From dbf62aef47075ee173ba601dd4096a417f1a4c75 Mon Sep 17 00:00:00 2001 From: codex Date: Mon, 19 Oct 2026 15:20:08 +0200 Subject: [PATCH] Record capped synthetic sizing comparison and remaining coverage --- ...sured-workload-sizing-and-weekly-review.md | 25 +++++++++++++++++++ 1 file changed, 25 insertions(+) diff --git a/workplans/CUST-WP-0071-measured-workload-sizing-and-weekly-review.md b/workplans/CUST-WP-0071-measured-workload-sizing-and-weekly-review.md index c17c7982..66943dae 100644 --- a/workplans/CUST-WP-0071-measured-workload-sizing-and-weekly-review.md +++ b/workplans/CUST-WP-0071-measured-workload-sizing-and-weekly-review.md @@ -468,3 +468,28 @@ that scenario is rejected as an allocation, not used to justify procurement. The cross-owner case is retained at `docs/changes/CUST-WP-0071/resource-control-review-handoff.json`, progress receipt `381ba7f7-35c6-49f4-8805-4503df060ccd`. No worker-injection service was enabled. + +## Existing-limit synthetic comparison — 2026-10-19 + +Vergabe `bfd1a2d` retains `container-comparison.json`, compressed raw samples, +the exact executed driver and the failed preflight beside the earlier evidence. +The isolated application container was inspected at 1000m CPU and 1Gi memory. +Across 180 seconds with the same two-user mix, 1552 ordinary operations and 516 +transfers completed without error. Ordinary p95 was 412ms; transfer p95 was +158ms. Measurement coverage was complete and no OOM kill occurred. Ten focused +tests cover failure cleanup and sampler failure; gaps now prevent conclusions. + +Mean application CPU was 940m; 1449 of 1807 CPU periods were throttled. Sampled +p95 of 1037m reflects sampling-period boundaries, not permission above the quota. +Peak cgroup memory was 426MiB including page cache, which is not comparable to +the earlier summed RSS. Separately sampled database CPU p95 was 60.63% of one +core. The first capped preflight failed before load because an internal Docker +bridge blocked its host-published database; cleanup and the corrected dedicated +bridge are retained, without claiming egress isolation. + +This meets the latency target for the tested local mix at the existing limit; +it supports no CPU-limit increase or lower memory limit. The 60m request remains +unvalidated under shared contention. T02 stays progress while the owner extends +the same configurable harness with edits, startup/restart and burst phases. +Actual deployed-release/hardware and shared-service observations remain required +for the final allocation review; attended human acceptance remains separate.