[HDFS-5769] Bootstrap a write pipeline from read-only replicas

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.3.0
• Component/s: hdfs-client, namenode
• Assignee: Unassigned
• Reporter: Eric Sirianni
• Created: 14/Jan/14 21:19
• Updated: 28/Jan/14 23:36

Description

HDFS-1606 added the capability for the DFSClient to add DataNodes to an existing write pipeline. HDFS-2832 and HDFS-5318 introduce read-only replicas to HDFS. This JIRA proposes using the HDFS-1606 addDatanode2ExistingPipeline() capability to enable recovering in scenarios where only read-only replicas are available (i.e. all read-write replicas are offline).

The following scenarios should be supported:

1. Recovering an initial write pipeline with repcount=1 (see HDFS-5434) when the datanode hosting the r/w replica fails (but r/o replicas are available)
2. Constructing an append pipeline for a block where the r/w replica is offline (but r/o replicas are available)

In both cases, a pipeline of length 0 can be "bootstrapped" to a pipeline of length 1 using available r/o replicas.