[
https://issues.apache.org/jira/browse/HUDI-5363?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17645527#comment-17645527
]
Alexey Kudinkin commented on HUDI-5363:
---------------------------------------
*Relationship of Parallelism and # of files created*
- Shuffle parallelism should be decoupled from the # of files
created (at least by default, in some explicit performance-oriented cases this
coupling might be alright)
- If single Spark partition writes into a single table partition, 1 file will
be
created* (depending on the data size)
- If single Spark partition writes into multiple physical partitions, number of
files created depends on the *ordering* of the records w/in the Spark partition
(this is the case for Iceberg/Delta, but not Hudi, since it keeps file-
writer open until done; for Hudi it could result in OOMs though)
> Remove default parallelism values for all ops
> ---------------------------------------------
>
> Key: HUDI-5363
> URL: https://issues.apache.org/jira/browse/HUDI-5363
> Project: Apache Hudi
> Issue Type: Bug
> Components: writer-core
> Affects Versions: 0.12.1
> Reporter: Alexey Kudinkin
> Assignee: Alexey Kudinkin
> Priority: Blocker
> Fix For: 0.13.0
>
>
> Currently, we always override the parallelism of the incoming datasets:
> # If user specified shuffle parallelism explicitly, we'd use it to override
> the original one
> # If user did NOT specify shuffle parallelism, we'd use default value of 200
> Second case is problematic: we're blindly overriding "natural" parallelism of
> the data (determined based on the source of the data) and replace it with
> static unrelated value.
> Instead, we should only be overriding the parallelism in following cases:
> # User provided an overriding value explicitly
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
