TL;DR
- Elastic 9.4 is GA and materially upgrades vector search: GPU‑accelerated indexing (cuVS) is GA, DiskBBQ gets faster (especially under restrictive filters), and new low‑bit quantization options arrive.
- Azure Cosmos DB added spherical quantization (public preview) to speed vector indexing while preserving search quality.
- Amazon OpenSearch Service expanded Cluster Insights across versions, improving at‑scale ops visibility for vector/hybrid clusters.
- Azure SDK for JavaScript (search‑documents 13.0.0) added a vector oversampling knob and richer knowledge‑source support—useful for recall/latency tuning.
- A new survey analyzes “write‑read decoupling” architectures for large‑scale search, with concrete implications for hybrid/vector systems.
Elastic 9.4 GA: faster vector search (DiskBBQ), GPU indexing (cuVS) now GA, more quantization options
- Key facts and current state of the topic
- Elastic 9.4 (May 5) is generally available. It: (a) promotes NVIDIA cuVS–based GPU‑accelerated vector indexing to GA, (b) improves DiskBBQ (Elasticsearch’s disk‑oriented vector index) with sizable latency drops under restrictive filters, and (c) adds new quantization bit‑widths (e.g., 2/4/7 bits) plus broader native‑code vector ops. (elastic.co)
- Important context and background information
- These lift indexing throughput (up to ~12× vs. CPU) and reduce search/verification cost, allowing deeper probes or more filters within the same SLA—useful for high‑QPS, filter‑heavy candidate stages feeding re‑rankers. (elastic.co)
- Recent developments or changes
- Elastic highlights ≥3× query‑latency improvement for restrictive filters and native‑code speedups for vector comparisons in 9.4’s “VectorDB enhancements.” Plan A/Bs vs. your current HNSW/BBQ settings. (elastic.co)
Azure Cosmos DB: spherical quantization (public preview) for faster vector indexing
- Key facts and current state of the topic
- Cosmos DB introduced “spherical” quantization (public preview) as a new option that can improve quantization time—translating to faster index builds and potential query‑time gains. (learn.microsoft.com)
- Important context and background information
- Quantization quality directly gates probe budgets and recall for compressed indexes; spherical variants aim to better preserve angular similarity while shrinking memory/compute. (learn.microsoft.com)
- Recent developments or changes
- Microsoft Learn’s performance tips now document spherical quantization (preview). Validate on your embeddings vs. existing PQ/BQ/BBQ baselines and measure end‑to‑end impact (build + search). (learn.microsoft.com)
Amazon OpenSearch Service: broader Cluster Insights for safer, faster ops
- Key facts and current state of the topic
- On May 5, AWS expanded Cluster Insights availability across OpenSearch (all versions) and Elasticsearch ≥6.8, surfacing proactive health/perf issues directly in the console and via EventBridge. (aws.amazon.com)
- Important context and background information
- For large, filter‑heavy vector/lexical stacks, clearer fleet‑level signals reduce upgrade risk and help hold p95/p99 during indexing/compaction or version moves.
- Recent developments or changes
- Use Cluster Insights to baseline vector index merges, memory‑optimized search warmups, and filter selectivity impacts before/after config changes; pair with 3.6 LTS planning where applicable. (aws.amazon.com)
Azure SDK (JS) search‑documents 13.0.0: vector oversampling and broader knowledge‑source support
- Key facts and current state of the topic
- The May 1 release adds a BaseVectorQuery.oversampling property for vector search (tuning recall/latency) and expands knowledge‑source support (searchIndex, Azure Blob, Indexed OneLake, web) for agentic/KB scenarios. (github.com)
- Important context and background information
- Oversampling helps avoid early‑cut recall loss in compressed or filtered pipelines by fetching more candidates before re‑rank; SDK‑level knobs make this tunable per query.
- Recent developments or changes
- If you run Azure AI Search or mixed Cosmos/Search stacks, expose oversampling as a traffic‑shaped parameter (e.g., by predicate selectivity) and monitor recall vs. p95. (github.com)
Write–read decoupling survey: architectural guidance for cloud‑scale retrieval
- Key facts and current state of the topic
- A May 2 survey analyzes modern search architectures (Elasticsearch, LinkedIn Galene, Uber Sia, Quickwit, Alibaba Havenask, Milvus, Vespa), arguing for “ScaleSearch” patterns that combine compute‑storage separation, full in‑memory indexing, and dedicated write nodes. (arxiv.org)
- Important context and background information
- The analysis covers hybrid (vector + lexical) and serverless patterns relevant to ad‑scale retrieval where continuous ingest, freshness, and predictable tails are hard.
- Recent developments or changes
- Treat as design input when evolving late‑interaction or hybrid stacks toward SSD/object‑store tiers and write/read isolation while keeping re‑ranking budgets tight. (arxiv.org)