Can Anna loss update when node failed before send merge request to other node？ about fluent HOT 3 CLOSED

Hello,

Thanks for the question. By default, each update (PUT) is acked when it is processed by one node, and so if this node fails before gossiping the update to other replicas, the update will be lost.

Persistence can be achieved via the put_all api that forces the PUT to be acked by all replicas before returning.

from fluent.

Hello,
Thanks for you answer.
After I read some code, I found gossiping period time is 10s by default.

// Define the gossip period (frequency)
#define PERIOD 10000000  // 10 seconds

During the period, only one replica's latest data can be seen by client, all other replica's data is not consistent with the replica which processed the PUT operation. This can be a problem for many application.

If client use put_all api, PUT performance can be a problem？

from fluent.

It is totally true that before the gossip, other replicas will serve stale versions to the clients, and there is actually a fundamental trade-off between staleness and performance in any coordination-free eventually consistent systems.

As you said, for certain applications, this staleness may not be tolerable. In this case we either have to reduce the gossip period (reduce the duration of staleness) or switch to a strongly consistent KVS (if the application cannot tolerate any staleness).

Interestingly, using put_all still introduces staleness as the PUT and GET are uncoordinated. put_all only increases the persistence of the updates.

from fluent.