[
https://issues.apache.org/jira/browse/FILEUPLOAD-368?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18042067#comment-18042067
]
David Campbell commented on FILEUPLOAD-368:
-------------------------------------------
Two graphs below from when the problem was happening when using
commons-fileupload 1.6.0... they're slightly skewed with their time so I've
lined them up based on time.
The first one shows the growth of /tmp disk utilisation over time, and the
second one shows garbage collection happening.
*When running 1.5, the first graph is completely different.*
When running 1.6.0 we're seeing:
# The uploaded files are not removed
# The File Descriptor counts for the process keep increasing (as reported by
lsof)
# The /tmp eventually fills up and then the application can't accept uploads,
gets errors
There are three servers and each server has two JVMs hosted on it, so there are
six JVMs.
!image-2025-12-02-13-24-56-498.png!
> 1.6.0 unlike 1.5: brought on an issue where temporary files and descriptors
> were not being freed
> -------------------------------------------------------------------------------------------------
>
> Key: FILEUPLOAD-368
> URL: https://issues.apache.org/jira/browse/FILEUPLOAD-368
> Project: Commons FileUpload
> Issue Type: Bug
> Affects Versions: 1.6.0
> Reporter: David Campbell
> Priority: Major
> Attachments: image-2025-12-02-13-22-04-101.png,
> image-2025-12-02-13-24-56-498.png
>
>
> An upgrade of commons-fileupload from 1.5 to 1.6.0 brought on a problem on
> high-load servers where uploaded files were not being deleted and the server
> file descriptor utilisation kept growing. On all server instances, file
> resource utilisation kept growing to the point where it was filling /tmp
> completely and then the server instances needed restarting to solve, as file
> space is not reclaimed even on file deletion on Unix servers when file
> descriptors are still open. The same issue occurred on all our server
> instances.
> We had a major production incident as a result.
> Rolled back to 1.5. That fixed the issue, problem completely gone.
> Don't know any further details of the whys at this point.
>
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
