[RATIS-1076] Should not shutdown and re-create grpc channel/stub in GrpcClientProtocolClient.

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Component/s: gRPC
• Labels: MultiRaft dense-storage
• Assignee: Glen Geng
• Reporter: Glen Geng
• Created: 25/Sep/20 05:12
• Updated: 11/Aug/23 19:03
• Resolved: 11/Aug/23 19:02

Description

according to reply from grpc community

https://github.com/grpc/grpc-java/issues/7449

https://github.com/grpc/grpc-java/issues/7442

A channel can be used throughout the lifetime of the application. There shouldn't be a case that the application needs to shut down and recreate it.

We would like to let StreamObserver.onError() to call ManagedChannel.resetConnectBackoff() as one of the recovery step: for non-backoff state channel, it would be a no-op; for backoff state channel, it will trigger an immediate re-connect to avoid restarted followers from timeout.

 

We've handled this issue for the ManagedChannel in GrpcServerProtocolClient in RATIS-1072. Just file this Jira to track the same problem in GrpcClientProtocolClient. There should not be any safety issues for GrpcClientProtocolClient, but since 

The overhead for creating a channel is big.

remove un-necessary shutdown and re-create grpc channel may improve performance of client operations.

 

BTW, ozone may suffer the same issue, since it also use ManagedChannel in:

org.apache.hadoop.ozone.container.replication.GrpcReplicationClient
org.apache.hadoop.hdds.scm.XceiverClientGrpc

Issue Links

• duplicates: RATIS-1072 Should not shutdown and re-create channel/stub in GrpcServerProtocolClient when StreamObserver::onError() is called.