CAP: HA and Consistency?

[AndrewWinterman]

In HA mode, Redis can lose writes when the primary failovers. It can lose writes on restarts even with persistence enabled.

How about DragonflyDB? Under what circumstances can it drop acknowledged writes? Is it a consistent datastore?

[chakaz]

Hi @AndrewWinterman
Dragonfly is an in-memory data store, like Redis, and as such it is not consistent. There is a very short period of time between when a master acknowledges a command, and until its replica(s) are updated with it. As such, if the master's machine / process terminate unexpectedly, it is possible that some mutations would not be seen on the replica, despite them having been acknowledged by the master.
Having said that, if the master remains alive, and the failover is done by promoting the replica, no commands should be lost.

[AndrewWinterman]

just to be clear, there's no straight line from "in-memory datastore" to "does not wait for replicas to confirm writes before acking", that was a design choice that's going to limit my ability to advocate for dragonflydb for locking/queueing usecases.

[chakaz]

Yes, you are of course right. My phrasing was improper.
To provide the best performance, at the moment, we do not wait for replicas to ACK commands before we commit them. We might add that as an option at some point, although it is currently not on our roadmap.