[
https://issues.apache.org/jira/browse/HIVE-14589?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sergey Shelukhin updated HIVE-14589:
------------------------------------
Description:
See HIVE-14574. (copied from the comment below) This basically creates the
nodes in ZK for "slots" in the cluster. The LLAPs try to take the lowest
available slot, starting from 0. Unlike worker-... nodes, the slots are reused,
which is the intent. The LLAPs are always sorted by the slot number for splits.
The idea is that as long as LLAP is running, it will retain the same position
in the ordering, regardless of other LLAPs restarting, without knowing about
each other, the predecessors location (if restarted in a different place), or
the total size of the cluster.
The restarting LLAPs may not take the same positions as their predecessors
(i.e. if two LLAPs restart they can swap slots) but it shouldn't matter because
they have lost their cache anyway.
I.e. if you have LLAPs with slots 1-2-3-4 and I nuke and restart 1, 2, and 4,
they will take whatever slots, but 3 will stay the 3rd and retain cache
locality.
This also handles size increase, as new LLAPs will always be added to the end
of the sequence, which is what consistent hashing needs.
One case it doesn't handle is permanent cluster size reduction. There will be a
permanent gap if LLAPs are removed that have the slots in the middle; until
some are restarted, it will result in misses
was:See HIVE-14574
> add consistent node replacement to LLAP for splits
> --------------------------------------------------
>
> Key: HIVE-14589
> URL: https://issues.apache.org/jira/browse/HIVE-14589
> Project: Hive
> Issue Type: Bug
> Reporter: Sergey Shelukhin
> Assignee: Sergey Shelukhin
> Attachments: HIVE-14589.01.patch, HIVE-14589.patch
>
>
> See HIVE-14574. (copied from the comment below) This basically creates the
> nodes in ZK for "slots" in the cluster. The LLAPs try to take the lowest
> available slot, starting from 0. Unlike worker-... nodes, the slots are
> reused, which is the intent. The LLAPs are always sorted by the slot number
> for splits.
> The idea is that as long as LLAP is running, it will retain the same position
> in the ordering, regardless of other LLAPs restarting, without knowing about
> each other, the predecessors location (if restarted in a different place), or
> the total size of the cluster.
> The restarting LLAPs may not take the same positions as their predecessors
> (i.e. if two LLAPs restart they can swap slots) but it shouldn't matter
> because they have lost their cache anyway.
> I.e. if you have LLAPs with slots 1-2-3-4 and I nuke and restart 1, 2, and 4,
> they will take whatever slots, but 3 will stay the 3rd and retain cache
> locality.
> This also handles size increase, as new LLAPs will always be added to the end
> of the sequence, which is what consistent hashing needs.
> One case it doesn't handle is permanent cluster size reduction. There will be
> a permanent gap if LLAPs are removed that have the slots in the middle; until
> some are restarted, it will result in misses
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
