AnthropicがClickHouseを使ってAI時代のオブザーバビリティをスケールさせる方法

Anthropic has established itself as a leader in developing large-scale language models like the Claude series, emphasizing safety and responsibility. The company's infrastructure is deeply integrated with observability, which plays a crucial role in both performance and protection. In response to the rapid increase in Claude's usage, particularly after the release of Claude 3.5, Anthropic faced challenges in scaling its observability systems to handle vast amounts of telemetry and logs. Seeking a better database solution, they chose ClickHouse for its ability to support real-time data ingestion, fast analysis, and scalable deployment within their secure environment. Although ClickHouse Cloud offered dynamic scaling and cost-effective storage, Anthropic required a custom deployment to meet its strict security standards, leading to a hybrid approach with ClickHouse Cloud's architecture adapted for their infrastructure. This setup, orchestrated via Kubernetes and monitored by Prometheus, enables Anthropic to maintain their security protocols while allowing engineers to focus on developing tools and models without being burdened by database maintenance. ClickHouse also supports Anthropic's move toward agent-driven analytics, facilitating advanced model training and enabling programmatic querying of metrics. This foundation allows Anthropic to continue advancing AI capabilities while maintaining rigorous safety measures.