Milvus in 2023: An Unprecedented Vector Database Amidst Tech Buzz

This post highlights Milvus's 2023 achievements, focusing on engineering advancements. Key technical contributions include: zero downtime during rolling upgrades (refined from v2.2.3), a 3x performance improvement in production environments (especially with filtered searches and streaming insert/search), a 5% recall improvement on the Beir dataset using hybrid search (dense + sparse embeddings) and a reranking model, and 10x memory saving on large datasets through disk-based indexing using memory-mapped files (MMap) introduced in Milvus 2.3.4. Other technical developments mentioned are dynamic schema (v2.2.9), Upsert, Range Search, and Cosine metrics (v2.3), Partition Key for multi-tenancy, and enhanced memory management, coroutine handling, and CPU optimization in v2.3.4 for B2B scenarios. The post also mentions the development of VectorDBBench for benchmarking and Pyanns for sparse embedding search.