Duo Zhang

Mind explaining more on why locating is so slow when the replica id does not exist?

I think we should try to check the replica id when we hit the RegionOfflineException, like what we have done in AsyncRpcRetryingCaller.

You can see how we implement timeout in AsyncRpcConnectionImpl... There is a timer, when we reach timeout, the timer task will call completeExceptionally method of the CompletableFuture, so the upper...

Have you guys find the way to use timer to implement the timeout? You can see the code in AsyncRpcRetryingCaller. And for ZKConnectionRegistry, you can pass the TIMER in AsyncConnectionImpl...

> > Here you just submit the async zk operation in the timer task, and then in the callback, you check reschedule the timer task and then you have a...

> Both approaches have their own drawbacks. I still wonder what we can do here for ConnectionRegistry timeouts, given that Zookeeper response might never come back for timer to be...

> > > Both approaches have their own drawbacks. I still wonder what we can do here for ConnectionRegistry timeouts, given that Zookeeper response might never come back for timer...

Interesting. Could you please upload a flame graph about this? I think @ndimiduk should have tested performance impact when implementing this feature. It is a bit strange why we didn't...

So the problem is allocation? What if we just store the region name some where so we do not need to construct it every time?

I think there are scenarios where we do not know how to retry? For example, we have finished a request to region, and when sending request to the next region,...