OceanBase Database Enterprise Edition V4.5.0|OceanBase Database|docs|OceanBase Documentations

V4.5.0

Version information

Release date: November 26, 2025
Version: V4.5.0
RPM version: oceanbase-4.5.0.0-100000172025112420

Version overview

OceanBase Database V4.5.0 is an agile iteration version designed for AI scenarios, featuring significant upgrades to its core AI capabilities. This version comprehensively optimizes IVF_FLAT and IVF_PQ indexes from four dimensions: construction, query, DML operations, and usability. It also introduces support for memory-based sparse vector indexes, delivering superior sparse vector retrieval performance. The hybrid search capabilities have been significantly enhanced, supporting multi-vector hybrid searches and allowing for the weighted fusion and re-ranking of query result sets. Additionally, leveraging its built-in embedding capabilities, the new version allows users to create semantic indexes based on text columns, simplifying the process of using vector indexes.

Furthermore, the new version supports the INTERVAL partitioning feature in Oracle-compatible mode, reducing the complexity of partition management. It also improves the performance and stability of the Index Skip Scan feature, which is now enabled by default.

Key features

Kernel enhancements

Index Skip Scan optimization

The new version enhances the performance and stability of the Index Skip Scan feature. Even if the execution plan mistakenly uses Index Skip Scan, the performance will not degrade significantly. Additionally, the Index Skip Scan feature is now enabled by default instead of being disabled by default. After the optimizer collects statistics, it can automatically generate an Index Skip Scan plan if the rules and cost conditions are met.
Support for Index Merge with JSON multi-valued indexes

By integrating JSON multi-valued indexes into the Index Merge framework, the new version allows for the merging of scan results from multiple indexes before accessing the table. This significantly reduces the number of table accesses, thereby improving the performance of filtering JSON columns with JSON multi-valued indexes using the member of, json_overlaps, and json_contains expressions.

Shared storage

The new version adds shared-storage-related parameters, including shared directories, private directories, upload, minor, major, and attach statistics.

Oracle compatibility

INTERVAL partitioning in Oracle-compatible mode

The new version implements the INTERVAL partitioning feature in Oracle-compatible mode. When new data exceeds the range of existing partitions, new partitions are automatically created, reducing the complexity of partition management.

AI feature enhancements

IVF vector index performance optimization

The new version comprehensively upgrades IVF/IVF_PQ indexes from four dimensions: construction, query, DML, and usability. In terms of construction, it optimizes index building time by parallelizing K-means and adapting the domain ID merge framework. For queries, it introduces an iterative filtering algorithm that intelligently selects pre/post filter strategies based on different filtering rates, ensuring result completeness while improving query performance. In DML, it optimizes incremental write efficiency through caching mechanisms. In terms of usability, it adds features such as index rebuild parameter modification, construction progress monitoring, and statement-level nprobes parameter dynamic configuration.
In-memory sparse vector index (experimental feature)

The new version supports in-memory sparse vector indexes. When sufficient resources are available, in-memory sparse vector indexes outperform disk-based ones, better meeting business needs for sparse vector retrieval. Users can create in-memory sparse vector indexes by specifying lib=vsag, type=sindi when creating a sparse vector index. It also supports pruning rules to enhance query performance.
Enhanced automatic partition rebuilding

Automatic partition rebuilding for vector indexes reduces memory usage. However, in scenarios with a large number of partitions and large data volumes per partition, the single-threaded model used for automatic partition rebuilding is inefficient. To better utilize cluster resources, the new version supports multi-threaded parallel construction for asynchronous indexes, with the number of threads specified by the vector_index_optimization_concurrency parameter.
Enhanced hybrid search

The new version enhances the hybrid search feature, supporting hybrid searches with multiple vectors and scalar conditions. It also allows for weighted fusion and re-ranking of query results from multiple vectors.
Full-text index performance optimization

The new version further optimizes performance for single full-text index topK pushdown scenarios using adaptive BMM/BMW pruning algorithms.
AI_Prompt and similarity expressions

The new version introduces the AI_Prompt function, which upgrades static text prompts to reusable, parameterizable function templates. This supports prompt reuse and dynamic combination with data, enabling efficient prompt engineering. The AI_Complete expression allows calling text generation models to process data tasks using prompts, such as sentiment analysis and summarization. It also supports three expressions for calculating vector similarity: inner_product_similarity, cosine_similarity, and l2_similarity. A value closer to 1 indicates greater similarity between two vectors.
Semantic index (experimental feature)

Semantic indexes simplify the process of using vector indexes by leveraging built-in embedding capabilities. This makes vector concepts transparent to users: users can directly write raw data to be stored, which is automatically embedded and indexed within OceanBase Database. During retrieval, users can directly use the raw data they want to query, which is automatically embedded and then queried against the vector index.

Performance improvements

Multi Get performance optimization

The new version optimizes Multi Get performance through the following measures:
- Columnar storage scans support columnar projections.
- The filter clause utilizes skip index prefixes for optimization.
- keep order scans support blockscan.
- Optimizations such as real-time statistics collection at the storage layer.
  
  These optimizations significantly improve table access performance, especially in scenarios with large data volumes. They also notably enhance the performance of wide table projections involving more than 80 columns, and improve scan performance in vectorized nest loop join scenarios.
Insert Up performance optimization

Under the premise of good data locality, where batch primary keys are relatively continuous, the new version merges multiple range point queries into a single range query, reducing the number of ranges and further improving table access performance. Additionally, the new version restructures the main process of the Insert Up operator, optimizes the conflict detection mechanism, and effectively enhances overall performance.

Compatibility changes

View changes

The following changes are introduced:

View	Change type	Description
DBA_TABLES	New	In MySQL-compatible mode, the DBA_TABLES view is added. It displays ordinary tables, indexes, and external tables in the current tenant that are not in the recycle bin, and shows their attribute information, such as whether automatic partition splitting is supported.
DBA_OB_KV_TTL_TASKS	Add column	The `scan_index` column is added to record the indexes scanned by the TTL task.
DBA_OB_KV_TTL_TASK_HISTORY	Add column	The `scan_index` column is added to record the indexes scanned by the TTL task.
CDB_OB_KV_TTL_TASKS	Add column	The `scan_index` column is added to record the indexes scanned by the TTL task.
CDB_OB_KV_TTL_TASK_HISTORY	Add column	The `scan_index` column is added to record the indexes scanned by the TTL task.
DBA_OB_USERS	Add column	The `PLUGIN` column is added to record the name of the plugin used for password hash calculation.
CDB_OB_USERS	Add column	The `PLUGIN` column is added to record the name of the plugin used for password hash calculation.
[G]V$OB_SQL_AUDIT	Add column	The `TX_TABLE_READ_CNT` column and the `OUTROW_LOB_CNT` column are added to record the number of times the transaction status table is queried and the number of Outrow LOBs read during the query, respectively.
[G]V$OB_UNITS	Add column	The `REPLICA_TYPE` column is added to record the replica type.

System variable changes

System variable	Change type	Description
plsql_can_transform_sql_to_assign	Change default value	The default value is changed from False to True.

Parameter changes

The following changes are introduced:

Parameter	Change type	Description
default_compress_func	Change value range	The value range is expanded to include the `zstd_1.5.7` option.
kv_transport_compress_func	Change value range	The value range is expanded to include the `zstd_1.5.7` option.
log_transport_compress_func	Change value range	The value range is expanded to include the `zstd_1.5.7` option.
log_storage_compress_func	Change value range	The value range is expanded to include the `zstd_1.5.7` option.
default_micro_block_format_version	Change default value	The default value is changed from 1 to 2, indicating that the new Flat format is enabled by default.
default_skip_index_level	New	A tenant-level parameter that specifies the default value of `skip_index_level` when a tenant creates a table.
ddl_high_thread_score	New	A tenant-level parameter that controls the number of DAG threads for executing DDL KV merge operations. The default value is 0, indicating that the database-defined value is used.
ob_enable_utl_tcp	New	A cluster-level parameter that controls whether the UTL_TCP system package feature is enabled.
vector_index_optimization_concurrency	New	A tenant-level parameter that controls the number of concurrent threads for each vector index background optimization task.
model_request_timeout	New	A tenant-level parameter that controls the maximum HTTP timeout for each request to a large model.
model_max_retries	New	A tenant-level parameter that controls the number of retries after an error occurs when accessing a large model.

Syntax changes

Syntax	Change description
Added AI_PROMPT expression	The AI_PROMPT expression organizes the input prompt template and the data to be filled in as a JSON object. Syntax format: AI_PROMPT('<template>', [ , , ... ]); template indicates a string template. Placeholders are represented in the form of "{0}". For example, what is the commonality between data{0} and data{1}? expr0 to exprN indicate parameter expressions, which must be varchar or JSON expressions. If the expression is of the varchar type, it indicates a simple string replacement. If the expression is of the JSON type, it indicates a complex type, such as an image. Only varchar is supported in this version. JSON is reserved for multimodal scenarios. The return value is a JSON object that contains the template and args fields. For more information, see the usage example.
Added AI_COMPLETE expression	The AI_COMPLETE expression calls a text generation model to complete tasks such as sentiment analysis and summarization based on the prompt. The prompt parameter allows input of the JSON type to adapt to the PROMPT_OBJECT generated by the AI_PROMPT expression. In this version, only the pure string form in the args field is supported. Syntax format: AI_COMPLETE(model_key, prompt [, model_parameters]) model_key indicates a predefined model in the database. This information is managed by the system table and includes the model's URL and type. prompt indicates the user's input prompt. The type can be text or JSON. The length is variable. The expression internally determines whether the type is varchar or text. If the type is JSON, it further checks whether it is in the prompt_object format. model_parameters specifies optional configurations for the large model provided by the large model service, such as temperature and top_n. The return value is the text output by the large model. The maximum output length of existing models is 100,000 tokens, which is approximately 400,000 characters. Therefore, the return value type is set to longtext to cover all model usages.
Added syntax for creating semantic indexes, including creating an index with a table and creating an index after the table is created	CREATE TABLE t (..., col varchar(xx), VECTOR [index \| key] index_name (col) WITH (distance=[inner_product \| L2 \| cosine], lib=[vsag], type=[hnsw], model=[model_name], dim=[dimension], sync_mode=[immediate \| manual \| async], sync_interval=[time_interval],...) CREATE VECTOR INDEX index_name ON tbl(doc) WITH (distance=[inner_product \| L2 \| cosine], lib=[vsag], type=[hnsw], model=[model_name], dim=[dimension], sync_mode=[immediate \| manual \| async], sync_interval=[time_interval],...)
Use the semantic_distance expression to query a semantic index	order by semantic_distance(column_name, 'query_text') [approximate\|approx] limit n Use the query text query_text to perform a vector query on the semantic index of the column specified by column_name and obtain the top n most similar rows.
Use the semantic_vector_distance expression to query a hybrid vector index	order by semantic_vector_distance(column_name, 'query_vector') [approximate\|approx] limit n Use the query vector query_vector to perform a vector query on the semantic index of the column specified by column_name and obtain the top n most similar rows.
Use the semantic_vector_distance expression to query a hybrid vector index	order by semantic_vector_distance(column_name, 'query_vector') limit n Use the query vector query_vector to query the semantic index of the column specified by column_name and obtain the top n closest rows as exact results.

Recommended versions of components and tools

The following table lists the recommended versions of components and tools for OceanBase Database V4.5.0.

Component	Version	Description
ODP	V4.3.5 BP1 V4.3.1 (LTS)	We recommend that you use V4.3.1 (LTS). If you need to use new features that are not supported in V4.3.1 (LTS), use V4.3.5 BP1.
oblogservice	V1.1.0	-
OCP	V4.4.0	-
ODC	V4.4.0	-
OBCDC	V4.5.0	-
OMS	V4.3.0 BP3	-
Binlog	V4.3.4	-
OCCI	V1.0.4	-
OBCI	V2.1.1 BP1	-
ECOB	V1.2.1	-
obclient	V2.2.11	-
libobclient	V2.2.11.1	-
OBJDBC	V2.4.16	-
ODBC	V2.0.9.6	-
obloader	V4.3.4.1	-

Upgrade notes

Non-shared storage mode

Smooth upgrades are supported from V4.4.0/V4.4.0 BP1/V4.4.1 to V4.5.0.
For environments without vector indexes, upgrades are supported from V4.3.0/V4.3.1/V4.3.2/V4.3.3/V4.3.4/V4.3.5/V4.3.5 BP1/V4.3.5 BP2 to V4.5.0. However, this applies only to POC and test environments, and there is a certain level of upgrade risk. If issues occur during the upgrade, the environment may need to be rebuilt.
For environments that use vector indexes, contact OceanBase technical support for an evaluation before upgrading from V4.3.0/V4.3.1/V4.3.2/V4.3.3/V4.3.4/V4.3.5/V4.3.5 BP1/V4.3.5 BP2 to V4.5.0.
Upgrades from V4.2.x or earlier to V4.5.0 are not supported.

Shared storage mode

Upgrades are supported from V4.4.0 BP1/V4.4.1/V4.4.1 hotfix1~3 to V4.5.0. Please note that upgrading a tenant with a single-replica deployment will result in service downtime.
Upgrades from V4.4.1 hotfix4 and V4.4.1 hotfix5 to V4.5.0 are not recommended.
Upgrades from V4.3.5 BP2 or V4.4.0 to V4.5.0 are not supported.
Before upgrading from V4.4.0 BP1 to V4.5.0, upgrade the corresponding logservice cluster to V1.1.0.