Skip to content

JavaVsJniWriteBenchmark: javaWrite duplicates javaWriteCascading; jniWrite includes Arrow conversion #477

Closed

Description

@dfa1

JavaVsJniWriteBenchmark no longer measures what its names say:

  • javaWrite uses WriteOptions.defaults(), which cascades to depth 3 since F32 two-decimal floats compress to 104% of raw (vortex-jni: 22%); F64 skips bitpacking under ALP #458, so it is the same configuration as javaWriteCascading. The cascading/plain size ratio it prints in teardown is meaningless now.
  • jniWrite builds Arrow vectors from the Java arrays inside the measured loop: ~280 ms of its ~1.18 s per write (async-profiler) is Arrow conversion, not Rust's encoder. That favors vortex-java in the comparison and should be stated next to the numbers (or the conversion hoisted into @Setup, if the API allows reusing exported batches).

Fix: drop or rename javaWrite (e.g. an explicit cascading(0) variant if first-match writes still matter), and document or remove the Arrow cost on the JNI side, in docs/explanation.md too.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions