LoginSign Up
← Back to all sparks
O

OceanBase

DEVOPS
Velocity6.3

Distributed relational database with native support for HTAP workloads

oceanbase.github.io ↗

OceanBase is rebuilding itself as a RAG backend without giving up the HTAP story

distributed-databasehtapvector-searchraghigh-availabilitymulti-branch
◆Current state
OceanBase Community Edition runs at least five branches in parallel — 4.2.5, 4.3.5, 4.4.1, 4.4.2, 4.6.0 and now 5.0.1 — with feature releases on the newest lines and hotfix trains keeping older LTS branches alive. The centre of gravity in this window is V4.6.0, which added a native SQL hybrid-retrieval interface fusing vector, full-text and scalar predicates in one query and reworked the execution framework behind it. The hotfix branches show where that work is straining: 4.4.1 and 4.3.5 patches are dominated by vector-index bugs — HNSW memory blowups, IVF cache leaks, index rebuilds hanging, planner misjudgements that skip the vector index entirely.
◆Where it's heading
Two threads run side by side and neither is being sacrificed. One is AI data infrastructure: hybrid search, sparse vectors, recall evaluation, Document AI and an explicitly named end-to-end RAG package, plus AI Functions that now handle images. The other is the HTAP and availability core: columnar replica routing with consistent reads, primary/standby strong sync at RPO=0 graduating from experimental to supported inside two releases, and V5.0.1 completing online conversion between row, columnar and hybrid storage formats. Note that publication dates on this feed do not match the stated release dates in the entries, so cadence read off timestamps alone will be misleading.
◆Prediction
The volume of vector-index defect fixes across three hotfix branches suggests the next releases on the 5.0 line will spend more effort stabilising hybrid search than extending it, with the RAG packaging work more likely to be documented and productised than expanded.

◆Recent moves

  1. 1d ago

    OceanBase 5.0.1 adds parallel DDL and online storage-format conversion

    Framed in its own release notes as a continuation of V4.6.0 rather than a new direction, despite the major version number. The substance is engine maturity: parallel DDL for adding and dropping columns and for partition operations, deferred online conversion from columnar and hybrid layouts back to row storage which completes the conversion matrix between all three formats, random-distribution tables, and Zone-level and subpartition table groups.

    View source ↗
  2. 21d ago

    OceanBase 4.4.2 BP2 makes RPO=0 standby sync officially supported

    Strong synchronous replication to a standby graduates from experimental in BP1 to formally supported here, with maximum-protection and maximum-availability modes that differ in whether the primary halts writes or degrades to async when the standby link fails. Promotion of an existing feature rather than new capability, but it is the piece that lets the LTS line be used where zero data loss is a hard requirement.

    View source ↗
  3. 25d ago

    v4.6.0_CE

    ⚡ SPARK

    This is the release the rest of the window orbits: V5.0.1 describes itself as its continuation, and the vector-index defects filling the 4.4.1 and 4.3.5 hotfix branches are the cost of the search work landing here. It marks the point where AI retrieval stops being a bolted-on index type and becomes part of the query engine.

    View source ↗
  4. 25d ago

    OceanBase 4.4.2 BP1 ships RPO=0 standby sync as experimental

    The experimental first cut of the strong-sync replication that BP2 later declares supported, including automatic degradation to async when the primary and standby fall too far apart and automatic re-engagement once the link is healthy again. Zone-level table groups also land here, giving partition placement a Zone-scoped control.

    View source ↗
  5. 3mo ago

    OceanBase 4.3.5 BP6 fixes vector index crashes and leaks

    A defect-only patch on an older LTS branch, and almost every item is a vector-index failure: rebuild core dumps, HNSW errors in latency-first mode, IVF memory exhaustion, a double free in the index adapter, data loss during parallel index completion. It shows the search work being backported into branches that predate it.

    View source ↗
  6. 4mo ago

    OceanBase 4.4.1 HF4 fixes tokenizer memory growth and vector plans

    Another vector- and full-text-focused hotfix: IK tokenizer memory blowup on very large documents, planner misjudgements that skipped the vector index and cost performance, and inconsistent HNSW_BQ results against brute-force search. The same defect cluster appearing across three branches is the clearest sign of how new this code is.

    View source ↗