This is a synthetic local comparison of the untouched published Server 2.4.26 Apache image (durableworkflow/server@sha256:d9d13157f91d5418ba74759bc1b32f1ba0fc649e165f6206faab55e0f4737a1e) with the local package-trimmed image (sha256:37a59f394ad300850ce35583e2e5c02f4d767e6fdca3a95dbe5ec70a44639e06) built from the exact Server 2.4.26 source and Workflow 2.2.18. The only Dockerfile change removes Node/Python and removes Git after pinned source verification. Draft PR #276 carries that change and the separate regression-test-container Python correction. No published candidate digest exists yet.
Host: x86_64, Linux 5.15.0-190-generic, Docker 27.5.1, four logical CPUs and 16,618,160,128 bytes RAM. Compose: scripts/perf/standard-workflow-profile.compose.yml plus scripts/perf/standard-workflow-fixed-envelope.compose.yml. Exact backend digests are in pins-baseline.env and pins-candidate.env. HTTP 1 CPU/1 GiB, queue worker 0.75 CPU/768 MiB, MySQL 1 CPU/1.5 GiB, Redis 0.25 CPU/256 MiB, each SDK worker 0.5 CPU/512 MiB, equal RAM and RAM+swap limits. Two SDK workers stayed running with unchanged restart counts in each measured window. The caller was a separate one-shot PHP SDK 2.1.5 container. The full ad hoc collector is run-window.sh; it calls the repository's scripts/perf/standard-workflow-offered.php.
The low window offered 0.5 workflow/s for 30 s (15 planned starts), with up to 60 s bounded drain. The high window offered 1.25 workflow/s for 60 s (75 starts), with up to 120 s bounded drain. Each image got 20 s of verified warmup after a switch. The high runs alternated candidate, published baseline, candidate. The additional low baseline repeats came after the high series, so low memory peaks carry that order effect. Latency is client start request beginning to the Server's closed_at, using the shared host clock. closed_during_offer_window excludes drain; completed_by_observation_deadline includes it. Every counted completion has exact result, completed status and ordered semantic history checked by the workload script. The high 60-request readiness probe series is separate and adds load.
| Load / image | In-window completions, each repeat | Total verified, each repeat | p95 latency range | Runtime CPU seconds range | Completed / runtime CPU-second range | HTTP peak MiB range |
|---|---|---|---|---|---|---|
| 0.5/s published | 14, 14, 14 / 15 | 15 / 15 | 3.107–4.185 s | 47.98–49.06 | 0.306–0.313 | 73.41–93.18 |
| 0.5/s trimmed | 14, 14, 14 / 15 | 15 / 15 | 3.377–3.724 s | 48.64–49.12 | 0.305–0.308 | 73.03–77.66 |
| 1.25/s published | 26, 28, 27 / 75 | 75 / 75 | 65.691–68.538 s | 185.11–188.99 | 0.397–0.405 | 91.54–99.00 |
| 1.25/s trimmed | 26, 27, 26 / 75 | 75 / 75 | 67.116–69.062 s | 186.45–188.45 | 0.398–0.402 | 94.90–95.23 |
Runtime CPU is the change in cgroup usage_usec summed across the six long-lived HTTP, queue, MySQL, Redis and two SDK worker containers from before the one-shot runner starts until it finishes drain and history verification. It excludes the one-shot caller and Docker sampler, and has no idle subtraction. This is a comparison unit, not a customer cost or a saturated-service guarantee. docker stats sampled memory about every two seconds; these observed peaks can miss a shorter peak. All 12 ordinary windows had zero skipped slots, late starts, workflow errors, poll errors, deadline backlog, swap and container restart changes. Under high load the accepted rate was below the 1.25/s offered rate and backlog drained within the bound; do not call 75/60 sustained throughput.
In the separate paired high-load readiness experiment, each image answered 60/60 /api/ready probes with HTTP 200. Published p50/p95/p99 was 1.120/1.376/1.682 s; trimmed was 1.114/1.473/1.586 s. The extra probes reduced in-window completions to 16/75 published and 17/75 trimmed. All 75 completed after drain with verified histories, zero swap/restarts and no failed jobs. The final disposable MySQL state had 710 completed runs, 2,130 completed tasks, no ready/leased tasks and zero failed jobs. Project containers, volumes and network were removed.
results-raw.jsonl preserves each complete workload JSON with its source filename. cgroup-raw.tsv, docker-stats-raw.jsonl and readiness-raw.tsv preserve individual counters/samples. source-sha256.txt checks the original local files. This is local evidence, not a published multi-architecture image qualification. Retain this artifact through 2026-12-28, then delete after the final report contains the reviewed results.