[ https://issues.apache.org/jira/browse/HDFS-10968?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Tsz Wo Nicholas Sze updated HDFS-10968: --------------------------------------- Hadoop Flags: Reviewed Description: For an EC block, it is possible we have enough internal blocks but without enough racks. The current reconstruction code calls {{BlockManager#isInNewRack}} to check if the target node can increase the total rack number for the case, which compares the target node's rack with source node racks: {code} for (DatanodeDescriptor src : srcs) { if (src.getNetworkLocation().equals(target.getNetworkLocation())) { return false; } } {code} However here the {{srcs}} may include a decommissioning node, in which case we should allow the target node to be in the same rack with it. For e.g., suppose we have 11 nodes: h1 ~ h11, which are located in racks r1, r1, r2, r2, r3, r3, r4, r4, r5, r5, r6, respectively. In case that an EC block has 9 live internal blocks on (h1~h8 + h11), and one internal block on h9 which is to be decommissioned. The current code will not choose h10 for reconstruction because isInNewRack thinks h10 is on the same rack with h9. was: For an EC block, it is possible we have enough internal blocks but without enough racks. The current reconstruction code calls {{BlockManager#isNewRack}} to check if the target node can increase the total rack number for the case, which compares the target node's rack with source node racks: {code} for (DatanodeDescriptor src : srcs) { if (src.getNetworkLocation().equals(target.getNetworkLocation())) { return false; } } {code} However here the {{srcs}} may include a decommissioning node, in which case we should allow the target node to be in the same rack with it. For e.g., suppose we have 11 nodes: h1 ~ h11, which are located in racks r1, r1, r2, r2, r3, r3, r4, r4, r5, r5, r6, respectively. In case that an EC block has 9 live internal blocks on (h1~h8 + h11), and one internal block on h9 which is to be decommissioned. The current code will not choose h10 for reconstruction because isNewRack thinks h10 is on the same rack with h9. Summary: BlockManager#isInNewRack should consider decommissioning nodes (was: BlockManager#isNewRack should consider decommissioning nodes) Good catch! +1 the patch looks good. > BlockManager#isInNewRack should consider decommissioning nodes > -------------------------------------------------------------- > > Key: HDFS-10968 > URL: https://issues.apache.org/jira/browse/HDFS-10968 > Project: Hadoop HDFS > Issue Type: Sub-task > Components: erasure-coding, namenode > Affects Versions: 3.0.0-alpha1 > Reporter: Jing Zhao > Assignee: Jing Zhao > Attachments: HDFS-10968.000.patch > > > For an EC block, it is possible we have enough internal blocks but without > enough racks. The current reconstruction code calls > {{BlockManager#isInNewRack}} to check if the target node can increase the > total rack number for the case, which compares the target node's rack with > source node racks: > {code} > for (DatanodeDescriptor src : srcs) { > if (src.getNetworkLocation().equals(target.getNetworkLocation())) { > return false; > } > } > {code} > However here the {{srcs}} may include a decommissioning node, in which case > we should allow the target node to be in the same rack with it. > For e.g., suppose we have 11 nodes: h1 ~ h11, which are located in racks r1, > r1, r2, r2, r3, r3, r4, r4, r5, r5, r6, respectively. In case that an EC > block has 9 live internal blocks on (h1~h8 + h11), and one internal block on > h9 which is to be decommissioned. The current code will not choose h10 for > reconstruction because isInNewRack thinks h10 is on the same rack with h9. -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org