何时该用 Liquid Clustering？湖仓布局策略基准测试
原标题：Article: When Should You Use Liquid Clustering? Benchmarking Lakehouse Layout Strategies
AI 摘要
文章基于 Delta Lake 上 5000 万条记录的受控基准测试，比较了分区、ZOrder 和 Liquid Clustering 三种湖仓存储布局在点查、时间范围聚合、多列选择性过滤和大聚合四类工作负载下的表现。结果显示 Liquid Clustering 在四类中的三类延迟最低，应作为查询模式多样或演进团队的默认选择；ZOrder 仅在过滤列稳定且与聚簇键精确匹配时更优，且当日期列作为次级键时时间范围查询比日期分区慢 78%。作者据此给出按工作负载选择布局的决策框架，并建议团队用自身数据实测后再定策略。
正文节选
Key Takeaways - No single storage layout is universally optimal for lakehouse analytics: The right choice is workload-dependent, because using Liquid Clustering, ZOrder, or Partitioning incorrectly can materially impact cost and latency. - ZOrder has a hidden failure mode. When a date column appears as a secondary or tertiary clustering key, time range queries perform significantly worse than on a simple date-partitioned table. - Bytes scanned is a misleading performance proxy. Liquid Clustering