milvus-io/milvus v2.6.24

milvus-2.6.24

on GitHub

v2.6.24

Release date: September 16, 2026

We are excited to announce the release of Milvus v2.6.24! This release focuses on resource efficiency — lower peak memory during segment loading, smarter force-merge compaction planning, and faster filtered search — along with important fixes for strong-consistency queries, replica channel balancing, and binlog import.

Improvements

• Reworked force-merge compaction planning with multi-round knapsack packing so merged segments track the configured target size more closely and ID preallocation no longer over-reserves (#52243)
• Refined segment loading resource control with batch-level memory estimation, bounded column-group temporary memory, and streaming V3 index loading, substantially lowering peak memory during segment load (#52670, #52787)
• Enabled delayed S3 file open by bumping milvus-storage, avoiding unnecessary object storage connections (#52788)
• Reduced log noise during group chunk creation by emitting a single summary line per chunk instead of one per field (#52900)
• Added optional per-request weights for RRF reranking in hybrid search, available through FunctionScore, legacy rank params, the RESTful API, and the Go client's WithWeights helper (#52910)
• Improved filtered search performance by eliminating an atomic refcount hotspot in the scalar filter evaluation path that accounted for roughly 48% of leaf CPU time (#53070)
• Added admission control that rejects RESTful DQL requests with HTTP 429 before body decoding when the proxy DQL queue is full, reducing proxy CPU usage under request floods (#53112)
• Reduced write amplification during Tantivy-based text and NGRAM index building, lowering disk I/O and index build cost (#53063)
• Added the configurable queryNode.segcore.interimIndex.growingBuildThreadRate parameter to allow multi-threaded interim index building on growing segments (#53034)
• Added idempotent broadcast support so that retried bulk import requests resolve to the existing task instead of creating duplicate imports (#53236)
• Reduced lock contention on the search path by caching per-chunk row counts during filter expression evaluation (#53240, #53248)
• Added a cluster-version gate that automatically enables write-before function materialization only after the whole cluster has been upgraded, keeping behavior consistent during rolling upgrades (#53262)

Bug fixes

• Fixed Woodpecker WAL ignoring the configured woodpecker.meta.prefix and writing metadata under the wrong etcd path, and added client-side append batching for higher small-batch concurrent write throughput (#50201)
• Fixed a deadlock during the switch to the streaming service that could leave DML requests hanging (#52293)
• Fixed SASL/SCRAM-SHA-256 authentication failures against Apache Kafka 4.x brokers by upgrading librdkafka to 2.6.1 (#52666)
• Fixed flush hanging forever when common.storage.useLoonFFI is enabled and minio.ssl.tlsMinVersion is left at its default value (#52731)
• Fixed an issue where all channels of a replica were loaded onto a single query node after scaling out, causing repeated out-of-memory kills and leaving the replica unserviceable (#53090)
• Fixed strong-consistency queries failing with channel tsafe stalled (error code 505) under heavy upsert workloads by isolating online write execution from segment loading on QueryNode (#53139)
• Fixed an issue where a frozen streaming node outside the primary resource group could be unfrozen and reassigned channels unexpectedly (#53230)
• Fixed an issue where C++ log files were written to the /tmp directory and could fill up disk space (#53292)
• Fixed binlog import failing when a nullable vector field had no binlog in the source segment (#53357, #53364)
• Fixed incorrect search results on embedding-list (vector array) fields when trailing empty lists were present, and upgraded Knowhere to v2.6.21 (#53496)
• Bumped Woodpecker to v0.1.44, fixing a duplicate-symbol link failure that broke GPU builds on the 2.6 branch (#53495)