• Sources: primary
  • Summary: Upstream PolarFS used a local fcntl lock, so two hosts could both mount the shared block device read-write. The new protocol writes a 512-byte ballot record per host, renews it every second, waits up to a 30-second lease plus clock skew allowance on takeover, and fences a node that cannot renew via abort, a SIGKILL timer, and an optional kernel watchdog. The post also lists bugs in the first version, such as skipping unreadable sectors and unsigned age math.
  • Why it matters: It is a worked example of fencing a single writer on shared storage, including the clock and failure-mode assumptions the guarantee depends on.