[
https://issues.apache.org/jira/browse/FLINK-30461?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Zhu Zhu updated FLINK-30461:
----------------------------
Component/s: Runtime / State Backends
> Some rocksdb sst files will remain forever
> ------------------------------------------
>
> Key: FLINK-30461
> URL: https://issues.apache.org/jira/browse/FLINK-30461
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Checkpointing, Runtime / State Backends
> Affects Versions: 1.16.0, 1.17.0, 1.15.3
> Reporter: Rui Fan
> Priority: Major
> Labels: pull-request-available
> Attachments: image-2022-12-20-18-45-32-948.png,
> image-2022-12-20-18-47-42-385.png
>
>
> In rocksdb incremental checkpoint mode, during file upload, if some files
> have been uploaded and some files have not been uploaded, the checkpoint is
> canceled due to checkpoint timeout at this time, and the uploaded files will
> remain.
>
> h2. Impact:
> The shared directory of a flink job has more than 1 million files. It
> exceeded the hdfs upper limit, causing new files not to be written.
> However only 50k files are available, the other 950k files should be cleaned
> up.
> !https://user-images.githubusercontent.com/38427477/207588272-dda7ba69-c84c-4372-aeb4-c54657b9b956.png|width=1962,height=364!
> h2. Root cause:
> If an exception is thrown during the checkpoint async phase, flink will clean
> up metaStateHandle, miscFiles and sstFiles.
> However, when all sst files are uploaded, they are added together to
> sstFiles. If some sst files have been uploaded and some sst files are still
> being uploaded, and the checkpoint is canceled due to checkpoint timeout at
> this time, all sst files will not be added to sstFiles. The uploaded sst will
> remain on hdfs.
> [code
> link|https://github.com/apache/flink/blob/49146cdec41467445de5fc81f100585142728bdf/flink-state-backends/flink-statebackend-rocksdb/src/main/java/org/apache/flink/contrib/streaming/state/snapshot/RocksIncrementalSnapshotStrategy.java#L328]
> h2. Solution:
> Using the CloseableRegistry as the tmpResourcesRegistry. If the async phase
> is failed, the tmpResourcesRegistry will cleanup these temporary resources.
>
> POC code:
> [https://github.com/1996fanrui/flink/commit/86a456b2bbdad6c032bf8e0bff71c4824abb3ce1]
>
>
> !image-2022-12-20-18-45-32-948.png|width=1114,height=442!
> !image-2022-12-20-18-47-42-385.png|width=1332,height=552!
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
