How Workers powers our internal maintenance scheduling pipeline

Cloudflare explains how building a maintenance scheduler on Cloudflare Workers helped de-risk critical operations by providing a centralized, automated view of the network state and enforcing safety constraints for maintenance across data centres. The system tackles challenges such as overlapping maintenance requests and customer-specific routing rules, ensuring at least one edge router remains online in a given area and preventing simultaneous data-centre outages for Aegis pools.

It uses a graph processing approach with typed associations to efficiently determine the impact of planned work, loading only data near each maintenance event to avoid loading the entire dataset into memory. A structured fetch pipeline with request deduplication, an LRU cache, and regionally cached GET requests reduced response sizes and improved throughput, achieving a ~99% cache hit rate while steering clear of subrequest limits.

For real‑time queries, the team uses Thanos to deliver metrics, and for long-term historical analysis they convert TSDB blocks to Apache Parquet to minimise I/O and enable fast, columnar reads from object storage. This combination, implemented at scale, has transformed how maintenance is scheduled and observed across Cloudflare’s global network.