[connortsui20]

Summary

Every CodSpeed report carries the same warnings. This removes the ones with a cause in this repository: "no flamegraph", "unable to measure", and all of "dominated by system calls" except real file I/O.

Changes

• Python 3.13 for the Python job. All 40 "no flamegraph" notes were the Python benchmarks. pytest-codspeed records Python frames only through the CPython perf trampoline, which exists from 3.12, and the job ran on 3.11. The 40 Python series moved once, because the interpreter changed and pytest-codspeed now runs each benchmark once before it measures.
• No allocator system calls in measured iterations. A simulated benchmark is one cold iteration, and the simulation leaves system calls out of the measure. mimalloc returned freed memory to the OS on a 10 ms timer that the slow simulation crosses, and glibc (used by zstd) trimmed and regrew its heap with brk. The simulation job now sets MIMALLOC_PURGE_DELAY=-1, MALLOC_TRIM_THRESHOLD_ and MALLOC_TOP_PAD_. With the setup fixes below, this took the system-call warning from 15 benchmarks to 2 Python ones on this PR, and fixed bench_compare_primitive[(10000, 2)].
• No one-time setup in measured iterations. ilike_contains and filter_run_end[len=4096_run=16_density=0.1] paid for process-wide setup: the first regex compilation and the session id interner, which read the CPU quota from cgroup files. Both bench binaries now run the operation once in main(), and the guide gains a rule against this mistake.
• filter_bool.rs drops the 1000-row size. It did under 1 µs of work, and about half of filter_powerlaw_by_mostly_false[1000]'s iteration was mimalloc setting up a fresh page after the previous, larger case. CodSpeed reported no execution time for it.

Not addressed: "Walltime results may be unreliable" and "Base and head ran on different runner types" come from running walltime on self-hosted machines instead of CodSpeed Macro Runners. Two Python scans still read a file inside the measured iteration, which is real I/O.

How the causes were found

The benchmarks were built with --cfg codspeed and run under valgrind --tool=callgrind --instr-atstart=no -v --trace-syscalls=yes. CodSpeed starts and stops callgrind instrumentation around each measured iteration, and -v logs those switches, so the system calls between them are exactly the ones inside a measured iteration. With the changes, none of the measured iterations in like, run_end_filter, filter_bool and single_encoding_throughput (116 in all) makes a system call.

Checks run: cargo clippy --locked -p vortex-array on the three changed benches (after rebasing onto #10703, which moved RunEnd into vortex-array) with -D warnings, cargo +nightly-2026-09-10 fmt --check, yamllint --strict, typos, git diff --check, and the callgrind traces above. Not run: the CodSpeed jobs (they run on this PR).

https://claude.ai/code/session_01KF3pMFx7U5ngk8BzQQFS2W

---

Generated by Claude Code

[codspeed]

Merging this PR will regress 7 benchmarks

3 warnings

⚠️ Walltime results may be unreliable

Some benchmarks used the Walltime instrument outside CodSpeed Macro Runners, which can make the results inconsistent.

Learn more about Macro Runners

See affected benchmarks

⚠️ 1 benchmark is dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

⚠️ Base and head ran on different runner types

Some benchmarks with significant performance changes were compared across different runner types, so the changes might not come from the code. Run base and head on the same runner type, or use CodSpeed Macro Runners for more stable results.

Learn more about Macro Runners

Open the report in CodSpeed to investigate

⚡ 14 improved benchmarks
❌ 7 regressed benchmarks
✅ 2408 untouched benchmarks
⏩ 126 skipped benchmarks¹
🗄️ 43 archived benchmarks run²

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Walltime	threshold_neon[8]	1.4 µs	2.6 µs	-45.99%
❌	Walltime	filter_neon[8]	922 ns	1,516 ns	-39.18%
❌	Walltime	threshold_neon[64]	2.6 µs	4.2 µs	-37.52%
❌	Walltime	filter_neon[64]	1.8 µs	2.7 µs	-34.65%
❌	CPU Simulation	chunked_dict_primitive_canonical_into[f64, (1000, 1000, 10)]	141 µs	203.9 µs	-30.82%
❌	Walltime	threshold_neon[80]	3.5 µs	4.7 µs	-26.06%
❌	Walltime	filter_neon[80]	2.5 µs	3.3 µs	-23.69%
⚡	CPU Simulation	filter_run_end[len=4096_run=16_density=0.1]	153.9 µs	53.4 µs	×2.9
⚡	CPU Simulation	density_sweep_random[0.001]	46.9 µs	19.3 µs	×2.4
⚡	CPU Simulation	filter_powerlaw_by_correlated_runs[10000]	29.4 µs	13.9 µs	×2.1
⚡	CPU Simulation	test_repeated_scan_scalar_at[int_float]	365.1 µs	307.8 µs	+18.65%
⚡	CPU Simulation	test_repeated_scan_scalar_at[int_str]	396.6 µs	335.2 µs	+18.34%
⚡	CPU Simulation	test_repeated_scan_scalar_at[int_str_float]	477.8 µs	410 µs	+16.52%
⚡	CPU Simulation	filter_powerlaw_by_power_law[250000]	182.9 µs	158.1 µs	+15.67%
⚡	Walltime	mul_u16_nonnull_avx512	12.5 µs	10.9 µs	+15.42%
⚡	Walltime	filter_avx2[80]	2 µs	1.7 µs	+14.38%
⚡	CPU Simulation	like_per_row_patterns	191.5 µs	168.1 µs	+13.94%
⚡	Walltime	subtract_shapes_avx512[(32768, PerRowPerRow)]	13.1 µs	11.6 µs	+13.03%
⚡	CPU Simulation	test_repeated_scan_filter[int]	2.7 ms	2.4 ms	+12.29%
⚡	CPU Simulation	test_repeated_scan_scalar_at[int]	235.4 µs	211.4 µs	+11.32%
...	...	...	...	...	...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ct/trusting-goodall-2kfoac (83d7d47) with develop (fb4a037)}

Footnotes

1. 126 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. 43 benchmarks were run, but are now archived. If they were deleted in another branch, consider rebasing to remove them from the report. Instead if they were added back, click here to restore them. ↩