TEngineDB-V: OLAP-Native Vector Search for Large-k Workloads at Tencent
This presentation explores TEngineDB-V, a groundbreaking system that reimagines vector search as a first-class analytical primitive within OLAP infrastructures. Designed to handle large-k workloads (k = 10³–10⁵) at billion-scale, the system introduces a segment-decoupled global indexing architecture, relational execution paradigm, and novel DPPQ quantization scheme. Deployed at Tencent for multi-modal analytics and LLM data curation, TEngineDB-V achieves up to 145× speedup over competitive systems while maintaining scalability to 10 billion vectors. The presentation unpacks the architectural innovations, query optimization strategies, and production performance that enable this new paradigm for similarity-driven analytics at internet scale.Script
Traditional vector search systems hit a wall when you ask for thousands or tens of thousands of nearest neighbors at billion-scale. TEngineDB-V from Tencent shatters that limitation by treating vector similarity not as an add-on index, but as a native analytical operation.
The core insight is architectural: instead of coupling vector indexes to every data segment, TEngineDB-V builds one global index expressed as sharded relational tables. This eliminates the scatter-gather execution that forces traditional systems to read and merge results from every segment, causing severe amplification when k reaches thousands.
TEngineDB-V decomposes vector search into standard relational operators: scan, project, top k, and join. The IVFPQ index itself becomes columnar tables that benefit from compression, multi-tier caching, and vectorized execution. This composability allows vector search to fuse seamlessly with filters, joins, and aggregates in a single optimized query plan.
When k increases from 20 to 1 million, competing systems like Milvus and DiskANN suffer catastrophic slowdowns. TEngineDB-V maintains sub-second to low single-digit second latency across that entire range, achieving 2.7 to 50 times speedup at k equals 20,000 and sustaining performance where graph-based and segment-coupled systems collapse under I/O and merge costs.
To preserve recall at large k under tight bit budgets, the authors introduce DPPQ: direction-aware progressive quantization. Each subvector is split into a norm and a unit direction, with the direction encoded using product quantization and hierarchical residual refinement applied across multiple epochs. This design outperforms both standard PQ and RaBitQ on recall without exploding storage costs.
In production at Tencent, TEngineDB-V handles 10 billion vectors with clusters managing over 100 terabytes of data and query latencies under 5 seconds. The system proves that rethinking vector search as a relational primitive unlocks internet-scale multi-modal analytics. Explore the full architecture and results at EmergentMind.com, where you can create videos like this one for any research paper.
