize..range.end as usize);
+                let column_handle = DynamicColumnHandle {
+                    file_slice,
+                    column_type,
+                    format_version: self.format_version,
+                };
+                Some((column_name, column_handle))
+            } else {
+                None
+            }
+        }))
+    }
+
+    // TODO Add unit tests
+    pub fn list_columns(&self) -> io::Result<Vec<(String, DynamicColumnHandle)>> {
+        Ok(self.iter_columns()?.collect())
+    }
+
+    pub async fn read_columns_async(
+        &self,
+        column_name: &str,
+    ) -> io::Result<Vec<DynamicColumnHandle>> {
+        let prefix = column_dictionary_prefix_for_column_name(column_name);
+        let stream = self
+            .column_dictionary
+            .prefix_range(prefix)
+            .into_stream_async()
+            .await?;
+        read_all_columns_in_stream(stream, &self.column_data, self.format_version)
+    }
+
+    /// Get all columns for the given column name.
+    ///
+    /// There can be more than one column associated to a given column name, provided they have
+    /// different types.
+    pub fn read_columns(&self, column_name: &str) -> io::Result<Vec<DynamicColumnHandle>> {
+        let prefix = column_dictionary_prefix_for_column_name(column_name);
+        let stream = self.column_dictionary.prefix_range(prefix).into_stream()?;
+        read_all_columns_in_stream(stream, &self.column_data, self.format_version)
+    }
+
+    pub async fn read_subpath_columns_async(
+        &self,
+        root_path: &str,
+    ) -> io::Result<Vec<DynamicColumnHandle>> {
+        let prefix = column_dictionary_prefix_for_subpath(root_path);
+        let stream = self
+            .column_dictionary
+            .prefix_range(prefix)
+            .into_stream_async()
+            .await?;
+        read_all_columns_in_stream(stream, &self.column_data, self.format_version)
+    }
+
+    /// Get all inner columns for a given JSON prefix, i.e columns for which the name starts
+    /// with the prefix then contain the [`JSON_PATH_SEGMENT_SEP`].
+    ///
+    /// There can be more than one column associated to each path within the JSON structure,
+    /// provided they have different types.
+    pub fn read_subpath_columns(&self, root_path: &str) -> io::Result<Vec<DynamicColumnHandle>> {
+        let prefix = column_dictionary_prefix_for_subpath(root_path);
+        let stream = self
+            .column_dictionary
+            .prefix_range(prefix.as_bytes())
+            .into_stream()?;
+        read_all_columns_in_stream(stream, &self.column_data, self.format_version)
+    }
+
+    /// Return the number of columns in the columnar.
+    pub fn num_columns(&self) -> usize {
+        self.column_dictionary.num_terms()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use common::json_path_writer::JSON_PATH_SEGMENT_SEP;
+
+    use crate::{ColumnType, ColumnarReader, ColumnarWriter};
+
+    #[test]
+    fn test_list_columns() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("col1", ColumnType::Str, false);
+        columnar_writer.record_column_type("col2", ColumnType::U64, false);
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(1, &mut buffer).unwrap();
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        let columns = columnar.list_columns().unwrap();
+        assert_eq!(columns.len(), 2);
+        assert_eq!(&columns[0].0, "col1");
+        assert_eq!(columns[0].1.column_type(), ColumnType::Str);
+        assert_eq!(&columns[1].0, "col2");
+        assert_eq!(columns[1].1.column_type(), ColumnType::U64);
+    }
+
+    #[test]
+    fn test_list_columns_strict_typing_prevents_coercion() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("count", ColumnType::U64, false);
+        columnar_writer.record_numerical(1, "count", 1u64);
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(2, &mut buffer).unwrap();
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        let columns = columnar.list_columns().unwrap();
+        assert_eq!(columns.len(), 1);
+        assert_eq!(&columns[0].0, "count");
+        assert_eq!(columns[0].1.column_type(), ColumnType::U64);
+    }
+
+    #[test]
+    fn test_read_columns() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("col", ColumnType::U64, false);
+        columnar_writer.record_numerical(1, "col", 1u64);
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(2, &mut buffer).unwrap();
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        {
+            let columns = columnar.read_columns("col").unwrap();
+            assert_eq!(columns.len(), 1);
+            assert_eq!(columns[0].column_type(), ColumnType::U64);
+        }
+        {
+            let columns = columnar.read_columns("other").unwrap();
+            assert_eq!(columns.len(), 0);
+        }
+    }
+
+    #[test]
+    fn test_read_subpath_columns() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_str(
+            0,
+            &format!("col1{}subcol1", JSON_PATH_SEGMENT_SEP as char),
+            "hello",
+        );
+        columnar_writer.record_numerical(
+            0,
+            &format!("col1{}subcol2", JSON_PATH_SEGMENT_SEP as char),
+            1i64,
+        );
+        columnar_writer.record_str(1, "col1", "hello");
+        columnar_writer.record_str(0, "col2", "hello");
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(2, &mut buffer).unwrap();
+
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        {
+            let columns = columnar.read_subpath_columns("col1").unwrap();
+            assert_eq!(columns.len(), 2);
+            assert_eq!(columns[0].column_type(), ColumnType::Str);
+            assert_eq!(columns[1].column_type(), ColumnType::I64);
+        }
+        {
+            let columns = columnar.read_subpath_columns("col1.subcol1").unwrap();
+            assert_eq!(columns.len(), 0);
+        }
+        {
+            let columns = columnar.read_subpath_columns("col2").unwrap();
+            assert_eq!(columns.len(), 0);
+        }
+        {
+            let columns = columnar.read_subpath_columns("other").unwrap();
+            assert_eq!(columns.len(), 0);
+        }
+    }
+
+    #[test]
+    #[should_panic(expected = "Input type forbidden")]
+    fn test_list_columns_strict_typing_panics_on_wrong_types() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("count", ColumnType::U64, false);
+        columnar_writer.record_numerical(1, "count", 1i64);
+    }
+}

columnar/src/columnar/writer/column_operation.rs

@@ -0,0 +1,359 @@
+use std::net::Ipv6Addr;
+
+use crate::dictionary::UnorderedId;
+use crate::utils::{place_bits, pop_first_byte, select_bits};
+use crate::value::NumericalValue;
+use crate::{InvalidData, NumericalType, RowId};
+
+/// When we build a columnar dataframe, we first just group
+/// all mutations per column, and appends them in append-only buffer
+/// in the stacker.
+///
+/// These ColumnOperation<T> are therefore serialize/deserialized
+/// in memory.
+///
+/// We represents all of these operations as `ColumnOperation`.
+#[derive(Eq, PartialEq, Debug, Clone, Copy)]
+pub(super) enum ColumnOperation<T> {
+    NewDoc(RowId),
+    Value(T),
+}
+
+#[derive(Copy, Clone, Eq, PartialEq, Debug)]
+struct ColumnOperationMetadata {
+    op_type: ColumnOperationType,
+    len: u8,
+}
+
+impl ColumnOperationMetadata {
+    fn to_code(self) -> u8 {
+        place_bits::<0, 6>(self.len) | place_bits::<6, 8>(self.op_type.to_code())
+    }
+
+    fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        let len = select_bits::<0, 6>(code);
+        let typ_code = select_bits::<6, 8>(code);
+        let column_type = ColumnOperationType::try_from_code(typ_code)?;
+        Ok(ColumnOperationMetadata {
+            op_type: column_type,
+            len,
+        })
+    }
+}
+
+#[derive(Copy, Clone, Eq, PartialEq, Debug)]
+#[repr(u8)]
+enum ColumnOperationType {
+    NewDoc = 0u8,
+    AddValue = 1u8,
+}
+
+impl ColumnOperationType {
+    pub fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        match code {
+            0 => Ok(Self::NewDoc),
+            1 => Ok(Self::AddValue),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+impl<V: SymbolValue> ColumnOperation<V> {
+    pub(super) fn serialize(self) -> impl AsRef<[u8]> {
+        let mut minibuf = MiniBuffer::default();
+        let column_op_metadata = match self {
+            ColumnOperation::NewDoc(new_doc) => {
+                let symbol_len = new_doc.serialize(&mut minibuf.bytes[1..]);
+                ColumnOperationMetadata {
+                    op_type: ColumnOperationType::NewDoc,
+                    len: symbol_len,
+                }
+            }
+            ColumnOperation::Value(val) => {
+                let symbol_len = val.serialize(&mut minibuf.bytes[1..]);
+                ColumnOperationMetadata {
+                    op_type: ColumnOperationType::AddValue,
+                    len: symbol_len,
+                }
+            }
+        };
+        minibuf.bytes[0] = column_op_metadata.to_code();
+        // +1 for the metadata
+        minibuf.len = 1 + column_op_metadata.len;
+        minibuf
+    }
+
+    /// Deserialize a column operation.
+    /// Returns None if the buffer is empty.
+    ///
+    /// Panics if the payload is invalid:
+    /// this deserialize method is meant to target in memory.
+    pub(super) fn deserialize(bytes: &mut &[u8]) -> Option<Self> {
+        let column_op_metadata_byte = pop_first_byte(bytes)?;
+        let column_op_metadata = ColumnOperationMetadata::try_from_code(column_op_metadata_byte)
+            .expect("Invalid op metadata byte");
+        let symbol_bytes: &[u8];
+        (symbol_bytes, *bytes) = bytes.split_at(column_op_metadata.len as usize);
+        match column_op_metadata.op_type {
+            ColumnOperationType::NewDoc => {
+                let new_doc = u32::deserialize(symbol_bytes);
+                Some(ColumnOperation::NewDoc(new_doc))
+            }
+            ColumnOperationType::AddValue => {
+                let value = V::deserialize(symbol_bytes);
+                Some(ColumnOperation::Value(value))
+            }
+        }
+    }
+}
+
+impl<T> From<T> for ColumnOperation<T> {
+    fn from(value: T) -> Self {
+        ColumnOperation::Value(value)
+    }
+}
+
+// Serialization trait very local to the writer.
+// As we write fast fields, we accumulate them in "in memory".
+// In order to limit memory usage, and in order
+// to benefit from the stacker, we do this by serialization our data
+// as "Symbols".
+pub(super) trait SymbolValue: Clone + Copy {
+    // Serializes the symbol into the given buffer.
+    // Returns the number of bytes written into the buffer.
+    /// # Panics
+    /// May not exceed 9bytes
+    fn serialize(self, buffer: &mut [u8]) -> u8;
+    // Panics if invalid
+    fn deserialize(bytes: &[u8]) -> Self;
+}
+
+impl SymbolValue for bool {
+    fn serialize(self, buffer: &mut [u8]) -> u8 {
+        buffer[0] = u8::from(self);
+        1u8
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        bytes[0] == 1u8
+    }
+}
+
+impl SymbolValue for Ipv6Addr {
+    fn serialize(self, buffer: &mut [u8]) -> u8 {
+        buffer[0..16].copy_from_slice(&self.octets());
+        16
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        let octets: [u8; 16] = bytes[0..16].try_into().unwrap();
+        Ipv6Addr::from(octets)
+    }
+}
+
+#[derive(Default)]
+struct MiniBuffer {
+    pub bytes: [u8; 17],
+    pub len: u8,
+}
+
+impl AsRef<[u8]> for MiniBuffer {
+    fn as_ref(&self) -> &[u8] {
+        &self.bytes[..self.len as usize]
+    }
+}
+
+impl SymbolValue for NumericalValue {
+    fn deserialize(mut bytes: &[u8]) -> Self {
+        let type_code = pop_first_byte(&mut bytes).unwrap();
+        let symbol_type = NumericalType::try_from_code(type_code).unwrap();
+        let mut octet: [u8; 8] = [0u8; 8];
+        octet[..bytes.len()].copy_from_slice(bytes);
+        match symbol_type {
+            NumericalType::U64 => {
+                let val: u64 = u64::from_le_bytes(octet);
+                NumericalValue::U64(val)
+            }
+            NumericalType::I64 => {
+                let encoded: u64 = u64::from_le_bytes(octet);
+                let val: i64 = decode_zig_zag(encoded);
+                NumericalValue::I64(val)
+            }
+            NumericalType::F64 => {
+                debug_assert_eq!(bytes.len(), 8);
+                let val: f64 = f64::from_le_bytes(octet);
+                NumericalValue::F64(val)
+            }
+        }
+    }
+
+    /// F64: Serialize with a fixed size of 9 bytes
+    /// U64: Serialize without leading zeroes
+    /// I64: ZigZag encoded and serialize without leading zeroes
+    fn serialize(self, output: &mut [u8]) -> u8 {
+        match self {
+            NumericalValue::F64(val) => {
+                output[0] = NumericalType::F64 as u8;
+                output[1..9].copy_from_slice(&val.to_le_bytes());
+                9u8
+            }
+            NumericalValue::U64(val) => {
+                let len = compute_num_bytes_for_u64(val) as u8;
+                output[0] = NumericalType::U64 as u8;
+                output[1..9].copy_from_slice(&val.to_le_bytes());
+                len + 1u8
+            }
+            NumericalValue::I64(val) => {
+                let zig_zag_encoded = encode_zig_zag(val);
+                let len = compute_num_bytes_for_u64(zig_zag_encoded) as u8;
+                output[0] = NumericalType::I64 as u8;
+                output[1..9].copy_from_slice(&zig_zag_encoded.to_le_bytes());
+                len + 1u8
+            }
+        }
+    }
+}
+
+impl SymbolValue for u32 {
+    fn serialize(self, output: &mut [u8]) -> u8 {
+        let len = compute_num_bytes_for_u64(self as u64);
+        output[0..4].copy_from_slice(&self.to_le_bytes());
+        len as u8
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        let mut quartet: [u8; 4] = [0u8; 4];
+        quartet[..bytes.len()].copy_from_slice(bytes);
+        u32::from_le_bytes(quartet)
+    }
+}
+
+impl SymbolValue for UnorderedId {
+    fn serialize(self, output: &mut [u8]) -> u8 {
+        self.0.serialize(output)
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        UnorderedId(u32::deserialize(bytes))
+    }
+}
+
+fn compute_num_bytes_for_u64(val: u64) -> usize {
+    let msb = (64u32 - val.leading_zeros()) as usize;
+    (msb + 7) / 8
+}
+
+fn encode_zig_zag(n: i64) -> u64 {
+    ((n << 1) ^ (n >> 63)) as u64
+}
+
+fn decode_zig_zag(n: u64) -> i64 {
+    ((n >> 1) as i64) ^ (-((n & 1) as i64))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[track_caller]
+    fn test_zig_zag_aux(val: i64) {
+        let encoded = super::encode_zig_zag(val);
+        assert_eq!(decode_zig_zag(encoded), val);
+        if let Some(abs_val) = val.checked_abs() {
+            let abs_val = abs_val as u64;
+            assert!(encoded <= abs_val * 2);
+        }
+    }
+
+    #[test]
+    fn test_zig_zag() {
+        assert_eq!(encode_zig_zag(0i64), 0u64);
+        assert_eq!(encode_zig_zag(-1i64), 1u64);
+        assert_eq!(encode_zig_zag(1i64), 2u64);
+        test_zig_zag_aux(0i64);
+        test_zig_zag_aux(i64::MIN);
+        test_zig_zag_aux(i64::MAX);
+    }
+
+    use proptest::prelude::any;
+    use proptest::proptest;
+
+    proptest! {
+        #[test]
+        fn test_proptest_zig_zag(val in any::<i64>()) {
+            test_zig_zag_aux(val);
+        }
+    }
+
+    #[test]
+    fn test_column_op_metadata_byte_serialization() {
+        for len in 0..=15 {
+            for op_type in [ColumnOperationType::AddValue, ColumnOperationType::NewDoc] {
+                let column_op_metadata = ColumnOperationMetadata { op_type, len };
+                let column_op_metadata_code = column_op_metadata.to_code();
+                let serdeser_metadata =
+                    ColumnOperationMetadata::try_from_code(column_op_metadata_code).unwrap();
+                assert_eq!(column_op_metadata, serdeser_metadata);
+            }
+        }
+    }
+
+    #[track_caller]
+    fn ser_deser_symbol(column_op: ColumnOperation<NumericalValue>) {
+        let buf = column_op.serialize();
+        let mut buffer = buf.as_ref().to_vec();
+        buffer.extend_from_slice(b"234234");
+        let mut bytes = &buffer[..];
+        let serdeser_symbol = ColumnOperation::deserialize(&mut bytes).unwrap();
+        assert_eq!(bytes.len() + buf.as_ref().len(), buffer.len());
+        assert_eq!(column_op, serdeser_symbol);
+    }
+
+    #[test]
+    fn test_compute_num_bytes_for_u64() {
+        assert_eq!(compute_num_bytes_for_u64(0), 0);
+        assert_eq!(compute_num_bytes_for_u64(1), 1);
+        assert_eq!(compute_num_bytes_for_u64(255), 1);
+        assert_eq!(compute_num_bytes_for_u64(256), 2);
+        assert_eq!(compute_num_bytes_for_u64((1 << 16) - 1), 2);
+        assert_eq!(compute_num_bytes_for_u64(1 << 16), 3);
+    }
+
+    #[test]
+    fn test_symbol_serialization() {
+        ser_deser_symbol(ColumnOperation::NewDoc(0));
+        ser_deser_symbol(ColumnOperation::NewDoc(3));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(0i64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(1i64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(257u64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(-257i64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(i64::MIN)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(0u64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(u64::MIN)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(u64::MAX)));
+    }
+
+    fn test_column_operation_unordered_aux(val: u32, expected_len: usize) {
+        let column_op = ColumnOperation::Value(UnorderedId(val));
+        let minibuf = column_op.serialize();
+        assert_eq!({ minibuf.as_ref().len() }, expected_len);
+        let mut buf = minibuf.as_ref().to_vec();
+        buf.extend_from_slice(&[2, 2, 2, 2, 2, 2]);
+        let mut cursor = &buf[..];
+        let column_op_serdeser: ColumnOperation<UnorderedId> =
+            ColumnOperation::deserialize(&mut cursor).unwrap();
+        assert_eq!(column_op_serdeser, ColumnOperation::Value(UnorderedId(val)));
+        assert_eq!(cursor.len() + expected_len, buf.len());
+    }
+
+    #[test]
+    fn test_column_operation_unordered() {
+        test_column_operation_unordered_aux(300u32, 3);
+        test_column_operation_unordered_aux(1u32, 2);
+        test_column_operation_unordered_aux(0u32, 1);
+    }
+}

columnar/src/columnar/writer/column_writers.rs

@@ -0,0 +1,340 @@
+use std::cmp::Ordering;
+
+use stacker::{ExpUnrolledLinkedList, MemoryArena};
+
+use crate::columnar::writer::column_operation::{ColumnOperation, SymbolValue};
+use crate::dictionary::{DictionaryBuilder, UnorderedId};
+use crate::{Cardinality, NumericalType, NumericalValue, RowId};
+
+#[derive(Copy, Clone, Debug, Eq, PartialEq)]
+#[repr(u8)]
+enum DocumentStep {
+    Same = 0,
+    Next = 1,
+    Skipped = 2,
+}
+
+#[inline(always)]
+fn delta_with_last_doc(last_doc_opt: Option<u32>, doc: u32) -> DocumentStep {
+    let expected_next_doc = last_doc_opt.map(|last_doc| last_doc + 1).unwrap_or(0u32);
+    match doc.cmp(&expected_next_doc) {
+        Ordering::Less => DocumentStep::Same,
+        Ordering::Equal => DocumentStep::Next,
+        Ordering::Greater => DocumentStep::Skipped,
+    }
+}
+
+#[derive(Copy, Clone, Default)]
+pub struct ColumnWriter {
+    // Detected cardinality of the column so far.
+    cardinality: Cardinality,
+    // Last document inserted.
+    // None if no doc has been added yet.
+    last_doc_opt: Option<u32>,
+    // Buffer containing the serialized values.
+    values: ExpUnrolledLinkedList,
+}
+
+impl ColumnWriter {
+    /// Returns an iterator over the Symbol that have been recorded
+    /// for the given column.
+    pub(super) fn operation_iterator<'a, V: SymbolValue>(
+        &self,
+        arena: &MemoryArena,
+        buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<V>> + 'a + use<'a, V> {
+        buffer.clear();
+        self.values.read_to_end(arena, buffer);
+        let mut cursor: &[u8] = &buffer[..];
+        std::iter::from_fn(move || ColumnOperation::deserialize(&mut cursor))
+    }
+
+    /// Records a change of the document being recorded.
+    ///
+    /// This function will also update the cardinality of the column
+    /// if necessary.
+    pub(super) fn record<S: SymbolValue>(&mut self, doc: RowId, value: S, arena: &mut MemoryArena) {
+        // Difference between `doc` and the last doc.
+        match delta_with_last_doc(self.last_doc_opt, doc) {
+            DocumentStep::Same => {
+                // This is the last encounterred document.
+                self.cardinality = Cardinality::Multivalued;
+            }
+            DocumentStep::Next => {
+                self.last_doc_opt = Some(doc);
+                self.write_symbol::<S>(ColumnOperation::NewDoc(doc), arena);
+            }
+            DocumentStep::Skipped => {
+                self.cardinality = self.cardinality.max(Cardinality::Optional);
+                self.last_doc_opt = Some(doc);
+                self.write_symbol::<S>(ColumnOperation::NewDoc(doc), arena);
+            }
+        }
+        self.write_symbol(ColumnOperation::Value(value), arena);
+    }
+
+    // Get the cardinality.
+    // The overall number of docs in the column is necessary to
+    // deal with the case where the all docs contain 1 value, except some documents
+    // at the end of the column.
+    pub(crate) fn get_cardinality(&self, num_docs: RowId) -> Cardinality {
+        match delta_with_last_doc(self.last_doc_opt, num_docs) {
+            DocumentStep::Same | DocumentStep::Next => self.cardinality,
+            DocumentStep::Skipped => self.cardinality.max(Cardinality::Optional),
+        }
+    }
+
+    /// Appends a new symbol to the `ColumnWriter`.
+    fn write_symbol<V: SymbolValue>(
+        &mut self,
+        column_operation: ColumnOperation<V>,
+        arena: &mut MemoryArena,
+    ) {
+        self.values
+            .writer(arena)
+            .extend_from_slice(column_operation.serialize().as_ref());
+    }
+}
+
+#[derive(Clone, Copy, Default)]
+pub(crate) struct NumericalColumnWriter {
+    compatible_numerical_types: CompatibleNumericalTypes,
+    column_writer: ColumnWriter,
+}
+
+impl NumericalColumnWriter {
+    pub fn force_numerical_type(&mut self, numerical_type: NumericalType) {
+        assert!(
+            self.compatible_numerical_types
+                .is_type_accepted(numerical_type)
+        );
+        self.compatible_numerical_types = CompatibleNumericalTypes::StaticType(numerical_type);
+    }
+}
+
+/// State used to store what types are still acceptable
+/// after having seen a set of numerical values.
+#[derive(Clone, Copy)]
+pub(crate) enum CompatibleNumericalTypes {
+    Dynamic {
+        all_values_within_i64_range: bool,
+        all_values_within_u64_range: bool,
+    },
+    StaticType(NumericalType),
+}
+
+impl Default for CompatibleNumericalTypes {
+    fn default() -> CompatibleNumericalTypes {
+        CompatibleNumericalTypes::Dynamic {
+            all_values_within_i64_range: true,
+            all_values_within_u64_range: true,
+        }
+    }
+}
+
+impl CompatibleNumericalTypes {
+    pub fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
+        match self {
+            CompatibleNumericalTypes::Dynamic {
+                all_values_within_i64_range,
+                all_values_within_u64_range,
+            } => match numerical_type {
+                NumericalType::I64 => *all_values_within_i64_range,
+                NumericalType::U64 => *all_values_within_u64_range,
+                NumericalType::F64 => true,
+            },
+            CompatibleNumericalTypes::StaticType(static_numerical_type) => {
+                *static_numerical_type == numerical_type
+            }
+        }
+    }
+
+    pub fn accept_value(&mut self, numerical_value: NumericalValue) {
+        match self {
+            CompatibleNumericalTypes::Dynamic {
+                all_values_within_i64_range,
+                all_values_within_u64_range,
+            } => match numerical_value {
+                NumericalValue::I64(val_i64) => {
+                    let value_within_u64_range = val_i64 >= 0i64;
+                    *all_values_within_u64_range &= value_within_u64_range;
+                }
+                NumericalValue::U64(val_u64) => {
+                    let value_within_i64_range = val_u64 < i64::MAX as u64;
+                    *all_values_within_i64_range &= value_within_i64_range;
+                }
+                NumericalValue::F64(_) => {
+                    *all_values_within_i64_range = false;
+                    *all_values_within_u64_range = false;
+                }
+            },
+            CompatibleNumericalTypes::StaticType(typ) => {
+                assert_eq!(
+                    numerical_value.numerical_type(),
+                    *typ,
+                    "Input type forbidden. This column has been forced to type {typ:?}, received \
+                     {numerical_value:?}"
+                );
+            }
+        }
+    }
+
+    pub fn to_numerical_type(self) -> NumericalType {
+        for numerical_type in [NumericalType::I64, NumericalType::U64] {
+            if self.is_type_accepted(numerical_type) {
+                return numerical_type;
+            }
+        }
+        NumericalType::F64
+    }
+}
+
+impl NumericalColumnWriter {
+    pub fn numerical_type(&self) -> NumericalType {
+        self.compatible_numerical_types.to_numerical_type()
+    }
+
+    pub fn cardinality(&self, num_docs: RowId) -> Cardinality {
+        self.column_writer.get_cardinality(num_docs)
+    }
+
+    pub fn record_numerical_value(
+        &mut self,
+        doc: RowId,
+        value: NumericalValue,
+        arena: &mut MemoryArena,
+    ) {
+        self.compatible_numerical_types.accept_value(value);
+        self.column_writer.record(doc, value, arena);
+    }
+
+    pub(super) fn operation_iterator<'a>(
+        self,
+        arena: &MemoryArena,
+        buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<NumericalValue>> + 'a + use<'a> {
+        self.column_writer.operation_iterator(arena, buffer)
+    }
+}
+
+#[derive(Copy, Clone)]
+pub(crate) struct StrOrBytesColumnWriter {
+    pub(crate) dictionary_id: u32,
+    pub(crate) column_writer: ColumnWriter,
+    // If true, when facing a multivalued cardinality,
+    // values associated to a given document will be sorted.
+    //
+    // This is useful for facets.
+    //
+    // If false, the order of appearance in the document will be
+    // observed.
+    pub(crate) sort_values_within_row: bool,
+}
+
+impl StrOrBytesColumnWriter {
+    pub(crate) fn with_dictionary_id(dictionary_id: u32) -> StrOrBytesColumnWriter {
+        StrOrBytesColumnWriter {
+            dictionary_id,
+            column_writer: Default::default(),
+            sort_values_within_row: false,
+        }
+    }
+
+    pub(crate) fn record_bytes(
+        &mut self,
+        doc: RowId,
+        bytes: &[u8],
+        dictionaries: &mut [DictionaryBuilder],
+        arena: &mut MemoryArena,
+    ) {
+        let unordered_id =
+            dictionaries[self.dictionary_id as usize].get_or_allocate_id(bytes, arena);
+        self.column_writer.record(doc, unordered_id, arena);
+    }
+
+    pub(super) fn operation_iterator<'a>(
+        &self,
+        arena: &MemoryArena,
+        byte_buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<UnorderedId>> + 'a + use<'a> {
+        self.column_writer.operation_iterator(arena, byte_buffer)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_delta_with_last_doc() {
+        assert_eq!(delta_with_last_doc(None, 0u32), DocumentStep::Next);
+        assert_eq!(delta_with_last_doc(None, 1u32), DocumentStep::Skipped);
+        assert_eq!(delta_with_last_doc(None, 2u32), DocumentStep::Skipped);
+        assert_eq!(delta_with_last_doc(Some(0u32), 0u32), DocumentStep::Same);
+        assert_eq!(delta_with_last_doc(Some(1u32), 1u32), DocumentStep::Same);
+        assert_eq!(delta_with_last_doc(Some(1u32), 2u32), DocumentStep::Next);
+        assert_eq!(delta_with_last_doc(Some(1u32), 3u32), DocumentStep::Skipped);
+        assert_eq!(delta_with_last_doc(Some(1u32), 4u32), DocumentStep::Skipped);
+    }
+
+    #[track_caller]
+    fn test_column_writer_coercion_iter_aux(
+        values: impl Iterator<Item = NumericalValue>,
+        expected_numerical_type: NumericalType,
+    ) {
+        let mut compatible_numerical_types = CompatibleNumericalTypes::default();
+        for value in values {
+            compatible_numerical_types.accept_value(value);
+        }
+        assert_eq!(
+            compatible_numerical_types.to_numerical_type(),
+            expected_numerical_type
+        );
+    }
+
+    #[track_caller]
+    fn test_column_writer_coercion_aux(
+        values: &[NumericalValue],
+        expected_numerical_type: NumericalType,
+    ) {
+        test_column_writer_coercion_iter_aux(values.iter().copied(), expected_numerical_type);
+        test_column_writer_coercion_iter_aux(values.iter().rev().copied(), expected_numerical_type);
+    }
+
+    #[test]
+    fn test_column_writer_coercion() {
+        test_column_writer_coercion_aux(&[], NumericalType::I64);
+        test_column_writer_coercion_aux(&[1i64.into()], NumericalType::I64);
+        test_column_writer_coercion_aux(&[1u64.into()], NumericalType::I64);
+        // We don't detect exact integer at the moment. We could!
+        test_column_writer_coercion_aux(&[1f64.into()], NumericalType::F64);
+        test_column_writer_coercion_aux(&[u64::MAX.into()], NumericalType::U64);
+        test_column_writer_coercion_aux(&[(i64::MAX as u64).into()], NumericalType::U64);
+        test_column_writer_coercion_aux(&[(1u64 << 63).into()], NumericalType::U64);
+        test_column_writer_coercion_aux(&[1i64.into(), 1u64.into()], NumericalType::I64);
+        test_column_writer_coercion_aux(&[u64::MAX.into(), (-1i64).into()], NumericalType::F64);
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_compatible_numerical_types_static_incompatible_type() {
+        let mut compatible_numerical_types =
+            CompatibleNumericalTypes::StaticType(NumericalType::U64);
+        compatible_numerical_types.accept_value(NumericalValue::I64(1i64));
+    }
+
+    #[test]
+    fn test_compatible_numerical_types_static_different_type_forbidden() {
+        let mut compatible_numerical_types =
+            CompatibleNumericalTypes::StaticType(NumericalType::U64);
+        compatible_numerical_types.accept_value(NumericalValue::U64(u64::MAX));
+    }
+
+    #[test]
+    fn test_compatible_numerical_types_static() {
+        for typ in [NumericalType::I64, NumericalType::I64, NumericalType::F64] {
+            let compatible_numerical_types = CompatibleNumericalTypes::StaticType(typ);
+            assert_eq!(compatible_numerical_types.to_numerical_type(), typ);
+        }
+    }
+}