Last 7 days
0
Features: 0
Changes: 0
Fixes: 0
Deprecations: 0
Milvus
Distributed vector database for similarity search and AI applications.
Latest Milvus changelog updates, official release notes, breaking changes, security patches, pricing changes, and developer reactions in one product feed.
Follow this Milvus release-notes page to spot useful features, risky migrations, noisy announcements, and source links before they hit your backlog.
Where Milvus updates are checked
Changes.Watch links back to official changelog and release-note sources so summaries stay easy to verify.
Related changelog channels
Use channels to follow groups of tools around a stack, workflow, or topic.
Milvus updates by type
Rolling windows show how many product updates landed in the last 7, 30, 90, and 365 days, grouped by existing changelog semantics.
0
3
7
39
Latest Milvus changelog updates
Release notes are pending and not yet available.
Milvus version 3.0.0 has been released.
Release note not yet published.
Enhanced query scheduling and batching with decoupled dispatch, added C++ thread‑pool activity metrics and Grafana monitoring, and sped up filter execution by skipping null bitmap processing
Added configurable concurrency for text tokenization, inline text index building, GPU Docker image upgrade, max array capacity config, S3 CRC32C support, RBAC role description metadata, standardized error handling, null predicate optimiz...
Added nullable vector fields and element‑level search on StructArray fields, allowing missing embeddings and fine‑grained vector queries
Added nullable vector column constructors and response decoding for all vector types, with Column.ValidCount support.
Added ARRAY APPEND and ARRAY REMOVE partial update operators for Array fields via gRPC and REST upsert APIs
Improved stability and performance with larger L0 compaction backlog limits, streaming node resource group isolation, fast proxy query failover, SIMD‑accelerated IN predicates, and parallel text index loading.
Added struct array vector sub‑field column types and EmbeddingList vector types for MAX SIM searches
Performance and observability upgrades: removed unused SuffixSnapshot layer, switched latency metrics to microseconds, separated internal upsert/delete requery metrics, reduced remote IO for BM25 IDF preload, added querycoord priority ov...
Improved stability and performance: faster MixCoord recovery, optimized query/filter execution, and accelerated unfiltered search on sealed segments
Added TruncateCollection, GetReplicateConfiguration, nullable pointer fields in row‑based APIs, and per‑cluster TLS support to the Go SDK
Added built‑in Google Gemini embedding support, configurable via API key (including Gemini Embedding 2).
Added GetReplicateConfiguration API and configurable TLS minimum version for object storage and CDC outbound connections.
Added new truncate API, RESTful search‑by‑pk, user‑specified warmup settings, LoadWithStrategyAsync, and sparse filtering support.
Critical security patch fixing CVE‑2026‑26190 authentication bypass on metrics port
Fixes critical security vulnerability CVE‑2026‑26190, adds KMS key revocation handling and secures metrics endpoint
Added security controls for the internal /expr endpoint to prevent remote expression execution vulnerabilities.
Add highlight scores, primary‑key search enhancements, and support for reopening segments on data/schema changes, plus storage version label metric and V2 format enforcement
Added automatic retry for object storage rate‑limit errors and configurable metadata batch processing to improve reliability
Added search result highlighting via text highlighter feature
Added QueryIterator for efficient pagination of large query results with configurable options
Upgraded Go to 1.24.11 and golang.org/x/crypto to patch high‑severity CVEs
Added /livez endpoint for Kubernetes liveness probes, GroupBy support on TIMESTAMPZ fields, mmap for JSON shredding, and DML request forwarding to improve resilience.
Added native Geometry and TIMESTAMPTZ data types with RTREE spatial index and timezone‑aware queries; STL‑SORT now supports VARCHAR and TIMESTAMPTZ.
Optimized term expression evaluation for large value sets, dramatically improving query performance.
Fixes a potential data loss issue in Milvus 2.5.x, urging users to upgrade
Fix critical CVE‑2025‑64513 security vulnerability
Critical security fix for CVE‑2025‑64513 – upgrade strongly recommended
Fix critical security vulnerability CVE‑2025‑64513 and encourage immediate upgrade
Added Struct data type for use within ARRAY, enabling array‑of‑vectors and more complex data modeling.
Updated Go to 1.24.6 and Arrow/Go to v17; added hybrid search requery policy config and granular flush target support in flushall operation.
Added Geospatial data type support with GIS functions, R‑Tree index, and AutoID insertion with primary keys.
Enabled writing primary key fields when AutoID is enabled and added manual compaction for L0 segments.
Improved performance and stability, notably removing compaction task timeouts and adding concurrency safeguards
Introduced major new features: JSON Shredding for faster dynamic field filtering, NGRAM indexing for LIKE queries, partial field upserts, Boost function, group‑by on JSON/dynamic fields, dynamic schema activation on existing collections,...
Adds POSIX-compatible remote storage support, model‑based rerankers, and dynamic SegCore parameter updates without restart.
Added search iterator v2 with limit and static param checks, plus support for function reranker and new index types (IvfRabitQ, MinHashLSH).
Enabled ARM SVE acceleration for bitset operations and added a direct I/O disk file writer, delivering notable performance gains.
Implemented SearchIteratorV2 in the Milvus client
Introduces Storage Format V2 with adaptive columnar layout, delivering up to 100x faster point lookups, drastically reduced file counts, and lower memory/I/O consumption.
Fixed multiple bugs across hybrid search, tokenizers, data retrieval, and segment handling to improve stability and correctness.
Fixed critical bugs that could corrupt metadata and cause data loss when renaming collections, along with many stability fixes (null bitmap alignment, text‑match, index leaks, etc.).
Prevent client panics on malformed Milvus search results
Added a separate chunk cache pool, auto-indexing for JSON fields, and local BM25 segment cache to boost performance and stability.
Exported milvusclient.annRequest for external use
Added ability to drop field properties and a cast function for JSON indexes
Added JSON index support for 'contains' expressions and new REST query consistency level options.
Fixed row‑based insertion failure with collection function output fields
Added multi-analyzer support with new tokenizers (Jieba, Lindera, ICU, Language Identifier) and a run‑analyzer API for dry‑run tokenization
Boosted search and load performance with optimized index format and SVE L2 metric support for FP16/NY functions.
Added support for skipping score normalization in the weighted re‑ranker and boosted JSON key stats building by batching documents
Add README and examples for the Go SDK client
Added support for JSON null/exists expressions, UTF‑8 validation for strings and VARCHAR fields, parsing sparse vectors from Parquet structs, and a configurable auto‑balancing trigger interval.
Added JSON Path Index feature enabling inverted indexes on dynamic and JSON columns to boost query performance.
Upgrade core toolchains: Go 1.22.7, Rust 1.83, and Etcd 3.5.18.
Added sub‑task pool for multi‑stage tasks and new auto‑balancing trigger interval plus QueryCoord balance status API to improve task handling and load balancing
Added support for up to 10K collections and 100K partitions per cluster
Performance and stability upgrades: configurable compaction intervals, reduced locking in DataCoord, larger metadata batch sizes, WalkWithPrefix for faster etcd ops, and memory reductions by removing unnecessary Bloom filters
Added PartitionKey isolation, Sparse Index with DAAT MaxScore, "is null" expression support, and customizable root privileges.
Added customizable root privileges and option to skip disk quota checks during bulk import.
Fixed numerous bugs affecting filters, concurrency, config updates, data loss, text match queries, sparse vector retrieval, column races, and bulk inserts with nullable/default fields.
Added tunable max VARCHAR length (up to 65,535) and enhanced expression type conversion.
Enabled mmap for sparse index, fixing a critical issue.
Fixed OOM memory estimation, privilege group expansion, and query node balancer issues.
Added expression templates to accelerate hybrid searches and extra metrics for deletion monitoring; restricted L0 file generation to specific partitions for PartitionKey deletion.
Added full‑text search with BM25, sparse vectors, bitmap indexes, nullable/default fields, and clustering compaction for faster queries
Added customizable privilege groups and a new Grant/Revoke API for finer‑grained RBAC control, plus built‑in groups for common scenarios
Added stability and performance improvements such as batch collection listing, removal of unnecessary segment cloning, and optimized describe/index operations
Version 2.4.16 is deprecated
Fixed critical deadlock in QueryNode shard client initialization and reverted unsupported bulk‑insert database support
Fixed critical collection info loss after snapshotKV garbage collection and multiple resource/goroutine leaks, enhancing stability
Fix critical issue where Milvus v2.4.13 could fail to retrieve collection information after a restart when all MetaKV snapshots were garbage collected
Added dynamic replica adjustment for loaded collections, enabling on‑the‑fly scaling without reload
Add trace ID propagation from client (cherry‑pick change)
Implemented graceful stop for flowgraph manager and added delete operation metrics, L0 forward policy, and empty sparse row support
Added static view for expressions and delete buffer quota logic to improve expression inspection and delete stability
Added upsert support for AutoID collections, beta field‑partial loading, and extensive RBAC enhancements including privilege groups and CDC messaging.
Fixed a critical bug that could return fewer results than the requested top‑k by excluding L0 segments from readable snapshots.
Introduced clustering compaction and decoupled it from the Shard DataNode, boosting query efficiency, fault tolerance and scalability.
Fixed critical data deletion bug and related retry issues, improving reliability of delete operations
Added a new metric to record proxy queue wait duration for improved observability
Fixed critical bugs causing panics, memory leaks, data loss, and incorrect query results across various components.
Optimized memory management and query performance by preserving fixed size memory, preventing slice growth during queries, and making the knowhere build pool size configurable on QueryNode.
Added sparse vector support (auto‑indexing, HNSW index, RESTful API) and RBAC for database describe/alter APIs.
Added RESTful API to stop components and enhanced delete‑by‑expression to respect rate limits based on matched data size
Fixed query iterator missing results during frequent upserts/deletes.
Fixed critical bulk insert stats log garbage collection issue, preserving data integrity.
Added support for sparse float vector bulk insert across binlog, JSON, and Parquet formats.
Strengthened REST v2 API authentication and optimized query result retrieval.
Added channel checkpoint information to the flush response.
Added Float16/BFloat16 vector support in bulk insert and enhanced sparse float vector search capabilities.
Added devops RESTful API for rolling upgrades and a task‑driven collection observer, plus configurable GC scan intervals and option to disable search optimization.
Added new capabilities including MinIO TLS support, AutoIndex for scalar fields, hybrid search refactoring, bitset‑based filtering acceleration, and enhanced import task handling with waiting and size limits.
Optimized index description with bulk retrieval and caching, and added new observability metrics (querycoord lag, entity stats)
Fixed critical data loss bug triggered by deleteBuf memory policy
Added TLS support for Kafka connections and enabled Milvus containers to run as non‑root for enhanced security
Fixed missing hybrid search results when using partition keys (issue 30607).
Fixed DiskANN construction failure (issue 30640)
Enhanced BulkInsert to support auto‑increment primary keys for VarChar and refactored its flush process to avoid hangs after node restarts.
Added RESTful support for array and JSON data types and accelerated index loading via concurrent methods
Added RBAC support for queries and alias, and restored MVCC to improve concurrent and phased query handling.
Added access logs (gRPC only), Parquet import, and binlog indexing on growing segments for faster searches and bulk data handling.
Improved stability and fault recovery with multiple bug fixes, including etcd health check, collection cleanup, and query coordination panics.
Added bulkinsert support for partitionkey and pure list JSON formats
Added pure list JSON support for bulk insert and direct plan construction for vector-output searches;
Introduced array datatype, complex delete expressions, binary metric types, float16 vector support, chunk cache, Tikv meta store, CDC MQ replication, bulk insert with partition keys, upsert API, gRPC metadata, and component IP configurat...
Restored SUBSTRUCTURE/SUPERSTRUCTURE binary metric types and added index info in GetSegmentInfo.
Reintroduced GPU support and added Arm64 images, plus a new NATS‑based experimental message queue and MMap capacity boost for larger datasets
Fixed cluster unavailability during rolling upgrades, enabling near‑zero‑downtime upgrades with the Kubernetes operator.
Fixed multiple bugs causing crashes, excessive storage usage, and incorrect query results, resolving performance degradation introduced in 2.2.12.
Introduced a high‑level RESTful API (sharing the gRPC port) with added endpoints for vector retrieval by ID, JSON ‘contains’ queries, bulk insert with partition keys, and GCS/OSS chunk manager support.
Fixed numerous crashes and panics across QueryNode, Session, RemoteChunkManager, and other components, improving stability.
Added RBAC protections for FlushAll and Database APIs, enhancing security.
Added JSON data type with dynamic schema support and partition keys, boosting flexible schemas and filter/search performance.
Fixed a critical bug causing a rootcoord panic when upgrading from v2.2.x to v2.2.7
Fixed multiple critical bugs causing panics, deadlocks, password validation errors, and query service unavailability.
Fixed DataCoord garbage-collection failures, rootcoord message backlog, and coordinator self-destruct bugs that caused system delays.
Added a First/Random replica selection policy to improve throughput when the primary replica fails
Introduces resource grouping for QueryNodes and a collection renaming API, adding flexible resource isolation and name management.
Introduced rolling upgrade via Helm and coordinator high‑availability (active‑standby) to enable zero‑downtime upgrades and sub‑30 s failover.
Fixed proxy cache cleanup and shard leader cache update errors that occurred after upgrading to 2.2.1
Added Pulsa tenant support, authentication, and TLS for etcd, enhancing security.
Added bulk insertion from files, query pagination, collection‑level TTL, quota limits, RBAC, DiskANN (beta) and data backup features.
Reduced scalar data memory usage and optimized InsertRecord memory consumption
Milvus version 2.1.3 released on 28 September 2022.
Fixed numerous bugs causing GC path errors, search hangs, flush panics, and performance regression with authentication enabled.
Fixed Query Node crash under high concurrency and garbage collector panics when parsing bad segment IDs
Adds new data types and deployment options: VARCHAR scalar, in‑memory replicas, embedded Milvus, Kafka (beta) and RESTful API (beta).
Fixed critical bugs causing collection load failures, query node crashes, and data coord panics
Added multi‑platform compilation (including macOS) and moved the Knowhere engine to its own repository; introduced automatic item expiration and mixed compaction logic as new features.
Changed default consistency level to Bounded Staleness, reducing search latency while keeping strong consistency available.
Added support for preload of specific partitions via load_collection().
Improved query performance in scenarios with no deleted entities.
Added ability to specify metric type for FLAT index searches.
Improved IVF search performance for large nq and nprobe values.
Fixed numerous stability and correctness bugs including result count mismatches, index build crashes, vector precision errors, and stress test hangs.
Resolved several critical bugs: missing index files after compaction, file descriptor leak when monitoring is enabled, server hangs under concurrent queries, user‑specified nlist not applied, and deleted vectors still appearing on GPU.
Updated Milvus configuration file for version 0.10.0 and upgraded FAISS library to 1.6.3.
Fixed duplicate data insertion after server reboot in multi‑partition scenarios.
Added startup checks for CPU, GPU driver, CUDA, log rotation, index‑building suspension, and multiple‑instance protection.
Added ANNOY index type and extended vector deletion support for multiple index types
Added substructure and superstructure distance metrics for the FLAT index, enabling chemical structure searches.
Added vector deletion, get‑by‑ID, flushing/compaction, and runtime configuration updates
Added CPU‑only and GPU Docker images, table partitioning APIs, and Mishards sharding middleware for scalable distributed deployment.
Optimized gRPC by removing compression and changing the generated API, doubling search result transmission speed
Fixed a hotfix where searches failed due to missing table file.
Fixed memory usage and resolved slowing vector search issues
Added IVFSQ8 Hybrid index and new Java SDK