You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Documentation — Public descriptions match the README: honest latency and CI thresholds, no stale throughput number. shard.yml, the Docker image, --help, and the installers now share that positioning.
Reporting — A refused connection is reported as connect_refused again when cryload is built with Crystal 1.21.1+. Crystal 1.21.1 (crystal-lang/crystal#17436) swapped the real connect error for a stale errno (connect: Operation now in progress), so refusals showed up as connect_failed. cryload patches the Unix event loops' connect until upstream ships the fix
Reporting — A connect that exceeds --timeout is reported as connect_timeout. The stdlib raises a bare IO::TimeoutError for it, so it used to land in read_timeout and connect_timeout was never emitted
CLI — Numeric flags reject nan, inf and out-of-range values with a config error (exit 2). --rate inf, -d inf or --rate 1e-10 used to crash with OverflowError, and --max-p99 nan crashed with JSON::Error after the whole run, all exiting 1 — the threshold-breach code. --rate now needs at least 1e-9 requests/sec, and huge finite limits (--max-p99 1e300) render instead of overflowing
Proxy — HTTPS targets through --proxy now honour --timeout on the tunnel (CONNECT reply, TLS handshake and every request; a stalled proxy used to block the worker forever), reconnect after the target closes a keep-alive connection instead of failing every later request as other, close the socket when CONNECT or the handshake fails, and support https:// proxies and IPv6 proxy literals. Tunnel phases: dns/connect describe the proxy hop (connect includes the CONNECT exchange), tls the handshake with the target
Proxy — Plain-HTTP targets through --proxy http://user:pass@... now send Proxy-Authorization; only the CONNECT tunnel used to carry the credentials
Accuracy — Transfer metrics count response bodies as received. cryload still sends Accept-Encoding: gzip, deflate by default, but no longer inflates the body, so transfer.* and "Total data" report compressed bytes for a compressing target, and the generator spends no CPU on zlib. An explicit -H Accept-Encoding still wins
Accuracy — The process exits with the exact code the report printed. The gates used to be evaluated again after the report was written, and in request-count mode the measurement window stayed open meanwhile, so a slow stdout reader could turn a printed verdict.exit_code: 0 into exit 1 under --min-rps. The window now closes when the last request is merged
Accuracy — Request-count workers flush their stats at least once a second, like duration workers. A worker whose -n share was below 250 requests stayed invisible until it finished, so progress and a Ctrl-C report of a slow run showed 0 requests
Toolchain — Release binaries, the Docker image and CI are built with Crystal 1.21.1 (was 1.21.0). It carries crystal-lang/crystal#17436, which fixes non-blocking connect on macOS 26.7: a refused connection was treated as established, so a localhost that resolves to ::1 first never fell back to 127.0.0.1. The minimum Crystal for building from source stays 1.21.0