till/managarten
1
0
Fork 0
mirror of https://github.com/Memo-2023/mana-monorepo.git synced 2026-05-14 22:21:10 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4
managarten/load-tests
History
Till JS 7f6b41654e test(load): k6 script for the unified apps/api server 
The pre-launch consolidation collapsed 17+ per-product backends into
the single Hono/Bun process at apps/api. That makes apps/api the
single point of failure for every authenticated module call the
unified Mana web app makes — a missing index, a hot-path allocation
in auth middleware, or rate-limiter contention degrades all 16
modules at once. The other scripts in load-tests/ already cover
mana-auth, mana-sync, mana-llm and the SvelteKit frontends, but
apps/api itself was unmeasured. This is that missing piece.

What it tests
-------------
A weighted mixed workload that walks the full middleware stack
(CORS → request logger → rate limit → auth → router → handler)
plus a representative range of handler shapes:

  25%  GET /health                            (no auth, baseline)
  20%  GET /api/v1/moodlit/presets            (auth + in-memory return)
  15%  GET /api/v1/chat/models                (auth + DB read)
  20%  POST /api/v1/calendar/events/expand    (auth + Zod + RRULE compute)
  12%  POST /api/v1/todo/compute/next-occurrence
                                              (auth + Zod + rrule lib)
   8%  POST /api/v1/todo/compute/validate     (auth + Zod + validation)

Deliberately no write endpoints — those would conflate write
amplification with API-server cost. The compute routes here all run
in <50ms warm; what we're measuring is the overhead the unified
server adds on top of pure handler work.

Per-route-class p95 budgets via tags:

  health      < 100ms
  authed_get  < 300ms
  authed_post < 500ms
  global      p95 < 500ms, p99 < 2s

Application-level error rate (4xx + 5xx + check failures) must stay
under 1% — exit code 1 otherwise, so it's CI-gateable.

Auth setup
----------
apps/api requires JWT on every /api/* route. setup() acquires a
token once before VUs start hammering and shares it for the run.
Three sources tried in order:

  1. $MANA_API_TOKEN  (CI passes a pre-minted token)
  2. login at $TEST_EMAIL / $TEST_PASSWORD
  3. register a fresh account on the fly

Bails with a clear error message if all three fail.

Load profile
------------
4 minute total: 30s warmup → 2m sustained @ 50 VUs → 1m peak @ 100 VUs
→ 30s cooldown. Override with --vus / --duration as usual.

Closes item #23 in docs/REFACTORING_AUDIT_2026_04.md.

Follow-ups not in this commit:
  - Wire into .github/workflows/daily-tests.yml (requires standing
    up the apps/api stack in the runner — bigger lift, separate PR)
  - Per-module thresholds once we have a few real runs and know
    where the natural baseline sits

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 12:27:54 +02:00
..
api.js test(load): k6 script for the unified apps/api server 2026-04-09 12:27:54 +02:00
auth-api.js feat: GPU offload, signup limit, load tests & capacity planning 2026-03-28 21:14:24 +01:00
llm-ollama.js feat: GPU offload, signup limit, load tests & capacity planning 2026-03-28 21:14:24 +01:00
README.md test(load): k6 script for the unified apps/api server 2026-04-09 12:27:54 +02:00
sync-websocket.js feat: GPU offload, signup limit, load tests & capacity planning 2026-03-28 21:14:24 +01:00
web-apps.js chore: rename mukke to music in infra, scripts, and CI/CD 2026-04-05 16:47:57 +02:00

README.md

Load Tests

k6-basierte Load Tests fuer die Mana-Infrastruktur.

Setup

# k6 installieren (macOS)
brew install k6

# WebSocket-Extension (fuer Sync-Tests)
# k6 hat WebSocket-Support eingebaut

Tests ausfuehren

# Gegen lokale Umgebung
k6 run load-tests/web-apps.js
k6 run load-tests/auth-api.js
k6 run load-tests/api.js
k6 run load-tests/sync-websocket.js
k6 run load-tests/llm-ollama.js

# api.js braucht ein gültiges JWT — entweder via $MANA_API_TOKEN
# oder es loggt sich mit den TEST_EMAIL/TEST_PASSWORD env vars ein
# (default: loadtest-api@mana.test / LoadTestApi123!).
k6 run -e MANA_API_TOKEN=eyJhbGc... load-tests/api.js

# Gegen Produktion (vorsichtig!)
k6 run -e BASE_URL=https://mana.how load-tests/web-apps.js

# Mit mehr/weniger Last
k6 run --vus 100 --duration 5m load-tests/web-apps.js

# JSON-Output fuer Grafana
k6 run --out json=results.json load-tests/web-apps.js

Test-Szenarien

Script Ziel Default VUs Dauer
web-apps.js SvelteKit Frontends (HTML-Responses) 10→50→10 5 min
auth-api.js Login, Register, Token Validation 5→20→5 4 min
api.js Unified apps/api Hono server (16 Module) — gemixte Workload mit Auth, Compute & Validation 10→50→100→0 4 min
sync-websocket.js mana-sync WebSocket Connections 10→30→10 5 min
llm-ollama.js Ollama Chat Completions 1→3→1 3 min

api.js Thresholds

Pro Route-Klasse hat das Script eigene p95-Budgets über tags:

Klasse Endpoints p95-Budget
health GET /health < 100ms
authed_get GET /api/v1/moodlit/presets, GET /api/v1/chat/models < 300ms
authed_post POST /api/v1/calendar/events/expand, POST /api/v1/todo/compute/* < 500ms
Global alle Requests aggregiert p95 < 500ms, p99 < 2s

Application-level error rate (4xx + 5xx + Check-Failures) muss unter 1% bleiben, sonst exit-code 1 → CI-Build bricht.

Metriken interpretieren

Metrik Gut Akzeptabel Schlecht
http_req_duration (p95) < 200ms < 1s > 2s
http_req_failed 0% < 1% > 5%
ws_connecting (p95) < 100ms < 500ms > 1s
iterations Steigend Stabil Fallend

Monitoring waehrend Tests

Grafana-Dashboard auf http://localhost:8080 (oder https://grafana.mana.how) beobachten:

  • Container CPU/RAM (cAdvisor)
  • PostgreSQL Connections
  • Redis Commands/sec
  • Netzwerk-Throughput