[
https://issues.apache.org/jira/browse/HUDI-4261?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Alexey Kudinkin updated HUDI-4261:
----------------------------------
Sprint: 2022/08/22, 2022/09/05, 2022/10/18 (was: 2022/08/22, 2022/09/05,
2022/10/18, 2022/11/15)
> OOM in bulk-insert when using "NONE" sort-mode for table w/ large # of
> partitions
> ---------------------------------------------------------------------------------
>
> Key: HUDI-4261
> URL: https://issues.apache.org/jira/browse/HUDI-4261
> Project: Apache Hudi
> Issue Type: Bug
> Reporter: Alexey Kudinkin
> Assignee: Alexey Kudinkin
> Priority: Blocker
> Fix For: 0.13.0
>
> Attachments: Screen Shot 2022-06-15 at 6.06.06 PM.png
>
>
> While experimenting w/ bulk-inserting i've stumbled upon an OOM failure when
> you do bulk-insert w/ sort-mode "NONE" for the table w/ large number of
> partitions (> 1000).
>
> This happens for the same reasons as HUDI-3883: every logical partition
> (let's say we have N of these, equal to shuffling-parallelism in Hudi)
> handled by Spark, (since no re-partitioning is done to align with the actual
> partition-column) will likely have a record from every physical partition on
> disk (let's say we have M of these). B/c of that every logical partition will
> be writing into every physical one.
> This will eventually produce
> # M * N files in the table
> # For every file in the table while writing Hudi will keep a "handle" in
> memory which in turn will hold full buffer worth of Parquet data (until
> flushed).
> This ultimately leads to an OOM.
>
> !Screen Shot 2022-06-15 at 6.06.06 PM.png!
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
