ClickHouse handles AI observability firehose

This marks a shift from monitoring code to monitoring every thought and action an AI system takes. A normal SaaS app might emit a few infrastructure metrics and request traces, but an agentic app can generate dozens or hundreds of events for one user task, including prompt inputs, model outputs, tool calls, retries, evaluator scores, token counts, and cost records. That turns each application into a log firehose, which strongly favors databases built for cheap storage and fast scans over huge append only streams.1 2 3 4 5

• ClickHouse was originally built for clickstream style data, huge numbers of time stamped events appended continuously, which is structurally similar to AI telemetry. Internal research ties that fit to adoption at AI heavy users like Anthropic and OpenAI, where prompts, traces, evals, and cost data are queried like web events rather than occasional application metrics.1 2 3
• The practical reason log volume explodes is that AI observability tracks the full workflow, not just whether a server stayed up. Langfuse stores traces, evaluations, prompt versions, and monitoring data in ClickHouse, and moved its tracing store from Postgres to ClickHouse before the acquisition. ClickHouse then bought Langfuse and HyperDX to add the workflow layer on top of the storage engine.4 6 7 8
• That volume changes the economics of observability. One operator interview described shrinking an OpenSearch cluster to one third the CPU and memory, extending retention from 7 to 30 days, and still paying about half as much after moving logs to ClickHouse. Another enterprise user said Datadog spend was 10 to 20 times higher at scale, which matters more as AI systems produce far more events per application.5 9 10

As AI apps move from prototypes to production, observability will look less like a sidecar to DevOps and more like a core data workload. That favors vendors that can bundle ingestion, storage, tracing, evaluation, and cost analysis in one stack. ClickHouse is moving in that direction, and higher event volume per app expands the market for that stack with every new agent shipped.2 4 6 8