ClickHouse Architecture Advisor

ClickHouse's own architecture advisor, an agent skill packaged as a Claude plugin in ClickHouse's agent-skills marketplace. It teaches your agent to design ClickHouse systems around your workload: it sizes up the workload first, works through ClickHouse's decision frameworks, and labels every recommendation by where it comes from. It complements ClickHouse's Best Practices skill, which checks schema and query rules; this one answers when, why and how. What it helps you decide - How to load data: direct batched inserts, async inserts, the Kafka engine with a materialized view, or a queue in front of ClickHouse - How to partition time-series data for retention and TTL, including when not to partition yet - How to enrich events: a dictionary, a runtime JOIN, denormalized columns or a refreshable materialized view - How to handle late-arriving data and mutable state with append-only tables, ReplacingMergeTree or CollapsingMergeTree instead of heavy mutations - When to pre-aggregate with incremental or refreshable materialized views, and when to query raw tables How it answers - It first identifies the workload: observability, security and SIEM, product analytics, IoT and telemetry, market data, or mixed OLAP with point lookups. - Each recommendation says what, why and how, links ClickHouse's docs, and gives a confidence level and a way to validate it: SQL, a metric or a smoke test. - Every recommendation is labelled official (backed by ClickHouse's docs), derived (follows from documented ClickHouse behaviour) or field (experience-based, flagged as a heuristic). - It comes with worked examples for high-throughput observability, SIEM, and real-time market surveillance. How to install it In Claude Code, run /plugin marketplace add ClickHouse/agent-skills, then /plugin install clickhouse-architecture-advisor@clickhouse-agent-skills. The skill also installs with npx skills add clickhouse/agent-skills, which offers ClickHouse's other skills too, or with clickhousectl skills. Good to know - It's guidance only: it doesn't connect to your database or run queries. - It works with open-source ClickHouse and ClickHouse Cloud, and targets ClickHouse 24.1 and later.

• “Design a ClickHouse architecture for 300K log events a second, with dashboards under 5 seconds fresh.”

  Copy

• “Should we load events with direct inserts, async inserts or the Kafka engine?”

  Copy

• “How should I partition this time-series table for retention?”

  Copy

• “Should this lookup be a JOIN, a dictionary or denormalized columns?”

  Copy

• “How do we handle late-arriving corrections and cancels without mutations?”

  Copy

• “Does this dashboard need a materialized view, or can it query the raw table?”