Weaviate 1.36 Release | Weaviate

What this post added

Introduces HFresh, a new disk-based vector index in technical preview, inspired by the SPFresh algorithm. HFresh divides vectors into postings stored on disk, using an in-memory HNSW index over centroids for efficient search. It features incremental rebalancing, uses HNSW for centroid indexing, and employs Rotational Quantization (RQ) for compression. Key tunable parameters include searchProbe, replicas, and maxPostingSizeKB. It supports cosine and l2-squared distance metrics and is suitable for large datasets, cost-sensitive deployments, and write-heavy workloads. Server-side Batching is now generally available, using a persistent connection and dynamic backpressure for optimized throughput. Object TTL is also generally available, allowing expiration of objects based on creation time, last update time, or a date property, with configurable deletion schedules and an option to filter expired objects. Async replication is now generally available with collection-level configuration via the new `asyncConfig` parameter, offering fine-grained control over data comparison and propagation between nodes.