[ https://issues.apache.org/jira/browse/HADOOP-8990?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13554818#comment-13554818 ]
Binglin Chang commented on HADOOP-8990: --------------------------------------- Thanks for all this, Sanjay. Let me summarize it to see if I fully understand it? The final wire format would be(I assume prefix is unified to 4 byte integer): request: |4byte total length|4byte header len|header|4byte request len|request| response: |4byte total length|4byte header len|header(with error string)|4byte response len(optional if error)|response(optional if error)| Still I prefer the most simple way. request: |4byte length|request(with request no type bytes)| response: |4byte length|response(with optional response no type bytes and optional error)| But if you choose the above solution, I'm also OK. What's the detail of avoiding buffer copy? I post in another thread that writedelimited to can't avoid buffer copy, and codedinputstream create a buffer to serialized, so unless we use one codeinputstream throughout code, it maybe be difficult otherwise. > Some minor issus in protobuf based ipc > -------------------------------------- > > Key: HADOOP-8990 > URL: https://issues.apache.org/jira/browse/HADOOP-8990 > Project: Hadoop Common > Issue Type: Improvement > Reporter: Binglin Chang > Priority: Minor > > 1. proto file naming > RpcPayloadHeader.proto include not only RpcPayLoadHeaderProto, but also > RpcResponseHeaderProto, which is irrelevant to the file name. > hadoop_rpc.proto only include HadoopRpcRequestProto, and the filename > "hadoop_rpc" is strange comparing to other .proto file names. > How about merge those two file into HadoopRpc.proto? > 2. proto class naming > In rpc request RpcPayloadHeaderProto includes callId, but in rpc response > callId is included in RpcResponseHeaderProto, and there is also > HadoopRpcRequestProto, this is just too confusing. > 3. The rpc system is not fully protobuf based, there are still some Writables: > RpcRequestWritable and RpcResponseWritable. > rpc response exception name and stack trace string. > And RpcRequestWritable uses protobuf style varint32 prefix, but > RpcResponseWritable uses int32 prefix, why this inconsistency? > Currently rpc request is splitted into length, PayLoadHeader and PayLoad, and > response into RpcResponseHeader, response and error message. > I think wrap request and response into single RequstProto and ResponseProto > is better, cause this gives a formal complete wire format definition, > or developer need to look into the source code and hard coding the > communication format. > These issues above make it very confusing and hard for developers to use > these rpc interfaces. > Some of these issues can be solved without breaking compatibility, but some > can not, but at least we need to know what will be changed and what will stay > stable? -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira