andygrove opened a new issue, #7296: URL: https://github.com/apache/datafusion-comet/issues/7296
### What is the problem the feature request solves? Part of #6834. Reading Comet's in-memory cache is JVM-only work before any data reaches native code. `ArrowCachedBatchSerializer` copies the cached body from heap to off-heap, inflates zstd through zstd-jni, runs Arrow's `VectorLoader`, wraps the vectors in `CometVector`s, and only then exports them over FFI. The write side is the same shape: column statistics and compression both run on the JVM. When the consumer is a native plan, none of that has to be JVM work. Handing the IPC bytes to native and decoding them with arrow-rs would remove the heap-to-off-heap copy, the JVM decompression and the intermediate vector allocation from every cached read, and the same applies to statistics and compression when a batch is cached. ### Describe the potential solution Add a native decode path for cached batches, used when the scan feeds a native operator, and measure it against the current path with the existing cache benchmark before deciding whether to move the write side as well. It fits alongside the prefetch work in #5487. The Spark-consumer case in #5485 still needs the JVM vectors, so the JVM path stays for it. ### Additional context Found in the cache audit (#6834). -- 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]
