On Wed, Oct 21, 2015 at 8:00 PM, Edward Capriolo <[email protected]> wrote: > There used to be this setting in zk. > > zookeeper.forceSync > Using Kafka I used to set this setting to false, The reason is I had a > number of clients writing offsets to zk and the disk system was always > sync-ing dirty pages. The system would be super high IO wait. > > I set this to false and disk usage basically dropped to 0. forceSync means use fsync for zk journal operations. Without this, writes to zookeeper are only going into the os page cache, so a crash of the node, could cause data loss on that node. Even worse, if a node crashes and reboots, it can accept two competing proposals and you can theoretically end up with a system with splitbrain. This is why durability is a requirement for all paxos like protocols.
-Ivan
