Retrieval at Scale | Drop for 2026-07-30

TL;DR

  • Milvus 3.0 is now GA, bringing “lake‑native” retrieval (zero‑copy External Collections, Snapshots, Storage V3) and multi‑vector ergonomics to production.
  • Weaviate shipped a string of 1.38.x performance/ops updates (MMR/Hybrid Diversity, LSM compaction, new batch filters) and opened 1.39.0‑rc.1 with Namespaces, Search REST API, and gRPC‑web at GA.
  • Elastic issued 9.4.4 security updates across Elasticsearch/Kibana—patch recommended for Lucene‑based vector/hybrid stacks.
  • New research extends ACORN‑1: RACORN‑1 targets low‑selectivity filtered ANN with adaptive fallbacks, reporting recall‑preserving speedups.

Milvus 3.0 GA: lake‑native retrieval enters production

  • Key facts and current state of the topic
    • Milvus 3.0.0 was tagged GA on July 29 and reported as generally available mid‑July. It introduces a lake‑native architecture: External Collections for zero‑copy queries over Parquet/Iceberg, Snapshots for MVCC‑style stable views, and a new manifest‑based Storage V3. Multi‑vector workflows (e.g., EmbList + on‑disk indexes) are a first‑class focus. (github.com)
  • Important context and background information
    • This continues the move from all‑in‑memory to SSD/object‑store tiers, improving freshness and cost at billion‑scale while keeping hybrid (vector + scalar/text) pipelines viable. (milvus.io)
  • Recent developments or changes
    • The GA tag (milvus‑3.0.0) is live; Milvus 2.6.x remains active for incremental upgrades. Plan canary tests for External Collections/Snapshots, and validate multi‑vector indexing + on‑disk search against production selectivities. (github.com)

Weaviate 1.39.0‑rc.1: Namespaces and new APIs (preview), plus 1.38.x hardening

  • Key facts and current state of the topic
    • A new RC (July 29) brings Namespaces (GA), Alter‑Schema reindex (GA), drop‑vector‑index (Preview), gRPC‑web (GA), and a Search REST API (GA). Recent 1.38.x patches improved Hybrid Diversity/MMR and LSM performance and added batch “ContainsAny/All/None” filters. (github.com)
  • Important context and background information
    • Namespaces simplify multi‑tenant/data‑domain isolation; MMR/Hybrid Diversity reduce near‑duplicate results without reindexing—useful in vector and hybrid pipelines. (github.com)
  • Recent developments or changes
    • Evaluate RC features in non‑prod; in prod, consider upgrading to ≥1.38.6 for MMR/compaction improvements and to 1.38.8 for the new batch filters when heavy filtering is common. (github.com)

Elastic Stack 9.4.4: security updates for Elasticsearch/Kibana

  • Key facts and current state of the topic
    • Elastic posted coordinated 9.4.4 security updates on July 21, 2026 across Elasticsearch and Kibana. Advisories list multiple CVEs and patched artifacts. (discuss.elastic.co)
  • Important context and background information
    • Many large‑scale retrieval stacks (filtered HNSW, BBQ/RaBitQ quantization, hybrid ranking) run on Elasticsearch/Lucene; timely patching avoids exposure and stability risks. (discuss.elastic.co)
  • Recent developments or changes
    • Upgrade affected clusters to 9.4.4 (or managed equivalents) and re‑run smoke tests for vector/lexical/hybrid query paths under typical filters. (discuss.elastic.co)

Weaviate 1.38.x: MMR/Hybrid Diversity and LSM compaction speedups

  • Key facts and current state of the topic
    • The July 16–29 patch series (1.38.5–1.38.8) improved LSM‑store performance, batch vectorization stability, Raft/replication, and added MMR/Hybrid Diversity and batch “ContainsAny/All/None” filters. (github.com)
  • Important context and background information
    • These updates target p95/p99 consistency and result quality (diversity) in high‑QPS, multi‑vector/hybrid deployments without retraining. (github.com)
  • Recent developments or changes
    • Consider enabling MMR/Hybrid Diversity where near‑dupes or topical redundancy are issues; monitor inverted‑index and vector‑verification CPU after upgrading. (github.com)

RACORN‑1: adaptive filtered‑ANN for low‑selectivity predicates

  • Key facts and current state of the topic
    • A new arXiv paper (July 1) extends ACORN‑1 for filtered HNSW by adding adaptive search fallback that turns filter‑failing nodes into transient bridges, aiming to preserve recall when filters are tight or data are skewed. (arxiv.org)
  • Important context and background information
    • Filtered ANN often hits recall cliffs under selective filters; RACORN‑1 proposes in‑place changes to traversal that can be integrated into existing ACORN‑style engines. (arxiv.org)
  • Recent developments or changes
    • Authors report recall‑preserving speedups on low‑selectivity workloads. If you rely on ACORN‑class traversal, prototype RACORN‑1‑style fallbacks and compare p95/p99 vs. strict post‑filter baselines. (arxiv.org)