[
https://issues.apache.org/jira/browse/FLINK-27338?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
luoyuxia updated FLINK-27338:
-----------------------------
Description:
Currently, for hive source, it'll use the hdfs block size configured with key
dfs.block.size in hdfs-site.xml as the max split size to split the files. The
default value is usually 128M/256M depending on configuration.
The strategy to split file is not reasonable for the number of splits tend to
be less so that can't make good use of the parallel computing.
What's more, when enable parallelism inference for hive source, it'll set the
parallelism of Hive source to the num of splits when it's not bigger than max
parallelism. So, it'll limit the source parallelism and could degrade the
perfermance.
To solve this problem, the idea is to calcuate a reasonable split size based on
files's total size, block size, default parallelism or parallelism configured
by user.
was:
Currently, for hive source, it'll use the hdfs block size configured as
dfs.block.size as the max split size to split the files. The default value is
usually 128M/256M depending on configuration.
The strategy to split file is not reasonable for each split will tend to be
larger so that can't make good use of the parallel computing .
What's more, when enable parallelism inference for hive source, it'll set the
parallelism of Hive source to the num of splits when it's not bigger than max
parallelism. So, it'll limit the source parallelism and could degrade the
perfermance.
To solve this problem, the idea is to calcuate a reasonable split size based on
files's total size, block size, default parallelism or parallelism configured
by user.
> Improve spliting file for Hive soure
> ------------------------------------
>
> Key: FLINK-27338
> URL: https://issues.apache.org/jira/browse/FLINK-27338
> Project: Flink
> Issue Type: Improvement
> Components: Connectors / Hive
> Reporter: luoyuxia
> Priority: Major
>
> Currently, for hive source, it'll use the hdfs block size configured with key
> dfs.block.size in hdfs-site.xml as the max split size to split the files. The
> default value is usually 128M/256M depending on configuration.
> The strategy to split file is not reasonable for the number of splits tend to
> be less so that can't make good use of the parallel computing.
> What's more, when enable parallelism inference for hive source, it'll set the
> parallelism of Hive source to the num of splits when it's not bigger than max
> parallelism. So, it'll limit the source parallelism and could degrade the
> perfermance.
> To solve this problem, the idea is to calcuate a reasonable split size based
> on files's total size, block size, default parallelism or parallelism
> configured by user.
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
