1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
name: Check and Test
on:
push:
branches:
- main
pull_request:
env:
CARGO_TERM_COLOR: always
CARGO_INCREMENTAL: 0
jobs:
test:
# Both architectures, because this crate's entire job is to characterise
# hardware. x86_64 and arm64 have different cache hierarchies and different
# memory latencies, and the working-set default is chosen against both — a
# change that behaves differently across them is exactly what must not ship.
strategy:
fail-fast: false
matrix:
os:
runs-on: ${{ matrix.os }}
steps:
- name: Checkout code
uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
# No toolchain action: rust-toolchain.toml declares the channel and the
# components, rustup is preinstalled on every runner, and the crate has no
# dependencies — so the pin stays the single source of truth and CI carries
# one third-party action instead of four.
- name: Unit and integration tests
run: cargo ci-test
msrv:
# Cargo.toml claims rust-version = "1.85" and CONTRIBUTING.md repeats it, but
# rust-toolchain.toml pins `stable` — so without this job nothing would
# notice the day the crate stops building on the version it advertises.
runs-on: ubuntu-24.04
steps:
- name: Checkout code
uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
- name: Build against the declared MSRV
run: |
set -euo pipefail
msrv="$(sed -n 's/^rust-version = "\(.*\)"/\1/p' Cargo.toml)"
test -n "$msrv"
echo "declared MSRV: $msrv"
rustup toolchain install "$msrv" --profile minimal
cargo "+$msrv" build --locked --all-features
lint:
# The three jobs cover deliberately different matrices. `test` and `probe`
# need both architectures because the cache hierarchies differ and that is
# what this crate characterises. `lint` runs on Linux and macOS because
# clippy and rustfmt are host-independent but the two runners have disagreed
# on toolchain minor version before. Do not narrow this matrix without
# updating the branch protection's required checks, which name each job by
# its matrix leg — a removed leg blocks every PR on a check that never runs.
strategy:
fail-fast: false
matrix:
os:
runs-on: ${{ matrix.os }}
steps:
- name: Checkout code
uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
- name: Formatting
run: cargo ci-fmt
- name: Clippy
run: cargo ci-lint
probe:
# Smoke the RELEASE binary. This job is the ONLY place the optimised build is
# exercised: `cargo test` links CARGO_BIN_EXE_tachyon against the debug
# binary, so the test suite structurally cannot catch the optimiser deleting
# the pointer chase (pure computation with an unused result).
#
# Note the symptom is inflation, not zero. `accesses` counts completed
# batches, so eliding the inner chase makes the outer loop nearly free and
# the count explodes (~1e12) while latency collapses toward zero. The
# ns_per_access bounds below are the assertion that catches it; an
# `accesses > 0` check would pass straight through it.
strategy:
fail-fast: false
matrix:
os:
runs-on: ${{ matrix.os }}
steps:
- name: Checkout code
uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
- name: Build release
run: cargo build --release --locked
- name: Probe reports real work
run: |
set -euo pipefail
out="$(./target/release/tachyon --seconds 2 --working-set-mb 32 --json)"
echo "$out"
python3 - "$out" <<'PY'
import json, sys
d = json.loads(sys.argv[1])
assert d["accesses"] > 0, f"no accesses recorded: {d}"
assert d["million_accesses_per_sec"] > 0, d
# THE load-bearing assertion. An uncontended DRAM round trip is
# ~60-120 ns. Far below means either the working set fit in cache or
# the chase was optimised away entirely (which drives latency toward
# 0 and inflates the access count); far above means something is very
# wrong with the host. Wide on purpose — a busy runner is allowed to
# be slow, and that is the tool working.
assert 20 < d["ns_per_access"] < 2000, f"latency implausible: {d}"
# Pin the field names: downstream runs join stored scores on these,
# so a rename silently breaks every record already written.
expected = {
"probe", "version", "million_accesses_per_sec", "ns_per_access",
"accesses", "elapsed_s", "threads", "working_set_bytes_per_thread",
}
assert set(d) == expected, f"JSON keys drifted: {sorted(set(d) ^ expected)}"
print(f"ok: {d['million_accesses_per_sec']} M/s, {d['ns_per_access']} ns")
PY