[
https://issues.apache.org/jira/browse/FLINK-40335?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mate Czagany updated FLINK-40335:
---------------------------------
Description:
h1. Summary
When using the hybrid shuffle mode, it's easy to run into this issue which is
caused by using a shared ByteBuffer.
TaskManagers first log exceptions from the index cache's removal listener
* BufferUnderflowException
* BufferOverflowException
* IllegalArgumentException: Negative position
Guava swallows these, and eventually terminate when a blocking shuffle I/O
thread hits
{code:java}
IOException: The spill file is corrupt: premature end of file as an uncaught
exception.{code}
This issue has been present since the introduction of the tiered storage (1.18)
h1. Proposed fix
Create one `ProducerMergedPartitionFileDataIndexRegionHelper` per
`ProducerMergedPartitionFileIndex` instead of sharing a singleton. The cost is
one 24-byte direct buffer per active index instead of one per JVM
was:
h1. Summary
When using the hybrid shuffle mode, it's easy to run into this issue which is
caused by using a shared ByteBuffer.
TaskManagers first log exceptions from the index cache's removal listener
* BufferUnderflowException
* BufferOverflowException
* IllegalArgumentException: Negative position
Guava swallows these, and eventually terminate when a blocking shuffle I/O
thread hits
```
IOException: The spill file is corrupt: premature end of file as an uncaught
exception.
```
This issue has been present since the introduction of the tiered storage (1.18)
h1. Proposed fix
Create one `ProducerMergedPartitionFileDataIndexRegionHelper` per
`ProducerMergedPartitionFileIndex` instead of sharing a singleton. The cost is
one 24-byte direct buffer per active index instead of one per JVM
> Hybrid shuffle index files can be corrupted by a ByteBuffer shared across
> result partitions
> -------------------------------------------------------------------------------------------
>
> Key: FLINK-40335
> URL: https://issues.apache.org/jira/browse/FLINK-40335
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Network
> Reporter: Mate Czagany
> Priority: Major
>
> h1. Summary
> When using the hybrid shuffle mode, it's easy to run into this issue which is
> caused by using a shared ByteBuffer.
> TaskManagers first log exceptions from the index cache's removal listener
> * BufferUnderflowException
> * BufferOverflowException
> * IllegalArgumentException: Negative position
> Guava swallows these, and eventually terminate when a blocking shuffle I/O
> thread hits
> {code:java}
> IOException: The spill file is corrupt: premature end of file as an uncaught
> exception.{code}
> This issue has been present since the introduction of the tiered storage
> (1.18)
> h1. Proposed fix
> Create one `ProducerMergedPartitionFileDataIndexRegionHelper` per
> `ProducerMergedPartitionFileIndex` instead of sharing a singleton. The cost
> is one 24-byte direct buffer per active index instead of one per JVM
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
