Load-Test Report — Telemetry Ingest

Status: measured (local harness)
Date: 2026-07-11T04:42:12.899Z
Target: http://127.0.0.1:3101/api/v1/telemetry/ingest
Topology: local spawn — STORE_BACKEND=memory, EVENT_BUS=memory, no Postgres/Redis (ingest + in-process pipeline)

Setup

Knob	Value
Tool	autocannon
Target rate	100 req/s
Duration	30s
Connections	20
Auth	RS256 partner JWT per request (unique jti + event_id)
Rate limits	Raised via INGEST_*_LIMIT / PREAUTH_IP_LIMIT for the soak

Results

Metric	Value
Requests sent	3000
Throughput (avg)	100.0 req/s
Latency p50	13.0 ms
Latency p90	29.0 ms
Latency p99	64.0 ms
Latency mean	15.2 ms
Latency max	122.0 ms
2xx	3000
non-2xx	0
errors / timeouts	0

Verdict

Pass — sub-second p99 under ~100 req/s on this topology. “Sub-second edge extraction” remains an SDK claim; this report measures gateway ingest + in-process pipeline latency.

What this does / does not prove

Proves	Does not prove
Gateway accept path under sustained load	On-device PDF/OCR/NER latency
JWT verify + pipeline + SSE fan-out (memory)	Multi-replica Redis Streams + Postgres + outbox under load
Rate-limit config can be raised for soak	Production Railway multi-AZ numbers

Re-run commands

# Baseline (memory topology — published numbers above)
pnpm --filter @waspada/gateway-api load-test

# Inline FMS delivery against a local echo server (no live bank webhook)
pnpm --filter @waspada/gateway-api load-test -- --mock-fms

# Optional: force mock FMS to return 503 at 20% to exercise retry/backoff
LOAD_TEST_MOCK_FMS_FAIL_RATE=0.2 pnpm --filter @waspada/gateway-api load-test -- --mock-fms

# Durable path (Redis + Postgres + outbox) — point at a staged gateway
LOAD_TEST_URL=https://… pnpm --filter @waspada/gateway-api load-test

Raw

See last-run.json in this directory for the full autocannon result object.