You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@hive.apache.org by "Mithun Radhakrishnan (JIRA)" <ji...@apache.org> on 2017/10/12 06:59:00 UTC
[jira] [Updated] (HIVE-17784) Make Tez AM's Queue headroom
calculation and nParallel tasks configurable.
[ https://issues.apache.org/jira/browse/HIVE-17784?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Mithun Radhakrishnan updated HIVE-17784:
----------------------------------------
Description:
Here's a couple of customizations we made at Yahoo with Hive Tez AMs:
# When calculating splits, {{HiveSplitGenerator}} takes the entire queue's capacity as available, and generates splits accordingly. While this greedy algorithm might be acceptable for exclusive queues, on a shared queue, greedy queries will hold other queries up. The algorithm that calculates the queue's headroom should be pluggable. The greedy version can be the default.
# {{TEZ_AM_VERTEX_MAX_TASK_CONCURRENCY}} and the AM's heap-size can be tuned separately from the AM's container size. We found that users who attempt to increase vertex concurrency tend to forget to bump AM memory/container sizes. It would be handier if those values were derived from the container size.
I'm combining these into a single patch, for easier review.
(Note to self: YHIVE-840)
was:
Here's a couple of customizations we made at Yahoo with Hive Tez AMs:
# When calculating splits, {{HiveSplitGenerator}} takes the entire queue's capacity as available, and generates splits accordingly. While this greedy algorithm might be acceptable for exclusive queues, on a shared queue, greedy queries will hold other queries up. The algorithm that calculates the queue's headroom should be pluggable. The greedy version can be the default.
# {{TEZ_AM_VERTEX_MAX_TASK_CONCURRENCY}} and the AM's heap-size can be tuned separately from the AM's container size. We found that users who attempt to increase vertex concurrency tend to forget to bump AM memory/container sizes. It would be handier if those values were derived from the container size.
I'm combining these into a single patch, for easier review.
> Make Tez AM's Queue headroom calculation and nParallel tasks configurable.
> --------------------------------------------------------------------------
>
> Key: HIVE-17784
> URL: https://issues.apache.org/jira/browse/HIVE-17784
> Project: Hive
> Issue Type: Bug
> Components: Query Planning, Tez
> Affects Versions: 2.2.0, 3.0.0
> Reporter: Mithun Radhakrishnan
> Assignee: Mithun Radhakrishnan
> Attachments: HIVE-17784.1.patch
>
>
> Here's a couple of customizations we made at Yahoo with Hive Tez AMs:
> # When calculating splits, {{HiveSplitGenerator}} takes the entire queue's capacity as available, and generates splits accordingly. While this greedy algorithm might be acceptable for exclusive queues, on a shared queue, greedy queries will hold other queries up. The algorithm that calculates the queue's headroom should be pluggable. The greedy version can be the default.
> # {{TEZ_AM_VERTEX_MAX_TASK_CONCURRENCY}} and the AM's heap-size can be tuned separately from the AM's container size. We found that users who attempt to increase vertex concurrency tend to forget to bump AM memory/container sizes. It would be handier if those values were derived from the container size.
> I'm combining these into a single patch, for easier review.
> (Note to self: YHIVE-840)
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)