Capacity Profile — OS-v1.0
Last verified: 2026-09-12. Engineering targets, NOT CERTIFIED CAPACITY / SLA. Owner: QA/platform. DEC-01/07/10/11 dan OP-01 harus resolved for final profile. Acuan: baseline, acceptance AC-22.
1. Workload wajib
| Parameter | Target profile |
|---|---|
| Concurrent | 500 active attempts bersamaan; bukan sekadar 500 registrations/jobs |
| Start burst | 500 authorized starts dalam 60 detik |
| Durasi | 120 menit per full run; dua full runs pada exact candidate |
| Questions | 100 mixed: 80 single-choice, 10 numeric, 10 essay |
| Autosave | Realistic edit/debounce 3–5s dengan per-question revisions dan idempotency; payload sesuai fixture |
| Heartbeat | 30s per client sebagai baseline load; actual policy dicatat |
| Submit | 500 submits dalam 30s; tambahan deadline auto-submit/reconnect wave |
| Headroom | 750 concurrent selama 15 menit, ukur saturation bukan janji kapasitas komersial |
| Kamera | Pre-exam + periodic/event still sesuai approved DEC-07, preview local, upload retry/failure |
| Proctor | Room command center fan-out, reconnect rebuild dan evidence access sesuai DEC-10 |
| Background | Finalization/export/retention/outbox contention diisolasi dari participant commands |
Actual event boleh berbeda, tetapi perubahan profile certification memerlukan persetujuan dan evidence baru. Windows camera/Zoom rehearsal diperlukan selain HTTP load.
2. SLO engineering
Autosave p95 ≤1s, p99 ≤3s; unexpected server 5xx <0.1%; nol ACK loss pada API/worker/Redis restart; nol duplicate authoritative receipt/grade akibat concurrent submission. Ukur validation/conflict/rate-limit/error secara terpisah, tidak menghapus 409/429/drop iterations dari laporan agar tampak hijau.
RPO ≤5m/RTO ≤60m pada site disaster adalah target berbeda, sesuai D09. Jangan mengklaim zero acknowledged loss untuk total infrastructure loss jika desain recovery tidak membuktikannya.
3. Metode pembuktian
Real HTTP melalui staging TLS/proxy/API/DB/workers/storage terisolasi, bukan in-process TestClient/thread jobs. Unique credentials/attempts, session windows dan form snapshots valid. Jangan loop peserta yang sudah submitted sebagai substitute active load.
Simpan exact FE/BE/migration/launcher/SEB config/policy/fixture identities; load generator capacity dan clocks; configured/actual VUs, starts/active duration, request counts/errors/timeouts/dropped iterations, latency per endpoint, DB locks/pool/IO, storage bytes/errors, worker backlog, ACK/receipt reconciliation dan raw result artifact.
Run start/save/reconnect/deadline/submit profiles, camera burst/failure retry storm, Redis/API/worker restart dan DB contention. Isolated load environment tidak boleh menghabiskan resource host Scola; topology/quota/exact targets disetujui sebelum tes.
4. Camera budget worksheet
Dengan N cameras, mean interval T detik, mean file B bytes dan K stills/participant: - Mean periodic uploads/s ≈ N/T; event/pre-exam/retry bursts dihitung terpisah. - Payload bandwidth ≈ N×B/T bytes/s, belum termasuk HTTP/TLS overhead. - Raw storage per event ≈ N×K×B; tambahkan version/thumbnail/export/retention overhead aktual. - Queue/client retry buffer harus bounded; saat outage jangan membanjiri answer connection pool.
Nilai T/B/K/TTL, worker pools dan minimum laptop ditetapkan dari approved policy/spike; tidak menggunakan bandwidth 100kbps atau pool 50+50 sebagai angka default universal. Zoom traffic terjadi di luar API tetapi berbagi akses internet peserta dan harus diuji operasional.
5. Kriteria gagal
ACK hilang, wrong scope/form, late answer diterima, duplicate receipt/grade, false saved UI, camera failure mengubah timer, missing evidence pada required capture policy tanpa incident handling, atau rollback/restore unsafe = FAIL meskipun latency memenuhi target.