I've run some benchmarking of sets with ReplicateTo.ONE enabled and found very poor performance (about 300 sets/s on two six-core and 20GB ram servers).
After some research it was found that the cause is replication delay (it is about 1 second) - the code in observePoll() in java client does memcached observes every X milliseconds (could be set up upon connection establishment) and Y times.
Tuning X and Y does not affect overall performance, because actual data replication takes around 1 second (so X = 100 and Y = 10, X = 10, Y = 100 is enough, but does not improve performance).
The question is:
Is there an ability to tune replication delay in Couchbase? Why this delay is about 1 second? Is it connected to number of updates in any way?
The main goal of minimizing replication delay is to lose less document updates upon node crash. At the moment, working at 20 000 sets/s (with ReplicateTo.ZERO) leads to 20000 document updates loss if the node is crashed.