coulomb/resource-control
5
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Review provisional Vergabe sizing against current node budget

Browse source 
Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a15374-f59f-7a31-86d2-d23b5d2ed181
This commit is contained in:
tegwick 2026-10-19 15:07:20 +02:00
parent 9c57b0d96b
commit 565b16e2d0
4 changed files with 118 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

41
docs/evidence/2026-10-19-custodian-allocation-review.md Normal file
View file

@ -0,0 +1,41 @@
# Initial allocation review — 2026-10-19
Provider: resource-control; beneficiary CUST-WP-0071-T03; receiving record
RESOURCE-WP-0008. This is an initial review, not an approved resize.
Sources: Custodian `8e67cfb`, `docs/evidence/2026-10-19-allocation-reconcile.json`
and its provenance; Vergabe `11f7c8a`,
`docs/evidence/2026-10-19-synthetic-sizing/`. Kubernetes target identity was
verified before the read-only capture at 12:50:41Z. A separate node description
agrees with 3605m scheduled CPU requests and 9856Mi memory requests.
| Reservation scenario | Scheduled CPU | Residual against 4000m |
| --- | ---: | ---: |
| Current snapshot, including Vergabe60m | 3605m | 395m |
| Replace Vergabe60m with experimental1950m | 5495m | -1495m |
The second row is an infeasible reservation scenario, not a proposal to raise
capacity or limits. Host and Kubernetes CPUs are the same four vCPUs and are
counted once. The 395m remainder is not guaranteed execution or rollout room.
STATE-WP-0091's passing preflight models its own release sequence only.
The synthetic run used two users/two workers,100 initial tenders,1MiB files,
0.2-second cycle think time,30s idle and180.889s work on an unconstrained local
Ryzen3500U host. It completed3028 operations without error, ordinary p95 227ms.
Aggregate app CPU p95 was1298m and sampled peak1917m. Tunable margins derive the
1950m request/2900m limit envelope; this is hardware/workload-specific and exceeds
the current production1000m CPU limit. Peak summed RSS194MiB includes shared-page
double counting; the calculated256Mi request/320Mi limit does not justify lowering
the existing1Gi limit without startup/restart and target-runtime evidence.
Current disposition: do not apply the unconstrained envelope. Retain the existing
explicitly accepted prototype settings pending the isolated1000m/1Gi comparison
and owner review. Passing a quota test would not establish that60m request protects
latency under production contention. Database demand remains a separate consumer.
No financial attribution, procurement or new spending decision follows.
Remaining budget inputs: thirteen grouped zero-request workloads; unknown namespace
owners for mas/playwright/rapp-qonto; representative shared-service and build-runner
demand; host/rollout margin; current contention and throttle evidence. Historical
portfolio utilization remains dated2026-08-11 and is not silently refreshed by this
reservation snapshot. RESOURCE-WP-0008 keeps the allocation review active.