VLDB 2027 Systems Paper (In Preparation)
Target: PVLDB Vol. 20, Industry Track
Working Title: Samyama: A Unified Graph-Vector Database with Cost-Based Query Planning and Late Materialization
Status: In preparation. Extends Paper 1 (arxiv:2603.08036) with deeper evaluation and VLDB-specific focus.
Strategy
The arxiv preprint covers everything broadly. For VLDB we focus and deepen on three core database contributions:
- Late materialization for property graphs (
NodeRef/EdgeRef) — 4× traversal speedup, 60% memory reduction - Graph-native cost-based planner — triple-level statistics, plan enumeration, direction reversal,
ExpandInto - Hybrid CSR adjacency — frozen tier + write buffer, two-phase bulk loading, billion-edge ingestion at single-server scale
De-emphasize: GPU acceleration, agentic enrichment, optimization solvers, RDF (mention as system features, don’t evaluate deeply).
Add: LDBC SNB comparison, billion-edge evaluation, cross-KG federation at scale.
Key New Results (since arxiv v2)
| Result | Impact |
|---|---|
| 11-KG mega-benchmark, 138M nodes / 1.22B edges, 500/500 pass on r7i.16xlarge | Scale + correctness credibility |
| Adjacency-aware aggregation (ADR-017) | Algorithmic contribution — closes degree-aggregation timeouts on PubMed-scale data |
| WCO TrieJoin for cyclic patterns | Algorithmic contribution |
Edge-arena removal (DS-07c) — COW edges via EdgeView | Engineering contribution — substantial memory savings at 1B-edge scale |
| MVCC: Snapshot Isolation + Read Committed with write-write conflict detection | Concurrency story |
| Rayon parallel algorithms (PageRank, LCC, CDLP) | Parallelism |
Manuscript sources, planning notes, and formatting TODO live in the private samyama-research repository under papers/paper-vldb-2027/.