Write-back
Cache policy: writes land only in cache, flushed later to the underlying media. Faster than write-through but risk: cache loss before flush = data loss. CPU L1/L2/L3 is write-back. SSD controllers and RAID cache too.
Dirty tracking: every cacheline is marked "modified" vs "clean" — only dirty lines need to be written back. MESI protocol for multi-CPU cache coherence. RAID controller without BBU should RUN write-through (safer), with BBU write-back (faster). SSD with PLP capacitor can write-back safely. Counter: application calls fsync() → cache flushed to disk. Power loss between write and fsync = application data lost from RAM but possibly already landed in flash. Classic OS debug: data corruption after power loss → check write-back configuration at every level.