Memory Leak for Pekko Shardedc cluster

[Susmit07]

Hellp Team

We were tying large file upoad 15-20gb using HDFS as source and S3 as sink with S3 multipartApi, but I am seeing leaking resources

One instance of java.util.zip.ZipFile$Source loaded by <system class loader> occupies 4,94,11,544 (55.80%) bytes. The instance is referenced by org.apache.pekko.dispatch.MonitorableThreadFactory$PekkoForkJoinWorkerThread @ 0x6c33b5548 cef-datasync-downloader-pekko.actor.default-dispatcher-6, loaded by jdk.internal.loader.ClassLoaders$AppClassLoader @ 0x6c34cda68.

ZiFileSource is tha major culprit here

Any suggestion elcome

[pjfanning]

A memory leak occurs when an app uses memory but does not release it when it is finished with that data. You have provided no evidence that the memory was leaked - just that a lot of memory was used. java.util.zip.ZipFile is a very memory intensive way to read zip data.
There is a java.util.zip.ZipInputStream class that can be used to stream through the zip data instead. It may be possible to support this approach but I haven't read the code yet.

[pjfanning]

Looking at this closer, the ZipFile is being used by the Java ClassLoader. Maybe you need to investigate the jars on your classpath to see which jars have massive files in them.

[raboof]

Indeed it would be good to understand where this ZipFile$Source is coming from / created.

This is the Eclipse Memory Analyzer (MAT), right? I'm not sure anymore how to read this - I see the URLJarFile which indeed suggests it'd be a good idea to see if there's any huge jars on the classpath, but it seems a bit odd that the 'Retained Heap' of none of the GC roots seems to be around 4,94,11,544 in size. Also that number seems oddly formatted?

It would be good to understand what this ZipFile$Source is used for, indeed.

[pjfanning]

@raboof that's how South Asians format numbers - they use lakhs and crores.

[raboof]

Interesting!

[Susmit07]

Exactly guys I am also a bit confused, explicitly not using ZipFile neither we are reading zip data its plain hdfs and s3 connectors reading parquet file, need to see the classpath