gandiva | Vectorized processing for Apache Arrow

Apache Arrow utilizes a pretty similar idea of Java off-heap storage as Apache Ignite does. For Apache Arrow it means that objects like IntVector don't actually store data in their on-heap layout. They just store a reference to a buffer containing an off-heap address of a physical representation. Technically it's a long offset pointing to a chunk of memory within JVM address space.

When you restart your JVM, address space changes. But in your Apache Ignite native persistence there's a record holding an old pointer. It leads to a SIGSEGV because it's not in the JVM address anymore (in fact it doesn't even exist after a restart).

You could use Apache Arrow serialization machinery to store data permanently in Apache Ignite or even somewhere else. But in fact after that you're going to lose Apache Arrow preciousness as a fast in-memory columnar store. It was initially designed to share off-heap data across multiple data-processing solutions.

Therefore I believe that technically it could be possible to leverage Apache Ignite binary storage format. In that case a custom BinarySerializer should be implemented. After that it would be possible to use it with the Apache Arrow vector classes.