[
https://issues.apache.org/jira/browse/HIVE-6455?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13912237#comment-13912237
]
Gunther Hagleitner commented on HIVE-6455:
------------------------------------------
This is cool. Still reviewing but some ideas:
- Instead of adding a column to the record to be used in the file sink, it'd be
cleaner (and faster) to use the key to determine new files. I believe that
could be achieved through startGroup/endGroup
- Looks like we'd end up duplicating partition column, bucket column, sort
columns in both key and value on the reduce sink. It might be possible to avoid
that, making the intermediate output smaller. Although I'm not sure this would
require additional changes to rebuild the row in the reduce task.
> Scalable dynamic partitioning and bucketing optimization
> --------------------------------------------------------
>
> Key: HIVE-6455
> URL: https://issues.apache.org/jira/browse/HIVE-6455
> Project: Hive
> Issue Type: New Feature
> Components: Query Processor
> Affects Versions: 0.13.0
> Reporter: Prasanth J
> Assignee: Prasanth J
> Labels: optimization
> Attachments: HIVE-6455.1.patch, HIVE-6455.1.patch, HIVE-6455.2.patch,
> HIVE-6455.3.patch, HIVE-6455.4.patch, HIVE-6455.4.patch, HIVE-6455.5.patch,
> HIVE-6455.6.patch, HIVE-6455.7.patch, HIVE-6455.8.patch
>
>
> The current implementation of dynamic partition works by keeping at least one
> record writer open per dynamic partition directory. In case of bucketing
> there can be multispray file writers which further adds up to the number of
> open record writers. The record writers of column oriented file format (like
> ORC, RCFile etc.) keeps some sort of in-memory buffers (value buffer or
> compression buffers) open all the time to buffer up the rows and compress
> them before flushing it to disk. Since these buffers are maintained per
> column basis the amount of constant memory that will required at runtime
> increases as the number of partitions and number of columns per partition
> increases. This often leads to OutOfMemory (OOM) exception in mappers or
> reducers depending on the number of open record writers. Users often tune the
> JVM heapsize (runtime memory) to get over such OOM issues.
> With this optimization, the dynamic partition columns and bucketing columns
> (in case of bucketed tables) are sorted before being fed to the reducers.
> Since the partitioning and bucketing columns are sorted, each reducers can
> keep only one record writer open at any time thereby reducing the memory
> pressure on the reducers. This optimization is highly scalable as the number
> of partition and number of columns per partition increases at the cost of
> sorting the columns.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
