V4.6.3 Hotfix2
Version information
- Release date: November 13, 2026
- Version: V4.6.3 Hotfix2
- RPM version:
- oceanbase.ai-4.6.3.0-100020012026111215
- oblogservice-1.3.3-52026101220
Version overview
This version primarily fixes issues related to Flight SQL write and session cleanup, correctness of full-text and hybrid search results, asynchronous index primary/standby switchover, and direct load rollback, enhancing the stability of queries, data import, and connection processing.
Bug fixes
- Flight SQL: Fixed an exception when writing a DECIMAL type parameter, and fixed the issue where the cleanup process would block when a request was canceled or a session was closed.
- Asynchronous index: Fixed the issue where incomplete recovery of index data after a primary/standby switchover caused query or major compaction exceptions.
- Full-text and hybrid search: Fixed the issue where the IK tokenizer omitted tokens when processing text in batches; fixed the issue where partition pruning caused query rows to be missed, aggregation on __score to crash, and abnormal column name case recognition when optional should conditions were present.
- Search index: Fixed the issue where string array comparison results were incorrect, and fixed the issue where a query reported an error when the candidate array was empty.
- Direct load: Fixed the issue where tasks waited for each other during the rollback process of incremental direct load, preventing rollback and DDL cleanup from completing.
- Metadata and connection stability: Fixed the issue where metadata for deleted log streams (LS) was not reclaimed, and fixed the issue where an OBServer crashed due to an excessively long client version string.
V4.6.3 Hotfix1
Version information
- Release date: November 06, 2026
- Version: V4.6.3 Hotfix1
- RPM version:
- oceanbase.ai-4.6.3.0-100010012026110611
- oblogservice-1.3.3-52026101220
Version overview
This version primarily improves query correctness and runtime stability for asynchronous indexes, vector search, and hybrid search. It fixes issues in scenarios such as index construction, refresh, major compaction, and primary/standby switchover, reducing the overhead of multi-partition refresh. It also enhances search index, vector index, and Flight SQL capabilities, and fixes problems related to LOB data reading, OBCDC synchronization, and shared storage.
Feature enhancements
- Search indexes support stored generated columns (
STORED): Search indexes can now be created on stored generated columns (STORED) usingCREATE TABLE,CREATE SEARCH INDEX, andALTER TABLE ADD SEARCH INDEX. Virtual generated columns (VIRTUAL) are not supported. For more information, see Search indexes. - Supports grouping and managing identical vectors: The HNSW series of indexes now supports the
duplicate_groupparameter, which reduces the impact of duplicate vectors on the graph structure and improves search recall in scenarios with many identical vectors. You can enable or disable this capability for existing indexes by rebuilding them in full; it will then no longer group only similar vectors. For more information, see HNSW series indexes. - Enhanced Flight SQL capability: Improves the Flight SQL feature and adds support for resource groups.
Bug fixes
- Asynchronous index construction and refresh: Fixed the issue where indexes were used for queries before initial construction was complete. Fixed issues with abnormal refresh configurations or statuses after index creation, table renaming, table replication, and offline DDL. Fixed the issue where indexes did not continue refreshing during long transactions, and manual refreshes reported errors.
- Asynchronous index primary/standby switchover and data migration: Fixed issues with abnormal index status reading, refresh retries, and query anomalies during primary/standby switchover and data migration, preventing merge tasks from continuing due to incorrect status judgments.
- Asynchronous index major compaction and recovery: Fixed issues with abnormal progress handling during major compaction retries and recovery, old major compaction results overwriting new progress, and continuous query errors after some major compactions, improving the stability of major compaction and fault recovery.
- Asynchronous index refresh performance and memory usage: Supports concurrent refresh by partition, reuses completed progress during retries to reduce duplicate processing, optimizes refresh overhead when no new data is added, and fixes abnormal memory usage and untimely release of temporary memory during refresh and retry processes.
- Hybrid search result accuracy: Fixed issues where excluded documents were incorrectly returned under nested
shouldandmust_notconditions, and some zero-score hit documents were missed. Fixed cross-data segment filtering that missed rows, nestedboolquery errors, and abnormal end-of-search status judgments. Fixed issues related to BM25 scoring statistics andtermsquery value conversion. - Hybrid search execution stability: Fixed memory allocation errors when searching for long strings in partitioned table JSON, and crash issues in specific
LIKEscalar filtering scenarios. Fixed parsing of thenum_candidatesparameter and execution issues for some batch queries inknnsearches. - Index update and DDL: Fixed abnormal index data updates during batch construction retries,
UPDATE, andUPDATE IGNOREscenarios, and index creation failures involving hidden columns. Fixed DDL taking a long time to complete in some failure scenarios, and reported memory allocation errors after creating indexes or performing direct load. - Vector index background tasks: Fixed the issue where background tasks repeatedly created databases, tables, or tenants after they entered the recycle bin, and incomplete resource release after task cancellation. Reduced unnecessary automatic rebuilding of empty HNSW tables while retaining user-initiated manual rebuilding capability.
- LOB reading and OBCDC synchronization: Fixed the issue where reading both historical and current versions of LOB data for the same SQL might return incorrect content, and some LOB write failures. Fixed the issue where OBCDC reported an error and stopped synchronization when performing offline DDL rebuilding for a vector index containing LOB data.
- Shared storage and recovery: Fixed metadata deletion, historical record cleanup, and abnormal status judgment issues during shared storage garbage collection. Fixed abnormal memory access after Tablet creation failed, and improper handling of checkpoint read errors during recovery.
- Log and SQL session stability: Supports compressing
libpalf.loglogs. Fixed log module crashes, excessive repeated log output in some asynchronous refresh scenarios, and continuous alert output due to abnormal remote SQL session processing.
V4.6.3
Version information
- Release date: October 16, 2026
- Version: V4.6.3
- RPM version:
- oceanbase.ai-4.6.3.0-100000032026101510
- oblogservice-1.3.3-52026101220
Version overview
OceanBase AI Database V4.6.3 is a feature-enhanced version for AI search and lakehouse data management. This version upgrades core capabilities in three main areas: vector and full-text search, storage engine, and lakehouse integration. It focuses on improving the performance, recall, and observability of vector and full-text search, enhancing access to hot and cold data, data branching, and Catalog synchronization capabilities, and improving the runtime isolation of Java/Python UDFs and the stability of CDC data synchronization. The specific enhancements are as follows:
- AI engine: Optimized the search performance of asynchronous search indexes and asynchronous full-text indexes to cover continuous incremental write scenarios; HNSW series and IVF series indexes now support L1 (Manhattan distance), enriching the selection of vector distance algorithms; added a single-vector search profile, which allows viewing execution details during filtering, index search, and incremental data processing stages, aiding in query time troubleshooting; supports online modification of HNSW series index parameters. Parameter changes take effect after rebuilding the backend partition, eliminating the need to delete and recreate the index first. In logical table scenarios, further improved the query performance stability, availability, and read/write consistency of IVF indexes after primary/standby switchover.
- Storage engine: Supports pruning the query range by partition hot/cold status using the
PARTITION TIERclause to reduce unnecessary data access. FORK TABLE/FORK DATABASE is extended to columnstore tables and optimizes fork performance for databases containing a large number of tables. Testing shows that a database with 1,000 tables can complete a single fork within 10 seconds, significantly shortening the preparation time for evaluation and test datasets. Also improved the stability of CDC. - Computing engine: Catalog synchronization now supports more DDL operations, partitioned tables, and manual synchronization, making the lakehouse integrated data management pipeline smoother; Java/Python UDFs support running in independent child processes, reducing the impact of external language exceptions on the OBServer main process.
Key features
AI engine
Enhanced performance and stability of asynchronous indexes (Experimental feature)
Optimized the search performance of asynchronous search indexes and asynchronous full-text indexes, covering scenarios with continuous incremental data writes. Key optimizations include single-token full-text search, search with multiple tokens requiring only a small number of top-ranked results (low TopK), phrase matching (
match_phrase), and range filtering on fields with many distinct values (high NDV Range). This reduces the overhead of repeated scans and candidate result processing, while improving the stability of incremental writes, major compactions, and fault recovery. Asynchronous indexes still use eventual consistency, making them suitable for businesses that can accept brief data visibility delays. For details, see Search indexes and Full-text search.Full direct load is supported for tables with search indexes or full-text indexes
When writing data to indexed tables in batches, you can improve import efficiency through full direct load. Heap tables with synchronous or asynchronous search indexes support full direct load via
INSERT INTO ... SELECTorLOAD DATA, but do not support incremental direct load. Tables with full-text indexes support full direct load viaINSERT INTO ... SELECTonly inFULLmode. For specific limitations, see Search indexes and Full-text search.View the profile of a single vector search
When a vector search SQL statement takes a long time or experiences occasional significant latency, you can use the
MONITORhint to collect execution information. Then, useDBMS_XPROFILE.DISPLAY_PROFILE()to view the time consumption and execution details of each internal stage of the vector search in a hierarchical manner, including scalar filtering, vector index search, and incremental data processing, to help identify the main cost. This capability covers all dense and sparse vector indexes currently supported. A single profile reflects only one execution of a specified query and does not provide summary metrics such as average time or P95 over multiple executions, nor historical trends. For details, see Index monitoring and maintenance.Online modification of HNSW series index parameters is supported
You can use
DBMS_VECTOR.ALTER_INDEX_PARAMSto adjust the build parameters and query parameters of HNSW series indexes online, to adapt to changes in data scale, recall rate, query latency, or resource usage requirements, without first dropping and recreating the index. When build parameters change, the system asynchronously rebuilds the index by partition in the background. Completion of the call does not mean all data segments have applied the new parameters; modifying only query parameters does not trigger a rebuild. You can view the actual index parameters used by each data segment in theSEGMENT_PARAMScolumn of the HNSW Segment Information view. For details, see ALTER_INDEX_PARAMS.Enhanced IVF index capability in logical table scenarios
In scenarios where multiple logical tables share a physical table, optimized the vector search for filtered IVF series indexes, enabling the query optimizer to more accurately select execution paths based on the data distribution of the current logical table partition, reducing query performance fluctuations. After a primary/standby switchover, the system asynchronously restores the IVF cluster center cache (excluding IVF_PQ), reducing the impact of cache restoration on frontend queries. Simultaneously, improved the incremental write processing for IVF indexes (excluding IVF_PQ), enhancing read/write consistency.
L1 distance is supported for HNSW and IVF series indexes
Supports using L1 (Manhattan distance) for vector index search. L1 distance is the sum of the absolute differences in each dimension between two vectors, and can be used as an option besides L2, inner product, and cosine distances based on business needs. Specify
distance=l1when creating the index, and usel1_distance()during search. Only supports FP32 general dense vectors in HNSW and IVF series indexes; quantized indexes or sparse vectors are not supported. For details, see Vector functions.
Storage engine
Query data by partition hot/cold status
When the same table stores both hot and cold data, you can use the
SELECT ... PARTITION TIERclause to specifyHOT,COLD,NOT_COLD, orNOT_HOTto query only partitions that meet the hot/cold status criteria. For example, online business operations can query only hot data, while archiving and analysis can query only cold data, to control the query scope and reduce unnecessary data access and resource overhead. If this clause is not specified, the query scope is not limited by hot/cold status. For details, see PARTITION TIER.Support for
ROWIDFor tables lacking a suitable primary key, you can first query
ROWID, then locate the data to be updated or deleted using this row identifier, without modifying the table structure. Supports queryingROWID, using table aliases to qualifyROWID, and locating rows byROWIDinUPDATEandDELETEoperations on a single-target table.ROWIDis used for row location from query to update or delete and should not be used as a long-term business identifier. For details, see ROWID pseudo-column.Reduced metadata memory overhead in large-scale table scenarios
Optimized the caching method for table structure metadata, loading user table metadata on demand and allowing infrequently accessed table metadata to be evicted from the cache, reducing memory overhead in large-scale table scenarios. Suitable for scenarios with a large number of tables where actual access is concentrated on a subset.
FORK TABLE/FORK DATABASE supports columnar storage tables
Data fork capability has been extended to columnar storage tables. Both row-based and columnar storage tables can create derived replicas via FORK TABLE / FORK DATABASE for data evaluation, testing, and analysis. Also optimized the performance of FORK DATABASE when containing a large number of tables, shortening the preparation time for database replicas. For details, see Overview of Fork.
Support for executing major compactions within a maintenance window
Supports configuring a maintenance window for tenant-level major compactions, facilitating scheduling major compactions during off-peak business hours and reducing resource contention. No new major compaction tasks are scheduled after the window ends, but ongoing tasks continue to complete. Major compactions not completed in the current round will continue with the original target in subsequent windows, and data that has already reached the target will not be processed repeatedly. Supports both automatic and manual triggering.
Computing engine
Enhanced catalog synchronization capability
Enhanced Catalog Sync to support synchronizing metadata such as databases and tables from OceanBase AI Database to Hive Metastore (HMS). This facilitates unified metadata management via Unified Catalog by Lakebase, and enables external computing engines to discover and access data within the database. This version supports more DDL operations and partitioned tables, and introduces manual synchronization capability to meet on-demand metadata synchronization requirements. For details, see Automatic catalog data synchronization and Manual catalog data synchronization.
Support for running Java/Python UDFs in independent child processes
Supports executing Java or Python UDFs in independent child processes (sandboxes), isolating them from the main OBServer process. This reduces the impact of external language runtime exceptions on database services. Added the
[G]V$OB_SANDBOX_PROCESSview to view child process status and resource usage.The Java UDF sandbox is disabled by default. It must be enabled under the system tenant using the cluster-level parameter
ob_enable_java_udf_sandboxand used with the corresponding version of the Java SDK. For configuration instructions, see ob_enable_java_udf_sandbox.External tables support skipping the UTF-8 file header BOM
Some tools generate UTF-8 files with a Byte Order Mark (BOM) at the file header. When reading from an external table, the BOM might be interpreted as content for the first column, causing data anomalies. This version supports explicitly setting
byteordermark=checkin theFORMATclause for CSV/text external tables. This detects and skips the BOM in the file header, improving compatibility with such files. If this attribute is not specified or set tonocheck, the BOM is not detected.
CDC
Enhanced data synchronization performance and stability
Optimized the processing efficiency of incremental data extraction and synchronization, reducing synchronization latency in scenarios with high-concurrency writes, cross-IDC extraction, and large transactions. Simultaneously reduced memory usage and CPU overhead of the CDC process, minimized latency fluctuations, and improved the stability of standalone and centralized CDC links.
Support for filtering change logs generated by TTL deletion
OBKV supports configuring whether to filter change logs generated by TTL deletion via
filter_ttl_delete_log, reducing the I/O pressure on LogService from such log synchronization. This feature is disabled by default. Once enabled, TTL-deletion changes are no longer output downstream via CDC, and the corresponding data downstream will not be automatically deleted through this link either, but this does not affect the normal execution of TTL tasks within the database. This feature requires use with liboblog V4.6.3 or later. For details, see OBKV-Table TTL.
Changes in product behavior
Parameter changes
Parameter |
Change type |
Effective level |
Description |
|---|---|---|---|
dtl_compress_func |
New | Tenant-level | Sets the compression algorithm used for data transmission across nodes during SQL parallel execution. Default value:lz4_1.0. Supportednone、lz4_1.0、snappy_1.0、zstd_1.3.8andzstd_1.5.7. The modification takes effect on newly initialized data transmission channels. This parameter does not take effect when message compression for parallel execution is not enabled. |
ob_enable_java_udf_sandbox |
New | Cluster-level | Controls whether to enable Java UDF sandbox mode. Default value:False. After it is enabled, Java UDFs run in independent child processes. |
shared_storage_data_check_thread_score |
New | Tenant-level | Sets the number of worker threads for the shared storage data existence inspection task, used to adjust the inspection concurrency resources. Value range:[0, 100]Default value:0indicates that the system default score is used, which does not mean inspection is disabled. |
sys_ckpt_thread_score |
New | Cluster-level | Sets the worker thread score for log stream checkpoint tasks in shared storage mode and controls the concurrency of checkpoint upload during the initial cluster creation. Value range:[0, 100]Default value:16; Set it to0Use the default score when16. |
View changes
View |
Change type |
Description |
|---|---|---|
DBA_JAVA_POLICY、USER_JAVA_POLICY |
New | Displays the Java permission policies for all users and the current user under the current tenant, including permission granting and restriction rules. |
DBA_OB_SHARED_MACRO_BLOCK_GC_STAT、CDB_OB_SHARED_MACRO_BLOCK_GC_STAT |
New | Displays the garbage collection statistics for the shared storage macroblock. |
GV$OB_TABLET_MACRO_CACHE_STAT、V$OB_TABLET_MACRO_CACHE_STAT |
New | Displays the number of macroblocks, data size, and cache ratio of tablets to facilitate checking the data caching status. |
GV$OB_SSTABLES、V$OB_SSTABLES、V$OB_SHARED_SSTABLES |
Add field | AddedMIN_MERGED_TRANS_VERSION、MAX_MERGED_TRANS_VERSIONColumn, which displays the major compaction transaction version information of the SSTable. |
DBA_OB_VECTOR_INDEX_TASKS、CDB_OB_VECTOR_INDEX_TASKS、DBA_OB_VECTOR_INDEX_TASK_HISTORY、CDB_OB_VECTOR_INDEX_TASK_HISTORY |
Extension Task Type | AddedALTER_INDEX_PARAMS、IVF_CACHE_REWARMRelated tasks, used to view background tasks for online modification of vector index parameters and IVF cache recovery. |
GV$OB_HNSW_INDEX_SEGMENT_INFO、V$OB_HNSW_INDEX_SEGMENT_INFO |
Add field | AddedSEGMENT_PARAMSColumn, showing the actual vector index parameters used by each data segment. |
Syntax changes
Features |
Description |
|---|---|
| Manually Synchronize Catalog | AddSYNC CATALOGThe statement supports synchronizing the metadata of all databases and tables, a specified database, or a specified table. Use it as follows:SYNC CATALOG *.*、SYNC CATALOG db_name.*andSYNC CATALOG db_name.table_name. |
| Query by Partition Hot/Cold Status | Added SELECT ... FROM table_name PARTITION [(partition_name [, partition_name] ...)] TIER ({HOT \ | COLD \ | NOT_COLD \ | NOT_HOT}) syntax, which allows you to filter partitions based on their hot or cold status. This clause applies to database tables but not to subquery aliases. TIER is a non-reserved keyword and can still be used as a regular table name, column name, or alias. |
ROWIDPseudo columns |
Queries SupportedROWID, referenced by table aliast.ROWID, and in theUPDATE、DELETEIn the statement, useROWIDLocates the row. |
INDEX_MERGE Hint |
Supported byCOLUMNS(...)Specify column, pass byINDEXES(...)Specify the index, or both.INDEXES(...)Can be used separately; no need to specify both at the same time.COLUMNS(...). |
System package changes
System package |
Description |
|---|---|
DBMS_JAVA |
Added the Java privilege management subroutine:GRANT_PERMISSION、RESTRICT_PERMISSION、REVOKE_PERMISSION、DISABLE_PERMISSION、ENABLE_PERMISSIONandDELETE_PERMISSION, which supports granting, restricting, revoking, disabling, enabling, and deleting privileges. Among these,GRANT_PERMISSIONandRESTRICT_PERMISSIONSupports returning the identifier of the permission record through output parameters. |
DBMS_VECTOR |
AddedALTER_INDEX_PARAMSSubprogram, which supports online modification of HNSW series index parameters and can use the optional parameteridx_parallel_buildSpecifies the degree of parallelism for building a physical index. |
Upgrade notes
- Direct upgrade from OceanBase Database to OceanBase AI Database is not supported.
- Upgrade from OceanBase AI Database V4.6.2.x to V4.6.3 is supported.
Recommended versions of components and tools
Component |
Version |
|---|---|
| ODP | 4.3.6.5 |
| oblogservice | 1.3.3 |
| OCP | 4.5.0 BP1 |
| ODC | 4.5.0 BP1 |
| OBCDC | 4.6.3 |
| Binlog | 4.3.6.9 |
| OBClient | 2.2.14 |
| JDBC | 2.4.18 |
| ODBC | 2.0.9.9 |
| OceanBase Loader and Dumper | ob-loader-dumper 4.3.7 |
| devdeps-java-extensions | 1.0.5 |
| DataStudio | 0.4.0 |
| OMS (private cloud) | 4.3.3 BP2 |
| obshell (private cloud) | 4.5.0 |
| OBD (Private Cloud) | 4.6.0 |
Considerations
- The default configuration template disables automatic load balancing. To enable it, you can do so manually.