milvus-2.5.4

v2.5.4

Release date: January 23, 2025

Milvus version	Python SDK version	Node.js SDK version	Java SDK version
2.5.4	2.5.4	2.5.4	2.5.4

We’re excited to announce the release of Milvus 2.5.4, which introduces key performance optimizations and new features such as PartitionKey isolation, Sparse Index with DAAT MaxScore, and enhanced locking mechanisms. A standout highlight of this release is its support for 10,000 collections and 1 million partitions, marking a major milestone for multi-tenant use cases. This version also addresses multiple bugs that improve overall stability and reliability, two of the critical bugs may cause data loss. We encourage you to upgrade or try out this latest release, and we look forward to your feedback in helping us continually refine Milvus!

Features

• Supports PartitionKey isolation to improve performance with multiple partition keys (#39245). For more information, refer to Use Partition Key.
• Sparse Index now supports DAAT MaxScore knowhere/#1015. For more information, refer to Sparse Vector.
• Adds support for is_null in expression (#38931)
• Root privileges can be customized (#39324)

Improvements

• Support 10K collections and 1million partitions in one cluster (#37630)
• Cached segments’ delta information to accelerate the Query Coordinator (#39349)
• Read metadata concurrently at the collection level to speed up failure recovery (#38900)
• Refined lock granularity in QueryNode (#39282), (#38907)
• Unified style by using CStatus to handle NewCollection CGO calls (#39303)
• Skipped generating the partition limiter if no partition is set (#38911)
• Added more RESTful API support (#38875) (#39425)
• Removed unnecessary Bloom Filters in QueryNode and DataNode to reduce memory usage (#38913)
• Speeded up data loading by accelerating task generation, scheduling, and execution in QueryCoord (#38905)
• Reduced locking in DataCoord to speed up load and insert operations (#38904)
• Added primary field names in SearchResult and QueryResults (#39222)
• Used both binlog size and index size as the disk quota throttling standard (#38844)
• Optimized memory usage for full-text search knowhere/#1011
• Added version control for scalar indexes (#39236)
• Improved the speed of fetching collection information from RootCoord by avoiding unnecessary copies (#38902)

Critial Bug fixs

• Fixed search failures for primary keys with indexes (#39390)
• Fixed potential data loss issue caused by restarting MixCoord and flushing concurrently (#39422)
• Fixed a delete failure triggered by improper concurrency between stats tasks and L0 compaction after MixCoord restarts (#39460)
• Fixed scalar inverted index incompatibility when upgrading from 2.4 to 2.5 (#39272)

Bug fixes

• Fixed slow query issues caused by coarse lock granularity during multi-column loading (#39255)
• Fixed an issue where using aliases could cause an iterator to traverse the wrong database (#39248)
• Fixed a resource group update failure when altering the database (#39356)
• Fixed a sporadic issue where the tantivy index could not delete index files during release (#39434)
• Fixed slow indexing caused by having too many threads (#39341)
• Fixed an issue preventing disk quota checks from being skipped during bulk import (#39319)
• Resolved freeze issues caused by too many message queue consumers by limiting concurrency (#38915)
• Fixed query timeouts caused by MixCoord restarts during large-scale compactions (#38926)
• Fixed channel imbalance issues caused by node downtime (#39200)
• Fixed an issue that could cause channel balance to become stuck. (#39160)
• Fixed an issue where RBAC custom group privilege level checks became ineffective (#39224)
• Fixed a failure to retrieve the number of rows in empty indexes (#39210)
• Fixed incorrect memory estimation for small segments (#38909)

milvus-2.4.21

v2.4.21

Release Date: January 22, 2025

Milvus version	Python SDK version	Java SDK version	Node.js SDK version
2.4.21	2.4.14	2.4.10	2.4.11

We’re excited to announce a new release of Milvus 2.4.21, featuring customizable root privileges, refined concurrency and locking mechanisms, and a host of bug fixes to enhance overall stability and reliability. We encourage you to upgrade or give it a try, and share your feedback to help us continually improve Milvus!

Features

• Root privileges can be customized (#39325)

Critical bug fixes

• Fixed potential data loss issue caused by restarting MixCoord and flushing concurrently (#39423)

Bug fixes

• Fixed an alias issue that could cause an iterator to traverse the wrong database (#39301)
• Resolved slow query problems caused by coarse lock granularity during multi-column loading (#39290)
• Fixed a resource group update failure when altering the database (#39357)
• Allowed skipping disk quota checks during bulk import (#39320)
• Supports Tantivy indexes built from multiple segments (#39127)
• Fixed incorrect memory estimation for small segments (#38910)
• Fixed ineffective RBAC custom group privilege checks (#39194)
• Resolved a server crash caused by passing an empty parameter to RestoreRBAC API (#39142)
• Fixed an error reporting issue when automatic refresh config operations failed (#39073)
• Fixed unrecognized expressions that used uppercase AND or OR (#38929)
• Resolved a partition drop failure that occurred after a load failure (#38871)

Improvements

• Unified style by using CStatus to handle NewCollection CGO calls (#39306)
• Unified style by using CStatus to handle NewCollection CGO calls (#39306)
• Refactored the delete MVCC function (#39258)
• Refined lock granularity in QueryNode (#38908)
• Added primary field names in SearchResult and QueryResults (#39223)
• Introduced a new Go package to manage proto (#39129)
• Cleaned up compaction tasks in compactionhandler (#38848)
• Accelerated bitset operations with SIMD (#39042)
• Optimized error handling for unloaded collection on QueryNode (#38920)
• Updated the default Etcd version to v3.5.16 (#38973)

milvus-2.5.3

v2.5.3

Release date: January 13, 2025

Milvus version	Python SDK version	Node.js SDK version	Java SDK version
2.5.3	2.5.3	2.5.3	2.5.4

Milvus 2.5.3 delivers critical bug fixes and performance enhancements to improve overall stability, reliability, and usability. This version refines concurrency handling, bolsters data indexing and retrieval, and updates several key components for a more robust user experience.

Bug fixes

• Fixed an issue where using an IN filter on a VARCHAR primary key could return empty results. (#39108)
• Fixed a concurrency problem between query and delete operations that could lead to incorrect results. (#39054)
• Fixed a failure caused by iterative filtering when an expr was empty in a query request. (#39034)
• Fixed an issue where a disk error during config updates led to the use of default config settings. (#39072)
• Fixed a potential loss of deleted data due to clustering compaction. (#39133)
• Fixed a broken text match query in growing data segments. (#39113)
• Fixed retrieval failures caused by the index not containing the original data for sparse vectors. (#39146)
• Fixed a possible column field race condition caused by concurrent querying and data loading. (#39152)
• Fixed bulk insert failures when nullable or default_value fields were not included in the data. (#39111)

Improvements

• Added a resource group API for the RESTful interface. (#39092)
• Optimized retrieve performance by leveraging bitset SIMD methods. (#39041)
• Used MVCC timestamp as the guarantee timestamp when specified. (#39019)
• Added missing delete metrics. (#38747)
• Updated Etcd to version v3.5.16. (#38969)
• Created a new Go package to manage protos.(#39128)

milvus-2.5.2

v2.5.2

Release date: January 3, 2025

Milvus version	Python SDK version	Node.js SDK version	Java SDK version
2.5.2	2.5.3	2.5.3	2.5.3

Milvus 2.5.2 supports modifying the maximum length for VARCHAR columns and resolves several critical issues related to concurrency, partition drops, and BM25 stats handling during import. We highly recommend upgrading to this version for improved stability and performance.

Improvements

• Generated disk usage logs only when the specified path does not exist. (#38822)
• Added a parameter for tuning the maximum VARCHAR length and restored the limit to 65,535. (#38883)
• Supported parameter type conversion for expressions. (#38782)

Bug fixes

• Fixed potential deadlocks in concurrency scenarios. (#38863)
• Generated the index_null_offset file only for fields that support null values. (#38834)
• Fixed the retrieve plan usage after free in the reduce phase. (#38841)
• Recognized expressions with capitalized AND and OR. (#38928)
• Allowed successful partition drops even if loading failed. (#38874)
• Fixed BM25 stats file registration issues during import. (#38881)

milvus-2.4.20

v2.4.20

Release Date: Janurary 2, 2025

Milvus version	Python SDK version	Java SDK version	Node.js SDK version
2.4.20	2.4.13	2.4.10	2.4.10

Milvus 2.4.20 addresses several critical issues, including fixing mmap for sparse index, resolving the failure to parse the correct database name when altering collection fields, and preventing deadlocks in multiple compaction tasks within the scheduler. Additionally, this version introduces an enhancement that allows adjustment of the system limit for maximum varchar length through a configuration item in the YAML file. We highly recommend upgrading to this version for improved stability and reliability.

Improvements

• Added param for tuning max varchar length (#38890)

Bug fixes

• Enabled mmap for sparse index (#38849)
• Fixed altercollectionfield interceptor dbname (#38664)
• Released compaction task lock when return function (#38857)
• Retrieve plan on heap was used after free when reduce (#38842)

milvus-2.5.1

v2.5.1

Release date: December 26, 2024

Milvus version	Python SDK version	Node.js SDK version	Java SDK version
2.5.1	2.5.2	2.5.2	2.5.2

Milvus 2.5.1 focuses on a series of bug fixes addressing memory loading, RBAC listings, query node balancing, and sealed segment indexing, while also improving the Web UI and interceptors. We highly recommend upgrading to 2.5.1 for enhanced stability and reliability.

Improvement

• Update web UI collection and query pages. (#38701)

Bug fixes

• Fixed OOM issues by adding a memory factor to loading estimations. (#38722)
• Fixed privilege group expansion when listing policies in RootCoord. (#38760)
• Fixed issues with listing privilege groups and collections. (#38738)
• Fixed the balancer to avoid repeatedly overloading the same query node. (#38724)
• Fixed unexpected balance tasks triggered after QueryCoord restarts. (#38725)
• Fixed load config updates not applying to loading collections. (#38737)
• Fixed zero read count during data import. (#38695)
• Fixed Unicode decoding for JSON keys in expressions. (#38653)
• Fixed interceptor DB name for alterCollectionField in 2.5. (#38663)
• Fixed empty index parameters for sealed segments when using BM25 brute force search. (#38752)

milvus-2.4.19

v2.4.19

Release Date: December 26, 2024

Milvus version	Python SDK version	Java SDK version	Node.js SDK version
2.4.19	2.4.13	2.4.9	2.4.9

Milvus 2.4.19 focuses on fixing issues in RBAC, Balancer, and loading processes, while also introducing several performance enhancements in areas such as PartitionKey Deletion and hybrid searching. We highly recommend upgrading to this version for increased stability and reliability.

Improvements

• Expression templates are introduced to accelerate hybrid searches. (#38624)
• Additional metrics are provided for improved deletion monitoring. (#38746)
• L0 file generation is restricted to specific partition for partitionKey deletion to reduce amplification. (#38232)

Bug fixes

• Fixed OOM issues by adding a memory factor to loading estimations. (#38721)
• Fixed privilege group expansion when listing policies in RootCoord. (#38759)
• Fixed access log retention, skipping empty log rotation. (#38661)
• Fixed the balancer to avoid repeatedly overloading the same query node. (#38720)
• Fixed issues with listing privilege groups and collections. (#38698)
• Fixed load config updates not applying to loading collections. (#38596)
• Fixed unexpected balance tasks triggered after QueryCoord restarts. (#38714)
• Fixed zero read count during data import. (#38696)
• Fixed Unicode decoding for JSON keys in expressions. (#38652)

milvus-2.5.0

v2.5.0

Release date: December 23, 2024

Milvus version	Python SDK version	Node.js SDK version	Java SDK version
2.5.0	2.5.1	2.5.2	2.5.2

Milvus 2.5.0 brings significant advancements to enhance usability, scalability, and performance for users dealing with vector search and large-scale data management. With this release, Milvus integrates powerful new features like term-based search, clustering compaction for optimized queries, and versatile support for sparse and dense vector search methods. Enhancements in cluster management, indexing, and data handling introduce new levels of flexibility and ease of use, making Milvus an even more robust and user-friendly vector database.

Key Features

Full Text Search

Milvus 2.5 supports full text search implemented with Sparse-BM25! This feature is an important complement to Milvus's strong semantic search capabilities, especially in scenarios involving rare words or technical terms. In previous versions, Milvus supported sparse vectors to assist with keyword search scenarios. These sparse vectors were generated outside of Milvus by neural models like SPLADEv2/BGE-M3 or statistical models such as the BM25 algorithm.

Powered by Tantivy, Milvus 2.5 has built-in analyzers and sparse vector extraction, extending the API from only receiving vectors as input to directly accepting text. BM25 statistical information is updated in real time as data is inserted, enhancing usability and accuracy. Additionally, sparse vectors based on approximate nearest neighbor (ANN) algorithms offer more powerful performance than standard keyword search systems.

For details, refer to Analyzer Overview and Full Text Search.

Cluster Management WebUI (Beta)

To better support massive data and rich features, Milvus's sophisticated design includes various dependencies, numerous node roles, complex data structures, and more. These aspects can pose challenges for usage and maintenance.

Milvus 2.5 introduces a built-in Cluster Management WebUI, reducing system maintenance difficulty by visualizing Milvus's complex runtime environment information. This includes details of databases and collections, segments, channels, dependencies, node health status, task information, slow queries, and more.

Text Match

Milvus 2.5 leverages analyzers and indexing from Tantivy for text preprocessing and index building, supporting precise natural language matching of text data based on specific terms. This feature is primarily used for filtered search to satisfy specific conditions and can incorporate scalar filtering to refine query results, allowing similarity searches within vectors that meet scalar criteria.

For details, refer to Analyzer Overview and Text Match.

Bitmap Index

A new scalar data index has been added to the Milvus family. The BitMap index uses an array of bits, equal in length to the number of rows, to represent the existence of values and accelerate searches.

Bitmap indexes have traditionally been effective for low-cardinality fields, which have a modest number of distinct values—for example, a column containing gender information with only two possible values: male and female.

For details, refer to Bitmap Index.

Nullable & Default Value

Milvus now supports setting nullable properties and default values for scalar fields other than the primary key field. For scalar fields marked as nullable=True, users can omit the field when inserting data; the system will treat it as a null value or default value (if set) without throwing an error.

Default values and nullable properties provide greater flexibility to Milvus. Users can utilize this feature for fields with uncertain values when creating collections. It also simplifies data migration from other database systems to Milvus, allowing for handling datasets containing null values while preserving original default value settings.

For details, refer to Nullable & Default Value.

Faiss-based HNSW SQ/PQ/PRQ

Through close collaboration with the Faiss community, the HNSW algorithm in Faiss has seen significant improvements in both functionality and performance. For considerations of stability and maintainability, Milvus 2.5 has officially migrated its support for HNSW from hnswlib to Faiss.

Based on Faiss, Milvus 2.5 supports multiple quantization methods on HNSW to meet the needs of different scenarios: SQ (Scalar Quantizers), PQ (Product Quantizer), and PRQ (Product Residual Quantizer). SQ and PQ are more common; SQ provides good query performance and build speed, while PQ offers better recall at the same compression ratio. Many vector databases commonly use binary quantization, which is a simple form of SQ quantization.

PRQ is a fusion of PQ and AQ (Additive Quantizer). Compared to PQ, it requires longer build times to deliver better recall, especially at high compression rates, saying binary compression.

Clustering Compaction (Beta)

Milvus 2.5 introduces Clustering Compaction to accelerate searches and reduce costs in large collections. By specifying a scalar field as a clustering key, data is redistributed by range to optimize storage and retrieval. Acting like a global index, this feature enables Milvus to efficiently prune data during queries based on clustering metadata, enhancing search performance when scalar filters are applied.

For details, refer to Clustering Compaction.

Other Features

Streaming Node (Beta)

Milvus 2.5 introduces a new component called the streaming node, which provides Write-Ahead Logging (WAL) services. This enables Milvus to achieve consensus before and after reading and writing channels, unlocking new features, functionalities, and optimizations. This feature is disabled by default in Milvus 2.5 and will be officially available in version 3.0.

IPv6 Support

Milvus now supports IPv6, allowing for expanded network connectivity and compatibility.

CSV Bulk Import

In addition to JSON and Parquet formats, Milvus now supports direct bulk import of data in CSV format.

Expression Templates for Query Acceleration

Milvus now supports expression templates, improving expression parsing efficiency, particularly in scenarios with complex expressions.

GroupBy Enhancements

• Customizable Group Size: Added support for specifying the number of entries returned for each group.
• Hybrid GroupBy Search: Supports hybrid GroupBy search based on multiple vector columns.

Iterator Enhancements

• MVCC Support: Users can now use iterators without being affected by subsequent data changes like inserts and deletions, thanks to Multi-Version Concurrency Control (MVCC).
• Persistent Cursor: Milvus now supports a persistent cursor for QueryIterator, enabling users to resume iteration from the last position after a Milvus restart without needing to restart the entire iteration process.

Improvements

Deletion Optimization

Improved the speed and reduced memory usage for large-scale deletions by optimizing lock usage and memory management.

Dependencies Upgrade

Upgraded to ETCD 3.5.16 and Pulsar 3.0.7 LTS, fixing existing CVEs and enhancing security. Note: The upgrade to Pulsar 3.x is not compatible with previous 2.x versions.

For users who already have a working Milvus deployment, you need to upgrade the ETCD and Pulsar components before you can use the new features and functions. For details, refer to Upgrade Pulsar from 2.x to 3.x

Local Storage V2

Introduced a new local file format in Milvus 2.5, enhancing loading and query efficiency for scalar data, reducing memory overhead, and laying the groundwork for future optimizations.

Expression Parsing Optimization

Improved expression parsing by implementing caching for repeated expressions, upgrading ANTLR, and optimizing the performance of NOT IN clauses.

Improved DDL Concurrency Performance

Optimized the concurrency performance of Data Definition Language (DDL) operations.

RESTful API Feature Alignment

Aligned the functionalities of the RESTful API with other SDKs for consistency.

Security & Configuration Updates

Supported TLS to secure inter-node communication in more complex or enterprise environments. For details, refer to Security Configuration.

Compaction Performance Enhancements

Removed maximum segment limitations in mixed compaction and now prioritizes smaller segments first, improving efficiency and speeding up queries on large or fragmented datasets.

Score-Based Channel Balancing

Introduced a policy that dynamically balances loads across channels, enhancing resource utilization and overall stability in large-scale deployments.

milvus-2.4.18

v2.4.18

Release Date: December 20, 2024

Milvus version	Python SDK version	Java SDK version	Node.js SDK version
2.4.18	2.4.12	2.4.9	2.4.9

Milvus 2.4.18 introduces customizable privilege groups and an enhanced Grant/Revoke API, significantly streamlining permission management. This version also brings a suite of built-in privilege groups for common operational scenarios, as well as a host of performance and stability improvements—ranging from CPU optimization to faster collection loading and more efficient indexing. In addition, critical bug fixes ensure a more resilient system, addressing issues like crash scenarios and synchronization failures. We recommend you upgrade to 2.4.18 to take advantage of these key enhancements and improved overall reliability.

Features

RBAC Enhancements:

• Introduces customizable privilege groups, allowing users to create, drop, list, and dynamically manage privileges (add/remove) within their defined groups.
• Includes a suite of built-in privilege groups for common operational scenarios:
  • Cluster-Level: ClusterReadOnly, ClusterReadWrite, ClusterAdmin
  • Database-Level: DatabaseReadOnly, DatabaseReadWrite, DatabaseAdmin
  • Collection-Level: CollectionReadOnly, CollectionReadWrite, CollectionAdmin
• Provides a new version of the Grant/Revoke API, enabling the use of these new interfaces without the need to specify an ObjectType.

Improvements

• Allow hyphens in partition names (#38474)
• Optimize CPU usage for health check requests (#35595)
• Support templates for expressions in RESTful APIs (#38161)
• Remove the limit on the number of load tasks per round (#38497)
• alterindex & altercollection now support modifying properties (#38361 #38111 #38421)
• alterdatabase supports deleting properties (#38450)
• Add detailed replica counts for resource groups (#38315)
• Support score-based balancing for channel policies (#38378)
• Add metrics to count the number of non-zero values/tokens in sparse searches (#38328)
• Remove the RPC layer of the coordinator when running in standalone or mixed mode (#38207)
• Add mmap file usage metrics (#38211)
• Support database requests in RESTful API (#38188)
• Enable rate limiting for RESTful V1 (#38190)
• Add collection ID to search request count metrics (#38144)
• Refine clustering compaction logs (#38102)
• Accelerate the collection loading process (#37841)
• Improve compaction performance by removing ParamTable lookups (#37882)
• Support retrying searches when topk is reduced and results are insufficient (#37093)
• Update Knowhere version (#38277)
  • Optimize sparse index and get ~10% performance improvement

Bug fixes

• Fixed a crash caused by growing-groupby (#38553)
• Fixed an issue where the SyncSegments RPC would always fail (#38032)
• Fixed an issue where sync tasks remained running after the DataNode had stopped (#38441)
• Fixed inaccurate general counts (#38525)
• Escaped prefixes before conducting searches in inverted indexes (#38425)
• Fixed an issue where roles could be dropped even though grants still existed (#38369)
• Fixed empty import task results (#38317)
• Fixed a DataNode issue where progress could stall at the writer buffer memory check (#38287)
• Fixed an issue that prevented the permission grant on the manualcompact API (#38168)
• Fixed inaccurate partition count metrics (#38073)
• Accelerated flushing speed by optimizing lock usage (#37897)
• Handled errors gracefully when the compaction queue is full (#37990)
• Optimized loading speed by separating the pool for target observation and collection loading (#37735)
• Fixed a crash caused by retrieving varchar data from a memory-mapped growing segment (#37995)
• Fixed an issue where channels could be accidentally released after balancing (#37940)

milvus-2.5.0-beta

v2.5.0-beta

Release date: November 26, 2024

Milvus version	Python SDK version	Node.js SDK version	Java SDK version
2.5.0-beta	2.5.0	2.5.0	2.5.0

Milvus 2.5.0-beta brings significant advancements to enhance usability, scalability, and performance for users dealing with vector search and large-scale data management. With this release, Milvus integrates powerful new features like term-based search, clustering compaction for optimized queries, and versatile support for sparse and dense vector search methods. Enhancements in cluster management, indexing, and data handling introduce new levels of flexibility and ease of use, making Milvus an even more robust and user-friendly vector database.

Key Features

Full Text Search

Milvus 2.5 supports full text search implemented with Sparse-BM25! This feature is an important complement to Milvus's strong semantic search capabilities, especially in scenarios involving rare words or technical terms. In previous versions, Milvus supported sparse vectors to assist with keyword search scenarios. These sparse vectors were generated outside of Milvus by neural models like SPLADEv2/BGE-M3 or statistical models such as the BM25 algorithm.

Milvus 2.5 has built-in tokenization and sparse vector extraction, extending the API from only receiving vectors as input to directly accepting text. BM25 statistical information is updated in real time as data is inserted, enhancing usability and accuracy. Additionally, sparse vectors based on approximate nearest neighbor (ANN) algorithms offer more powerful performance than standard keyword search systems.

For details, refer to Full Text Search.

Cluster Management WebUI (Beta)

To better support massive data and rich features, Milvus's sophisticated design includes various dependencies, numerous node roles, complex data structures, and more. These aspects can pose challenges for usage and maintenance.

Milvus 2.5 introduces a built-in Cluster Management WebUI, reducing system maintenance difficulty by visualizing Milvus's complex runtime environment information. This includes details of databases and collections, segments, channels, dependencies, node health status, task information, slow queries, and more.

Text Match

Milvus 2.5 leverages analyzers and indexing from Tantivy for text preprocessing and index building, supporting precise natural language matching of text data based on specific terms. This feature is primarily used for filtered search to satisfy specific conditions and can incorporate scalar filtering to refine query results, allowing similarity searches within vectors that meet scalar criteria.

For details, refer to Keyword Match.

Bitmap Index

A new scalar data index has been added to the Milvus family. The BitMap index uses an array of bits, equal in length to the number of rows, to represent the existence of values and accelerate searches.

Bitmap indexes have traditionally been effective for low-cardinality fields, which have a modest number of distinct values—for example, a column containing gender information with only two possible values: male and female.

For details, refer to Bitmap Index.

Nullable & Default Value

Milvus now supports setting nullable properties and default values for scalar fields other than the primary key field. For scalar fields marked as nullable=True, users can omit the field when inserting data; the system will treat it as a null value or default value (if set) without throwing an error.

Default values and nullable properties provide greater flexibility to Milvus. Users can utilize this feature for fields with uncertain values when creating collections. It also simplifies data migration from other database systems to Milvus, allowing for handling datasets containing null values while preserving original default value settings.

For details, refer to Default Value & Null.

Faiss-based HNSW SQ/PQ/PRQ

Through close collaboration with the Faiss community, the HNSW algorithm in Faiss has seen significant improvements in both functionality and performance. For considerations of stability and maintainability, Milvus 2.5 has officially migrated its support for HNSW from hnswlib to Faiss.

Based on Faiss, Milvus 2.5 supports multiple quantization methods on HNSW to meet the needs of different scenarios: SQ (Scalar Quantizers), PQ (Product Quantizer), and PRQ (Product Residual Quantizer). SQ and PQ are more common; SQ provides good query performance and build speed, while PQ offers better recall at the same compression ratio. Many vector databases commonly use binary quantization, which is a simple form of SQ quantization.

PRQ is a fusion of PQ and AQ (Additive Quantizer). Compared to PQ, it requires longer build times to deliver better recall, especially at high compression rates, saying binary compression.

Clustering Compaction (Beta)

Milvus 2.5 introduces Clustering Compaction to accelerate searches and reduce costs in large collections. By specifying a scalar field as a clustering key, data is redistributed by range to optimize storage and retrieval. Acting like a global index, this feature enables Milvus to efficiently prune data during queries based on clustering metadata, enhancing search performance when scalar filters are applied.

For details, refer to Clustering Compaction.

Other Features

Streaming Node (Beta)

Milvus 2.5 introduces a new component called the streaming node, which provides Write-Ahead Logging (WAL) services. This enables Milvus to achieve consensus before and after reading and writing channels, unlocking new features, functionalities, and optimizations. This feature is disabled by default in Milvus 2.5 and will be officially available in version 3.0.

IPv6 Support

Milvus now supports IPv6, allowing for expanded network connectivity and compatibility.

CSV Bulk Import

In addition to JSON and Parquet formats, Milvus now supports direct bulk import of data in CSV format.

Expression Templates for Query Acceleration

Milvus now supports expression eemplates, improving expression parsing efficiency, particularly in scenarios with complex expressions.

GroupBy Enhancements

• Customizable Group Size: Added support for specifying the number of entries returned for each group.
• Hybrid GroupBy Search: Supports hybrid GroupBy search based on multiple vector columns.

Iterator Enhancements

• MVCC Support: Users can now use iterators without being affected by subsequent data changes like inserts and deletions, thanks to Multi-Version Concurrency Control (MVCC).
• Persistent Cursor: Milvus now supports a persistent cursor for QueryIterator, enabling users to resume iteration from the last position after a Milvus restart without needing to restart the entire iteration process.

Improvements

Deletion Optimization

Improved the speed and reduced memory usage for large-scale deletions by optimizing lock usage and memory management.

Dependencies Upgrade

Upgraded to ETCD 3.5.16 and Pulsar 3.0.7 LTS, fixing existing CVEs and enhancing security. Note: The upgrade to Pulsar 3.x is not compatible with previous 2.x versions.

Local Storage V2

Introduced a new local file format in Milvus 2.5, enhancing loading and query efficiency for scalar data, reducing memory overhead, and laying the groundwork for future optimizations.

Expression Parsing Optimization

Improved expression parsing by implementing caching for repeated expressions, upgrading ANTLR, and optimizing the performance of NOT IN clauses.

Improved DDL Concurrency Performance

Optimized the concurrency performance of Data Definition Language (DDL) operations.

RESTful API Feature Alignment

Aligned the functionalities of the RESTful API with other SDKs for consistency.