Cloudflare Town Lake unified data platform processes 91K billing queries monthly; AI agent Skipper standardizes analytics

Cloudflare detailed Town Lake, its internal unified data platform built atop Trino, Iceberg, R2, and DataHub, which consolidates access to operational, billing, security, and business data previously fragmented across PostgreSQL databases, ClickHouse clusters, Kafka streams, BigQuery datasets, and object storage. Billing workloads accounted for 53% of platform queries (~91,760 billing-related queries from 324 employees in a measured period), revealing the architectural priority of unsampled, exact data for revenue accounting rather than approximate analytics for dashboards.

Town Lake paired with Skipper, an AI-powered analytics agent that enables natural language queries across the platform. Cloudflare found that simplifying AI prompts improved accuracy, while consolidating overlapping tools reduced incorrect model selections. Incorporating SQL transformation logic and data lineage into agent context further improved semantic understanding beyond raw schema metadata. The platform addresses longstanding friction: previously, answering a single question like 'How many domains signed up today in the Top 100 by traffic?' required knowing which system to query, what credentials to use, what language to write, and whether the data was sampled or stale.

For infrastructure teams building AI on production data, Cloudflare's architecture signals a critical lesson: billing and revenue-critical workloads cannot tolerate downsampling or sampling approximations, driving separate high-fidelity analytics pipelines. The integration of AI agents as a natural-language interface over governed data—with automated PII detection and default-deny table access—previews the data-ops stack that production AI systems require. Teams migrating multi-source telemetry to foundation models should study Town Lake's classification-first, permission-second governance model.