Running Storage at Facebook
summary
Facebook moves blob storage from HDFS to a Warm Storage system that supports mixed encodings and a draining helper service that picks which hosts to evacuate for planned maintenance. Claims: 3× replication versus Reed‑Solomon (6,9), (10,14), (9,18) trade CPU and read‑repair amplification for 1.4–2× storage overhead and higher loss tolerance. Operators tune a reconstruction rate to meet SLAs; draining speeds planned work but not unplanned failures. A misconfigured encoding plus poor monitoring once caused capacity to run short.
AI-written summary · about