[
https://issues.apache.org/jira/browse/JENA-1615?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16643280#comment-16643280
]
Damien Obrist commented on JENA-1615:
-------------------------------------
h3. Investigation
[^open_files_after_compaction_before_gc.png] shows duplicate file descriptors
for some of the files. Analyzing the Java heap and playing with garbage
collection, I saw that the first garbage collection after a compaction releases
some of these file descriptors. This is expected as
[BlockAccessMapped#_close|https://github.com/apache/jena/blob/3d456654feb2cf7617a85a5245c80b827900076f/jena-db/jena-dboe-base/src/main/java/org/apache/jena/dboe/base/file/BlockAccessMapped.java#L249]
dereferences its {{MappedByteBuffers}}, whose file descriptors are
subsequently closed by
[FileChannelImpl#Unmapper|http://hg.openjdk.java.net/jdk8/jdk8/jdk/file/687fd7c7986d/src/share/classes/sun/nio/ch/FileChannelImpl.java#l784].
[^open_files_after_compaction_after_gc.png] shows that after garbage collection
there is still an open file descriptor for each TDB file. It seems these are
the {{FileChannelImpl}}'s [file descriptor
instances|http://hg.openjdk.java.net/jdk8/jdk8/jdk/file/687fd7c7986d/src/share/classes/sun/nio/ch/FileChannelImpl.java#l49],
which seem to be
[duplicated|http://hg.openjdk.java.net/jdk8/jdk8/jdk/file/687fd7c7986d/src/share/classes/sun/nio/ch/FileChannelImpl.java#l900]
when creating the memory mapping. The {{FileChannelImpl}} objects get
[garbage-collected|https://github.com/apache/jena/blob/3f371dfa952f4af8c2f2511cf4f36e82a56f5789/jena-db/jena-dboe-base/src/main/java/org/apache/jena/dboe/base/file/BlockAccessBase.java#L169],
but their file descriptors are never closed.
h3. Proposed solution
Based on my investigation, I think the {{BlockAccessBase}}'s {{FileChannel}}
instances need to be closed before they are dereferenced. This would seem to be
consistent with what is done in TDB1's
[BlockAccessBase#_close|https://github.com/apache/jena/blob/0d3928eaf449e7b375038a892a6c9c3b0dd05908/jena-tdb/src/main/java/org/apache/jena/tdb/base/file/BlockAccessBase.java#L152].
I have created a pull request with the proposed fix:
[https://github.com/apache/jena/pull/481]
The change seems to fix the issue: using the mentioned reproduction steps with
a patched Jena SNAPSHOT shows that all (but two) file descriptors are closed
after the garbage collection:
[^open_files_after_compaction_after_gc_with_fix.png]
Please let me know what you think and thanks in advance for your feedback!
> Compaction leaks file descriptors
> ---------------------------------
>
> Key: JENA-1615
> URL: https://issues.apache.org/jira/browse/JENA-1615
> Project: Apache Jena
> Issue Type: Bug
> Components: Core, TDB2
> Affects Versions: Jena 3.8.0
> Environment: I reproduced the issue on the following environments:
> * OS / Java:
> ** MacOS 10.13.5
> Java 1.8.0_161 (Oracle)
> ** Debian 9.5
> Java 1.8.0_181 (OpenJDK)
> * Jena version 3.8.0
> * TDB2 mode: mapped
> Reporter: Damien Obrist
> Priority: Major
> Attachments: open_files_after_compaction_after_gc.png,
> open_files_after_compaction_after_gc_with_fix.png,
> open_files_after_compaction_before_gc.png, open_files_before_compaction.png
>
>
> h3. Context
> I'm using a TDB2 dataset in a long-running Scala application, in which the
> dataset gets compacted regularly. After compactions, the application removes
> the {{Data-xxxx}} folder of the previous generation. However, the
> corresponding disk space isn't properly returned back to the OS, but is still
> reported as being used by {{df}}. Indeed, {{lsof}} shows that the application
> keeps open file descriptors that point to the old generation's files. Only
> stopping / restarting the JVM frees the disk space for good.
> h3. Reproduction steps
> * Connect to an existing TDB2 dataset
> {code}
> val dataset = TDB2Factory.connectDataset("sample"){code}
> * Check open files
> [^open_files_before_compaction.png]
> * Compact the dataset
> {code}DatabaseMgr.compact(dataset.asDatasetGraph){code}
> * Check open files (before garbage collection)
> [^open_files_after_compaction_before_gc.png]
> * Check open files (after garbage collection)
> [^open_files_after_compaction_after_gc.png]
> The last sceenshot shows that, even after garbage collection, there are still
> open file descriptors pointing to the old generation {{Data-0001}}.
> h3. Impact
> Depending on how disk usage is being reported, this can be quite problematic.
> In our case, we're running on an OpenShift infrastructure with limited
> storage. After only a handful of compactions, the storage is considered full
> and cannot be used anymore.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
