AI doesn’t always generate perfect ClickHouse schemas (yet)

When using LLMs (Large Language Models) to design ClickHouse tables for real-time event analytics, users may encounter several pitfalls if they rely solely on AI-generated schemas without human validation. The text highlights common mistakes such as inappropriate partitioning, overuse of custom codecs, unnecessary projections, and mismanaged JSON columns, which can lead to inefficiencies and performance issues at scale. It emphasizes the importance of starting with simple schemas, understanding the rationale behind AI-generated decisions, and adding complexity only when justified by actual workload measurements. The text also advises consulting human experts for complex scenarios and large-scale operations, noting that while LLMs are helpful for getting started, human insight is crucial for nuanced, high-stakes decisions. As AI tools continue to improve, collaboration between AI and human expertise will become increasingly important in database design and operation.