feat(java): support session

[jackye1995]

This allows Java to also pass in a Session shared across datasets similar to python and rust. Session can then be used for engine side caching implementation in Spark and Trino

[github-actions]

Code Review: feat(java): support session

P0 Issues

Memory leak in Session.fromHandle() (java/src/main/java/org/lance/Session.java:687-690)

When Dataset.session() is called, it invokes nativeGetSessionHandle() which calls handle_from_session() in Rust. This function creates a new Box<Arc<LanceSession>> every time:

pub fn handle_from_session(session: Arc<LanceSession>) -> jlong {
    let boxed: Box<Arc<LanceSession>> = Box::new(session);
    Box::into_raw(boxed) as jlong
}

The Java side then wraps this in a new Session object via Session.fromHandle(handle). However, this creates a new handle that is never released unless the user explicitly closes the returned Session object. The Dataset.session() method caches this internally, but:

1. If session() is called and the returned Session is not closed, the handle leaks
2. The cached session in Dataset is never closed when the dataset is closed

Recommendation: Either:

• Document that users must close the session returned by Dataset.session() (breaking typical Java patterns)
• Have Dataset.close() also close the cached session if it was created from a handle
• Change the design to track whether the Session owns its handle or is just referencing it

P1 Issues

Default cache size mismatch (java/src/main/java/org/lance/Session.java:589-593)

public static final long DEFAULT_INDEX_CACHE_SIZE_BYTES = 1L * 1024 * 1024 * 1024;  // 1 GiB
public static final long DEFAULT_METADATA_CACHE_SIZE_BYTES = 256L * 1024 * 1024;    // 256 MiB

But Rust defaults (rust/lance/src/dataset.rs:141-145):

pub const DEFAULT_INDEX_CACHE_SIZE: usize = 6 * 1024 * 1024 * 1024;    // 6 GiB
pub const DEFAULT_METADATA_CACHE_SIZE: usize = 1024 * 1024 * 1024;      // 1 GiB

And ReadOptions.Builder defaults (java/src/main/java/org/lance/ReadOptions.java:103-104):

private long indexCacheSizeBytes = 6L * 1024 * 1024 * 1024;    // 6 GiB
private long metadataCacheSizeBytes = 1024L * 1024 * 1024;      // 1 GiB

The Session defaults are inconsistent with both Rust and ReadOptions. Consider aligning them.

---

The test coverage is good, and the overall approach of exposing Session to Java is sound. The main concern is the potential memory leak from handles created via Dataset.session().

[majin1102]

LGTM, left one non-blocking comment

[majin1102]

This might cause dangling references without write lock. That said, Rust's structures are shared and close() seems to work fine. Up to you.

[jackye1995]

good point, I updated to just load it at build time