Translytical Processing via DB-OS Co-designed Buffer: Cross-Engine Isolation and Tunable Update Visibility for HTAP

Single-node HTAP increasingly co-locates a row-store OLTP engine with a vectorized OLAP engine over the same logical database. In practice, however, systems either share a monolithic buffer cache, often incurring contention, cache pollution, and unpredictable interference, or decouple engines via log shipping and materialization pipelines, trading freshness for operational isolation. We present V ista, a DB–OS co-designed buffer virtualization framework that explores a middle ground via segment-granular sharing. Vista organizes buffer memory into huge-page-backed virtual segments: OLTP executes over active segments, which are periodically sealed into immutable regions and remapped read-only into the OLAP process. This design enforces physical isolation between ongoing OLTP writes and OLAP scans while exposing recent updates with bounded and tunable visibility lag determined by the sealing policy. A user-level library exports snapshot metadata at sealing boundaries and bridges the row/page–tto–tvector mismatch via on-the-fly translation with an opportunistic format cache. We implement V ista in PostgreSQL and integrate it with DuckDB. We quantify isolation–tfreshness tradeoffs under HTAP and show that Vista reduces the interference and improves query latency, achieving up to 12.8× speedup over state-of-the-art baselines.