An open-source monitoring system with a dimensional data model, flexible query language, efficient time series database and modern alerting approach.

Prometheus stands as a cornerstone in the realm of modern observability, providing a robust toolkit for monitoring and alerting in dynamic, cloud-native environments. Delving into Prometheus, readers can uncover the intricacies of time-series data collection, metrics instrumentation, and scalable monitoring architectures. With Prometheus, readers gain the tools and knowledge needed to ensure the reliability, availability, and performance of their distributed systems.

Prometheus

Prometheus v3.5.0 introduces a `use-uncached-io` feature flag (Linux-only) that uses direct I/O for compaction chunk writes, bypassing the page cache. This reduces Prometheus's page cache footprint by 20–50% as measured by `container_memory_cache`, making memory usage more predictable and reducing pressure on shared cache. The implementation required a custom `directIOWriter` to satisfy direct I/O's strict alignment requirements (file offset, buffer address, and transfer size must align to the storage device's logical sector size). Future work includes extending direct I/O to index files and WAL writes, experimenting with the newer `RWF_DONTCACHE` flag in Linux v6.14, and adding support for non-Linux platforms.

Uncached I/O in Prometheus

Building more confidence around directIOWriter