HyperStreamDB is a ultra-fast, Rust-native hybrid vector database built directly for Iceberg tables. Harness GPU-accelerated SIMD kernels and intelligent query optimization to build real-time AI search at petabyte scale.
use hyperstreamdb::{Database, Query, Metric};
// 1. Connect to Iceberg catalog directly
let db = Database::connect("thrift://localhost:9083").await?;
// 2. Build hybrid query with automatic pushdowns
let results = db.query("kb_embeddings")
.vector_search("embedding", &[0.12, -0.43, 0.87])
.metric(Metric::Cosine)
.filter("category = 'commercial' AND active = true")
.limit(5)
.explain() // Returns visual plan with CUDA execution
.execute()
.await?;
println!("Query completed in: {:?}", results.latency());Stop paying the "vector tax". Move away from proprietary, isolated vector databases that force complex ETL pipelines and duplicate your core business datasets.
Utilize hand-optimized CUDA backends and SIMD assembly pipelines (AVX-512, NEON) to search millions of high-dimensional vectors in sub-millisecond latencies.
Search vectors directly in Iceberg V2/V3 datasets. Write metadata caches to your catalog and leverage partition and row-group pruning for highly targeted lookups.
Automatically pushes scalar predicate filters down below the vector search layer, executing combined filters efficiently rather than scanning the entire space post-query.
Rest assured with a native Write-Ahead Log (WAL), crash-consistent state recovery, TLS transport encryption, and granular role-based access control (RBAC).
Deploy integrations out of the box for Apache Spark, Trino, and native Python runtimes, unifying batch ingestion and real-time inference in a single database.
By avoiding proprietary schemas and using parquet-backed column-optimized storage, HyperStreamDB drops vector DB operations costs to an absolute fraction of alternatives.
Comparing search latencies and infrastructure footprint under heavy loads (Cohere 1M vector dataset, 768 dimensions).
How HyperStreamDB embeds into your data lakehouse without duplicating data.
Standard data file assets stored in cloud object storage (S3/MinIO) in Parquet format.
Performs limit pushdowns, predicate evaluation, metadata caching, and row-group skipping.
Applies parallel quantized distance calculation directly on pruned partitions via CUDA.
Choose the perfect tier for your workloads, from open source developer instances to scale-out enterprise architectures.
Speak to an engineer today to learn how HyperStreamDB can integrate into your existing Spark, Trino, and Iceberg pipelines.