vinothchandar commented on code in PR #19309:
URL: https://github.com/apache/hudi/pull/19309#discussion_r3799866142


##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents

Review Comment:
   Small thing. but the ToC makes it easy to follow. thanks



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)

Review Comment:
   Let's add a top-level `## Limitations` section to this RFC, tracking what's 
knowingly not addressed in this first cut — so we can land the MVP and iterate 
with eyes open. 
   
   First entry: **VECTOR schema evolution**. What happens when the indexed 
column's dimension or element type changes 



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)
+
+- ANN families beyond IVF + RaBitQ (e.g. HNSW, DiskANN).
+- Filtered search (arbitrary predicate + kNN) as a first-class planned 
operation.
+- Time-travel-consistent index reads for historical snapshots.
+- Native non-Spark generation construction, GPU encoding, and 
workload-specific auto-tuning.
+
+### 1.3 Alternatives considered
+
+- **One index record per vector** is simple but creates billions of MDT 
records and excessive
+  write, compaction, and scan amplification; posting blocks preserve MDT 
ownership while
+  amortizing that overhead.
+- **Dedicated index files in the table** permit specialized layouts but 
introduce a second
+  commit, cleaning, and snapshot protocol. MDT records reuse Hudi's existing 
transaction and
+  table-service machinery.
+- **External sidecar indexes or vector databases** may offer richer serving 
features but lose
+  atomic Hudi snapshot semantics and require another storage system. They 
remain valid when
+  independent serving infrastructure is desired.
+- **Native ANN libraries** improve local kernels but do not define durable 
object-store layout,
+  multi-writer maintenance, or snapshot visibility. They may be used behind 
the interfaces in
+  this RFC without changing the persisted contract.
+
+---
+
+## 2. Architecture
+
+![RFC-109 architecture overview](diagrams/01-architecture-overview.svg)
+
+The design splits responsibilities the way Hudi already does between the data 
table and the
+metadata table:
+
+```text
+DATA TABLE (parquet/orc)                METADATA TABLE (vector_index partition)
+  authoritative vectors + payload  ←──   the index: centroids, quantizer, 
posting blocks,
+  read only for final re-ranking          cluster manifests, generation 
manifest

Review Comment:
   Readability: "generation" is used here (and throughout §2–§3) but only gets 
defined in §4.5. Add a one-liner up front — e.g. "a *generation* is an 
immutable, atomically-published version of the entire index (centroids + 
quantizer + routing + postings)" — so first-time readers aren't guessing until 
§4.5.



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)
+
+- ANN families beyond IVF + RaBitQ (e.g. HNSW, DiskANN).
+- Filtered search (arbitrary predicate + kNN) as a first-class planned 
operation.
+- Time-travel-consistent index reads for historical snapshots.
+- Native non-Spark generation construction, GPU encoding, and 
workload-specific auto-tuning.
+
+### 1.3 Alternatives considered
+
+- **One index record per vector** is simple but creates billions of MDT 
records and excessive
+  write, compaction, and scan amplification; posting blocks preserve MDT 
ownership while
+  amortizing that overhead.
+- **Dedicated index files in the table** permit specialized layouts but 
introduce a second
+  commit, cleaning, and snapshot protocol. MDT records reuse Hudi's existing 
transaction and
+  table-service machinery.
+- **External sidecar indexes or vector databases** may offer richer serving 
features but lose
+  atomic Hudi snapshot semantics and require another storage system. They 
remain valid when
+  independent serving infrastructure is desired.
+- **Native ANN libraries** improve local kernels but do not define durable 
object-store layout,
+  multi-writer maintenance, or snapshot visibility. They may be used behind 
the interfaces in
+  this RFC without changing the persisted contract.
+
+---
+
+## 2. Architecture
+
+![RFC-109 architecture overview](diagrams/01-architecture-overview.svg)
+
+The design splits responsibilities the way Hudi already does between the data 
table and the
+metadata table:
+
+```text
+DATA TABLE (parquet/orc)                METADATA TABLE (vector_index partition)
+  authoritative vectors + payload  ←──   the index: centroids, quantizer, 
posting blocks,
+  read only for final re-ranking          cluster manifests, generation 
manifest
+                                          read for candidate generation
+```
+
+Each vector index is one MDT partition. RFC-109 adds no table properties, 
common writer-
+dispatch changes, timeline-semantics changes, or behavior visible to 
non-vector readers or
+writers. Every consistency mechanism is either a record in the vector-index 
MDT partition or
+logic in the RFC-owned vector indexer and query planner. Creating an index:
+
+```sql
+CREATE INDEX embedding_idx
+ON products
+USING VECTOR (embedding)
+OPTIONS (
+  'vector.metric'      = 'cosine',
+  'vector.quantizer'   = 'IVF_RABITQ',
+  'vector.num_clusters'= '4096'
+);
+```
+
+creates:
+
+```text
+.hoodie/metadata/vector_index_embedding_idx/
+```
+
+and does not add generated columns to the base-table schema. RFC-109 consumes 
rather than
+redefines the RFC-99 vector contract: the source must be a top-level Hudi
+`VECTOR(D[, elementType])`. The table schema is authoritative for `D` and 
element type;
+index definitions and generation manifests repeat them only for integrity 
validation.
+
+The current RFC-99 storage backing is fixed-width bytes: Avro `FIXED` and 
Parquet
+`FIXED_LEN_BYTE_ARRAY(D × elementWidth)`. Engine adapters expose idiomatic 
values—Spark uses
+an annotated `ArrayType(FloatType|DoubleType|ByteType)`—and convert at the 
storage boundary.
+This fixed width also avoids Parquet LIST repetition-level traversal during 
positional exact
+fetch. A plain `ARRAY<FLOAT>` is not implicitly indexable and requires 
explicit migration or
+backfill to `VECTOR(D)`; index creation neither reinterprets nor rewrites it.
+
+The query path uses the MDT first to discover candidates, then reads 
base-table vectors for
+exact re-ranking:
+
+```text
+query vector
+  → compare to centroids, pick nprobe clusters          (in-memory, ms)
+  → MDT prefix-scan those clusters' posting blocks       (targeted range reads)
+  → two-pass RaBitQ scoring, keep refineFactor·K best    (bit math + error 
bounds)
+  → validate candidate freshness via Record Level Index  (batched point 
lookups)
+  → fetch ONLY those rows from the base table by position (page-level reads)
+  → exact distance on real vectors → final top-K
+```
+
+---
+
+## 3. IVF + RaBitQ Index Algorithm
+
+Every practical ANN index answers two questions: **where to look** (avoid 
scanning
+everything) and **how to compare cheaply** (avoid full-precision math on what 
is scanned).
+
+### 3.1 Where to look: IVF routing
+
+Inverted File (IVF) indexing clusters vectors with KMeans into `numClusters` 
groups (e.g.
+~4K–64K). Each vector belongs to its nearest centroid. A query compares 
against the centroids
+only (thousands, not billions), selects the `nprobe` nearest clusters, and 
scans only those
+clusters' entries. `nprobe` is the recall dial.
+
+IVF is the right fit for a lakehouse-resident index because a cluster's 
entries can be stored
+**contiguously**, which maps directly onto sorted key ranges in Hudi's MDT 
(§4). Graph
+indexes (HNSW) give excellent in-memory recall but require random traversal of 
the whole
+graph, which fights columnar, immutable, object-store storage.
+
+### 3.2 How to compare cheaply: RaBitQ quantization
+
+Inside a probed cluster there are still thousands of full vectors. 
Quantization stores a
+small *code* per vector plus a few correction scalars, so most comparisons run 
on compressed
+codes and only the best few hundred candidates are re-checked against real 
vectors.
+
+RaBitQ is chosen over scalar (SQ), product (PQ), and plain binary quantization 
for four
+reasons:
+
+1. **Unbiased estimator with a provable per-vector error bound.** Each code 
carries scalars
+   that turn a cheap bit-level dot product into an *unbiased* estimate of the 
true distance,
+   plus a bound on how wrong it can be. The bound enables **safe pruning**: 
skip a vector
+   only when even its best plausible distance cannot make top-K.
+2. **No codebooks.** RaBitQ needs only a random rotation (a seed) and the 
centroids — both
+   tiny, both versioned in the index metadata. Nothing to retrain when data 
drifts.
+3. **Tunable precision.** B = 1 bit/dim is a fast coarse filter; B = 4 
bits/dim gives
+   near-SQ quality at ~8× less space. Both are used in a two-pass scan (§6.2).
+4. **Metric-flexible.** One stored code serves L2, cosine, and dot-product; 
the metric is
+   applied at query time.
+
+The four ideas, precisely:
+
+- **Residual.** Store each vector as its difference from its centroid, `r = v 
− c`.
+  Residuals are small and centered, which is what lets few bits go far.
+- **Rotation.** Apply one fixed random orthonormal rotation `R` (derived from 
a per-generation
+  seed) to everything first: `x = R·r`. This spreads information evenly across 
dimensions,
+  which is what makes the error bound hold for any data distribution. Only the 
seed is stored.
+- **Code.** Quantize `x` to B bits per dimension, stored as **bit planes** — 
plane 0 holds
+  bit 0 of every dimension, plane 1 bit 1, etc. The top plane alone is a 1-bit 
sign sketch.
+  Scoring a plane against the (transformed) query is `AND`/`XOR` + `popcount` 
— a few CPU
+  instructions per 64 dimensions.
+- **Factors.** A handful of small scalars per vector (residual norm, two 
rescale factors, an
+  error term, a centroid correction) that convert plane math into an unbiased 
distance
+  estimate plus its confidence interval.
+
+At query time the query vector is transformed the same way (`R·(q − c)` per 
probed cluster),
+and the per-plane popcounts combine with the stored factors into the estimate 
and its bound.
+The estimate builds the shortlist; exact base-table distances produce the 
final ranking (§6).
+
+### 3.3 Why this fits Hudi
+
+- Centroids are small enough to load at planning time: `K × D` floats (~12 MB 
for K=4096,
+  D=768).
+- Codes are compact: a 1B × 128-dim float table's raw vectors are ~512 GB; the 
RaBitQ index
+  including keys and locators is ~136 GB, and the scanned portion per query is 
tens of MB.
+- Posting keys are prefix-scannable by generation, cluster, and shard (§4).
+- Quantizer state is stable: a seed and centroids, no per-generation learned 
codebook.
+- Exact re-ranking preserves correctness for returned candidates.
+
+---
+
+## 4. Metadata Table Storage Model: the Posting Block
+
+This section describes the core storage contribution of this RFC.
+
+### 4.1 The posting block
+
+A naive index would write one MDT record per indexed vector. At a billion rows 
that is a
+billion MDT records per generation — prohibitive to write, compact, scan, and 
clean.
+
+Instead, the index sorts entries by `(cluster, fileGroup, rowPosition)` and 
packs
+~1–4K vectors' worth of codes and metadata into a single MDT record — a 
**posting block** —

Review Comment:
   Minor: let's define "posting" for the average reader when it first appears — 
it's inverted-index jargon (a cluster's posting list = the set of entries 
assigned to it) and many Hudi devs won't have that context. one line here would 
do it



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)
+
+- ANN families beyond IVF + RaBitQ (e.g. HNSW, DiskANN).
+- Filtered search (arbitrary predicate + kNN) as a first-class planned 
operation.
+- Time-travel-consistent index reads for historical snapshots.
+- Native non-Spark generation construction, GPU encoding, and 
workload-specific auto-tuning.
+
+### 1.3 Alternatives considered
+
+- **One index record per vector** is simple but creates billions of MDT 
records and excessive
+  write, compaction, and scan amplification; posting blocks preserve MDT 
ownership while
+  amortizing that overhead.
+- **Dedicated index files in the table** permit specialized layouts but 
introduce a second
+  commit, cleaning, and snapshot protocol. MDT records reuse Hudi's existing 
transaction and
+  table-service machinery.
+- **External sidecar indexes or vector databases** may offer richer serving 
features but lose
+  atomic Hudi snapshot semantics and require another storage system. They 
remain valid when
+  independent serving infrastructure is desired.
+- **Native ANN libraries** improve local kernels but do not define durable 
object-store layout,
+  multi-writer maintenance, or snapshot visibility. They may be used behind 
the interfaces in
+  this RFC without changing the persisted contract.
+
+---
+
+## 2. Architecture
+
+![RFC-109 architecture overview](diagrams/01-architecture-overview.svg)
+
+The design splits responsibilities the way Hudi already does between the data 
table and the
+metadata table:
+
+```text
+DATA TABLE (parquet/orc)                METADATA TABLE (vector_index partition)
+  authoritative vectors + payload  ←──   the index: centroids, quantizer, 
posting blocks,
+  read only for final re-ranking          cluster manifests, generation 
manifest
+                                          read for candidate generation
+```
+
+Each vector index is one MDT partition. RFC-109 adds no table properties, 
common writer-
+dispatch changes, timeline-semantics changes, or behavior visible to 
non-vector readers or
+writers. Every consistency mechanism is either a record in the vector-index 
MDT partition or
+logic in the RFC-owned vector indexer and query planner. Creating an index:
+
+```sql
+CREATE INDEX embedding_idx
+ON products
+USING VECTOR (embedding)
+OPTIONS (
+  'vector.metric'      = 'cosine',
+  'vector.quantizer'   = 'IVF_RABITQ',
+  'vector.num_clusters'= '4096'
+);
+```
+
+creates:
+
+```text
+.hoodie/metadata/vector_index_embedding_idx/
+```
+
+and does not add generated columns to the base-table schema. RFC-109 consumes 
rather than
+redefines the RFC-99 vector contract: the source must be a top-level Hudi
+`VECTOR(D[, elementType])`. The table schema is authoritative for `D` and 
element type;
+index definitions and generation manifests repeat them only for integrity 
validation.
+
+The current RFC-99 storage backing is fixed-width bytes: Avro `FIXED` and 
Parquet
+`FIXED_LEN_BYTE_ARRAY(D × elementWidth)`. Engine adapters expose idiomatic 
values—Spark uses
+an annotated `ArrayType(FloatType|DoubleType|ByteType)`—and convert at the 
storage boundary.
+This fixed width also avoids Parquet LIST repetition-level traversal during 
positional exact
+fetch. A plain `ARRAY<FLOAT>` is not implicitly indexable and requires 
explicit migration or
+backfill to `VECTOR(D)`; index creation neither reinterprets nor rewrites it.
+
+The query path uses the MDT first to discover candidates, then reads 
base-table vectors for
+exact re-ranking:
+
+```text
+query vector
+  → compare to centroids, pick nprobe clusters          (in-memory, ms)
+  → MDT prefix-scan those clusters' posting blocks       (targeted range reads)
+  → two-pass RaBitQ scoring, keep refineFactor·K best    (bit math + error 
bounds)
+  → validate candidate freshness via Record Level Index  (batched point 
lookups)
+  → fetch ONLY those rows from the base table by position (page-level reads)
+  → exact distance on real vectors → final top-K
+```
+
+---
+
+## 3. IVF + RaBitQ Index Algorithm
+
+Every practical ANN index answers two questions: **where to look** (avoid 
scanning
+everything) and **how to compare cheaply** (avoid full-precision math on what 
is scanned).
+
+### 3.1 Where to look: IVF routing
+
+Inverted File (IVF) indexing clusters vectors with KMeans into `numClusters` 
groups (e.g.
+~4K–64K). Each vector belongs to its nearest centroid. A query compares 
against the centroids
+only (thousands, not billions), selects the `nprobe` nearest clusters, and 
scans only those
+clusters' entries. `nprobe` is the recall dial.
+
+IVF is the right fit for a lakehouse-resident index because a cluster's 
entries can be stored
+**contiguously**, which maps directly onto sorted key ranges in Hudi's MDT 
(§4). Graph
+indexes (HNSW) give excellent in-memory recall but require random traversal of 
the whole
+graph, which fights columnar, immutable, object-store storage.
+
+### 3.2 How to compare cheaply: RaBitQ quantization
+
+Inside a probed cluster there are still thousands of full vectors. 
Quantization stores a
+small *code* per vector plus a few correction scalars, so most comparisons run 
on compressed
+codes and only the best few hundred candidates are re-checked against real 
vectors.
+
+RaBitQ is chosen over scalar (SQ), product (PQ), and plain binary quantization 
for four
+reasons:
+
+1. **Unbiased estimator with a provable per-vector error bound.** Each code 
carries scalars
+   that turn a cheap bit-level dot product into an *unbiased* estimate of the 
true distance,
+   plus a bound on how wrong it can be. The bound enables **safe pruning**: 
skip a vector
+   only when even its best plausible distance cannot make top-K.
+2. **No codebooks.** RaBitQ needs only a random rotation (a seed) and the 
centroids — both
+   tiny, both versioned in the index metadata. Nothing to retrain when data 
drifts.
+3. **Tunable precision.** B = 1 bit/dim is a fast coarse filter; B = 4 
bits/dim gives
+   near-SQ quality at ~8× less space. Both are used in a two-pass scan (§6.2).
+4. **Metric-flexible.** One stored code serves L2, cosine, and dot-product; 
the metric is
+   applied at query time.
+
+The four ideas, precisely:
+
+- **Residual.** Store each vector as its difference from its centroid, `r = v 
− c`.
+  Residuals are small and centered, which is what lets few bits go far.
+- **Rotation.** Apply one fixed random orthonormal rotation `R` (derived from 
a per-generation
+  seed) to everything first: `x = R·r`. This spreads information evenly across 
dimensions,
+  which is what makes the error bound hold for any data distribution. Only the 
seed is stored.
+- **Code.** Quantize `x` to B bits per dimension, stored as **bit planes** — 
plane 0 holds
+  bit 0 of every dimension, plane 1 bit 1, etc. The top plane alone is a 1-bit 
sign sketch.
+  Scoring a plane against the (transformed) query is `AND`/`XOR` + `popcount` 
— a few CPU
+  instructions per 64 dimensions.
+- **Factors.** A handful of small scalars per vector (residual norm, two 
rescale factors, an
+  error term, a centroid correction) that convert plane math into an unbiased 
distance
+  estimate plus its confidence interval.
+
+At query time the query vector is transformed the same way (`R·(q − c)` per 
probed cluster),
+and the per-plane popcounts combine with the stored factors into the estimate 
and its bound.
+The estimate builds the shortlist; exact base-table distances produce the 
final ranking (§6).
+
+### 3.3 Why this fits Hudi
+
+- Centroids are small enough to load at planning time: `K × D` floats (~12 MB 
for K=4096,
+  D=768).
+- Codes are compact: a 1B × 128-dim float table's raw vectors are ~512 GB; the 
RaBitQ index
+  including keys and locators is ~136 GB, and the scanned portion per query is 
tens of MB.
+- Posting keys are prefix-scannable by generation, cluster, and shard (§4).
+- Quantizer state is stable: a seed and centroids, no per-generation learned 
codebook.
+- Exact re-ranking preserves correctness for returned candidates.
+
+---
+
+## 4. Metadata Table Storage Model: the Posting Block
+
+This section describes the core storage contribution of this RFC.
+
+### 4.1 The posting block
+
+A naive index would write one MDT record per indexed vector. At a billion rows 
that is a
+billion MDT records per generation — prohibitive to write, compact, scan, and 
clean.
+
+Instead, the index sorts entries by `(cluster, fileGroup, rowPosition)` and 
packs
+~1–4K vectors' worth of codes and metadata into a single MDT record — a 
**posting block** —
+laid out column-wise (structure-of-arrays) so each scan pass touches only the 
columns it
+needs:
+
+```text
+POSTING BLOCK (~512 KB target, one MDT record)
+┌───────────────────────────────────────────────┐
+│ S1  sign planes        ← pass 1 touches this   │
+│ S2  extra bit planes   ← pass 2, survivors     │
+│ S3  factor arrays      ← both passes           │
+│ S4  row locators       ← finalists only        │
+│ S5  dictionaries (file groups, partitions)     │

Review Comment:
   Since the partition dictionary already lives in the posting block, partition 
pushdown (bounding the ANN search to a subset of table partitions) looks within 
reach — filter candidates against the dictionary during the scan pass. Fine to 
defer for v1, but let's call it out as a known gap in the Limitations section 
given partition predicates are the most common filter in practice?



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)
+
+- ANN families beyond IVF + RaBitQ (e.g. HNSW, DiskANN).
+- Filtered search (arbitrary predicate + kNN) as a first-class planned 
operation.
+- Time-travel-consistent index reads for historical snapshots.
+- Native non-Spark generation construction, GPU encoding, and 
workload-specific auto-tuning.
+
+### 1.3 Alternatives considered
+
+- **One index record per vector** is simple but creates billions of MDT 
records and excessive
+  write, compaction, and scan amplification; posting blocks preserve MDT 
ownership while
+  amortizing that overhead.
+- **Dedicated index files in the table** permit specialized layouts but 
introduce a second
+  commit, cleaning, and snapshot protocol. MDT records reuse Hudi's existing 
transaction and
+  table-service machinery.
+- **External sidecar indexes or vector databases** may offer richer serving 
features but lose
+  atomic Hudi snapshot semantics and require another storage system. They 
remain valid when
+  independent serving infrastructure is desired.
+- **Native ANN libraries** improve local kernels but do not define durable 
object-store layout,
+  multi-writer maintenance, or snapshot visibility. They may be used behind 
the interfaces in
+  this RFC without changing the persisted contract.
+
+---
+
+## 2. Architecture
+
+![RFC-109 architecture overview](diagrams/01-architecture-overview.svg)
+
+The design splits responsibilities the way Hudi already does between the data 
table and the
+metadata table:
+
+```text
+DATA TABLE (parquet/orc)                METADATA TABLE (vector_index partition)
+  authoritative vectors + payload  ←──   the index: centroids, quantizer, 
posting blocks,
+  read only for final re-ranking          cluster manifests, generation 
manifest
+                                          read for candidate generation
+```
+
+Each vector index is one MDT partition. RFC-109 adds no table properties, 
common writer-
+dispatch changes, timeline-semantics changes, or behavior visible to 
non-vector readers or
+writers. Every consistency mechanism is either a record in the vector-index 
MDT partition or
+logic in the RFC-owned vector indexer and query planner. Creating an index:
+
+```sql
+CREATE INDEX embedding_idx
+ON products
+USING VECTOR (embedding)
+OPTIONS (
+  'vector.metric'      = 'cosine',
+  'vector.quantizer'   = 'IVF_RABITQ',
+  'vector.num_clusters'= '4096'
+);
+```
+
+creates:
+
+```text
+.hoodie/metadata/vector_index_embedding_idx/
+```
+
+and does not add generated columns to the base-table schema. RFC-109 consumes 
rather than
+redefines the RFC-99 vector contract: the source must be a top-level Hudi
+`VECTOR(D[, elementType])`. The table schema is authoritative for `D` and 
element type;
+index definitions and generation manifests repeat them only for integrity 
validation.
+
+The current RFC-99 storage backing is fixed-width bytes: Avro `FIXED` and 
Parquet
+`FIXED_LEN_BYTE_ARRAY(D × elementWidth)`. Engine adapters expose idiomatic 
values—Spark uses
+an annotated `ArrayType(FloatType|DoubleType|ByteType)`—and convert at the 
storage boundary.
+This fixed width also avoids Parquet LIST repetition-level traversal during 
positional exact
+fetch. A plain `ARRAY<FLOAT>` is not implicitly indexable and requires 
explicit migration or
+backfill to `VECTOR(D)`; index creation neither reinterprets nor rewrites it.
+
+The query path uses the MDT first to discover candidates, then reads 
base-table vectors for
+exact re-ranking:
+
+```text
+query vector
+  → compare to centroids, pick nprobe clusters          (in-memory, ms)
+  → MDT prefix-scan those clusters' posting blocks       (targeted range reads)
+  → two-pass RaBitQ scoring, keep refineFactor·K best    (bit math + error 
bounds)
+  → validate candidate freshness via Record Level Index  (batched point 
lookups)
+  → fetch ONLY those rows from the base table by position (page-level reads)
+  → exact distance on real vectors → final top-K
+```
+
+---
+
+## 3. IVF + RaBitQ Index Algorithm
+
+Every practical ANN index answers two questions: **where to look** (avoid 
scanning
+everything) and **how to compare cheaply** (avoid full-precision math on what 
is scanned).
+
+### 3.1 Where to look: IVF routing
+
+Inverted File (IVF) indexing clusters vectors with KMeans into `numClusters` 
groups (e.g.
+~4K–64K). Each vector belongs to its nearest centroid. A query compares 
against the centroids
+only (thousands, not billions), selects the `nprobe` nearest clusters, and 
scans only those
+clusters' entries. `nprobe` is the recall dial.
+
+IVF is the right fit for a lakehouse-resident index because a cluster's 
entries can be stored
+**contiguously**, which maps directly onto sorted key ranges in Hudi's MDT 
(§4). Graph
+indexes (HNSW) give excellent in-memory recall but require random traversal of 
the whole
+graph, which fights columnar, immutable, object-store storage.
+
+### 3.2 How to compare cheaply: RaBitQ quantization
+
+Inside a probed cluster there are still thousands of full vectors. 
Quantization stores a
+small *code* per vector plus a few correction scalars, so most comparisons run 
on compressed
+codes and only the best few hundred candidates are re-checked against real 
vectors.
+
+RaBitQ is chosen over scalar (SQ), product (PQ), and plain binary quantization 
for four
+reasons:
+
+1. **Unbiased estimator with a provable per-vector error bound.** Each code 
carries scalars
+   that turn a cheap bit-level dot product into an *unbiased* estimate of the 
true distance,
+   plus a bound on how wrong it can be. The bound enables **safe pruning**: 
skip a vector
+   only when even its best plausible distance cannot make top-K.
+2. **No codebooks.** RaBitQ needs only a random rotation (a seed) and the 
centroids — both
+   tiny, both versioned in the index metadata. Nothing to retrain when data 
drifts.
+3. **Tunable precision.** B = 1 bit/dim is a fast coarse filter; B = 4 
bits/dim gives
+   near-SQ quality at ~8× less space. Both are used in a two-pass scan (§6.2).
+4. **Metric-flexible.** One stored code serves L2, cosine, and dot-product; 
the metric is
+   applied at query time.
+
+The four ideas, precisely:
+
+- **Residual.** Store each vector as its difference from its centroid, `r = v 
− c`.
+  Residuals are small and centered, which is what lets few bits go far.
+- **Rotation.** Apply one fixed random orthonormal rotation `R` (derived from 
a per-generation
+  seed) to everything first: `x = R·r`. This spreads information evenly across 
dimensions,
+  which is what makes the error bound hold for any data distribution. Only the 
seed is stored.
+- **Code.** Quantize `x` to B bits per dimension, stored as **bit planes** — 
plane 0 holds
+  bit 0 of every dimension, plane 1 bit 1, etc. The top plane alone is a 1-bit 
sign sketch.
+  Scoring a plane against the (transformed) query is `AND`/`XOR` + `popcount` 
— a few CPU
+  instructions per 64 dimensions.
+- **Factors.** A handful of small scalars per vector (residual norm, two 
rescale factors, an
+  error term, a centroid correction) that convert plane math into an unbiased 
distance
+  estimate plus its confidence interval.
+
+At query time the query vector is transformed the same way (`R·(q − c)` per 
probed cluster),
+and the per-plane popcounts combine with the stored factors into the estimate 
and its bound.
+The estimate builds the shortlist; exact base-table distances produce the 
final ranking (§6).
+
+### 3.3 Why this fits Hudi
+
+- Centroids are small enough to load at planning time: `K × D` floats (~12 MB 
for K=4096,
+  D=768).
+- Codes are compact: a 1B × 128-dim float table's raw vectors are ~512 GB; the 
RaBitQ index
+  including keys and locators is ~136 GB, and the scanned portion per query is 
tens of MB.
+- Posting keys are prefix-scannable by generation, cluster, and shard (§4).
+- Quantizer state is stable: a seed and centroids, no per-generation learned 
codebook.
+- Exact re-ranking preserves correctness for returned candidates.
+
+---
+
+## 4. Metadata Table Storage Model: the Posting Block
+
+This section describes the core storage contribution of this RFC.
+
+### 4.1 The posting block
+
+A naive index would write one MDT record per indexed vector. At a billion rows 
that is a
+billion MDT records per generation — prohibitive to write, compact, scan, and 
clean.
+
+Instead, the index sorts entries by `(cluster, fileGroup, rowPosition)` and 
packs
+~1–4K vectors' worth of codes and metadata into a single MDT record — a 
**posting block** —
+laid out column-wise (structure-of-arrays) so each scan pass touches only the 
columns it
+needs:
+
+```text
+POSTING BLOCK (~512 KB target, one MDT record)
+┌───────────────────────────────────────────────┐
+│ S1  sign planes        ← pass 1 touches this   │
+│ S2  extra bit planes   ← pass 2, survivors     │
+│ S3  factor arrays      ← both passes           │
+│ S4  row locators       ← finalists only        │
+│ S5  dictionaries (file groups, partitions)     │
+│ S6  record keys        ← finalists only        │
+└───────────────────────────────────────────────┘
+key:  0x10 | generation | clusterId | shardId | blockId
+```
+
+Consequences:
+
+- **~1000× fewer MDT records.** One block record replaces ~1–4K per-vector 
records, cutting
+  write amplification, compaction cost, and cleaner load by three orders of 
magnitude.
+- **Contiguous cluster scans.** The binary key scheme makes "scan cluster 
12345" a single
+  contiguous HFile range read rather than thousands of point lookups.
+- **Pay only for promise.** Column-wise layout means pass 1 reads only sign 
planes + factors;
+  only survivors touch extra planes; only finalists touch locators and keys 
(§6.2).
+
+### 4.2 Row families
+
+The `vector_index_<name>` partition holds several record families under one 
binary-sorted key
+scheme, so one prefix scan of a cluster returns its blocks and any fresh 
deltas together:
+
+| Key family | Cardinality | Purpose |
+|---|---:|---|
+| `__manifest__` | 1 | Active generation pointer and persisted-format version. 
|
+| `M\|<generation>` | generations | Generation state, vector schema, bootstrap 
baseline, verified-contiguous frontier, chunk counts/checksums, and quantizer 
metadata. |
+| `T\|<generation>\|<chunk>` | chunks per generation | Size-bounded chunks of 
the serialized `K × D` centroid matrix. |
+| `C\|<generation>\|<cluster>` | K per generation | Cluster manifest: routing 
version, shard count, vector count, candidate file groups, and counters. |
+| `P\|<generation>\|<cluster>\|<shard>\|<blockId>` | blocks per generation | 
**Posting block** (packed codes, factors, locators, keys). |
+| `P\|...\|<DELTA>` | deltas | Small per-record delta records appended between 
compactions. |
+| `F\|<generation>\|<dataInstant>` | data writes after baseline | Atomic proof 
that the generation incorporated one source data-write instant. |
+
+### 4.3 Posting shards
+
+Large clusters are split into posting shards so one hot cluster does not 
become one oversized

Review Comment:
   Terminology: "shard" is overloaded — we already use file groups/buckets 
elsewhere and "shard" in the RLI context means something different. Add a 
clarifying line that a posting shard is purely an intra-cluster key-range split 
of the MDT prefix (nothing to do with RLI sharding or data-table bucketing). 
Consider renaming to something like "sub-range" to avoid the collision.



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs

Review Comment:
   this was key observation, from your runs. Thanks for being thorough. Can you 
dump any research numbers into an Appendix. So the data lives with the repo



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)
+
+- ANN families beyond IVF + RaBitQ (e.g. HNSW, DiskANN).
+- Filtered search (arbitrary predicate + kNN) as a first-class planned 
operation.
+- Time-travel-consistent index reads for historical snapshots.
+- Native non-Spark generation construction, GPU encoding, and 
workload-specific auto-tuning.
+
+### 1.3 Alternatives considered
+
+- **One index record per vector** is simple but creates billions of MDT 
records and excessive
+  write, compaction, and scan amplification; posting blocks preserve MDT 
ownership while
+  amortizing that overhead.
+- **Dedicated index files in the table** permit specialized layouts but 
introduce a second
+  commit, cleaning, and snapshot protocol. MDT records reuse Hudi's existing 
transaction and
+  table-service machinery.
+- **External sidecar indexes or vector databases** may offer richer serving 
features but lose
+  atomic Hudi snapshot semantics and require another storage system. They 
remain valid when
+  independent serving infrastructure is desired.
+- **Native ANN libraries** improve local kernels but do not define durable 
object-store layout,
+  multi-writer maintenance, or snapshot visibility. They may be used behind 
the interfaces in
+  this RFC without changing the persisted contract.
+
+---
+
+## 2. Architecture
+
+![RFC-109 architecture overview](diagrams/01-architecture-overview.svg)
+
+The design splits responsibilities the way Hudi already does between the data 
table and the
+metadata table:
+
+```text
+DATA TABLE (parquet/orc)                METADATA TABLE (vector_index partition)
+  authoritative vectors + payload  ←──   the index: centroids, quantizer, 
posting blocks,
+  read only for final re-ranking          cluster manifests, generation 
manifest
+                                          read for candidate generation
+```
+
+Each vector index is one MDT partition. RFC-109 adds no table properties, 
common writer-
+dispatch changes, timeline-semantics changes, or behavior visible to 
non-vector readers or
+writers. Every consistency mechanism is either a record in the vector-index 
MDT partition or
+logic in the RFC-owned vector indexer and query planner. Creating an index:
+
+```sql
+CREATE INDEX embedding_idx
+ON products
+USING VECTOR (embedding)
+OPTIONS (
+  'vector.metric'      = 'cosine',
+  'vector.quantizer'   = 'IVF_RABITQ',
+  'vector.num_clusters'= '4096'
+);
+```
+
+creates:
+
+```text
+.hoodie/metadata/vector_index_embedding_idx/
+```
+
+and does not add generated columns to the base-table schema. RFC-109 consumes 
rather than
+redefines the RFC-99 vector contract: the source must be a top-level Hudi
+`VECTOR(D[, elementType])`. The table schema is authoritative for `D` and 
element type;
+index definitions and generation manifests repeat them only for integrity 
validation.
+
+The current RFC-99 storage backing is fixed-width bytes: Avro `FIXED` and 
Parquet
+`FIXED_LEN_BYTE_ARRAY(D × elementWidth)`. Engine adapters expose idiomatic 
values—Spark uses
+an annotated `ArrayType(FloatType|DoubleType|ByteType)`—and convert at the 
storage boundary.
+This fixed width also avoids Parquet LIST repetition-level traversal during 
positional exact
+fetch. A plain `ARRAY<FLOAT>` is not implicitly indexable and requires 
explicit migration or
+backfill to `VECTOR(D)`; index creation neither reinterprets nor rewrites it.
+
+The query path uses the MDT first to discover candidates, then reads 
base-table vectors for
+exact re-ranking:
+
+```text
+query vector
+  → compare to centroids, pick nprobe clusters          (in-memory, ms)
+  → MDT prefix-scan those clusters' posting blocks       (targeted range reads)
+  → two-pass RaBitQ scoring, keep refineFactor·K best    (bit math + error 
bounds)
+  → validate candidate freshness via Record Level Index  (batched point 
lookups)
+  → fetch ONLY those rows from the base table by position (page-level reads)
+  → exact distance on real vectors → final top-K
+```
+
+---
+
+## 3. IVF + RaBitQ Index Algorithm
+
+Every practical ANN index answers two questions: **where to look** (avoid 
scanning
+everything) and **how to compare cheaply** (avoid full-precision math on what 
is scanned).
+
+### 3.1 Where to look: IVF routing
+
+Inverted File (IVF) indexing clusters vectors with KMeans into `numClusters` 
groups (e.g.
+~4K–64K). Each vector belongs to its nearest centroid. A query compares 
against the centroids
+only (thousands, not billions), selects the `nprobe` nearest clusters, and 
scans only those
+clusters' entries. `nprobe` is the recall dial.
+
+IVF is the right fit for a lakehouse-resident index because a cluster's 
entries can be stored
+**contiguously**, which maps directly onto sorted key ranges in Hudi's MDT 
(§4). Graph
+indexes (HNSW) give excellent in-memory recall but require random traversal of 
the whole
+graph, which fights columnar, immutable, object-store storage.
+
+### 3.2 How to compare cheaply: RaBitQ quantization
+
+Inside a probed cluster there are still thousands of full vectors. 
Quantization stores a
+small *code* per vector plus a few correction scalars, so most comparisons run 
on compressed
+codes and only the best few hundred candidates are re-checked against real 
vectors.
+
+RaBitQ is chosen over scalar (SQ), product (PQ), and plain binary quantization 
for four
+reasons:
+
+1. **Unbiased estimator with a provable per-vector error bound.** Each code 
carries scalars
+   that turn a cheap bit-level dot product into an *unbiased* estimate of the 
true distance,
+   plus a bound on how wrong it can be. The bound enables **safe pruning**: 
skip a vector
+   only when even its best plausible distance cannot make top-K.
+2. **No codebooks.** RaBitQ needs only a random rotation (a seed) and the 
centroids — both
+   tiny, both versioned in the index metadata. Nothing to retrain when data 
drifts.
+3. **Tunable precision.** B = 1 bit/dim is a fast coarse filter; B = 4 
bits/dim gives
+   near-SQ quality at ~8× less space. Both are used in a two-pass scan (§6.2).
+4. **Metric-flexible.** One stored code serves L2, cosine, and dot-product; 
the metric is
+   applied at query time.
+
+The four ideas, precisely:
+
+- **Residual.** Store each vector as its difference from its centroid, `r = v 
− c`.
+  Residuals are small and centered, which is what lets few bits go far.
+- **Rotation.** Apply one fixed random orthonormal rotation `R` (derived from 
a per-generation
+  seed) to everything first: `x = R·r`. This spreads information evenly across 
dimensions,
+  which is what makes the error bound hold for any data distribution. Only the 
seed is stored.
+- **Code.** Quantize `x` to B bits per dimension, stored as **bit planes** — 
plane 0 holds
+  bit 0 of every dimension, plane 1 bit 1, etc. The top plane alone is a 1-bit 
sign sketch.
+  Scoring a plane against the (transformed) query is `AND`/`XOR` + `popcount` 
— a few CPU
+  instructions per 64 dimensions.
+- **Factors.** A handful of small scalars per vector (residual norm, two 
rescale factors, an
+  error term, a centroid correction) that convert plane math into an unbiased 
distance
+  estimate plus its confidence interval.
+
+At query time the query vector is transformed the same way (`R·(q − c)` per 
probed cluster),
+and the per-plane popcounts combine with the stored factors into the estimate 
and its bound.
+The estimate builds the shortlist; exact base-table distances produce the 
final ranking (§6).
+
+### 3.3 Why this fits Hudi
+
+- Centroids are small enough to load at planning time: `K × D` floats (~12 MB 
for K=4096,
+  D=768).
+- Codes are compact: a 1B × 128-dim float table's raw vectors are ~512 GB; the 
RaBitQ index
+  including keys and locators is ~136 GB, and the scanned portion per query is 
tens of MB.
+- Posting keys are prefix-scannable by generation, cluster, and shard (§4).
+- Quantizer state is stable: a seed and centroids, no per-generation learned 
codebook.
+- Exact re-ranking preserves correctness for returned candidates.
+
+---
+
+## 4. Metadata Table Storage Model: the Posting Block
+
+This section describes the core storage contribution of this RFC.
+
+### 4.1 The posting block
+
+A naive index would write one MDT record per indexed vector. At a billion rows 
that is a
+billion MDT records per generation — prohibitive to write, compact, scan, and 
clean.
+
+Instead, the index sorts entries by `(cluster, fileGroup, rowPosition)` and 
packs
+~1–4K vectors' worth of codes and metadata into a single MDT record — a 
**posting block** —
+laid out column-wise (structure-of-arrays) so each scan pass touches only the 
columns it
+needs:
+
+```text
+POSTING BLOCK (~512 KB target, one MDT record)
+┌───────────────────────────────────────────────┐
+│ S1  sign planes        ← pass 1 touches this   │
+│ S2  extra bit planes   ← pass 2, survivors     │
+│ S3  factor arrays      ← both passes           │
+│ S4  row locators       ← finalists only        │
+│ S5  dictionaries (file groups, partitions)     │
+│ S6  record keys        ← finalists only        │
+└───────────────────────────────────────────────┘
+key:  0x10 | generation | clusterId | shardId | blockId
+```
+
+Consequences:
+
+- **~1000× fewer MDT records.** One block record replaces ~1–4K per-vector 
records, cutting
+  write amplification, compaction cost, and cleaner load by three orders of 
magnitude.
+- **Contiguous cluster scans.** The binary key scheme makes "scan cluster 
12345" a single
+  contiguous HFile range read rather than thousands of point lookups.
+- **Pay only for promise.** Column-wise layout means pass 1 reads only sign 
planes + factors;
+  only survivors touch extra planes; only finalists touch locators and keys 
(§6.2).
+
+### 4.2 Row families
+
+The `vector_index_<name>` partition holds several record families under one 
binary-sorted key
+scheme, so one prefix scan of a cluster returns its blocks and any fresh 
deltas together:
+
+| Key family | Cardinality | Purpose |
+|---|---:|---|
+| `__manifest__` | 1 | Active generation pointer and persisted-format version. 
|
+| `M\|<generation>` | generations | Generation state, vector schema, bootstrap 
baseline, verified-contiguous frontier, chunk counts/checksums, and quantizer 
metadata. |
+| `T\|<generation>\|<chunk>` | chunks per generation | Size-bounded chunks of 
the serialized `K × D` centroid matrix. |
+| `C\|<generation>\|<cluster>` | K per generation | Cluster manifest: routing 
version, shard count, vector count, candidate file groups, and counters. |
+| `P\|<generation>\|<cluster>\|<shard>\|<blockId>` | blocks per generation | 
**Posting block** (packed codes, factors, locators, keys). |
+| `P\|...\|<DELTA>` | deltas | Small per-record delta records appended between 
compactions. |
+| `F\|<generation>\|<dataInstant>` | data writes after baseline | Atomic proof 
that the generation incorporated one source data-write instant. |
+
+### 4.3 Posting shards
+
+Large clusters are split into posting shards so one hot cluster does not 
become one oversized
+prefix range. The cluster manifest stores `shardCount` and `routingVersion`; 
writers compute
+`shardId = hash(record_key) % shardCount`. Changing `shardCount` changes every 
key's mapping,
+so maintenance must rewrite the whole cluster under a new routing version and 
publish the
+manifest change atomically with that rewrite. Large remaps use a new 
generation rather than
+mixing routing versions in one cluster range. Within a shard, entries are 
packed by `blockId`.
+
+### 4.4 Delta records
+
+Between compactions, per-commit vector writes append small **delta records** 
(`blockId`
+marked `DELTA`) at the end of the same cluster key range. Because they share 
the prefix, a
+single cluster prefix scan sees packed blocks and fresh deltas in one pass 
(§6, §7).
+
+### 4.5 Generation model
+
+A generation is a consistent set of centroid, quantizer, cluster, and 
posting-block metadata:
+
+```text
+__manifest__      -> active generation id + persisted-format version
+M|<gen>           -> generation metadata + schema + verified frontier + chunk 
integrity
+T|<gen>|...       -> size-bounded centroid chunks
+C|<gen>|...       -> cluster manifests
+P|<gen>|...       -> posting blocks + deltas
+F|<gen>|...       -> incorporated data-write markers
+```
+
+The builder allocates a generation id once and writes deterministic keys, so 
retrying a
+partially completed build overwrites the same records. A `BUILDING` generation 
is invisible
+to readers; bootstrap validates expected chunks, checksums, cluster/block 
counts, schema, and
+memory budgets before one MDT commit changes it to `ACTIVE` and flips 
`__manifest__`.
+Abandoned or invalid `BUILDING` generations are never activated and may be 
garbage-collected.
+Old generations become `RETIRED` and remain until no retained snapshot or 
pinned reader needs
+them. Generation is the only independent version axis: centroid, quantizer, 
routing, and
+posting changes that cannot be published atomically in place create a new 
generation.
+
+---
+
+## 5. Bootstrap and Write Path
+
+![RFC-109 bootstrap, write, and maintenance 
path](diagrams/03-write-maintenance.svg)
+
+### 5.1 Spark bootstrap
+
+Bootstrap and full rebuild are Spark-only in v1. They build an invisible 
generation from one
+pinned table snapshot:
+
+```text
+1. Read latest file slices; extract key, partition, file group, base instant, 
row position,
+   and vector bytes; train IVF centroids from a bounded sample.
+2. Validate K × D × elementWidth and driver/executor memory budgets before 
materialization.
+3. Broadcast centroids; assign each vector, encode RaBitQ metadata, sort by
+   (cluster, fileGroup, rowPosition), and pack posting blocks.
+4. Write M|, size-bounded T| centroid chunks, C|, and P| records under 
deterministic keys.
+5. Validate chunk counts/checksums and posting counts, then atomically 
activate the generation.
+```
+
+The training sample satisfies percentage and per-cluster floors, for example
+`min(N, max(1M, 256*K, min(10M, 0.5%–1% of N)))`. Chunking keeps individual 
MDT records
+bounded; readers reconstruct the matrix only after validating manifest 
integrity metadata.
+
+### 5.2 Incremental inserts and vector updates
+
+For each relevant row, the metadata writer's vector hook computes the cluster, 
shard, RaBitQ
+code, factors, and base-table location and appends a posting delta. If an 
update moves a
+record, delta supersession and query-time RLI arbitration hide the stale entry 
(§6.3).
+The hook also emits the commit's freshness marker in the same MDT commit as 
its vector deltas.
+This uses the existing per-data-commit `Indexer.buildUpdate` dispatch; 
`VectorIndexer` returns
+an `F|...` record even without deltas, so its update is never empty and needs 
no new seam.
+
+### 5.3 Non-vector updates and deletes
+
+A non-vector update can retain its code and placement; a moved locator is 
re-resolved through
+RLI and refreshed by maintenance before the old slice is cleaned. Deletes 
require no posting
+delta because RLI arbitration removes deleted candidates. The hook still emits 
one freshness
+marker when a data write produces no vector records—including schema-only or 
allowed-empty
+commits—so an empty update cannot create a permanent false-stale gap.
+
+### 5.4 Engine support
+
+Snapshot extraction and centroid training are engine-specific; routing, 
encoding, record
+construction, freshness markers, and publication are common contracts. Spark 
supplies
+bootstrap/rebuild in v1. Feature-aware Spark, Flink, and Java writers may 
maintain an active
+generation through the common metadata-writer hook; unaware writers keep 
standard MDT
+layer-2 skip semantics, with their commits detected by the query-time 
freshness gate (§9).
+
+### 5.5 Multi-writer rebuild and catch-up
+
+Rebuilds are replay-only: writers continue updating the active generation and 
never dual-write

Review Comment:
   Concurrency behavior is currently scattered across this section and §9.1, 
and it's hard to convince ourselves it's complete. Can we add a table spelling 
out the scenarios across OCC and NBCC explicitly? 
   
   Rows to cover : (a) two OCC writers, both vector-aware; (b) OCC writer + 
concurrent rebuild; (c) NBCC/multi-writer deltacommits with vector deltas 
landing out of order; (d) vector-aware + legacy writer interleaved; (e) rebuild 
racing clustering/replacecommit; (f) async MDT compaction racing the vector 
hook. For each: what wins, what the marker frontier does, and what the reader 
observes (fresh / stale-gap / FAIL). Even if the answers all fall out of 
TransactionManager + the metadata lock + markers, having the matrix in one 
place will save every reviewer re-deriving it.
   
   
   I am happy to take this on myself as well and push a commit. please lmk
   



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)
+
+- ANN families beyond IVF + RaBitQ (e.g. HNSW, DiskANN).
+- Filtered search (arbitrary predicate + kNN) as a first-class planned 
operation.
+- Time-travel-consistent index reads for historical snapshots.
+- Native non-Spark generation construction, GPU encoding, and 
workload-specific auto-tuning.
+
+### 1.3 Alternatives considered
+
+- **One index record per vector** is simple but creates billions of MDT 
records and excessive
+  write, compaction, and scan amplification; posting blocks preserve MDT 
ownership while
+  amortizing that overhead.
+- **Dedicated index files in the table** permit specialized layouts but 
introduce a second
+  commit, cleaning, and snapshot protocol. MDT records reuse Hudi's existing 
transaction and
+  table-service machinery.
+- **External sidecar indexes or vector databases** may offer richer serving 
features but lose
+  atomic Hudi snapshot semantics and require another storage system. They 
remain valid when
+  independent serving infrastructure is desired.
+- **Native ANN libraries** improve local kernels but do not define durable 
object-store layout,
+  multi-writer maintenance, or snapshot visibility. They may be used behind 
the interfaces in
+  this RFC without changing the persisted contract.
+
+---
+
+## 2. Architecture
+
+![RFC-109 architecture overview](diagrams/01-architecture-overview.svg)
+
+The design splits responsibilities the way Hudi already does between the data 
table and the
+metadata table:
+
+```text
+DATA TABLE (parquet/orc)                METADATA TABLE (vector_index partition)
+  authoritative vectors + payload  ←──   the index: centroids, quantizer, 
posting blocks,
+  read only for final re-ranking          cluster manifests, generation 
manifest
+                                          read for candidate generation
+```
+
+Each vector index is one MDT partition. RFC-109 adds no table properties, 
common writer-
+dispatch changes, timeline-semantics changes, or behavior visible to 
non-vector readers or
+writers. Every consistency mechanism is either a record in the vector-index 
MDT partition or
+logic in the RFC-owned vector indexer and query planner. Creating an index:
+
+```sql
+CREATE INDEX embedding_idx
+ON products
+USING VECTOR (embedding)
+OPTIONS (
+  'vector.metric'      = 'cosine',
+  'vector.quantizer'   = 'IVF_RABITQ',
+  'vector.num_clusters'= '4096'
+);
+```
+
+creates:
+
+```text
+.hoodie/metadata/vector_index_embedding_idx/
+```
+
+and does not add generated columns to the base-table schema. RFC-109 consumes 
rather than
+redefines the RFC-99 vector contract: the source must be a top-level Hudi
+`VECTOR(D[, elementType])`. The table schema is authoritative for `D` and 
element type;
+index definitions and generation manifests repeat them only for integrity 
validation.
+
+The current RFC-99 storage backing is fixed-width bytes: Avro `FIXED` and 
Parquet
+`FIXED_LEN_BYTE_ARRAY(D × elementWidth)`. Engine adapters expose idiomatic 
values—Spark uses
+an annotated `ArrayType(FloatType|DoubleType|ByteType)`—and convert at the 
storage boundary.
+This fixed width also avoids Parquet LIST repetition-level traversal during 
positional exact
+fetch. A plain `ARRAY<FLOAT>` is not implicitly indexable and requires 
explicit migration or
+backfill to `VECTOR(D)`; index creation neither reinterprets nor rewrites it.
+
+The query path uses the MDT first to discover candidates, then reads 
base-table vectors for
+exact re-ranking:
+
+```text
+query vector
+  → compare to centroids, pick nprobe clusters          (in-memory, ms)
+  → MDT prefix-scan those clusters' posting blocks       (targeted range reads)
+  → two-pass RaBitQ scoring, keep refineFactor·K best    (bit math + error 
bounds)
+  → validate candidate freshness via Record Level Index  (batched point 
lookups)
+  → fetch ONLY those rows from the base table by position (page-level reads)
+  → exact distance on real vectors → final top-K
+```
+
+---
+
+## 3. IVF + RaBitQ Index Algorithm
+
+Every practical ANN index answers two questions: **where to look** (avoid 
scanning
+everything) and **how to compare cheaply** (avoid full-precision math on what 
is scanned).
+
+### 3.1 Where to look: IVF routing
+
+Inverted File (IVF) indexing clusters vectors with KMeans into `numClusters` 
groups (e.g.
+~4K–64K). Each vector belongs to its nearest centroid. A query compares 
against the centroids
+only (thousands, not billions), selects the `nprobe` nearest clusters, and 
scans only those
+clusters' entries. `nprobe` is the recall dial.
+
+IVF is the right fit for a lakehouse-resident index because a cluster's 
entries can be stored
+**contiguously**, which maps directly onto sorted key ranges in Hudi's MDT 
(§4). Graph
+indexes (HNSW) give excellent in-memory recall but require random traversal of 
the whole
+graph, which fights columnar, immutable, object-store storage.
+
+### 3.2 How to compare cheaply: RaBitQ quantization
+
+Inside a probed cluster there are still thousands of full vectors. 
Quantization stores a
+small *code* per vector plus a few correction scalars, so most comparisons run 
on compressed
+codes and only the best few hundred candidates are re-checked against real 
vectors.
+
+RaBitQ is chosen over scalar (SQ), product (PQ), and plain binary quantization 
for four
+reasons:
+
+1. **Unbiased estimator with a provable per-vector error bound.** Each code 
carries scalars
+   that turn a cheap bit-level dot product into an *unbiased* estimate of the 
true distance,
+   plus a bound on how wrong it can be. The bound enables **safe pruning**: 
skip a vector
+   only when even its best plausible distance cannot make top-K.
+2. **No codebooks.** RaBitQ needs only a random rotation (a seed) and the 
centroids — both
+   tiny, both versioned in the index metadata. Nothing to retrain when data 
drifts.
+3. **Tunable precision.** B = 1 bit/dim is a fast coarse filter; B = 4 
bits/dim gives
+   near-SQ quality at ~8× less space. Both are used in a two-pass scan (§6.2).
+4. **Metric-flexible.** One stored code serves L2, cosine, and dot-product; 
the metric is
+   applied at query time.
+
+The four ideas, precisely:
+
+- **Residual.** Store each vector as its difference from its centroid, `r = v 
− c`.
+  Residuals are small and centered, which is what lets few bits go far.
+- **Rotation.** Apply one fixed random orthonormal rotation `R` (derived from 
a per-generation
+  seed) to everything first: `x = R·r`. This spreads information evenly across 
dimensions,
+  which is what makes the error bound hold for any data distribution. Only the 
seed is stored.
+- **Code.** Quantize `x` to B bits per dimension, stored as **bit planes** — 
plane 0 holds
+  bit 0 of every dimension, plane 1 bit 1, etc. The top plane alone is a 1-bit 
sign sketch.
+  Scoring a plane against the (transformed) query is `AND`/`XOR` + `popcount` 
— a few CPU
+  instructions per 64 dimensions.
+- **Factors.** A handful of small scalars per vector (residual norm, two 
rescale factors, an
+  error term, a centroid correction) that convert plane math into an unbiased 
distance
+  estimate plus its confidence interval.
+
+At query time the query vector is transformed the same way (`R·(q − c)` per 
probed cluster),
+and the per-plane popcounts combine with the stored factors into the estimate 
and its bound.
+The estimate builds the shortlist; exact base-table distances produce the 
final ranking (§6).
+
+### 3.3 Why this fits Hudi
+
+- Centroids are small enough to load at planning time: `K × D` floats (~12 MB 
for K=4096,
+  D=768).
+- Codes are compact: a 1B × 128-dim float table's raw vectors are ~512 GB; the 
RaBitQ index
+  including keys and locators is ~136 GB, and the scanned portion per query is 
tens of MB.
+- Posting keys are prefix-scannable by generation, cluster, and shard (§4).
+- Quantizer state is stable: a seed and centroids, no per-generation learned 
codebook.
+- Exact re-ranking preserves correctness for returned candidates.
+
+---
+
+## 4. Metadata Table Storage Model: the Posting Block
+
+This section describes the core storage contribution of this RFC.
+
+### 4.1 The posting block
+
+A naive index would write one MDT record per indexed vector. At a billion rows 
that is a
+billion MDT records per generation — prohibitive to write, compact, scan, and 
clean.
+
+Instead, the index sorts entries by `(cluster, fileGroup, rowPosition)` and 
packs
+~1–4K vectors' worth of codes and metadata into a single MDT record — a 
**posting block** —
+laid out column-wise (structure-of-arrays) so each scan pass touches only the 
columns it
+needs:
+
+```text
+POSTING BLOCK (~512 KB target, one MDT record)
+┌───────────────────────────────────────────────┐
+│ S1  sign planes        ← pass 1 touches this   │
+│ S2  extra bit planes   ← pass 2, survivors     │
+│ S3  factor arrays      ← both passes           │
+│ S4  row locators       ← finalists only        │
+│ S5  dictionaries (file groups, partitions)     │
+│ S6  record keys        ← finalists only        │
+└───────────────────────────────────────────────┘
+key:  0x10 | generation | clusterId | shardId | blockId
+```
+
+Consequences:
+
+- **~1000× fewer MDT records.** One block record replaces ~1–4K per-vector 
records, cutting
+  write amplification, compaction cost, and cleaner load by three orders of 
magnitude.
+- **Contiguous cluster scans.** The binary key scheme makes "scan cluster 
12345" a single
+  contiguous HFile range read rather than thousands of point lookups.
+- **Pay only for promise.** Column-wise layout means pass 1 reads only sign 
planes + factors;
+  only survivors touch extra planes; only finalists touch locators and keys 
(§6.2).
+
+### 4.2 Row families
+
+The `vector_index_<name>` partition holds several record families under one 
binary-sorted key
+scheme, so one prefix scan of a cluster returns its blocks and any fresh 
deltas together:
+
+| Key family | Cardinality | Purpose |
+|---|---:|---|
+| `__manifest__` | 1 | Active generation pointer and persisted-format version. 
|
+| `M\|<generation>` | generations | Generation state, vector schema, bootstrap 
baseline, verified-contiguous frontier, chunk counts/checksums, and quantizer 
metadata. |
+| `T\|<generation>\|<chunk>` | chunks per generation | Size-bounded chunks of 
the serialized `K × D` centroid matrix. |
+| `C\|<generation>\|<cluster>` | K per generation | Cluster manifest: routing 
version, shard count, vector count, candidate file groups, and counters. |
+| `P\|<generation>\|<cluster>\|<shard>\|<blockId>` | blocks per generation | 
**Posting block** (packed codes, factors, locators, keys). |
+| `P\|...\|<DELTA>` | deltas | Small per-record delta records appended between 
compactions. |
+| `F\|<generation>\|<dataInstant>` | data writes after baseline | Atomic proof 
that the generation incorporated one source data-write instant. |
+
+### 4.3 Posting shards
+
+Large clusters are split into posting shards so one hot cluster does not 
become one oversized
+prefix range. The cluster manifest stores `shardCount` and `routingVersion`; 
writers compute
+`shardId = hash(record_key) % shardCount`. Changing `shardCount` changes every 
key's mapping,
+so maintenance must rewrite the whole cluster under a new routing version and 
publish the
+manifest change atomically with that rewrite. Large remaps use a new 
generation rather than
+mixing routing versions in one cluster range. Within a shard, entries are 
packed by `blockId`.
+
+### 4.4 Delta records
+
+Between compactions, per-commit vector writes append small **delta records** 
(`blockId`
+marked `DELTA`) at the end of the same cluster key range. Because they share 
the prefix, a
+single cluster prefix scan sees packed blocks and fresh deltas in one pass 
(§6, §7).
+
+### 4.5 Generation model
+
+A generation is a consistent set of centroid, quantizer, cluster, and 
posting-block metadata:
+
+```text
+__manifest__      -> active generation id + persisted-format version
+M|<gen>           -> generation metadata + schema + verified frontier + chunk 
integrity
+T|<gen>|...       -> size-bounded centroid chunks
+C|<gen>|...       -> cluster manifests
+P|<gen>|...       -> posting blocks + deltas
+F|<gen>|...       -> incorporated data-write markers
+```
+
+The builder allocates a generation id once and writes deterministic keys, so 
retrying a
+partially completed build overwrites the same records. A `BUILDING` generation 
is invisible
+to readers; bootstrap validates expected chunks, checksums, cluster/block 
counts, schema, and
+memory budgets before one MDT commit changes it to `ACTIVE` and flips 
`__manifest__`.
+Abandoned or invalid `BUILDING` generations are never activated and may be 
garbage-collected.
+Old generations become `RETIRED` and remain until no retained snapshot or 
pinned reader needs
+them. Generation is the only independent version axis: centroid, quantizer, 
routing, and
+posting changes that cannot be published atomically in place create a new 
generation.
+
+---
+
+## 5. Bootstrap and Write Path
+
+![RFC-109 bootstrap, write, and maintenance 
path](diagrams/03-write-maintenance.svg)
+
+### 5.1 Spark bootstrap
+
+Bootstrap and full rebuild are Spark-only in v1. They build an invisible 
generation from one
+pinned table snapshot:
+
+```text
+1. Read latest file slices; extract key, partition, file group, base instant, 
row position,
+   and vector bytes; train IVF centroids from a bounded sample.
+2. Validate K × D × elementWidth and driver/executor memory budgets before 
materialization.
+3. Broadcast centroids; assign each vector, encode RaBitQ metadata, sort by
+   (cluster, fileGroup, rowPosition), and pack posting blocks.
+4. Write M|, size-bounded T| centroid chunks, C|, and P| records under 
deterministic keys.
+5. Validate chunk counts/checksums and posting counts, then atomically 
activate the generation.
+```
+
+The training sample satisfies percentage and per-cluster floors, for example
+`min(N, max(1M, 256*K, min(10M, 0.5%–1% of N)))`. Chunking keeps individual 
MDT records
+bounded; readers reconstruct the matrix only after validating manifest 
integrity metadata.
+
+### 5.2 Incremental inserts and vector updates
+
+For each relevant row, the metadata writer's vector hook computes the cluster, 
shard, RaBitQ
+code, factors, and base-table location and appends a posting delta. If an 
update moves a
+record, delta supersession and query-time RLI arbitration hide the stale entry 
(§6.3).
+The hook also emits the commit's freshness marker in the same MDT commit as 
its vector deltas.
+This uses the existing per-data-commit `Indexer.buildUpdate` dispatch; 
`VectorIndexer` returns
+an `F|...` record even without deltas, so its update is never empty and needs 
no new seam.
+
+### 5.3 Non-vector updates and deletes
+
+A non-vector update can retain its code and placement; a moved locator is 
re-resolved through

Review Comment:
   This section assumes we can tell a "non-vector update" apart from a vector 
update — but how do we actually detect that the vector column is unchanged?  
Needs fleshing out or a way to safely deal with it with slower perf (treat as 
vector update when in doubt?).



##########
rfc/rfc-109/rfc-109.md:
##########
@@ -0,0 +1,714 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one or more
+  contributor license agreements.  See the NOTICE file distributed with
+  this work for additional information regarding copyright ownership.
+  The ASF licenses this file to You under the Apache License, Version 2.0
+  (the "License"); you may not use this file except in compliance with
+  the License.  You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+-->
+
+# RFC-109: Native Vector Search Support in Apache Hudi
+
+## Proposers
+
+@chrevanthreddy
+
+## Approvers
+
+- TBD
+
+## Status
+
+Umbrella issue: 
[apache/hudi#19094](https://github.com/apache/hudi/issues/19094)
+
+Related: [apache/hudi#18676](https://github.com/apache/hudi/issues/18676)
+
+State: UNDER REVIEW
+
+---
+
+## Table of Contents
+
+- [Abstract](#abstract)
+- [1. Goals and Non-Goals](#1-goals-and-non-goals)
+- [2. Architecture](#2-architecture)
+- [3. IVF + RaBitQ Index Algorithm](#3-ivf--rabitq-index-algorithm)
+- [4. Metadata Table Storage Model: the Posting 
Block](#4-metadata-table-storage-model-the-posting-block)
+- [5. Bootstrap and Write Path](#5-bootstrap-and-write-path)
+- [6. Read Path](#6-read-path)
+- [7. Maintenance, Rebalancing, and 
Cleaner](#7-maintenance-rebalancing-and-cleaner)
+- [8. Spark API Surface](#8-spark-api-surface)
+- [9. Correctness, Compatibility, and 
Freshness](#9-correctness-compatibility-and-freshness)
+- [10. Test Plan](#10-test-plan)
+- [11. Benchmark Evidence, Rollout, and MVP 
Scope](#11-benchmark-evidence-rollout-and-mvp-scope)
+- [12. References](#12-references)
+
+---
+
+## Abstract
+
+This RFC proposes native approximate nearest-neighbor (ANN) vector search in 
Apache Hudi.
+Tables increasingly carry embedding columns represented by Hudi's 
fixed-dimension
+`VECTOR(D[, elementType])` logical type, and users want to ask *"find the K 
rows most
+similar to this query vector"* — for semantic search, recommendations, RAG, and
+deduplication — without copying data into a separate vector database.
+
+Today the only option on a Hudi table is a brute-force scan: read every 
vector, compute
+every distance. That is correct but scales linearly with table size (tens of 
seconds at a
+billion rows). This RFC adds an index so that vector queries read only a 
small, targeted
+fraction of the index and the table, return results with high recall, and stay
+transactionally consistent with the table under upserts and deletes — all with 
**no new
+storage system**. The index lives in the Hudi Metadata Table (MDT), like 
Hudi's existing
+record-level and secondary indexes, and is maintained by the same table 
services.
+
+The design combines three well-understood pieces — IVF clustering, RaBitQ 
quantization, and
+exact re-ranking — with one storage innovation that makes them practical on an 
immutable,
+columnar, object-store-resident lakehouse:
+
+> **The posting block.** Instead of one MDT record per indexed vector, the 
index packs
+> ~1–4K vectors into a single MDT record laid out column-wise 
(structure-of-arrays), keyed
+> so that one IVF cluster forms one contiguous, prefix-scannable key range. 
This reduces MDT
+> record count by roughly three orders of magnitude, turns "scan a cluster" 
into a single
+> contiguous range read, and lets a query touch only the columns a given scan 
pass needs.
+
+The base table remains the source of truth for exact vector values. The MDT 
stores only
+routing, pruning, and approximate-scoring metadata; final ranking always reads 
exact vectors
+from the base table.
+
+Artifacted measurements on a **1-billion-row, 128-dimensional table** show 
approximate-only
+recall@10 of 0.824, 0.843, and 0.845 at `nprobe=16`, `32`, and `64`; exact 
reranking reaches
+0.983 at `nprobe=32` with `refineFactor=50` (§11). These modes serve different 
goals:
+approximate-only provides the latency floor, while exact reranking is the 
recommended
+quality mode.
+
+---
+
+## 1. Goals and Non-Goals
+
+### 1.1 Goals
+
+1. Keep authoritative vector values in a top-level base-table `VECTOR(D[, 
elementType])` column.
+2. Store the vector index in the MDT, maintained by Hudi metadata-table 
commits, compaction,
+   and cleaning — no hidden or generated columns in base-table files.
+3. Make candidate discovery cheap and targeted: probe a few clusters, scan 
contiguous key
+   ranges, score on compressed codes with a provable pruning bound.
+4. Make results trustworthy: approximate math selects candidates; **exact** 
distance on
+   base-table vectors ranks them.
+5. Stay transactionally consistent: snapshot-pinned reads, correct behavior 
under inserts,
+   updates, deletes, and clustering.
+6. Be engine-neutral in design; Spark is the first implementation.
+7. Maintain the index incrementally (no global rebuild for normal churn) and 
support
+   versioned, zero-downtime rebuilds.
+
+### 1.2 Non-Goals (initial landing)
+
+- ANN families beyond IVF + RaBitQ (e.g. HNSW, DiskANN).
+- Filtered search (arbitrary predicate + kNN) as a first-class planned 
operation.
+- Time-travel-consistent index reads for historical snapshots.
+- Native non-Spark generation construction, GPU encoding, and 
workload-specific auto-tuning.
+
+### 1.3 Alternatives considered
+
+- **One index record per vector** is simple but creates billions of MDT 
records and excessive
+  write, compaction, and scan amplification; posting blocks preserve MDT 
ownership while
+  amortizing that overhead.
+- **Dedicated index files in the table** permit specialized layouts but 
introduce a second
+  commit, cleaning, and snapshot protocol. MDT records reuse Hudi's existing 
transaction and
+  table-service machinery.
+- **External sidecar indexes or vector databases** may offer richer serving 
features but lose
+  atomic Hudi snapshot semantics and require another storage system. They 
remain valid when
+  independent serving infrastructure is desired.
+- **Native ANN libraries** improve local kernels but do not define durable 
object-store layout,
+  multi-writer maintenance, or snapshot visibility. They may be used behind 
the interfaces in
+  this RFC without changing the persisted contract.
+
+---
+
+## 2. Architecture
+
+![RFC-109 architecture overview](diagrams/01-architecture-overview.svg)
+
+The design splits responsibilities the way Hudi already does between the data 
table and the
+metadata table:
+
+```text
+DATA TABLE (parquet/orc)                METADATA TABLE (vector_index partition)
+  authoritative vectors + payload  ←──   the index: centroids, quantizer, 
posting blocks,
+  read only for final re-ranking          cluster manifests, generation 
manifest
+                                          read for candidate generation
+```
+
+Each vector index is one MDT partition. RFC-109 adds no table properties, 
common writer-
+dispatch changes, timeline-semantics changes, or behavior visible to 
non-vector readers or
+writers. Every consistency mechanism is either a record in the vector-index 
MDT partition or
+logic in the RFC-owned vector indexer and query planner. Creating an index:
+
+```sql
+CREATE INDEX embedding_idx
+ON products
+USING VECTOR (embedding)
+OPTIONS (
+  'vector.metric'      = 'cosine',
+  'vector.quantizer'   = 'IVF_RABITQ',
+  'vector.num_clusters'= '4096'
+);
+```
+
+creates:
+
+```text
+.hoodie/metadata/vector_index_embedding_idx/
+```
+
+and does not add generated columns to the base-table schema. RFC-109 consumes 
rather than
+redefines the RFC-99 vector contract: the source must be a top-level Hudi
+`VECTOR(D[, elementType])`. The table schema is authoritative for `D` and 
element type;
+index definitions and generation manifests repeat them only for integrity 
validation.
+
+The current RFC-99 storage backing is fixed-width bytes: Avro `FIXED` and 
Parquet
+`FIXED_LEN_BYTE_ARRAY(D × elementWidth)`. Engine adapters expose idiomatic 
values—Spark uses
+an annotated `ArrayType(FloatType|DoubleType|ByteType)`—and convert at the 
storage boundary.
+This fixed width also avoids Parquet LIST repetition-level traversal during 
positional exact
+fetch. A plain `ARRAY<FLOAT>` is not implicitly indexable and requires 
explicit migration or
+backfill to `VECTOR(D)`; index creation neither reinterprets nor rewrites it.
+
+The query path uses the MDT first to discover candidates, then reads 
base-table vectors for
+exact re-ranking:
+
+```text
+query vector
+  → compare to centroids, pick nprobe clusters          (in-memory, ms)
+  → MDT prefix-scan those clusters' posting blocks       (targeted range reads)
+  → two-pass RaBitQ scoring, keep refineFactor·K best    (bit math + error 
bounds)
+  → validate candidate freshness via Record Level Index  (batched point 
lookups)
+  → fetch ONLY those rows from the base table by position (page-level reads)
+  → exact distance on real vectors → final top-K
+```
+
+---
+
+## 3. IVF + RaBitQ Index Algorithm
+
+Every practical ANN index answers two questions: **where to look** (avoid 
scanning
+everything) and **how to compare cheaply** (avoid full-precision math on what 
is scanned).
+
+### 3.1 Where to look: IVF routing
+
+Inverted File (IVF) indexing clusters vectors with KMeans into `numClusters` 
groups (e.g.
+~4K–64K). Each vector belongs to its nearest centroid. A query compares 
against the centroids
+only (thousands, not billions), selects the `nprobe` nearest clusters, and 
scans only those
+clusters' entries. `nprobe` is the recall dial.
+
+IVF is the right fit for a lakehouse-resident index because a cluster's 
entries can be stored
+**contiguously**, which maps directly onto sorted key ranges in Hudi's MDT 
(§4). Graph
+indexes (HNSW) give excellent in-memory recall but require random traversal of 
the whole
+graph, which fights columnar, immutable, object-store storage.
+
+### 3.2 How to compare cheaply: RaBitQ quantization
+
+Inside a probed cluster there are still thousands of full vectors. 
Quantization stores a
+small *code* per vector plus a few correction scalars, so most comparisons run 
on compressed
+codes and only the best few hundred candidates are re-checked against real 
vectors.
+
+RaBitQ is chosen over scalar (SQ), product (PQ), and plain binary quantization 
for four
+reasons:
+
+1. **Unbiased estimator with a provable per-vector error bound.** Each code 
carries scalars
+   that turn a cheap bit-level dot product into an *unbiased* estimate of the 
true distance,
+   plus a bound on how wrong it can be. The bound enables **safe pruning**: 
skip a vector
+   only when even its best plausible distance cannot make top-K.
+2. **No codebooks.** RaBitQ needs only a random rotation (a seed) and the 
centroids — both
+   tiny, both versioned in the index metadata. Nothing to retrain when data 
drifts.
+3. **Tunable precision.** B = 1 bit/dim is a fast coarse filter; B = 4 
bits/dim gives
+   near-SQ quality at ~8× less space. Both are used in a two-pass scan (§6.2).
+4. **Metric-flexible.** One stored code serves L2, cosine, and dot-product; 
the metric is
+   applied at query time.
+
+The four ideas, precisely:
+
+- **Residual.** Store each vector as its difference from its centroid, `r = v 
− c`.
+  Residuals are small and centered, which is what lets few bits go far.
+- **Rotation.** Apply one fixed random orthonormal rotation `R` (derived from 
a per-generation
+  seed) to everything first: `x = R·r`. This spreads information evenly across 
dimensions,
+  which is what makes the error bound hold for any data distribution. Only the 
seed is stored.
+- **Code.** Quantize `x` to B bits per dimension, stored as **bit planes** — 
plane 0 holds
+  bit 0 of every dimension, plane 1 bit 1, etc. The top plane alone is a 1-bit 
sign sketch.
+  Scoring a plane against the (transformed) query is `AND`/`XOR` + `popcount` 
— a few CPU
+  instructions per 64 dimensions.
+- **Factors.** A handful of small scalars per vector (residual norm, two 
rescale factors, an
+  error term, a centroid correction) that convert plane math into an unbiased 
distance
+  estimate plus its confidence interval.
+
+At query time the query vector is transformed the same way (`R·(q − c)` per 
probed cluster),
+and the per-plane popcounts combine with the stored factors into the estimate 
and its bound.
+The estimate builds the shortlist; exact base-table distances produce the 
final ranking (§6).
+
+### 3.3 Why this fits Hudi
+
+- Centroids are small enough to load at planning time: `K × D` floats (~12 MB 
for K=4096,
+  D=768).
+- Codes are compact: a 1B × 128-dim float table's raw vectors are ~512 GB; the 
RaBitQ index
+  including keys and locators is ~136 GB, and the scanned portion per query is 
tens of MB.
+- Posting keys are prefix-scannable by generation, cluster, and shard (§4).
+- Quantizer state is stable: a seed and centroids, no per-generation learned 
codebook.
+- Exact re-ranking preserves correctness for returned candidates.
+
+---
+
+## 4. Metadata Table Storage Model: the Posting Block
+
+This section describes the core storage contribution of this RFC.
+
+### 4.1 The posting block
+
+A naive index would write one MDT record per indexed vector. At a billion rows 
that is a
+billion MDT records per generation — prohibitive to write, compact, scan, and 
clean.
+
+Instead, the index sorts entries by `(cluster, fileGroup, rowPosition)` and 
packs
+~1–4K vectors' worth of codes and metadata into a single MDT record — a 
**posting block** —
+laid out column-wise (structure-of-arrays) so each scan pass touches only the 
columns it
+needs:
+
+```text
+POSTING BLOCK (~512 KB target, one MDT record)
+┌───────────────────────────────────────────────┐
+│ S1  sign planes        ← pass 1 touches this   │
+│ S2  extra bit planes   ← pass 2, survivors     │
+│ S3  factor arrays      ← both passes           │
+│ S4  row locators       ← finalists only        │
+│ S5  dictionaries (file groups, partitions)     │
+│ S6  record keys        ← finalists only        │
+└───────────────────────────────────────────────┘
+key:  0x10 | generation | clusterId | shardId | blockId
+```
+
+Consequences:
+
+- **~1000× fewer MDT records.** One block record replaces ~1–4K per-vector 
records, cutting
+  write amplification, compaction cost, and cleaner load by three orders of 
magnitude.
+- **Contiguous cluster scans.** The binary key scheme makes "scan cluster 
12345" a single
+  contiguous HFile range read rather than thousands of point lookups.
+- **Pay only for promise.** Column-wise layout means pass 1 reads only sign 
planes + factors;
+  only survivors touch extra planes; only finalists touch locators and keys 
(§6.2).
+
+### 4.2 Row families
+
+The `vector_index_<name>` partition holds several record families under one 
binary-sorted key
+scheme, so one prefix scan of a cluster returns its blocks and any fresh 
deltas together:
+
+| Key family | Cardinality | Purpose |
+|---|---:|---|
+| `__manifest__` | 1 | Active generation pointer and persisted-format version. 
|
+| `M\|<generation>` | generations | Generation state, vector schema, bootstrap 
baseline, verified-contiguous frontier, chunk counts/checksums, and quantizer 
metadata. |
+| `T\|<generation>\|<chunk>` | chunks per generation | Size-bounded chunks of 
the serialized `K × D` centroid matrix. |
+| `C\|<generation>\|<cluster>` | K per generation | Cluster manifest: routing 
version, shard count, vector count, candidate file groups, and counters. |
+| `P\|<generation>\|<cluster>\|<shard>\|<blockId>` | blocks per generation | 
**Posting block** (packed codes, factors, locators, keys). |
+| `P\|...\|<DELTA>` | deltas | Small per-record delta records appended between 
compactions. |
+| `F\|<generation>\|<dataInstant>` | data writes after baseline | Atomic proof 
that the generation incorporated one source data-write instant. |
+
+### 4.3 Posting shards
+
+Large clusters are split into posting shards so one hot cluster does not 
become one oversized
+prefix range. The cluster manifest stores `shardCount` and `routingVersion`; 
writers compute
+`shardId = hash(record_key) % shardCount`. Changing `shardCount` changes every 
key's mapping,
+so maintenance must rewrite the whole cluster under a new routing version and 
publish the
+manifest change atomically with that rewrite. Large remaps use a new 
generation rather than
+mixing routing versions in one cluster range. Within a shard, entries are 
packed by `blockId`.
+
+### 4.4 Delta records
+
+Between compactions, per-commit vector writes append small **delta records** 
(`blockId`
+marked `DELTA`) at the end of the same cluster key range. Because they share 
the prefix, a
+single cluster prefix scan sees packed blocks and fresh deltas in one pass 
(§6, §7).
+
+### 4.5 Generation model
+
+A generation is a consistent set of centroid, quantizer, cluster, and 
posting-block metadata:
+
+```text
+__manifest__      -> active generation id + persisted-format version
+M|<gen>           -> generation metadata + schema + verified frontier + chunk 
integrity
+T|<gen>|...       -> size-bounded centroid chunks
+C|<gen>|...       -> cluster manifests
+P|<gen>|...       -> posting blocks + deltas
+F|<gen>|...       -> incorporated data-write markers
+```
+
+The builder allocates a generation id once and writes deterministic keys, so 
retrying a
+partially completed build overwrites the same records. A `BUILDING` generation 
is invisible
+to readers; bootstrap validates expected chunks, checksums, cluster/block 
counts, schema, and
+memory budgets before one MDT commit changes it to `ACTIVE` and flips 
`__manifest__`.
+Abandoned or invalid `BUILDING` generations are never activated and may be 
garbage-collected.
+Old generations become `RETIRED` and remain until no retained snapshot or 
pinned reader needs
+them. Generation is the only independent version axis: centroid, quantizer, 
routing, and
+posting changes that cannot be published atomically in place create a new 
generation.
+
+---
+
+## 5. Bootstrap and Write Path
+
+![RFC-109 bootstrap, write, and maintenance 
path](diagrams/03-write-maintenance.svg)
+
+### 5.1 Spark bootstrap
+
+Bootstrap and full rebuild are Spark-only in v1. They build an invisible 
generation from one
+pinned table snapshot:
+
+```text
+1. Read latest file slices; extract key, partition, file group, base instant, 
row position,
+   and vector bytes; train IVF centroids from a bounded sample.
+2. Validate K × D × elementWidth and driver/executor memory budgets before 
materialization.
+3. Broadcast centroids; assign each vector, encode RaBitQ metadata, sort by
+   (cluster, fileGroup, rowPosition), and pack posting blocks.
+4. Write M|, size-bounded T| centroid chunks, C|, and P| records under 
deterministic keys.
+5. Validate chunk counts/checksums and posting counts, then atomically 
activate the generation.
+```
+
+The training sample satisfies percentage and per-cluster floors, for example
+`min(N, max(1M, 256*K, min(10M, 0.5%–1% of N)))`. Chunking keeps individual 
MDT records
+bounded; readers reconstruct the matrix only after validating manifest 
integrity metadata.
+
+### 5.2 Incremental inserts and vector updates
+
+For each relevant row, the metadata writer's vector hook computes the cluster, 
shard, RaBitQ
+code, factors, and base-table location and appends a posting delta. If an 
update moves a
+record, delta supersession and query-time RLI arbitration hide the stale entry 
(§6.3).
+The hook also emits the commit's freshness marker in the same MDT commit as 
its vector deltas.
+This uses the existing per-data-commit `Indexer.buildUpdate` dispatch; 
`VectorIndexer` returns
+an `F|...` record even without deltas, so its update is never empty and needs 
no new seam.
+
+### 5.3 Non-vector updates and deletes
+
+A non-vector update can retain its code and placement; a moved locator is 
re-resolved through
+RLI and refreshed by maintenance before the old slice is cleaned. Deletes 
require no posting
+delta because RLI arbitration removes deleted candidates. The hook still emits 
one freshness
+marker when a data write produces no vector records—including schema-only or 
allowed-empty
+commits—so an empty update cannot create a permanent false-stale gap.
+
+### 5.4 Engine support
+
+Snapshot extraction and centroid training are engine-specific; routing, 
encoding, record
+construction, freshness markers, and publication are common contracts. Spark 
supplies
+bootstrap/rebuild in v1. Feature-aware Spark, Flink, and Java writers may 
maintain an active
+generation through the common metadata-writer hook; unaware writers keep 
standard MDT
+layer-2 skip semantics, with their commits detected by the query-time 
freshness gate (§9).
+
+### 5.5 Multi-writer rebuild and catch-up
+
+Rebuilds are replay-only: writers continue updating the active generation and 
never dual-write
+to a `BUILDING` generation. After constructing at baseline `T_boot`, the 
builder replays
+completed data writes in timeline order outside the coordination lock. Replay 
uses the active
+timeline and, when repairing a gap older than its retained window, the 
existing archived-
+timeline APIs. Each replay writes the new generation's deltas and 
source-instant marker
+atomically. When the gap is bounded,
+`TransactionManager` and the metadata-index lock select `T_cut`, perform a 
final micro-catch-up,
+verify contiguous coverage, and publish the new generation in one MDT commit. 
If replay cannot
+converge within its round/time budget, the attempt exits without activation; 
it never holds an
+unbounded lock or publishes partial state.
+
+---
+
+## 6. Read Path
+
+![RFC-109 vector search read path](diagrams/02-read-path.svg)
+
+### 6.1 Query planning
+
+```text
+1. Pin the table snapshot and validate its vector schema against the active 
manifest.
+2. Enforce the generation's marker frontier at the pinned data-write instant 
(§9).
+3. Load and validate generation metadata, centroid chunks, quantizer, and 
cluster manifests.
+4. Probe centroids, select top-nprobe clusters, and resolve shard 
prefixes/file groups.
+```
+
+### 6.2 Two-pass scan over posting blocks
+
+For each selected `(cluster, shard)`, prefix-scan 
`P|<gen>|<cluster>|<shard>|*` (blocks +
+deltas) and score column-wise:
+
+```text
+Pass 1 (bound + prune): read S1 sign planes + S3 factors. For every vector 
compute an
+        optimistic bound — the best distance it could possibly achieve. Skip 
vectors whose
+        best case cannot beat the current K-th candidate (typically 85–95% 
pruned).
+Pass 2 (refine):        for survivors, read S2 extra bit planes and compute 
the full
+        multibit unbiased estimate; keep the refineFactor·K best.
+Finalize:               only the surviving finalists touch S4 locators and S6 
keys to learn
+        where they live and who they are.
+```
+
+Cost is proportional to promise, not to table size. Delta records encountered 
in the same
+scan are scored identically and supersede matching packed entries.
+
+### 6.3 Freshness arbitration and exact re-rank
+
+The Record Level Index (RLI) is a hard prerequisite. Vector-index creation 
rejects a table
+without RLI, and the vector planner resolves the dependency again for every 
query. If RLI was
+subsequently dropped by a path unaware of vector indexes, planning fails with 
the missing
+dependency and remediation rather than serving results. RFC-109 does not 
change the common
+index-drop path. Posting locators are hints; finalist keys are validated by 
batched RLI lookup
+on the pinned snapshot. Moved keys are re-resolved and deleted keys are 
removed.
+
+Candidate generation retains a surplus before arbitration. If stale/deleted 
candidates leave
+fewer than K live rows, staged continuation scans the next candidates/ranges 
and repeats RLI
+arbitration until K live rows are available or the request budget is 
exhausted. Exact mode
+then positionally reads those vectors and ranks by exact distance. Budget 
exhaustion invokes
+exact-scan fallback or fails explicitly; it never silently returns fewer than 
K as a complete
+answer. `indexLagInstants`, when present, is marker-frontier lag as defined in 
§9.
+
+### 6.4 Batched queries
+
+For a relation of query vectors, the plan shares MDT work: load generation 
metadata once,
+encode all queries, probe per query, group by selected `(cluster, shard)`, 
scan each range
+once while maintaining a per-query top-R heap, then read the union of 
candidate rows and
+exact re-rank per query.
+
+### 6.5 Fallback
+
+Missing, incompatible, stale, or budget-exhausted index state must not produce 
a silently
+incorrect answer. Per policy, Hudi fails, emits a warning with 
`indexLagInstants`, or bypasses
+the index for an exact table scan. Exact-rerank defaults to `FAIL` for stale 
state; callers
+that require strict availability may configure exact fallback.
+
+---
+
+## 7. Maintenance, Rebalancing, and Cleaner
+
+Vectors change. Index health is kept without global retraining or re-encoding 
through three
+tiers of increasing rarity (LIRE — Local Incremental Rebalancing).

Review Comment:
   Is LIRE in v1? §11 scopes the MVP to create/drop/rebuild + §7.1/§7.2-style 
maintenance, but this section fully specs §7.3 split/merge as well. Let's mark 
clearly here what ships in the first cut and note in the Limitations section 
that until then, sustained skew/drift is handled by full generation rebuild 
only?



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to