[jira] [Comment Edited] (HIVE-14574) use consistent hashing for LLAP consistent splits to alleviate impact from cluster changes
[ https://issues.apache.org/jira/browse/HIVE-14574?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15427233#comment-15427233 ] Prasanth Jayachandran edited comment on HIVE-14574 at 8/18/16 9:55 PM: --- The 3 byte difference will still be there based on what split strategy is chosen. If a big file is chosen by ETL split strategy the first split will start from 3 offset. If chosen by BI split strategy the first split will start from 0. My fix was related to inconsistently choosing strategies based on AM cache being on or off. was (Author: prasanth_j): The 3 byte difference will still be there based on what split strategy is chose. If a big file is chosen by ETL split strategy the first split will start from 3 offset. If chosen by BI split strategy the first split will start from 0. My fix was related to inconsistently choosing strategies based on AM cache being on or off. > use consistent hashing for LLAP consistent splits to alleviate impact from > cluster changes > -- > > Key: HIVE-14574 > URL: https://issues.apache.org/jira/browse/HIVE-14574 > Project: Hive > Issue Type: Bug >Reporter: Sergey Shelukhin >Assignee: Sergey Shelukhin > Attachments: HIVE-14574.patch > > -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Comment Edited] (HIVE-14574) use consistent hashing for LLAP consistent splits to alleviate impact from cluster changes
[ https://issues.apache.org/jira/browse/HIVE-14574?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15427319#comment-15427319 ] Sergey Shelukhin edited comment on HIVE-14574 at 8/18/16 10:59 PM: --- We can easily achieve unique IDs by taking ZK node name (which is unique and sequential). However, as the new node is re-added to the tail on every restart, it throws everything off. What we want conceptually is that restarted nodes go into the same position in the order as the nodes they replaced, but that is difficult to achieve (or impossible, I am not sure we have such a concept with Slider). We can just have sequential numbers in ZK to take, with every registering node fighting for the lowest number. I wonder if there's already a primitive for that in curator ;) That way the replacement nodes take the place of the nodes that died, most of the time, and leave the running ones undisturbed for most of the time. We can also assume we usually restart in the same place and order by the first time there was LLAP on that particular node for that cluster, then by name. was (Author: sershe): We can easily achieve unique IDs by taking ZK node name (which is unique and sequential). However, as the new node is re-added to the tail on every restart, it throws everything off. What we want conceptually is that restarted nodes go into the same position in the order as the nodes they replaced, but that is difficult to achieve (or impossible, I am not sure we have such a concept with Slider). We can just have sequential numbers in ZK to take, with every registering node fighting for the lowest number. I wonder if there's already a primitive for that in curator ;) We can also assume we usually restart in the same place and order by the first time there was LLAP on that particular node for that cluster, then by name. > use consistent hashing for LLAP consistent splits to alleviate impact from > cluster changes > -- > > Key: HIVE-14574 > URL: https://issues.apache.org/jira/browse/HIVE-14574 > Project: Hive > Issue Type: Bug >Reporter: Sergey Shelukhin >Assignee: Sergey Shelukhin > Attachments: HIVE-14574.patch > > -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Comment Edited] (HIVE-14574) use consistent hashing for LLAP consistent splits to alleviate impact from cluster changes
[ https://issues.apache.org/jira/browse/HIVE-14574?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15431860#comment-15431860 ] Sergey Shelukhin edited comment on HIVE-14574 at 8/23/16 12:30 AM: --- This patch actually breaks unique Ids cause it is no longer propagated to clients... I'm fixing this in the other JIRA and may just resolve this as dup of that, or port it back into this patch. The reason is that worker-x is also unique (for debug purposes too) so there's no point in the guid. Is there? was (Author: sershe): This patch actually breaks unique Ids cause it is no longer propagated to clients... I'm fixing this in the other JIRA and may just resolve this as dup of that, or port it back into that patch. The reason is that worker-x is also unique (for debug purposes too) so there's no point in the guid. Is there? > use consistent hashing for LLAP consistent splits to alleviate impact from > cluster changes > -- > > Key: HIVE-14574 > URL: https://issues.apache.org/jira/browse/HIVE-14574 > Project: Hive > Issue Type: Bug >Reporter: Sergey Shelukhin >Assignee: Sergey Shelukhin > Attachments: HIVE-14574.01.patch, HIVE-14574.patch > > -- This message was sent by Atlassian JIRA (v6.3.4#6332)