TTVI: A Two-Tier Vector Index for Low-WAL Approximate Nearest Neighbor Search in Databases
IEEE Access · 2026 · 14 · 114336-114346
DOI: 10.1109/access.2026.3689333 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ibrar Ahmed. Die Arbeit nennt 59 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: specific grant from funding agencies in the public, commercial, or not-for-profit sectors..
Abstract (Originalsprache)
Vector similarity search in relational databases underpins retrieval-augmented generation and semantic search, but indexes like HNSW and IVFFlat in pgvector produce high write-ahead log (WAL) volume because every insert touches many index pages. The Two-Tier Vector Index (TTVI) uses a coarse partition table for routing, per-partition refinement structures for search, and an insert buffer that a background merge flushes into partition-sized index units. A formal WAL cost model shows that each merge produces at most one WAL record per affected partition regardless of how many inserts were buffered; amortized per-insert WAL cost falls as batch size grows. The paper details build and search algorithms, complexity and space, supported distance metrics, and integration with the PostgreSQL access-method API. On synthetic 768- and 1536-dimensional data, TTVI cuts WAL volume by 40–70% for 10K–100K insert batches while keeping recall and query latency in line with IVFFlat and reaching 4– $8\times $ higher insert throughput. The design targets insert-heavy workloads where WAL or replication lag is the bottleneck.
Förderung
- specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Zitieren
APA
Ahmed, I. (2026). TTVI: A Two-Tier Vector Index for Low-WAL Approximate Nearest Neighbor Search in Databases. IEEE Access, 14, 114336-114346. https://doi.org/10.1109/access.2026.3689333
Vancouver
Ahmed I. TTVI: A Two-Tier Vector Index for Low-WAL Approximate Nearest Neighbor Search in Databases. IEEE Access. 2026;14:114336-114346. doi:10.1109/access.2026.3689333
BibTeX
@article{Ahmed2026,
author = {Ahmed, Ibrar},
title = {TTVI: A Two-Tier Vector Index for Low-WAL Approximate Nearest Neighbor Search in Databases},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {114336--114346},
doi = {10.1109/access.2026.3689333}
}
EUVIMED