Load Testing

k6 scripts for smoke, sustained load, and stress testing — targeting the governance pipeline, dashboard, and escalation endpoints. Integrated into the deployment pipeline as a gate between staging and production.

Three Test Profiles

Smoke

1 virtual user, 1 iteration. Validates every endpoint responds correctly — health, pipeline (all verdict paths), shadow mode, dashboard, and escalation. Any failure aborts. Runs first in CI.

Load

Three concurrent scenarios: pipeline submissions ramping to 8 req/s, dashboard polling with 10 virtual users, and escalation queries at 2 virtual users. Measures p95 latency under sustained realistic traffic for 7.5 minutes.

Stress

Ramps pipeline traffic to 50 req/s (100 VUs) and dashboard to 80 VUs. Finds the breaking point, measures when 429s appear, and verifies the system recovers after load drops. Expects degradation but not collapse.

Running Locally

Install k6, then start the API in open mode:

# Start API (no auth, no rate limiting)
GAAS_OPEN_MODE=true GAAS_RATE_LIMIT_ENABLED=false \
  uvicorn packages.api.main:app

# Run smoke test
k6 run tests/load/smoke.js

# Run load test with JSON export
k6 run tests/load/load.js --out json=results.json

# Run stress test
k6 run tests/load/stress.js

Override the target URL for remote environments:

K6_BASE_URL=https://staging.example.com K6_API_KEY=gsk_... \
  k6 run tests/load/smoke.js

Payload Mix

The load and stress tests select intents from a weighted random pool that exercises all verdict paths through the governance pipeline:

Scenario	Pipeline Path	Weight
Clean low-risk	Full pipeline → APPROVE	50%
PCI regulated data	Policy fast-fail → BLOCK	15%
High financial ($75k)	Triggers deliberation → ESCALATE	15%
Minor data (COPPA)	Policy fast-fail → BLOCK	10%
Audit tampering	Policy fast-fail → BLOCK	10%

Deliberation coverage. The high-financial scenario triggers multi-agent deliberation — the most expensive code path and the primary target for latency measurement.

Pass/Fail Thresholds

Endpoint Group	Metric	Threshold
Pipeline (live & shadow)	p95 latency	< 2,000ms
Dashboard	p95 latency	< 500ms
Escalation	p95 latency	< 500ms
Health / Metrics	p95 latency	< 200ms
Global	Error rate	< 5%
Stress (pipeline)	Success rate	≥ 70%

Custom Metrics

Beyond standard k6 HTTP metrics, the load tests track governance-specific measurements:

Metric	Type	Source
`gaas_pipeline_latency_ms`	Trend	Extracted from `X-GaaS-Pipeline-Latency-Ms` response header — server-side pipeline execution time
`gaas_verdicts_approve`	Rate	Fraction of decisions that result in approve or approve_modified
`gaas_error_rate`	Rate	Application-level errors (non-200 on expected-success requests)
`gaas_rate_limit_429s`	Counter	Rate-limit rejections during stress tests

CI Integration

A k6 smoke run executes automatically as part of the staging gate. Every merge to main deploys to a real staging environment; the gate must pass before that build becomes eligible for promotion to production. The GitHub Actions workflow:

Builds both container images and pushes them to Amazon ECR, addressed by digest
Runs database migrations and deploys those exact digests to the staging cluster
Runs the smoke suite against the deployed service, then a k6 smoke run (1 virtual user, 1 iteration)
Tags the digests as promotable only if every check passes

The gate runs a smoke check, not a load test. The heavier load and stress profiles described on this page are run deliberately, not on every merge. What the gate runs is a single-iteration k6 smoke pass — its value is that it exercises the real system, not that it applies load. It runs against the deployed service with authentication and rate limiting on: k6 targets staging through the load balancer using a real organisation API key, so a run exercises the ALB, the container, Postgres and the full authentication path — not a throwaway container with its protections switched off. Correctness thresholds (request-failure rate and check success) are identical to production and are never relaxed; only latency ceilings differ, because staging runs on half of production's CPU and memory.

Promotion never rebuilds. Production receives the byte-identical image digests that passed the staging gate. A rebuild would produce different bytes from the ones that were tested, which would make the gate decorative.

Environment Variables

Variable	Default	Description
`K6_BASE_URL`	`http://localhost:8000`	Target API base URL
`K6_API_KEY`	(empty)	API key for authenticated environments
`GAAS_OPEN_MODE`	`false`	Set to `true` on the API server to skip authentication
`GAAS_RATE_LIMIT_ENABLED`	`true`	Set to `false` on the API server to disable rate limiting

Observability & Alerting — Prometheus metrics that k6 results correlate with
Intent Declaration API — the pipeline endpoints under test
Conversational Dashboard — the dashboard endpoints under test