[feature](lance) support lance catalog in doris by zhangstar333 · Pull Request #65730 · apache/doris

zhangstar333 · 2026-07-16T14:27:21Z

What problem does this PR solve?

Issue #66340
Problem Summary:
doc: apache/doris-website#4024

This PR introduces read support for the Lance format in Doris, including:

Querying Lance datasets through a Lance Catalog.
Pushing supported predicates down to Lance to reduce the amount of data scanned.
Reading Lance datasets directly through S3 and Local TVFs.
Performing vector similarity search through the vector_search() TVF, with the distance exposed through the _distance column.
Lance is currently read-only and uses File Scanner V2. Writing to Lance is not supported yet.

  -- Create a Lance Catalog
  CREATE CATALOG lance_catalog PROPERTIES (
      "type" = "lance",
      "lance.catalog.type" = "filesystem",
      "warehouse" = "s3://warehouse/lance",
      "s3.endpoint" = "http://127.0.0.1:9000",
      "s3.access_key" = "admin",
      "s3.secret_key" = "password",
      "s3.region" = "us-east-1",
      "use_path_style" = "true"
  );

  -- Regular query; supported WHERE predicates are pushed down to Lance
  SELECT user_id, user_name
  FROM lance_catalog.doris.user_profiles
  WHERE engagement_score >= 90;

  -- Read a Lance dataset directly through the S3 TVF
  SELECT user_id, user_name
  FROM s3(
      "uri" = "s3://warehouse/lance/doris/user_profiles.lance",
      "format" = "lance",
      "s3.endpoint" = "http://127.0.0.1:9000",
      "s3.access_key" = "admin",
      "s3.secret_key" = "password",
      "s3.region" = "us-east-1",
      "use_path_style" = "true"
  );

  -- Perform a vector similarity search
  SELECT row_id, label, _distance
  FROM vector_search(
      "table" = "lance_catalog.doris.vector_search",
      "column" = "embedding",
      "query_vector" = "[0,0,0,0]",
      "top_k" = "5",
      "metric" = "l2"
  )
  ORDER BY _distance, row_id;

None

Check List (For Author)

Test
- Regression test
- Unit Test
- Manual test (add detailed scripts or steps below)
- No need to test or manual test. Explain why:
  - This is a refactor/code format and no logic has been changed.
  - Previous test can cover this change.
  - No code files have been changed.
  - Other reason
Behavior changed:
- No.
- Yes.
Does this need documentation?
- No.
- Yes.

Check List (For Reviewer who merge this PR)

Confirm the release note
Confirm test cases
Confirm document
Add branch pick label

hello-stephen · 2026-07-16T14:27:27Z

Thank you for your contribution to Apache Doris.
Don't know what should be done next? See How to process your PR.

Please clearly describe your PR:

What problem was fixed (it's best to include specific error reporting information). How it was fixed.
Which behaviors were modified. What was the previous behavior, what is it now, why was it modified, and what possible impacts might there be.
What features were added. Why was this function added?
Which code was refactored and why was this part of the code refactored?
Which functions were optimized and what is the difference before and after the optimization?

zhangstar333 · 2026-07-31T07:37:40Z

/review

Gabriel39

I found several security and correctness issues that should be addressed before merging.

github-actions · 2026-07-31T09:24:11Z

Codex automated review failed and did not complete.

Error: Review step was failure (possibly timeout or cancelled)
Workflow run: https://github.com/apache/doris/actions/runs/30613481239

Please inspect the workflow logs and rerun the review after the underlying issue is resolved.

github-actions

This adds a substantial Lance integration, but the current scan path still has blocking correctness, memory-safety, process-safety, security-display, snapshot-consistency, placement, and resource-governance defects. The inline comments identify thirteen distinct issues that remain after deduplicating against the four live comments already on the PR. Review was static-only as required; no build or test commands were run.

Gabriel39 · 2026-07-31T10:11:01Z

在这一版实现里，我们要把Postfilter改成Lance SDK 原生 Postfilter
另外，测试覆盖要把各种索引类型都加上

zhangstar333 · 2026-08-04T08:19:07Z

run buildall

zhangstar333 · 2026-08-04T11:33:21Z

run buildall

hello-stephen · 2026-08-04T13:07:30Z

Cloud UT Coverage Report

Increment line coverage 🎉

Increment coverage report
Complete coverage report

Category	Coverage
Function Coverage	77.62% (1904/2453)
Line Coverage	64.53% (34057/52774)
Region Coverage	64.49% (17216/26694)
Branch Coverage	54.03% (9220/17064)

zhangstar333 · 2026-08-04T15:45:46Z

run buildall

hello-stephen · 2026-08-04T16:52:21Z

Cloud UT Coverage Report

Increment line coverage 🎉

Increment coverage report
Complete coverage report

Category	Coverage
Function Coverage	77.62% (1904/2453)
Line Coverage	64.45% (34012/52774)
Region Coverage	64.46% (17206/26694)
Branch Coverage	53.96% (9207/17064)

hello-stephen · 2026-08-04T19:19:43Z

BE Regression && UT Coverage Report

Increment line coverage 69.09% (847/1226) 🎉

Increment coverage report
Complete coverage report

Category	Coverage
Function Coverage	73.67% (30364/41215)
Line Coverage	57.62% (333867/579401)
Region Coverage	54.49% (277771/509733)
Branch Coverage	55.46% (124428/224358)

hello-stephen · 2026-08-04T19:33:15Z

FE Regression Coverage Report

Increment line coverage 65.66% (1017/1549) 🎉
Increment coverage report
Complete coverage report

zhangstar333 · 2026-08-05T00:50:26Z

run buildall

hello-stephen · 2026-08-05T10:56:50Z

BE Regression && UT Coverage Report

Increment line coverage 69.31% (856/1235) 🎉

Increment coverage report
Complete coverage report

Category	Coverage
Function Coverage	73.67% (30386/41244)
Line Coverage	57.57% (333689/579583)
Region Coverage	54.50% (277884/509860)
Branch Coverage	55.43% (124402/224419)

…pache#66495) The vector_search() query path is algorithm-agnostic, but until now only flat search was actually exercised: the fixture from apache#65730 never created the vector index its tests claimed to cover (the referenced create_vector_search_index.py was never committed), so Lance silently fell back to flat KNN and the "IVF_PQ" regression proved nothing about indexes. This change adds verified IVF_FLAT query coverage and repairs the IVF_PQ fixture: - Add lance_build_preinstalled_catalog.py, an offline fixture generator pinned to pylance 4.0.1 (matching lance-c v0.1.2 / lance-rs 4.0.1). It builds the Directory V2 catalog through the namespace API (all_types re-registered, doris namespace, one indexed table per algorithm: vector_search_ivf_flat and vector_search_ivf_pq), then self-checks: Lance plans contain ANNSubIndex/ANNIvfPartition, indexes cover both fragments, IVF_FLAT full-probe equals flat search exactly, and the row-256 partition-boundary discriminator holds per table. The generated catalog (404K) is committed; the manifest is compacted back to one version/one data file. - Deterministic data: 1024 rows, two fragments, 16-dim Float32 with embedding[j] = (row_id - 1) + j, so a query equal to row r's vector has exact squared L2 distance 16 * (n - r)^2 to row n - integer ladder, tie-free at the endpoints, exact in Float32. - Rewrite test_lance_vector_search to target the real IVF_PQ table and add test_lance_vector_search_ivf_flat. Both suites carry a silent-fallback discriminator: querying the row-256 boundary with nprobes=1 must differ from the flat result (a pipeline that ignores the index returns exactly the flat rows and fails the assertion). IVF_FLAT additionally asserts indexed == flat programmatically (its algorithm guarantee); IVF_PQ freezes refine_factor-reranked results in goldens without claiming algorithm guarantees. - Remove run07_create_vector_types.sql and the stale golden output. Cross-version compatibility of the generated manifest was verified against lance-java 4.0.0 (lance-spark-bundle 0.4.0, runtime table registration) and lance-java 9.1.0-beta.3 (FE Directory Namespace client): both list, describe, register and deregister tables against it correctly. The remaining algorithms (IVF_SQ, IVF_HNSW_FLAT, IVF_HNSW_SQ, IVF_HNSW_PQ) follow the same pattern in follow-up PRs. Draft until the .out goldens are regenerated by a full docker regression run, which also provides the FE -> BE -> lance-c end-to-end evidence. Part of apache#66495. Claude-Session: https://claude.ai/code/session_014jKR13f5jdqs8NFyEuCnAN

### What problem does this PR solve? Issue Number: apache#66497 Related PR: apache#65730 Problem Summary: Lance external tables can contain logical vector and scalar indexes, but SHOW INDEX currently only handles internal OlapTable metadata and returns an empty result for Lance tables. Read authoritative logical index descriptions from one latest Directory Dataset snapshot, resolve field IDs from the same snapshot schema, and expose deterministic bounded metadata through the existing 13-column SHOW INDEX result without changing internal-table behavior. ### Release note Support SHOW INDEX for Lance filesystem/Directory Catalog tables. Lance REST index inspection remains unsupported in this phase. ### Check List (For Author) - Test: Unit Test and Regression test - Behavior changed: Yes. SHOW INDEX now returns logical indexes for Lance Directory tables and explicitly rejects Lance REST catalogs. - Does this need documentation: Yes. Document the supported Directory-only scope and REST limitation.

### What problem does this PR solve? Issue Number: N/A Related PR: #65730, #66581 Problem Summary: Ordinary Lance scans currently read **every row of a fragment** even when the query only needs the first N rows (e.g. `SELECT ... LIMIT 10`). Lance applies its own LIMIT *after* the scanner's filter, so the query LIMIT can be forwarded to each fragment scanner and let it stop early, cutting IO and decode cost. **How it is fixed** - `thrift`: add an optional `TLanceFileDesc.limit`. - `FE` (`LanceScanNode`): push the query limit into each fragment split via `canPushDownLimit()`, and surface `lanceLimit` in the explain output. - `BE` (`lance_reader`): forward it to the scanner through `lance_scanner_set_limit` for ordinary scans; vector search keeps its own `top_k` limit. **Correctness** The limit is pushed **only when all predicates are already pushed into Lance** (no residual Doris conjunct). Otherwise Doris still re-filters the returned rows, and truncating a fragment early could drop valid results. `OFFSET` needs no special handling: Nereids' `SplitLimit` rewrites `Limit(limit, offset)` into a global `Limit(limit, offset)` over a local `Limit(limit + offset, 0)`, and that local bound is what reaches the scan node. So `getLimit()` already includes the offset; each fragment fetches up to `limit + offset` rows and the upper global LIMIT still applies the offset and the final bound. Per-fragment truncation is therefore always safe. **Behavior change** Query results are unchanged. Only the number of rows scanned per fragment is reduced for LIMIT queries; the explain output shows an extra `lanceLimit=N` line when the limit is pushed. ### Release note Push down LIMIT into Lance fragment scanners to reduce the rows scanned for `LIMIT` / `LIMIT ... OFFSET` queries over Lance tables. ### Check List (For Author) - Test - [x] Unit Test (`LanceThriftContractTest` covers the limit round-trip and the no-limit case) - [ ] Manual test — `SELECT * FROM <lance_tbl> LIMIT 10` returns 10 rows and `EXPLAIN` shows `lanceLimit=10`; a query with a non-pushable predicate keeps the limit out of the scan - Behavior changed: - [x] No. - Does this need documentation? - [x] No.

### What problem does this PR solve? Issue Number: apache#66497 Related PR: apache#65730 Problem Summary: Lance external tables can contain logical vector and scalar indexes, but SHOW INDEX currently only handles internal OlapTable metadata and returns an empty result for Lance tables. Read authoritative logical index descriptions from one latest Directory Dataset snapshot, resolve field IDs from the same snapshot schema, and expose deterministic bounded metadata through the existing 13-column SHOW INDEX result without changing internal-table behavior. ### Release note Support SHOW INDEX for Lance filesystem/Directory Catalog tables. Lance REST index inspection remains unsupported in this phase. ### Check List (For Author) - Test: Unit Test and Regression test - Behavior changed: Yes. SHOW INDEX now returns logical indexes for Lance Directory tables and explicitly rejects Lance REST catalogs. - Does this need documentation: Yes. Document the supported Directory-only scope and REST limitation.

### What problem does this PR solve? Issue Number: #66497 Related PR: #65730 Lance external tables can contain logical vector and scalar indexes, but `SHOW INDEX` currently only handles internal `OlapTable` metadata and returns an empty result for Lance tables. This change reads authoritative logical index descriptions from one latest Lance Directory Dataset snapshot, resolves field IDs against the schema from that same snapshot, and exposes deterministic, bounded metadata through the existing 13-column `SHOW INDEX` result without changing internal-table behavior. ### What changed? - Add immutable Lance logical-index metadata and a loader backed by one latest Dataset snapshot. - Support `SHOW INDEX`, `SHOW INDEXES`, `SHOW KEY`, and `SHOW KEYS` for Lance filesystem/Directory Catalog tables. - Preserve existing internal-table behavior and permission-check ordering (privilege check runs before any catalog initialization). - Return a fixed unsupported error for Lance REST catalogs in this phase. - Sanitize Dataset URI, storage credential, and REST secret values from index metadata errors. - Defensively reject non-Lance tables resolved from a Lance catalog instead of failing on a raw cast. - Add unit and regression coverage for normalization, formatting, permissions, aliases, empty-index tables, and error handling. ### Design walkthrough (diagrams) `SHOW INDEX` for Lance Directory tables is a small read pipeline whose main concern is object ownership across the JNI boundary: SDK objects (`Dataset`, `IndexDescription`, Arrow allocators) never escape the loader, and the command layer only sees immutable `LanceLogicalIndex` values. Layered view — three new classes (`LanceLogicalIndex`, `LanceIndexMetadataLoader`, `LanceMetadataReadExecutor`) and three touch points (`ShowIndexCommand`, `LanceExternalTable`, `LanceExternalCatalog`): ```mermaid flowchart TB subgraph CMD["Command layer · Nereids"] A["ShowIndexCommand analyze() privilege check (:83) handleShowIndex() dispatch (:111) buildLanceRows() row mapping (:159)"] end subgraph EXT["External table layer"] B["LanceExternalTable loadIndexMetadata() (:69) — thin delegate"] end subgraph CAT["Catalog layer · caller thread"] C["LanceExternalCatalog isRestCatalogConfigured() (:172) early REST rejection resolveTableAccess() (:371) namespace resolution"] C2["sanitizedRootCauseMessage() (:473) URI / credential / token redaction"] end subgraph BND["Execution boundary · lance-metadata-read pool"] D["LanceMetadataReadExecutor (:50) 4 concurrent · 16 queued · deadline ≤ 60s"] E["LanceIndexMetadataLoader load() (:82) → normalize() (:206)"] end subgraph NAT["native · JNI"] F["Dataset.open(uri, latest) schema + describeIndices from one snapshot"] end G["LanceLogicalIndex immutable value object · pure-Java boundary"] A --> B --> C --> D --> E --> F E --> G C -.->|any failure| C2 C2 -.->|redacted exception only| A ``` Request timeline and resource ownership — namespace resolution stays on the caller thread (it owns the catalog's shared namespace/allocator), the JNI read runs on a bounded worker with a task-owned allocator, and a timed-out caller never cancels the in-flight native call: ```mermaid sequenceDiagram autonumber participant U as User participant C as Caller thread ShowIndexCommand participant K as Caller thread LanceExternalCatalog participant X as Executor boundary participant W as Worker thread lance-metadata-read participant J as Lance JNI Dataset U->>C: SHOW INDEX FROM lance_db.t C->>C: analyze() → checkTblPriv(SHOW) Note over C: privilege check runs before any catalog initialization C->>K: isRestCatalogConfigured() Note over K: reads normalized properties only, no namespace init — REST catalogs rejected here with a fixed error C->>K: resolveTableAccess() Note over K: shared namespace / allocator owned by the caller thread, so catalog close stays safe K-->>C: ResolvedTableAccess(uri, options) C->>X: execute(task) X->>W: submit (wrapped task checks deadline first) W->>W: expired in queue? fail before entering JNI W->>J: open(latest) + schema + describeIndices Note over W,J: task-owned RootAllocator (256MB) — catalog close cannot release it early J-->>W: IndexDescription + LanceSchema W-->>X: immutable index list X-->>C: future.get(remaining deadline) Note over C,X: timeout / interrupt never cancels the Future — the worker stays sole owner of native resources C->>C: buildLanceRows() → 13 columns C-->>U: ShowResultSet ``` Loader internals — one snapshot feeds both the schema mapping and the index descriptions; normalization is fail-closed and produces deterministic, bounded output: ```mermaid flowchart LR O["Dataset.open(uri, latest) index cache 0 · metadata cache 64MB schema and indexes from one snapshot"] O --> S["getLanceSchema().fields()"] O --> L["listIndexes() physical entries ≤ 16384"] subgraph SB["schema branch"] S --> DFS["buildFieldNamesById() (:148) DFS depth ≤ 64 · fields ≤ 16384 field id → dotted path, quoted when needed"] end subgraph IB["index branch"] L --> FL["drop system indexes __lance_frag_reuse · __lance_mem_wal"] FL --> DD["dedupe by logical name ≤ 256"] DD --> CR["describeUserIndexes() (:95) per-name describeIndices(hasName) must return exactly 1, name must match"] end DFS --> V CR --> V subgraph NB["normalize() (:206) · fail-closed"] V["known, unique field ids columns ≤ 64 · names ≤ 16KB total"] --> P["details JSON allowlist TreeMap-sorted · ≤ 400B"] P --> R["exact duplicate names rejected sorted by (name, position)"] end R --> OUT["LanceLogicalIndex list immutable · deterministic"] ``` ### Fail-closed semantics (intentional) Three strict behaviors are deliberate and locked by unit tests — the command reports an error rather than showing potentially misleading metadata: - An index referencing an unknown or nested field ID fails the whole `SHOW INDEX` (`testRejectsUnknownNestedDuplicateNullAndEmptyFieldIds`). - One malformed index entry (invalid details JSON, oversize fields, duplicate name) fails the whole result instead of silently dropping rows (`testRejectsMalformedAndNonObjectJsonWithoutEchoingInput`). - Index names differing only by case are preserved as distinct entries; only exact duplicates are rejected (`testRejectsExactDuplicateNameButPreservesCaseOnlyNames`). ### User/developer impact Users can inspect logical Lance indexes through the standard 13-column `SHOW INDEX` interface for filesystem/Directory catalogs. Lance REST index inspection remains unsupported. ### Release note Support `SHOW INDEX` for Lance filesystem/Directory Catalog tables. ### Check List (For Author) - [x] Unit tests: 28 tests passed - [x] Full FE build: all 27 modules passed - [x] Lance regression test: generation and normal verification modes passed - [x] `git diff --check` - [x] Behavior changed: `SHOW INDEX` now returns logical indexes for Lance Directory tables and explicitly rejects Lance REST catalogs - [x] Documentation needed: document the supported Directory-only scope and REST limitation

Issue Number: close apache#66496 Related PR: apache#65730 Problem Summary: Lance schema discovery rejected Blob v2, Arrow JSON, BFloat16, Arrow null, and Arrow duration fields as unsupported. Add validated FE and BE mappings, materialize Blob payloads through lance-c, normalize BFloat16 values to FLOAT, preserve duration counts as BIGINT, and materialize Arrow null values safely. Support reading Lance Blob v2, Arrow JSON, BFloat16, null, and duration columns. - Test: Unit Test (added; not run locally at user request) - Behavior changed: Yes (the newly supported Lance columns are queryable) - Does this need documentation: Yes (type matrix is documented in the regression suite; user documentation follow-up is required)

zhangstar333 requested a review from yiguolei as a code owner July 16, 2026 14:27

zhangstar333 force-pushed the branch-4.1-lance-catalog branch 2 times, most recently from ce3b540 to f760bd0 Compare July 30, 2026 12:25

zhangstar333 changed the title ~~Branch 4.1 test~~ [feature](lance) support lance catalog in doris Jul 30, 2026

zhangstar333 force-pushed the branch-4.1-lance-catalog branch from 4dd1bee to eb379a8 Compare July 31, 2026 07:17

morningman mentioned this pull request Jul 31, 2026

[Tracking] Support Lance Catalog #66322

Closed

22 tasks

Gabriel39 reviewed Jul 31, 2026

View reviewed changes

github-actions Bot previously requested changes Jul 31, 2026

View reviewed changes

Gabriel39 mentioned this pull request Jul 31, 2026

[Tracking] Lance feature matrix for Doris 4.2 #66340

Open

85 tasks

zhangstar333 force-pushed the branch-4.1-lance-catalog branch from eb379a8 to 56d7904 Compare August 3, 2026 15:08

Gabriel39 added the dev/5.0.x label Aug 4, 2026

zhangstar333 force-pushed the branch-4.1-lance-catalog branch from c9a4827 to 70516cf Compare August 4, 2026 11:32

zhangstar333 added 6 commits August 5, 2026 08:55

thrift

b43b4f3

be code about lance reader

d724df4

be ut test

47cf8a9

docker lance

38ab475

update format v2

8a919a3

update

19da132

Gabriel39 merged commit e3289c1 into apache:branch-4.1 Aug 5, 2026
27 of 34 checks passed

FANNG1 mentioned this pull request Aug 6, 2026

[test](lance) Add query coverage for the remaining ANN index types and vector element types #66512

Merged

Gabriel39 mentioned this pull request Aug 6, 2026

[Feature](lance) Add namespace and table DDL support #66516

Open

14 tasks

Jay-ju mentioned this pull request Aug 10, 2026

[opt](lance) push down LIMIT into Lance fragment scanners #66608

Merged

4 tasks

u70b3 mentioned this pull request Aug 11, 2026

[feature](lance) Support SHOW INDEX for Lance Directory tables #66637

Merged

6 tasks

FANNG1 mentioned this pull request Aug 14, 2026

[Bug](lance) REST catalog against Gravitino: namespace-vended credentials are dropped, and the root database is not listable #66772

Open

3 tasks

Jay-ju mentioned this pull request Aug 23, 2026

[feature](lance) push COUNT(*) down to Lance dataset metadata #66999

Open

Uh oh!

Conversation

zhangstar333 commented Jul 16, 2026 • edited by morningman Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What problem does this PR solve?

Check List (For Author)

Check List (For Reviewer who merge this PR)

Uh oh!

hello-stephen commented Jul 16, 2026

Uh oh!

zhangstar333 commented Jul 31, 2026

Uh oh!

Gabriel39 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Jul 31, 2026

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Gabriel39 commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zhangstar333 commented Aug 4, 2026

Uh oh!

zhangstar333 commented Aug 4, 2026

Uh oh!

hello-stephen commented Aug 4, 2026

Cloud UT Coverage Report

Uh oh!

zhangstar333 commented Aug 4, 2026

Uh oh!

hello-stephen commented Aug 4, 2026

Cloud UT Coverage Report

Uh oh!

hello-stephen commented Aug 4, 2026

BE Regression && UT Coverage Report

Uh oh!

hello-stephen commented Aug 4, 2026

FE Regression Coverage Report

Uh oh!

zhangstar333 commented Aug 5, 2026

Uh oh!

Uh oh!

hello-stephen commented Aug 5, 2026

BE Regression && UT Coverage Report

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

zhangstar333 commented Jul 16, 2026 •

edited by morningman

Loading

Gabriel39 commented Jul 31, 2026 •

edited

Loading