Latest Results
bench(vortex-geo): benchmark envelope end-to-end at the scalar-fn level
Replace the write-path shape comparison (its numbers are preserved in
that commit's message) with an end-to-end benchmark of
vortex.geo.envelope: the lazy ScalarFnArray executed to Canonical
through a geo session, over harness-built columns. Cases cross nesting
depth (Point, MultiPoint 32 pts/row, MultiPolygon 2 polygons x 2 rings
x 8 vertices) with validity (non-nullable, ~10% periodic nulls, ~50%
pseudo-random nulls). A _test-harness feature plus a self
dev-dependency exposes test_harness to benches, the vortex-array
pattern.
Medians on an Apple M5 Max, 131072 rows per case: point ~0.8us at any
validity (the zero-copy fast path makes execute O(1), so this is
dispatch overhead), multipoint 1.39ms, multipolygon 2.17ms
non-nullable / 2.11ms periodic nulls / 1.37ms random nulls (null rows
are empty in storage and cost nothing). Row count scales linearly:
throughput per row is identical at 1M rows.
Signed-off-by: Nemo Yu <zyu379@wisc.edu> Latest Branches
0%
0%
-14%
© 2026 CodSpeed Technology