Bug Report

What did you do?

We have a cache to manage the mapping between URL and gRPC connection:

But if the DNS has a delay, the gRPC connection won't be re-established. If the pod IP can be reused, there is a possibility that the request is sent to the wrong place. We have encountered this issue in the serverless environment, which causes the tso request to be sent to the tso server, which is not the primary. And the client keeps trying at the same connection, which is wrong until the TiDB is restarted.