Push-tier benchmark¶

bench_push_tier.rs measures parsing plus per-field insertion into TableBuilder, while leaving out row finalization and Arrow export.

Why this shape¶

The inline parser reads UTF-8 TSV-like records. Each tab-separated token must look like key=value; values stay borrowed from the input. The PushOnly sink delegates field insertion to TableBuilder but calls advance_row() at row end. That deliberately skips finish_row() work such as null filling, dirty-mask handling, and filter checks. The result isolates the cost of resolving and pushing fields.

One warmup runs before seven measured runs. The page reports median, best, worst, coefficient of variation, rows, MB/s, and nanoseconds per field. The row estimate comes from newline count, and field count comes from the first line, so malformed or heterogeneous input can make the derived rate less meaningful.

Run it¶

cargo run --release -p rypipe-core --example bench_push_tier -- path/to/file.tsv

For hardware counters, build once and pass the executable to perf stat on Linux:

perf stat -e cycles,instructions,branch-misses,L1-dcache-load-misses \
  target/release/examples/bench_push_tier path/to/file.tsv

Use the same file, build, and machine when comparing changes. This is a push phase measurement. It does not measure finish_row(), Arrow conversion, parallel scheduling, or source I/O setup.

Source: bench_push_tier.rs.