内核 列存引擎 HTAP —— WAL 实时同步、优化器自动路由行/列
- 一句话
- 行存 WAL 实时同步并自动维护列存副本,优化器按代价自动选择走行存还是列存,读流量经读池单端点自动分流——PG 生态内最省事的 HTAP。Row-store WAL is synced in real time and a columnar replica is maintained automatically; the optimizer routes each query by cost to row or column store, with read traffic fanned out through a single read-pool endpoint — the easiest HTAP in the PG ecosystem.
- 窄场景
- PG 交易库上直接跑分析查询(运营报表、实时看板),不想引入 TiFlash/ClickHouse 第二套系统;已在 GCP 上。Running analytical queries (ops reports, real-time dashboards) directly on a PG transactional database without introducing a second system like TiFlash/ClickHouse; already on GCP.
- 机制
- AlloyDB 把 PG 的 WAL 流同时喂给列存引擎,自动维护与行存一致的列式副本,用户无感;优化器按查询特征(点查 vs 大扫描聚合)自动路由;读池提供单一端点,读请求自动分流。这是"存储层 HTAP"路线:列存是内嵌加速而非独立引擎——轻,但天花板也低。AlloyDB feeds the PG WAL stream into the columnar engine as well, maintaining a consistent columnar replica with zero user effort; the optimizer routes by query shape (point lookup vs. large scan+aggregation); the read pool exposes one endpoint and fans reads out automatically. This is the "storage-layer HTAP" route: columnar is an embedded accelerator, not a standalone engine — lightweight, but with a lower ceiling.
- 生产验证
- thebuild.com 2026 独立评测肯定列存引擎是 AlloyDB 相对 PG 和 Aurora 的真实优势;同时祛魅:Omni 本地版只是"带 Google 优势的 PG fork"(https://thebuild.com/blog/managed-postgres-examined-google-alloydb-for-postgresql/)。厂商"4x 事务、100x 分析"口径无独立复现。thebuild.com's 2026 independent review affirms the columnar engine as a genuine AlloyDB advantage over PG and Aurora — while deflating the narrative: the on-prem Omni edition is merely "a PG fork with Google advantages" (https://thebuild.com/blog/managed-postgres-examined-google-alloydb-for-postgresql/). The vendor's "4x transactional, 100x analytical" claims have no independent reproduction.
- 竞品差距
- Aurora 无原生列存(并行查询弱一档);TiDB TiFlash 是独立 MPP 节点,大规模分析扩展更强但架构更重;PG 自建列存靠扩展/外部系统。"PG 语法 + 无感列存加速"在 GCP 生态内无对手,但它不是数仓,别拿它当 BigQuery 用。Aurora has no native columnar store (parallel queries a tier weaker); TiDB TiFlash is an independent MPP node — more scalable for large analytics but heavier; DIY PG columnar needs extensions/external systems. "PG syntax + invisible columnar acceleration" has no rival inside the GCP ecosystem — but it is not a warehouse; don't use it as BigQuery.
- 证据等级
- 独立评测(2026,机制与相对优势);厂商性能数字未独立复现Independent review (2026, mechanism + relative advantage); vendor performance figures not independently reproduced
- 最后核验
- 2026-10-012026-10-01