[HADOOP-9786] RetryInvocationHandler#isRpcInvocation should support ProtocolTranslator

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.0.0-alpha1
• Fix Version/s: 2.1.0-beta
• Hadoop Flags: Reviewed
• Assignee: Jing Zhao
• Reporter: Jing Zhao
• Created: 29/Jul/13 17:43
• Updated: 12/May/16 18:26
• Resolved: 30/Jul/13 05:58

Description

Currently the RetryInvocationHandler uses the same RPC ids (i.e., clientId + callId) only when the invocation is a RPC invocation. To check whether an invocation is RPC, RetryInvocationHandler#isRpcInvocation directly apply Proxy#isProxyClass on RetryInvocation#currentProxy. However, if currentProxy is an instance of ProtocolTranslator (e.g., ClientNamenodeProtocolTranslatorPB), the real dynamically-generated proxy object is contained within currentProxy and needs to be retrieved by calling ProtocolTranslator#getUnderlyingProxyObject. Failing to recognize a RPC invocation can cause a retry request to have different "clientId + callId" with its initial call, and fail to hit the corresponding retry cache entry in the NameNode side.

Issue Links

• is related to: HDFS-4942 Add retry cache support in Namenode