"exhibits Stale Read, Lost Update, and other forms of G-single in healthy clusters, when no faults occur. Transactions may (apparently) modify data in the interval between a single transaction’s reads and writes; read-modify-write patterns, like those used in many ORMs, are likely unsafe."
innodb storage engine prefetches and caches itself with default 16k record size
increase writes by disabling disk flushing with innodbflushlogattrxcommit=0 and syncbinlog=0. ZFS disable all explicit disk flushing to any one file system by setting sync=disabled so it flushes every ~5sec