v1.4.10

Bump celeris dependency to v1.5.4 for the v1.5.4 cross-repo benchmark release set. No loadgen behavior change.

loadgen v1.4.9

loadgen v1.4.9

Resilience fixes that let the full benchmark matrix survive servers it previously DNF'd on. Each is the loadgen side of a column that produced zero requests.

• h1: backoff-paced reconnect on read EOF — a Connection: close server (or one closing mid-response under churn-close) surfaces as a read EOF on the status/header line, not just the write. The read paths now reconnect + back off like the write path, instead of spinning read-EOFs with no pacing. (drogon churn-close: 0 → healthy.)
• h2: honor server SETTINGS_MAX_FRAME_SIZE + send-window flow control — POST bodies > 16384 B were sent as one oversized DATA frame (FRAME_SIZE_ERROR) and bodies > the 65535 initial window overran flow control. Now split at the server's frame size and paced against the connection + per-stream send windows, replenished from WINDOW_UPDATE. (post-64k-h2 across the h2 columns.)
• h2: re-dial connections the server closes/GOAWAYs mid-cell — the h2 client dialed once and never recovered a torn-down conn; a server that GOAWAYs periodically (hypercorn) left the slot dead and every request hot-looped. Each conn is now a re-dialable slot with single-flight + backoff. (fastapi-h2: ~1.1B-error hot loop → healthy.)

Verified on x/net 0.56.0 incl. -race, with regression tests for the flow-control split and the reconnect path, plus a live hypercorn h2c repro.

loadgen v1.4.8

loadgen v1.4.8

Robustness + measurement-integrity fixes surfaced by the v3.9 celeris bench post-mortem.

Changes

• Dial backoff + never-connected fail-fast. Failed connects now use bounded jittered exponential backoff; a cell that never establishes a stream aborts fast (instead of a ~386k-dials/sec hot loop against a dead server). Kills the 34.7M-error dead-port cells.
• Calibrated warmup→measure handoff. The full worker set warms up and hands the converged rate into the measured window, so saturation cells start at the knee with zero t=0 error burst — no samples discarded.
• Typed ConnectErrors counter (dial/handshake failures) distinct from the total, surfaced in Result + per-second timeseries.
• Warmup visibility: warmup request/error/connect-error counts in Result + a stderr summary.
• Version self-report fixed (was a stale 1.4.5 constant).

v1.4.7 — celeris v1.4.15

Changes

• Bump github.com/goceleris/celeris v1.4.14 → v1.4.15 (io_uring churn-close memory-leak fix, celeris#314).

loadgen uses celeris only as a library for its integration testserver; no behavioral change to the load generator. Verified: build + vet + go test -race ./... + the live-celeris integration matrix.

Full Changelog: v1.4.6...v1.4.7

v1.4.6 — celeris v1.4.14 + Go 1.26.4 toolchain

Changes

• Bump github.com/goceleris/celeris v1.4.10 → v1.4.14 (HTTP/2 connection-level send flow-control fix + stream-accounting fixes from v1.4.13, metrics dep refresh from v1.4.14).
• Toolchain: Go 1.26.4 (go.mod + CI), golangci-lint v2.12 — aligns the celeris ecosystem on one toolchain.
• Test robustness: widen the per-bucket error-accounting tolerance in TestBenchmarkerTimeseriesP99AndErrors to remove a load-sensitive -race flake (approximation bound only).

Full Changelog: v1.4.5...v1.4.6

v1.4.5 — streaming load modes (WS/SSE) + per-bucket time-series

Extends loadgen beyond plain request/response, all driven through the existing Client interface (worker/latency/timeseries pipeline reused unchanged).

Added

• Streaming load modes — Config.Mode + -mode CLI flag:
  • ws-echo / ws-large-echo / ws-hub — RFC 6455 WebSocket client (masked client frames, full extended-length encoding; ws-large-echo round-trips 64 KiB via the 64-bit length path; ws-hub blocks per call until the server's next broadcast).
  • sse-fanout — SSE client; each delivered data: event = one request (inter-event delivery latency).
  • Mutually exclusive with HTTP2 / Mix / H2CUpgrade.
• Per-bucket P99 + errors in the time-series — each TimeseriesPoint now carries p99_ms (windowed HDR histogram, reset every 1 s) and errors (per-window delta) beside t/rps. The cumulative histogram and final Result.Latency percentiles are unchanged — the window histogram is additive.
• Version constant (1.4.5) stamped into Result.LoadgenVersion.

Binaries

linux/darwin × amd64/arm64, built CGO_ENABLED=0 -trimpath and attached below.

Consumed by goceleris/probatorium for the WS/SSE benchmark scenarios and over-run tail-latency charts.

v1.4.4 — rated mode + HdrHistogram + federation + recv-Q + CPU sampler

Wave 11 release. All four v1.4.4 milestone issues land in this version.

Added

• Rated mode (-rate <constant>, Config.RateRPS) — scheduled-send benchmarker with Gil-Tene-style coordinated-omission correction. Latency interval starts at the intended dispatch deadline. Saturation mode (-rate 0) unchanged.
  Closes #48.

• HdrHistogram emission — Result.Latency.Histogram carries a V2-compressed payload of the merged per-worker hdrhistograms. Existing P50/P75/.../P9999 summary fields preserved. Downstream tools (probatorium, perfmatrix) can re-merge across hosts/arches/git refs without re-running.
  Closes #49.

• EPOLLIN backlog probe (-recvq-probe) — every 5 s, sample /proc/<pid>/net/sockstat plus per-conn recv-Q; latch Result.LoadgenRecvQHigh when median exceeds 64 KB sustained 10 s. Linux-only; Darwin/Windows stub keeps the field zero.
  Closes #50.

• 1 Hz self-CPU sampler (-cpu-monitor) — Result.CpuPctP95 exposes 95th-percentile loadgen CPU for "loadgen-bound" detection.

• Multi-host loadgen federation (-peer, -sidecar) — coordinator + sidecar mode; per-host HdrHistograms streamed back and merged client-side; one unified Result for the whole fleet.
  Closes #51.

Changed

• cmd/loadgen/main.go exposes -rate, -peer, -sidecar, -out, -cpu-monitor, -recvq-probe directly.
• go.mod bumped: github.com/goceleris/celeris v1.4.1 → v1.4.3, golang.org/x/net v0.53.0 → v0.54.0 (supersedes dependabot #47).
• golang.org/x/net/http2/h2c deprecation warnings suppressed via //nolint:staticcheck on the three test scaffolds that depend on it as a reference implementation; migration to http.Server.Protocols tracked for v1.5.0.

Compatibility

• Wire-compatible with celeris ≥ v1.4.3.
• Config / Result shapes are additive — old fields preserved.
• Probatorium consumes this release directly. celeris v1.4.4 will bump its test/perfmatrix submodule to v1.4.4 via goceleris/celeris#275.

Full Changelog: v1.4.3...v1.4.4

v1.4.3

chore(deps): bump go.mod + CI Go to 1.26.3 to absorb stdlib CVE fixes (#45)

v1.4.2

Revert of v1.4.0 Unblocker (#42) and v1.4.1 h1 Close fix (#43) — both introduced subtler problems than they solved. Restores the pre-v1.4.0 Run/warmup contract.

v1.4.1

fix(h1): Unblock now closes the conn rather than SetDeadline (#43)