V4.6.2.2
Version information
- Release date: September 23, 2026
- Version: V4.6.2.2
- RPM version:
- oceanbase.ai-4.6.2.2-100000052026092215
- oblogservice-1.3.3-42026092110
- oceanbase-cdc-4.6.2.2-62026092216
Version overview
OceanBase AI Database V4.6.2.2 is a stability-enhanced version. Its core highlight is support for OBCDC synchronization with LOB-type data. The first OBCDC version compatible with OceanBase AI Database has also been released, enabling data to be synchronized downstream via OBCDC and thus establishing a link for data flow downstream. This version fixes multiple issues related to migration, FORK TABLE, GC, and hybrid search capabilities, and introduces two new FORK TABLE capabilities and exact KNN for non-vector indexes:
- Supports forking tables with a search index (
search_index). Thesearch_indexis retained with the table and is independently maintained between the source and target tables. - Supports the
FORK ... WITH READ LOCKclause: During the fork, the source table is readable but not writable. Writing can resume after the SQL returns. - Supports exact KNN for non-vector indexes: On heap tables (HEAP) containing vector columns, exact nearest neighbor searches can be performed using the
HYBRID_SEARCHclause without creating HNSW/IVF series vector indexes.
Key features
LOB-type data supported for OBCDC
In previous versions, OBCDC could not synchronize data changes containing LOB columns, making it difficult for multimodal large-object data within the database (such as document content, metadata, multimedia information, etc.) to flow downstream. Starting from this version, OBCDC supports synchronizing table changes containing LOB columns, ensuring the complete and consistent synchronization of data. The specific features are described as follows:
- In
FULLmode, unmodified LOB columns can be correctly synchronized, ensuring the integrity of downstream data. - Appending streaming-written LOB data is currently not supported; when such writes occur, OBCDC will proactively report an error to avoid inconsistent data. It is recommended to configure tables that use appending streaming writing for LOBs to the
tb_black_listexclusion list.
This is suitable for real-time synchronization of multimodal, semi-structured data from OceanBase AI Database to downstream data warehouses, search engines, or subscription systems via OBCDC.
FORK TABLE enhancements
FORK TABLE has been enhanced with two capabilities to ensure fork results are correct while retaining search capabilities:
- Supports forking tables with a search index (
search_index): Thesearch_indexand its index auxiliary table are completely retained during the fork, decoupling and independently maintaining the source and target tables. After the fork returns, the target table's index can be used and supports hybrid search. - Supports the
FORK ... WITH READ LOCKclause: During the fork, the source table is readable but not writable (uncommitted write transactions must wait, and new write requests are blocked). Writing to the source table can resume after the SQL returns. If the fork fails abnormally, the read lock is still released normally, ensuring the fork result is correct.
Both syntaxes can be used concurrently on the same source table without interfering with each other.
Exact KNN for non-vector indexes supported
Previously, the knn query in hybrid search relied heavily on HNSW/IVF series vector indexes; without these indexes, vector search could not be used. This version supports performing exact nearest neighbor searches directly through HYBRID_SEARCH on heap tables without creating any vector indexes. The specific features are described as follows:
- Supports
cosine,l2, andipdistance algorithms, as well as search methods such as scalar and full-text filtering, RRF, and weighted fusion. - The syntax is the same as when indexes are present, and the system automatically selects the access path.
This is suitable for scenarios such as small-set searches after strong filtering (e.g., thousands of rows filtered by user_id), maintenance-free small tables, requirement for exact results, and full-text + vector fusion. It can eliminate the overhead of index construction and maintenance while ensuring accurate results.
Upgrade notes
- Upgrading from OceanBase Database to OceanBase AI Database is not supported.
- Upgrading from OceanBase AI Database V4.6.0.1 to OceanBase AI Database V4.6.2 is not supported.
- Upgrading from OceanBase AI Database V4.6.2 or OceanBase AI Database V4.6.2.1 to OceanBase AI Database V4.6.2.2 is supported.
Recommended versions of components and tools
Component |
Version |
Description |
|---|---|---|
| ODP | ODP4.3.6.5 | |
| oblogservice | oblogservice 1.3.3 | |
| OCP | OCP 4.5.0 BP1 | |
| ODC | ODC 4.5.0 | |
| OBCDC | OBCDC 4.6.2.2 | |
| OBClient | OBClient 2.2.14 | |
| JDBC | JDBC 2.4.18 | |
| ODBC | ODBC 2.0.9.9 | |
| OceanBase Loader and Dumper | ob-loader-dumper 4.3.7 | |
| devdeps-java-extensions | 1.0.3 | |
| DataStudio | 0.4.0 | |
| OMS (private cloud) | 4.3.2 BP3 | Incremental data pull from OceanBase AI Database is not currently supported. |
| obshell (private cloud) | obshell 4.5.0 | |
| OBD | 4.6.0 |
V4.6.2.1
Version information
- Release date: September 10, 2026
- Version: V4.6.2.1
- RPM version:
- oceanbase.ai-4.6.2.1-100000092026090921
- oblogservice-1.3.2-92026090716
Version overview
OceanBase AI Database V4.6.2.1 enhances capabilities for AI search, model integration, and lakehouse read/write operations, enabling you to perform more AI and data analytics work directly within the database. The specific enhancements include:
- AI engine:
- AI search: Full-text support for multilingual analysis; semantic indexing now covers images and large-scale text post-build; hybrid search allows for finer control over recall tightness; vector search is more stable after large-scale recall and node migration.
- Model integration: One-step integration with mainstream models using vendor configurations, with gateway-based multi-endpoint routing and fusing to reduce the impact of rate limiting and single points of failure.
- Computing engine: Provides high-speed batch writes for the DataFrame ecosystem; path queries such as
JSON_EXTRACTcan be faster without modifying the SQL. - Storage engine: Asynchronous indexing reduces write pressure; multiple logical tables can be created in the same Namespace for associated queries; LOB support for append writes, with capacity increased to 4 TB.
It also supports querying Paimon data lake tables via SQL. The Log Service can add or remove zones as needed, and single-replica deployments can recover service after a zone failure.
Key features
AI engine
Full-text index supports multi-language text analyzers (Analyzers)
Provides more appropriate tokenization and analysis for texts in English, Thai, Vietnamese, Indonesian, Malay, and other languages. Built-in Analyzers include
standard,english,thai,vietnamese,indonesian, andmalay. ThestandardAnalyzer is suitable for mixed-language text. You can also use a custom JSON analysis pipeline. For details, see Full-text search.More flexible
minimum_should_matchin hybrid searchControls the minimum number of matching optional conditions required for a match. In addition to integers, negative numbers, percentages, and combined formats (such as
"75%","3<90%", and"2<-25% 9<-3") are supported. This parameter takes effect inmatch,multi_match,query_string, andbool(should) scenarios, facilitating parameter tuning between achieving a wider recall and higher precision. For details, see HYBRID_SEARCH.Batch embedding supported for semantic indexes created later
When creating a semantic index for a large table, the model is called via HTTP one record at a time by default, which is time-consuming and costly. You can set
ai_service_tiertoBATCHto use the vendor's batch interface for bulk embedding. This capability applies only to full creation after an existing table exists and does not support creation during table creation. For details, see Semantic search and AI model interaction (Batch File).Semantic index supports image URLs
Supports storing image URLs in
VARCHARcolumns. When creating an index, specifycontent_type=image, and the multimodal embedding model automatically vectorizes it. During queries, you can still pass in the original URL or text, supporting image search by image or text. Suitable for scenarios such as product image search, content moderation, and multimodal RAG. For details, see Semantic search.AI functions are easier to integrate and support images
Uses vendor configurations (Provider), model profiles (Model Profile), and gateways (Gateway) to manage connection credentials, call parameters, and multi-endpoint routing separately. Built-in default configurations from major vendors are provided; generally, you only need to fill in the API Key to call. Also supports multimodality for images: use
AI_PROMPTto mix images with text,AI_COMPLETEfor image understanding, andAI_EMBEDto generate image vectors. For details, see AI model registration.Vector index background tasks are more controllable and easier to debug
Background tasks such as index rebuilding and incremental segment merging are uniformly scheduled, with priorities and concurrency levels allocated to reduce tenant resource preemption. You can view task types, status, and progress through system views, or manually trigger tasks using
DBMS_VECTOR.TRIGGER.Async_TASKto reduce inspection and troubleshooting costs. For details, see Vector index maintenance.Vector indexes are searchable immediately after migration
After node migration or replication is complete, HNSW vector indexes can provide queries without temporary loading, avoiding significant query latency jitter after migration.
Vector Search Iterator (large-scale similar recall)
HNSW approximate search returns up to 16,384 similar results at a time. For larger-scale recall, you can use the Search Iterator to fetch data in batches, breaking through the single-top-K limit. For details, see HNSW series indexes.
Model fault tolerance and circuit breaking
When AI functions or semantic indexes access external models, issues such as API Key throttling, brief model instance failures, or network timeouts often occur. Multiple endpoints (with different Keys, vendors, or model instances) can be placed under one gateway, with traffic routed among healthy endpoints based on weight. If an endpoint fails consecutively, it is automatically disconnected and switched to another available endpoint, reducing the impact of single-point failures on business operations. For details, see AI model registration.
Search indexes support nested queries (Nested Query)
By default, JSON object arrays are flattened by path, which loses field associations within the same object and may lead to false positives in hybrid searches. When creating a search index, declare nested paths using
NESTED_PATHSto require multiple conditions to match the same array element. For details, see Search index.
Computing engine
Faster JSON path query
Path queries such as
JSON_EXTRACTare the most commonly used operations on JSON. This version optimizes the JSON read path, benefiting the vast majority of JSON functions without requiring SQL changes. Internal testing shows that response times for high-frequency functions likeJSON_EXTRACThave been reduced by an average of 1/3 to 2/3.Arrow Flight SQL for high-speed writing
When writing data in batches from clients such as Spark Connectors or Python DataFrames, row-by-row SQL processing often cannot match their throughput. This version supports pushing data in batches using the columnar format over the Apache Arrow Flight SQL protocol. Writing still uses existing SQL transactions, index maintenance, and constraint checks. It supports append, upsert, update, and direct load operations, covering types like
VECTORandARRAY, making it suitable for large-scale imports and high-speed streaming writes.
Storage engine
Asynchronous indexes (Experimental feature)
Synchronous index writes are immediately visible and provide strong query consistency, but they require synchronous index maintenance during write, which is resource-intensive. This version supports adding the
ASYNCkeyword when creating full-text/search indexes: tables are written to as usual, while indexes are periodically built in the background and are typically visible within seconds; queries return eventual consistency, allowing hybrid searches with synchronous indexes. You can view refresh tasks through views and manually trigger major compactions usingDBMS_ASYNC_INDEX.FORCEMERGE. Choose asynchronous indexes when write throughput is prioritized and brief visibility delays are acceptable; use synchronous indexes when consistency is prioritized. For details, see Full-text search and Search indexes.Multiple logical tables in a single Namespace
Multiple business tables can share underlying physical tables, avoiding the need to create numerous physical tables for each application. Previously, a session could access only one logical table; this version allows creating multiple logical tables under the same Namespace and supports JOINs and regular DML within the same Namespace. Combined with the collection/app isolation of json-table-sdk, it is suitable for multi-table association scenarios like flash applications. Cross-Namespace JOINs are not currently supported.
Macroblock cache integrated with KVCache
In shared storage, data was stored as macroblocks in object storage, and query performance relied on local caches. This version unifies the management of macroblock memory caches under KVCache and adds asynchronous persistence to disk: hot macroblocks are written to local disk cache in the background. The read order is memory cache → local disk cache → shared storage; if the cache misses or a write fails, data is fetched from the origin, without affecting data correctness. Usage remains unchanged, which can reduce repeated accesses to object storage and make reads faster.
LOB support for direct load and offline DDL
Tables containing large objects (LOBs) can be rebuilt using direct load and offline DDL. Batch import and table rebuilding no longer need to bypass LOB columns. Currently, full direct load and incremental baseline are supported; incremental direct load is not supported. The lifecycle, commit, and release of old and new data remain atomically consistent. For details, see LOB types.
LOB support for append-only writing
LONGTEXTandLONGBLOBare no longer limited to a 512 MB cap, with capacity increased to 4 TB − 1, while significantly reducing memory usage during writes (for example, appending an 11 GB LOB in 8 MB chunks requires only about 16 MB of memory per request). You can useLOB_WRITEAPPEND(lob, amount, buffer)to append data in batches within a single rowUPDATE, eliminating the need to rewrite the entire segment; after writing, you can useSUBSTR()for streaming reads. Rollback and continuous appending are supported within a transaction. Suitable for streaming write scenarios involving extremely large text and multimodal objects. For details, see LOB types and LOB_WRITEAPPEND function.Dynamic partitioning support for string time columns
When time is stored as a string, dynamic partitioning can also be used as the
LIST COLUMNSpartitioning key. The partition value format is automatically inferred byTIME_UNIT, and the strategy is consistent withRANGE COLUMNS. For details, see Create a dynamic partitioned table.
Log Service
Log Service supports dynamic addition and removal of zones
When disaster recovery requirements, deployment environments, or log write performance change, you can add or remove zones for the Log Service, flexibly choosing single-IDC or multi-IDC deployment. Topological elasticity adjustments can be achieved without rebuilding the Log Service.
Data Lake
Support for querying Paimon tables
You can directly query Apache Paimon format data lake tables using SQL. Access is provided through HMS Catalog, supporting Parquet and ORC formats, with automatic optimization such as predicate pushdown and partition pruning. It supports parallel reading and thread pool management for accelerating data lake queries. For details, see Read Paimon tables via plugins.
Disaster Recovery
Single-replica cross-zone disaster recovery
In single-replica deployment, if the entire zone fails, all tenant services can be restored in the standby zone. The recovery process starts from shared storage, requiring no movement of business data, making it suitable for retaining cross-zone recovery capability while controlling costs. For details, see Disaster recovery operations in single-replica scenarios.
Upgrade notes
- OceanBase Database and OceanBase AI Database are two independent products. They do not support direct upgrade in either direction; only logical migration is supported.
- Upgrading from OceanBase AI Database V4.6.2 to V4.6.2.1 is supported.
Recommended versions of components and tools
Component |
Version |
Note |
|---|---|---|
| ODP | ODP4.3.6.5 | |
| oblogservice | oblogservice 1.3.2 | |
| OCP | OCP 4.5.0 BP1 | |
| ODC | ODC 4.5.0 | |
| OBClient | OBClient 2.2.14 | |
| JDBC | JDBC 2.4.18 | |
| ODBC | ODBC 2.0.9.9 | |
| OceanBase Loader and Dumper | ob-loader-dumper 4.3.7 | |
| devdeps-java-extensions | 1.0.3 | |
| OMS (private cloud) | 4.3.2 BP3 | Incremental data pull from OceanBase AI Database is not currently supported. |
| obshell (private cloud) | obshell 4.5.0 | |
| OBD | 4.6.0 |
V4.6.2
Version information
- Release date: August 10, 2026
- Version: V4.6.2
- RPM version:
- oceanbase.ai-4.6.2.0-100000162026080717
- oblogservice-1.3.1-112026080715
Version overview
OceanBase AI Database is a new-generation, multimodel, cold-hot layered, storage-compute separated database. It is tailored for AI data scenarios on top of OceanBase's large-scale production-proven distributed foundation. Evolving from OceanBase V4.6.0, it supports only the storage-compute separated deployment model and operates in MySQL-compatible mode. This version supports online upgrades to later releases. Key features include:
- For agent evaluation scenarios, it supports Fork table/database (experimental feature).
- Supports Catalog Sync: actively synchronizes DDL operations such as table and database creation from the internal catalog to an external HMS, facilitating data integration and reading by data platforms.
- Hybrid search capability: supports reranking refinement and parallel execution within partitions.
Key features
Enhanced LOB capabilities
This version enhances LOB capabilities to better support multimodel scenarios and facilitate compatibility upgrades in future versions. Note that tables with LOBs do not support CDC, direct load, or offline DDL. For details, see LOB types.
Fork Table/Database (Experimental feature)
Adds the Fork Table/Database feature, which allows you to quickly snapshot user tables or databases and create derived copies of these objects. Initially, the copies share underlying data, but read and write operations become isolated afterward. This feature is suitable for agent evaluation scenarios: you can quickly fork an evaluation dataset from a production dataset based on production data, achieving data isolation and reproducibility of evaluations, and avoiding interference between multiple evaluation tasks. For details, see Overview of Fork.
Enhanced logical table capabilities (Experimental feature)
Monitoring, O&M, and throttling
Supports monitoring and resource management at the namespace level. For monitoring, you can view statistics such as row count and storage usage, as well as background collection task information. For resource management, it supports throttling QPS, TPS, and data bandwidth, and allows setting upper limits for row count, capacity, and concurrent thread slots. O&M personnel can configure this via
DBMS_LOGIC_TABLE.SET_NAMESPACE_RESOURCE_LIMIT.JSON Search performance optimization
Optimized the processing path for JSON Search in logical tables during compilation and execution phases, improving query performance.
Logical table warming
Query access refreshes the heat of the corresponding namespace; the system can pull data from hot namespaces into local cache and evict cold namespaces' local cache to free up space.
For specific details on logical table capabilities in the current version, see Agent hybrid search scenario.
Catalog Sync
OceanBase supports connecting to external HMS / Iceberg Catalogs for metadata reading. Building upon this, OceanBase AI Database further provides Catalog Sync: actively synchronizes DDL operations such as table and database creation from the internal catalog to an external HMS, enabling tables written to OceanBase AI Database to be discovered and queried by third-party engines like Spark. A typical use case is: after creating the relevant database and table in OceanBase AI Database, synchronize the metadata to Hive Metastore, facilitating unified reading of OceanBase AI Database tables from Hive Metastore and Iceberg tables from HMS by data platforms, forming a data integration and reading solution. Currently, it supports unidirectional synchronization from internal to external HMS. The parameter catalog_sync_target controls whether to synchronize to the target Catalog, and this applies only to DDL operations occurring after it is enabled. For details, see Catalog data synchronization.
Catalog external table sampling
Supports sampling statistics for catalog external tables (such as external tables from data sources like Hive or Iceberg). You can use the DBMS_STATS.GATHER_CATALOG_TABLE_STATS procedure to collect statistics for external tables, enabling the optimizer to generate better execution plans and improve performance in cross-source federated queries. For more information, see Statistics.
Medium-scale parallelism optimization
Optimized parallel execution efficiency in medium-scale parallelism (DOP approximately 2000–6000) scenarios, improving AP query throughput by about 15%–30%.
Expression enhancements
Added and fixed multiple built-in functions, including:
- New expressions:
NCHR,NEW_TIME,INSTR,ROUND,randCanonical,parseDateTime - New aggregate functions:
avgWeighted,string_agg - Optimized expression performance:
json_extract,json.seek,json_length,json_type,ifnull,space,extract,date_format
SQL execution optimization
Multiple operator and execution path optimizations:
- HashDistinct dynamically selects between hash or sort deduplication strategies based on data volume.
- Partition Sort reduces sorting costs by leveraging partition order.
- Added vectorized paths for operators such as Window Functions and CASE WHEN.
- Enhanced aggregation pushdown, covering more scenarios where it can be pushed down to the storage layer for execution.
- Enhanced column pruning, capable of precisely pruning unused columns in complex view nesting scenarios.
- Optimized operator memory control and spilling to disk, reducing the risk of sudden memory spikes.
ARM AP performance optimization
Enhanced vectorization and memory access optimization for ARM platforms. AP core operator performance on ARM platforms is improved by about 20%–40%, with overall TPC-H performance approaching x86 levels.
Hybrid search supports rerank refinement
Hybrid search supports merging multiple recall results from the rough ranking phase through RRF, MinMax normalization, or weighted fusion. OceanBase AI Database provides a rerank parameter in the hybrid search syntax, allowing the system to proceed to the refinement phase after rough ranking fusion: it calls an external AI reranking model to perform a semantic secondary sorting of the top candidate documents, further enhancing the relevance of the top results. For more information, see Hybrid search.
Parallel execution within hybrid search partitions
Hybrid search is executed serially by default; when data volumes are large, a single query can be slow. This version supports parallel execution within partitions to reduce query response time (RT), using two methods:
- Multiple-path parallelism: Multiple recall paths (multiple
knnpaths + full-textquerypath) within the same query statement are executed in parallel by different worker threads, with the overall RT approaching that of the longest path. - Path-level parallelism: A single
querysearch path is executed in parallel, including recall paths such as full-text, scalar, and JSON type filtering. After splitting by data range, multiple worker threads scan different intervals in parallel, further reducing the RT of a single path.
Both methods can be used simultaneously. This capability is disabled by default and requires setting the tenant-level parameter _enable_hybrid_search_parallel_execution = true to enable it. Enabling it may increase CPU/memory usage; it is recommended to evaluate in a resource-adequate environment before enabling. For more information, see Hybrid search.
Hybrid search syntax extension
This version extends the hybrid search SQL syntax to meet requirements such as custom result set size, fuzzy matching, bucket aggregation, field folding, and custom sorting:
- Result set size (
num_candidates): Controls the scale of the KNN candidate set (affects recall and latency; the final TopK is still determined byk, etc.). - Wildcard filtering (
wildcard): The current version only supports filtering based on JSON paths and regular string columns. - Aggregation (
aggs.terms/aggs.cardinality): Supports terms grouping statistics and cardinality estimation. - Collapse (
collapse): Deduplicates search results by the specified field, returning only the best-matching document per group. - Custom sorting (
sort): Supports sorting by any table field and provides control over the position of NULL values. - Boost weight (
boost): Scalar filter conditions can participate in scoring during queries, and score weights can be specified using the ES-compatibleboostparameter. term/terms: Added support for querying array fields.
For details, see Hybrid search.
Changes in product behavior
Object storage access protocol changes
Features |
Note |
|---|---|
| Object Storage Access Protocol | Access to object storage via the OSS protocol is not supported. The S3 protocol must be used instead. |
Upgrade notes
- Upgrading from OceanBase Database to OceanBase AI Database is not supported.
- This version supports online upgrades to subsequent release versions.
Recommended versions of components and tools
Component |
Version |
|---|---|
| ODP | ODP 4.3.6.5 |
| oblogservice | oblogservice 1.3.1 |
| OCP | OCP 4.5.0 |
| ODC | ODC 4.4.3 BP2 |
| OBClient | OBClient 2.2.13 |
| JDBC | JDBC 2.4.18 |
| ODBC | ODBC 2.0.9.9 |
| OceanBase Loader and Dumper | ob-loader-dumper 4.3.6 |
| devdeps-java-extensions | 1.0.3 |
| OMS | 4.3.2 BP3 |
| OBD | 4.5.0 |
Considerations
- Tables with LOBs do not support direct load / offline DDL.
- The default configuration template disables automatic load balancing; if needed, it can be manually enabled.
- Physical restore is recommended to be performed on an empty cluster.