Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Won't Fix
-
None
-
None
-
None
-
None
Description
In HDFS-5318, each datanode that can reach a (read only) block reports itself as a valid location for the block. While accurate, this increases (redundant) block report traffic and- without partitioning on the backend- may return an overwhelming number of replica locations for each block.
Instead, a DN could report only that the shared storage is reachable. The contents of the storage could be reported separately/synthetically to the block manager, which can collapse all instances into a single storage. A subset of locations- closest to the client, etc.- can be returned, rather than all possible locations.