If the master isn’t configured with sync_binlog, it might not have flushed its last several binary log events to disk before crashing. The replication I/O thread may, therefore, have been in the middle of reading from an event that never made it to disk. When the master restarts, the replica will reconnect and try to read that event again, but the master will respond by telling it that there’s no such binlog offset. The binlog dump process is typically almost instantaneous, so this is not uncommon.