peterxcli opened a new pull request, #5027: URL: https://github.com/apache/datafusion-comet/pull/5027
## Which issue does this PR close? Closes #4174. ## Rationale for this change `CometScalaUDFCodegen`, currently the only `CometUDF` implementation, creates output Arrow vectors from the process-wide `CometArrowAllocator`. These JVM allocations bypass Spark’s task memory accounting. FFI input vectors only wrap native-owned buffers that are already accounted by Comet, so they must not be charged again. Output buffers, however, are created on the JVM and may remain alive through the FFI release callback after the Java vector closes. ## What changes are included in this PR? - Add a per-task Arrow child allocator backed by a non-spillable Spark `MemoryConsumer`. - Pass the allocator through the generic `CometUDF` interface and codegen output path. - Keep imported native buffers on the root allocator to avoid double-accounting. - Release Spark accounting with the task while retaining allocator state until the final FFI buffer release. ## How are these changes tested? Added `CometUdfBridgeTest` covering allocation accounting, FFI retention, task completion, and final cleanup. The focused test passes with Spark 4.1 and Spark 3.5/Scala 2.12. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
