[HDFS-1567] DFSClient should retry reading from all datanodes in round robin fashion

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: hdfs-client
• Assignee: Unassigned
• Reporter: Dhruba Borthakur
• Created: 04/Jan/11 06:42
• Updated: 04/Jan/11 06:46

Description

In the current implementation, the DFSClient retries the same datanode a few times (for reading) before marking the datanode as "dead" and moving on to trying the read-request from the next replica. This introduces additional delay when the first candidate datanode was really out-of-service. Instead, it might be beneficial (for performance) to make the DFSClient retry all datanode replicas one after another (while keeping the total number of retries per datanode the same as now); if all the replicas were tried unsuccessfully once, then loop back to retrying the first one and so on.