Details
-
Wish
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
-
None
Description
Kafka can achieve ~750K/s records for async replication. We can see that it's highly performant for eventual consistency model.
Some might argue that ZooKeeper is a strong consistency model. Nonetheless there is nuance – ZooKeeper can read stale data. Let's say that stale data is "consistent data". ZK can do async replication and provide only consistent data.
This can be easily achieved by MVCC database design. There might be other benefits too, e.g. watcher can now have some kind of version and reconnection won't incur data loss; multi can rollback easily to older version.
However, this requires change no less than rewriting ZK. Just raise this topic up and see what people think.