Event pipelines, rate limits and health tracking with honest failure modes. When a dependency goes down, callers know exactly where they stand. Personal study of the Open Gateway telemetry problem. Not Telefónica code. Not production traffic.
~90ADAPTERS ON THE WATCHLIST
4COUNTRIES REPORTING
3×DOWN → 1ALERT PER EPISODE, PINNED BY TESTS
100%IDEMPOTENT INSERTS AT THE HUB
System diagrams
FIG. 1 — KAFKA PIPELINE: gateway → Kafka → hub → Oracle, with a live SSE dashboard tapped off the hub.FIG. 2 — TOKEN BUCKET: every quota decision is one atomic operation; if Redis falls over, the local fallback answers.
The 45-second demo
3×DOWN → 1 alert. Personal study of the Open Gateway telemetry problem — not Telefónica code, not production traffic.
The floor’s work
F22026
Kafka Adapter Telemetry
Event-driven adapter telemetry: a Kafka gateway + hub pipeline with idempotent Oracle persistence, per-adapter health tracking, alerting, and a live SSE mission-control dashboard.
The problem. Fleets of ~90 API adapters across 4 countries need operational visibility: who is DOWN, when did it start, and what already alerted. Built as a personal study of the Open Gateway telemetry problem, end to end.
Highlights
Hexagonal per service: gateway publishes transit profiles (low/moderate/high/overload), hub persists idempotently by event id.
Alert rule pinned by tests: 3 consecutive DOWN → exactly 1 alert per episode.
DLT with retries absorbs duplicate and corrupted JSON without blocking the lane.
demo.sh proves the whole loop from zero: idempotent counts, single alert, DLT routing.
10/10 tasks delivered with Spec-Driven Development; 3 ADRs; diagrams auto-published to GitHub Pages in CI.
Rate-limiting toolkit for LLM gateways: atomic token-bucket quotas, pluggable stores, honest failure modes — with a Javalin gateway demo exposing an OpenAI-style endpoint.
The problem. LLM inference is expensive: unbounded traffic is a DoS on your own wallet. Most limiters race under concurrency or hard-fail when Redis does.
Highlights
Every quota decision computed atomically — never a read-modify-write that slips a token past the limit.
On Redis failure: graceful local fallback + loud metrics, not mass 5xx.
Contract test suite shared by in-memory and Redis implementations, plus concurrency and parity tests.
Standard X-RateLimit-* headers and correct 429 + Retry-After.