See kernel logs below. Prometheus, while generating high I/O load, hits the container memcg limit and is OOM killed. ext4 immediately reports "Error loading buddy information" as a result of a failed ext4_mb_load_buddy() call from ext4_discard_preallocations(), presumably running due to a Prometheus file descriptor being closed. That report is made via ext4_error(), causing a journal abort and read-only remount of the root FS. I haven't done an exhaustive audit, but it doesn't appear that any allocations under ext4_mb_load_buddy() should be accountable to the memcg.

This has happened on 4.7.3-coreos-r2 and 4.8.15-coreos. The root filesystem is in EBS. The problem is sporadic but recurring.