OceanBase v4.4.1 CE: AI Functions, Vector Search, and Performance Gains

by

Overview

OceanBase Database Community Edition (CE) has released version V4.4.1_CE on October 24, 2025. The update focuses heavily on AI integration, vector search capabilities, and performance tuning for both OLTP and OLAP workloads. However, the release notes explicitly state that V4.4.1 only supports new cluster installations and does not support upgrading from earlier versions.

Upgrade and Compatibility

Users must perform a fresh installation for V4.4.1. This is a significant change from previous versions which typically supported in-place upgrades. The release also removes the third-party dependency on cos-c-sdk due to stability issues and the deprecation of append write operations by the underlying storage service.

Key Features and Enhancements

AI and Vector Search

V4.4.1 introduces a suite of AI capabilities directly within the SQL engine:

  • AI Function Service: Users can now call text generation, embedding, and reranking models via SQL expressions using new functions like AI_COMPLETE, AI_EMBED, and AI_RERANK. This is managed via the DBMS_AI_SERVICE package.
  • Hybrid Search: The database now supports combining vector-based semantic search with keyword search using full-text indexes, improving accuracy for complex queries.
  • Vector Index Optimizations: Significant improvements have been made to vector search performance on ARM architectures using NEON-based vectorization. The IVF index has been optimized with better caching and parallelized construction, and the HNSW index has seen performance gains in small-scale scenarios.

Kernel and Storage Enhancements

  • Locality Changes: The database now supports reducing locality from 2F (two replicas) to 1F (single replica) using ALTER TENANT, allowing for tenants in single-replica mode.
  • Flat Row Storage: A new storage format is introduced to improve read/write performance and space efficiency in sparse row scenarios, controlled by the micro_block_format_version attribute.
  • ODPS Storage API: Support for the ODPS Storage API has been added to external tables, offering better read performance than the legacy Tunnel API by avoiding session overhead.

Performance Optimizations

OceanBase reports substantial performance improvements across various workloads:

  • UDF Performance: UDF execution has been optimized through vectorized execution and result caching, with reported improvements of 15% to 26% in TPC-DAS test cases.
  • PS Cursors: Prepared statement cursors have been optimized to operate in a dual-threaded, non-streaming mode, yielding performance boosts of approximately 77% for fetching the first N rows and 90% for fetching all rows.
  • Startup Acceleration: In standalone mode, the total time for observer startup, bootstrap, and tenant creation has been optimized to 8 seconds.
  • Batch DML: True batch interfaces at the storage layer have been implemented, significantly improving performance for bulk insert, delete, and insert on duplicate key update operations.

Compatibility and Tools

  • UDF Support: Both Java and Python User-Defined Functions are now supported in MySQL-compatible mode, allowing for greater flexibility and integration with existing ecosystems.
  • External Tables: Enhanced support for JDBC-compatible data sources, Iceberg data lake formats, and HDFS paths with Kerberos authentication.
  • Tool Versions: The release recommends specific versions for supporting tools: ODP V4.3.5, OCP V4.3.6_BP1, OBD V4.0.0, OBCDC V4.4.1, and OMS V4.2.11.

Breaking Changes and Configuration

Vector Memory Configuration: The parameter ob_vector_memory_limit_percentage has changed behavior. Setting it to 0 now enables an adaptive mode (40% for tenants with <8GB memory, 50% for >8GB) instead of disabling the feature.

System Variables: The default value for PLSQL_OPTIMIZE_LEVEL has been changed from 2 to 0 to accelerate PL compilation.

Considerations

Users should note that full-text indexes are limited to partial indexes (global indexes are not supported), and vector indexes are also partial. Creating IVF indexes immediately after table creation is not recommended; it is better to create them after data import. Additionally, Iceberg V3 and Avro files are not supported.