[HDFS-211] Support for options within the Datanode Transfer protocol

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Assignee: Unassigned
• Reporter: George Porter
• Created: 05/Sep/08 23:34
• Updated: 18/Jul/14 22:25
• Resolved: 18/Jul/14 22:25

Description

This proposal is to extend the DataNode Transfer protocol (described in DataNode.java lines 916 - 995) to support "options", in the spirit of IP or TCP options. This should make this protocol more extensible, allowing the client to include metadata along with commands. This would support efforts to include end-to-end and causal tracing into Hadoop, and hopefully other efforts as well.

Options should have a type, and be of variable length. It should be possible to include multiple options along with each datanode command. The option should apply to both the command and any data that is part of the command. If the datanode does not understand a given option, it should ignore it. Options should be sent end-to-end through intermediate datanodes, if necessary. For example, if an OP_WRITE_BLOCK command is pipelined through several machines, the options should be sent along the pipeline. Nodes along the pipeline may modify the options.

BTW, If HADOOP-4005 (concrete datanode protocol) is implemented, then it should solve this problem by simply letting the user add state to the concrete protocol class.

Issue Links

• relates to: HDFS-400 Make the Data transfer streaming protocol concrete