04 Touring the Stack: One Checkout, Traced Through Every Container

Touring the Stack: One Checkout, Traced Through Every Container

Part 4 of a series on observability for microservices. Part 3 built the stack; this post drives it. Every command below is copy-pasteable against the companion repo's stack/ once it's running. Series index.

📦 GitHub: https://github.com/geekchow/O11y-Micro-Service

Reading a pipeline diagram only gets you so far. This post pushes one HTTP request through a live 13-container stack and finds its footprint in every component --- trace, metrics, and logs --- with real commands.

We'll deliberately trigger a declined checkout : any amountCents ending in 7 gets rejected by the payment service. Error traces are ideal tour subjects because the Collector's tail-sampling policy always keeps them --- there's no chance our request gets sampled away before we can find it.

Step 0 --- send the request

bash 复制代码

curl -s -X POST localhost:8080/checkout -H 'Content-Type: application/json' \
  -d '{"cartId":"tour","amountCents":47,"tenant":"acme"}'
# → {"orderId":"<UUID>","status":"declined"}     amount ends in 7 ⇒ declined

Note the orderId in the response --- it's our join key into the logs, from which we'll pull the trace_id.

Step 1 --- what the JVMs just did, invisibly

The five services already did their work without any code watching: Tomcat and RestClient auto-instrumentation created spans, the manual checkout span ran, and a WARN log line fired in both payment and gateway. There's nothing to see inside the services --- by design, that work happens off to the side, buffered and shipped asynchronously so it never blocks the response the customer sees.

Step 2 --- Loki has the trace_id first

bash 复制代码

ORDER=<paste the orderId>
curl -s -G 'localhost:3100/loki/api/v1/query_range' \
  --data-urlencode "query={service_name=\"payment\"} |= \"$ORDER\"" \
  --data-urlencode 'limit=1' \
  | python3 -c "import json,sys; s=json.load(sys.stdin)['data']['result'][0]['stream']; print('trace_id:', s['trace_id'])"

The trace_id comes back as structured metadata , not as text inside the log line --- the OTel Logback appender stamped it because the WARN happened inside the span's active Context. This is the mechanism, not a coincidence: log.warn(...) never mentions trace IDs anywhere in the application code.

Step 3 --- Tempo has the whole story

bash 复制代码

TID=<paste the trace_id>
curl -s "localhost:3200/api/traces/$TID" | python3 -c "
import json,sys
t=json.load(sys.stdin)
for b in t['batches']:
    svc=[a['value']['stringValue'] for a in b['resource']['attributes'] if a['key']=='service.name'][0]
    for ss in b.get('scopeSpans',[]):
        for s in ss['spans']: print(f\"{svc:10s} {s['name']}\")"

Expect spans from four services --- gateway (including the manual checkout span), auth, cart → inventory, and payment --- all under one trace_id. This is context propagation, made visible after the fact. This trace exists in Tempo because the Collector's status_code: ERROR policy matched it; a healthy sibling checkout only has a 25% chance of surviving the same tail-sampling filter in this demo config (1% in a production-like ratio).

Step 4 --- it's already a number in Prometheus

bash 复制代码

open 'http://localhost:9090/graph?g0.expr=sum%20by%20(service_name%2C%20status_code)%20(rate(traces_span_metrics_calls_total%7Bspan_kind%3D%22SPAN_KIND_SERVER%22%7D%5B2m%5D))'

Your checkout is one anonymous increment inside traces_span_metrics_calls_total{status_code="STATUS_CODE_ERROR"}, produced by the spanmetrics connector from the same spans you just saw in Tempo --- but generated before tail sampling ran. Individual identity is gone; the rate is preserved. That's the metrics trade-off, mechanized in one connector.

Step 5 --- watch the pipeline count your spans

bash 复制代码

# agent tier accepted them...
curl -s localhost:8888/metrics | grep -E '^otelcol_receiver_accepted_spans'
# ...gateway tier accepted them, and exported FEWER than it accepted:
curl -s localhost:8889/metrics | grep -E '^otelcol_(receiver_accepted|exporter_sent)_spans'

The gap between the gateway's accepted and sent span counts is the tail sampler discarding healthy traces --- the only place in the whole stack where telemetry is deliberately thrown away. Watch it live at the zpages debug endpoint: http://localhost:55680/debug/pipelinez.

Step 6 --- close the loop in Grafana

Open http://localhost:3000 → Checkout --- RED dashboard:

Error-rate panel --- your decline shows up in the red series (the metric view).
p99 panel → click an exemplar dot --- the trace opens directly (metric → trace pivot).
In the waterfall, click the payment span → "Logs for this span" --- your WARN line appears (trace → log pivot).
In the log line's details, click trace_id → "View trace" --- you're back at the trace (log → trace pivot --- full circle).

None of those three pivots involve a plugin or a glue service. Each one is a single provisioning key in Grafana's datasource config:

Click	Wiring key	What it does
exemplar dot → trace	`exemplarTraceIdDestinations` on the Prometheus datasource	reads the exemplar's `trace_id` label, opens it in Tempo
span → its logs	`tracesToLogsV2.query` on the Tempo datasource	templated Loki query: `{service_name=...}
log line → trace	`derivedFields` on the Loki datasource	lifts the `trace_id` metadata into a View trace link

The pivots exist because every signal carries the same trace_id --- Grafana just needs to be told which field holds it in each store.

What one request became

One HTTP call turned into: roughly ten spans converging in Tempo under a single trace_id, one increment across three Prometheus counters plus an exemplar pointing back at the trace, and two WARN lines in Loki carrying that same trace_id as structured metadata --- while both Collectors visibly counted, enriched, and (for other, healthier traces) culled data along the way. Every container did exactly the one job it owns, and you just watched each of them do it.

The next post takes this further: reproducing a real production incident on purpose, by flipping one environment variable, and watching the entire detection-to-rollback loop fire on your own machine.

➡️ Next: Part 5 --- Reproducing a Production Incident on Purpose