TVA Cuts Temporal Graph Query Latency 9.9x by Decoupling Version Metadata

9.9x lower temporal query latency and 2.2x less storage overhead — that's what TVA, a new temporal graph storage system, delivers over state-of-the-art systems. For anyone running real-time analytics on evolving graphs (think fraud detection on transaction networks or time-series knowledge graphs), this is the number that matters.

The Problem with Existing Temporal Graph Storage

Temporal graphs capture how vertices, edges, and their properties change over time. Standard graph storage systems treat time as an afterthought — they either lack native temporal support or bury versioning in bloated indexes that crater under real-time query loads. Neighborhood scans across multiple versions become a gauntlet of random I/O and redundant metadata lookups.

How TVA Separates Version Metadata from Actual Data

TVA introduces a multi-version storage architecture that decouples version metadata from the property values themselves. For each vertex or edge, version metadata (timestamps, validity ranges) lives in compact structures while the actual property data sits in a separate, contiguous region. That design lets the system locate valid versions for a given time point with a single dereference — no scanning through piles of irrelevant data.

Two tailored data structures make this work: a temporal table and an enhanced hopscotch-based hash table. The temporal table organizes version metadata for adjacent vertices and edges so that a neighborhood scan from a start vertex doesn't require random I/O for each neighbor's metadata. Hopscotch hashing further compacts the metadata, fitting more entries per cache line.

Then there's the version-skipping strategy: during repeated neighborhood scans over multiple vertices, TVA reuses temporal information from the prior scan. That eliminates redundant metadata lookups across scans, a common bottleneck in batch temporal queries.

Benchmark Results and Real-World Impact

Evaluated against state-of-the-art temporal graph storage systems, TVA hits 9.9x lower temporal query latency and 2.2x lower storage overhead. Those aren't cherry-picked numbers — they come from a standard benchmark suite covering diverse query patterns. The system works on unmodified temporal graph workloads, meaning no migration cost for existing analytics pipelines.

Real-time analytics on temporal graphs — continuous graph queries over sliding time windows, time-travel queries, version-diff computation — become practical without throwing hardware at the problem. TVA's architecture points the way to treating time as a first-class dimension in graph storage, not a hack bolted on after the fact. Expect this design to influence future commercial graph databases aiming for sub-millisecond temporal queries.

Source: TVA: A Version-aware Temporal Graph Storage System for Real-time Analytics
Domain: arxiv.org