I put OpenTelemetry + SigNoz on my 6-microservice Spring Boot app — and my 6GB laptop nearly gave up

A single click in my app — "create ticket" — quietly touches four services and crosses a Kafka topic before anything reaches the user's inbox. When that click got slow, I had no idea which hop to blame. Running docker logs on six containers, one at a time, is not observability. It's guessing with extra steps.

So over one long weekend I finally fixed it: I added real distributed tracing to my microservices with SigNoz and OpenTelemetry — without touching a single line of my application code. Here's exactly how it went, including the part where my laptop tapped out.

The app: six services, zero visibility

The project is a side build of mine called TicketFlow — a small support-ticket system, but structured the "real" way:

• an API Gateway (JWT auth + routing)
• Eureka for service discovery
• Auth, User, and Ticket services, each with its own Postgres database
• a Notification service that emails users
• an Angular 17 frontend
• Kafka connecting the ticket flow to notifications

It looks clean in a diagram. But the "create ticket" request travels Gateway → Ticket service → Kafka → Notification service, and I had no way to actually see that journey. If it took four seconds, I couldn't tell you where those seconds went.

Why the OpenTelemetry Java agent

The thing that sold me was the OpenTelemetry Java agent. You attach one .jar with -javaagent, and it auto-instruments Spring MVC, Spring Cloud Gateway, JDBC (Postgres), and — the part I cared about most — Kafka producers and consumers. No SDK wiring in every service, no annotations. For six services, that's the difference between an evening and a week.

SigNoz is the backend that receives it all and lets you look at your traces, metrics, and logs in one place.

Self-hosting SigNoz (in WSL, the right way)

I'm on Windows, so everything runs in WSL2 (Ubuntu). First lesson, learned the annoying way: run Docker Engine inside WSL, not Docker Desktop. SigNoz's ClickHouse Keeper doesn't play nice with the Desktop backend on Windows.

Second lesson: the old install.sh most tutorials mention is deprecated now. The current self-host path is Foundry:

curl -fsSL https://signoz.io/foundry.sh | bash

A tiny casting.yaml:

apiVersion: v1alpha1
kind: Installation
metadata:
  name: signoz
spec:
  deployment:
    flavor: compose
    mode: docker

Then:

foundryctl cast -f casting.yaml

A few minutes of image pulls later, SigNoz was live on http://localhost:8080, with its collector listening on :4317.

SigNoz up and running — logs, traces, and metrics ingestion all active.

Instrumenting all six services with zero code changes

Because my services run as containers, I didn't open a single Java file. I downloaded the agent once into the project:

mkdir -p otel
wget -O otel/opentelemetry-javaagent.jar \
  https://github.com/open-telemetry/opentelemetry-java-instrumentation/releases/latest/download/opentelemetry-javaagent.jar

Then, in docker-compose.yml, I merged this into each Java service's existing environment: block, changing OTEL_SERVICE_NAME for each one:

    environment:
      # ...my existing env vars...
      JAVA_TOOL_OPTIONS: "-javaagent:/otel/opentelemetry-javaagent.jar"
      OTEL_SERVICE_NAME: "ticket-service"
      OTEL_EXPORTER_OTLP_ENDPOINT: "http://host.docker.internal:4317"
      OTEL_EXPORTER_OTLP_PROTOCOL: "grpc"
      OTEL_LOGS_EXPORTER: "otlp"
    volumes:
      - ./otel/opentelemetry-javaagent.jar:/otel/opentelemetry-javaagent.jar
    extra_hosts:
      - "host.docker.internal:host-gateway"

That extra_hosts line is the bit that trips people up: my app containers sit on their own Docker network, but SigNoz's collector is published on the host at :4317. host.docker.internal:host-gateway is how a WSL container reaches back to the host.

One conflict I hit immediately: my API Gateway published host port 8080 — the exact port SigNoz's UI uses. They can't both bind it, so the gateway refused to start. One-line fix: map the gateway to 8090:8080 instead. Internal routing (Angular → gateway over the Docker network) uses container names, so nothing else broke — I just hit the gateway at localhost:8090 after that.

The part where my laptop said no

First docker compose up --build, and the Angular build died on a TypeScript error — Cannot find name 'BuiltinIteratorReturn', a newer @types/node wanting a newer TypeScript than my project pins. A failed build aborts the whole up, so nothing started. I pinned @types/node and set skipLibCheck: true to get moving.

Then the real wall. I tried to boot all nine JVMs at once plus SigNoz's ClickHouse stack — on a laptop with about 6 GB of RAM total, where WSL was quietly capped at ~2.8 GB. The VM started swapping, then swap-died, and froze so hard the terminal stopped responding.

Not gonna lie — terminal frozen, laptop fans screaming, well past midnight — I seriously thought about giving up and going back to grepping logs the old way. But quitting over a memory setting felt too dumb, so I pushed on.

The fix was three things:

1. Heap caps — -Xmx256m on each app's JAVA_TOOL_OPTIONS, plus modest KAFKA_HEAP_OPTS for Kafka and Zookeeper.
2. More room for WSL — a C:\Users\<me>\.wslconfig file:

   [wsl2]
   memory=4GB
   swap=6GB

then wsl --shutdown to apply it.

1. A staggered bring-up instead of a boot storm:

   docker compose up -d postgres zookeeper kafka
   docker compose up -d eureka-server          # wait ~30s until healthy
   docker compose up -d auth-service user-service ticket-service notification-service api-gateway

After that, all five Spring Boot apps started, registered with Eureka, and the gateway reported {"status":"UP"}.

The payoff: watching a request cross Kafka

I generated some real traffic through the gateway — register, login, create a ticket — and opened SigNoz. All five services showed up right away.

All five services reporting in. Notice the multi-second P99 latency — that was my first hint the environment, not the code, was the bottleneck.

Then came the moment that made the whole exercise worth it — the trace for one create-ticket call.

I'd sketched this exact request flow on paper more times than I can count. Seeing it actually render — every hop in order, with real timings attached — was oddly satisfying, like the whiteboard diagram had finally come to life.

One trace, ten spans: api-gateway → ticket-service → INSERT ticket_db → ticket-events publish (Kafka producer) → ticket-events process on notification-service (Kafka consumer).

This is the feature I like most in SigNoz: the trace context propagates across the Kafka topic. The producer span in the ticket service and the consumer span in the notification service are stitched into the same trace — even though those two services never call each other and only meet through a topic named ticket-events. Two fully decoupled services, one continuous story. And the Postgres INSERT and SELECT calls sit right there as child spans, so I can see the DB time too.

There was a bonus insight hiding in the timings. The API call itself returned in about 3.6 seconds — but the trace didn't stop there. The notification service picked up the Kafka event and kept working for another ~10 seconds, long after the user already had their response. Because SigNoz keeps both halves in a single trace, I could actually see that asynchronous background work happening — something that's completely invisible if you only stare at the API's response time. That's the whole promise of Kafka — fire an event and move on — finally made visible.

It flagged failures cleanly too. A mistyped login showed up as its own trace, marked 401, with the error landing exactly on the auth service's Postgres lookup — no log-diving required.

A failed login, caught as a 401 — SigNoz pointed straight at where it broke.

What the traces actually taught me

Here's the twist. My requests were slow — multi-second spans everywhere. My first instinct was "my code is doing something dumb." But the traces told a different story: the time wasn't in my business logic. It was the environment — memory pressure and swapping dragging everything down. On a properly-resourced box, those same spans would collapse to milliseconds.

That's the real lesson of distributed tracing: it doesn't lie about where time goes. Without it, I'd have spent an evening "optimizing" ticket-service code that was never the problem. With it, one look at the waterfall pointed straight at the actual bottleneck.

A few things I'm taking away:

• The OpenTelemetry Java agent is the fastest way to instrument an existing Spring Boot fleet — Spring, JDBC, and Kafka, all with zero code changes.
• Kafka context propagation turns a pile of decoupled services back into one readable request flow. In an event-driven system, this alone justifies tracing.
• Self-hosting SigNoz is realistic, but budget your RAM honestly. On a small machine, cap the JVM heaps and give WSL room, or it will swap-die.
• Observability is a mindset shift: stop guessing where the time went, and go look.

If you're building anything where more than two services talk to each other, this is worth an afternoon. Mine fought back a little because of the hardware — but watching that one trace stitch itself across Kafka made it completely worth it.

---

Built with Spring Boot, Kafka, Angular, PostgreSQL, OpenTelemetry, and SigNoz.