v0.32.0

k6 → native YAML converter

perfscale convert k6 script.js writes a schema-validated test.yaml +
config.yaml pair from a real JavaScript AST. It converts static HTTP calls,
payloads/headers/environment templates, status/body checks, sleep and groups,
fixed/staged/ramping-arrival load, and thresholds (including abortOnFail).
Dynamic JavaScript, xk6 modules and options without an exact native equivalent
fail with a source line instead of being silently dropped. Diagnostics now also
show the column and source snippet, underline the unsupported construct, and
suggest a concrete fix (if: alternatives, static values, before:/after:,
or keeping the script on the k6 runner). Existing outputs are protected unless
--force is passed; scripts outside the static subset continue to run through
perfscale run --k6.

Arrival: coordinated-omission correction (co_correct)

arrival.co_correct: true gives zrk/wrk2-style latency for open-model runs:
a permit that finds the pool saturated is queued instead of dropped, so a
stall shows up as latency in engine_arrival_iteration_ms (measured from the
scheduled start) rather than vanishing into dropped_iterations, and the
dispatcher paces with sub-millisecond precision. Against HttpArena's actix at
10k req/s the p50 matches zrk (0.43 ms vs 1.69 ms in the default mode). The
default stays k6-compatible drop-on-saturation; permits still queued at the
deadline now count as dropped in both modes.

Faster engine on multi-core machines

VUs now run on one single-threaded event loop per HTTP client shard (the
gcannon/zrk model) instead of the work-stealing runtime, which kept a VU and
its connection on different threads and paid a cross-thread wakeup per
request. The CLI also switches to mimalloc and caches parsed URLs. On Linux
CI: +9–10% RPS at 200–256 VUs and −13% CPU at saturation (2.14× k6).
PERFSCALE_VU_RUNTIME=shared restores the old executor. Peak RSS is higher
(~60 MiB vs ~29 MiB at 10 VUs) — mimalloc keeps per-thread segments.

Native HTTP/1.1 fast path

std/http@v1 requests to http:// and https:// on the default pool now go
through a native HTTP/1.1 client: one buffer out and the response parsed in
the VU's own task. There is no hyper connection task, no channel hop and no
tower layers. Static steps reuse their assembled request, lean outputs skip
body decoding when nothing reads the response, and labeled metrics are
recorded without allocations. On Linux CI at 10 VUs, perfscale (quiet) now
does 144k req/s at 10.5 µs CPU per request (zrk: 142k at 13.0 µs). At 256 VUs
it reaches 205k req/s (0.94× zrk), and HTTPS does 117k req/s (k6: 62k).
co_correct pacing moved to a dedicated thread (CPU 39.9 s → 11.8 s at
10k req/s). PERFSCALE_HTTP_CLIENT=reqwest turns the fast path off.

Fix: generated values repeated across iterations

Without seed:, consecutive generators could share one random stream, so
${uuid} and @std/random values repeated between steps of consecutive
iterations — a create-by-unique-name test (e.g. Kubernetes resources) hit
409 Conflict from the second iteration on. Unseeded runs now hash their
seeds like seeded ones.

• Benchmarks: zrk (HttpArena's open-model, CO-corrected generator) joins
  the throughput/saturation tables in --closed mode, and a new latency
  suite compares fixed-rate, CO-corrected latency: zrk -R vs perfscale
  co_correct vs the default drop-on-saturation arrival (LAT_RATE,
  LAT_CONNS, LAT_DURATION).
• Benchmarks: gcannon — HttpArena's io_uring HTTP/1.1 load generator —
  joins the throughput and saturation tables as a competitor engine (APPS
  default now includes gcannon; Linux only, built from pinned tags against
  liburing 2.9 by the bench workflow). Same fairness rules as the other
  rows: one keep-alive connection per VU, no pipelining.

JMeter: JTL metrics, -J properties, thresholds gates

perfscale run --jmeter now writes each run's samples to a private CSV
results.jtl and parses it incrementally into the shared metrics store:
the summary gains full http_req_duration percentiles (p50/p90/p95/p99),
per-JMeter-label series (bounded cardinality), and
http_req_latency/http_req_connect/data_received/data_sent. Plans are
parameterizable from the CLI with repeatable -Jkey=value /
--jmeter-property key=value, and -c config.yaml now applies report:
(end-of-run and during-run reporters) plus after: std/thresholds@v1
gates — a violated severity: fail gate exits non-zero. A zero-sample plan
still falls back to the console summary = translation; a malformed results
file fails the run instead of silently passing. Native-only config (load
shape, before:, other after: actions, abort_on_fail) is rejected with
a clear error rather than silently ignored.

What's Changed

• docs: escape the pipe in classic|native — it broke the site's MDX by @vitalics in #41
• fix(docker): retry the k6 download in install-runners.sh by @vitalics in #42
• feat: add JTL metrics and properties to JMeter by @vitalics in #43
• feat(bench): gcannon (HttpArena's io_uring generator) as a competitor engine by @vitalics in #44
• fix(generate): unseeded generators repeated values across messages by @vitalics in #46
• perf(engine): thread-per-core VU executor, mimalloc, cached URL parsing by @vitalics in #48
• feat: arrival co_correct (zrk-style CO correction) + zrk in bench by @vitalics in #45
• Feat/bench gcannon by @vitalics in #49
• feat(cli): convert static k6 scripts to native YAML by @vitalics in #47
• fix(cli): explain k6 conversion errors with source hints by @vitalics in #51
• perf: native HTTP/1.1 fast path, lean outputs, co_correct pacer thread by @vitalics in #52
• Release v0.32.0 by @vitalics in #50

Full Changelog: v0.31.1...v0.32.0