instrument-distributed-tracing
정보
이 스킬은 개발자들이 분산 추적을 위해 OpenTelemetry로 애플리케이션을 계측하는 데 도움을 줍니다. 자동 및 수동 계측, 컨텍스트 전파, Jaeger나 Tempo 같은 백엔드와의 통합을 다룹니다. 지연 시간 문제를 디버깅하고, 마이크로서비스 간 요청 흐름을 이해하며, 추적을 로그 및 메트릭과 연관시키는 데 활용하세요.
빠른 설치
Claude Code
추천npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/instrument-distributed-tracingClaude Code에서 이 명령을 복사하여 붙여넣어 스킬을 설치하세요
문서
Instrument Distributed Tracing
Wire OpenTelemetry. Track requests cross microservices. Find slow spots.
When Use
- Debug latency cross many services
- Follow request flow, see service deps
- Spot slow DB queries, slow API calls inside transaction
- Tie traces to logs and metrics for root cause
- Measure end-to-end latency: user req to response
- Migrate old tracing (Zipkin, Jaeger) to OpenTelemetry
- Prove SLO compliance via latency percentiles
Inputs
- Required: List of services to instrument (languages, frameworks)
- Required: Backend choice (Jaeger, Tempo, Zipkin, vendor SaaS)
- Optional: Existing instrumentation libs (OpenTracing, Zipkin)
- Optional: Sampling strategy (percent, rate limit)
- Optional: Custom span attrs for business metadata
Steps
See Extended Examples for complete configuration files and templates.
Step 1: Stand Up Backend
Deploy Jaeger or Grafana Tempo. Receive, store traces.
Option A: Jaeger all-in-one (dev/test):
# docker-compose.yml
version: '3.8'
services:
jaeger:
image: jaegertracing/all-in-one:1.51
ports:
- "5775:5775/udp" # Zipkin compact thrift
- "6831:6831/udp" # Jaeger compact thrift
- "6832:6832/udp" # Jaeger binary thrift
- "5778:5778" # Serve configs
- "16686:16686" # Jaeger UI
- "14268:14268" # Jaeger HTTP thrift
- "14250:14250" # Jaeger GRPC
- "9411:9411" # Zipkin compatible endpoint
environment:
- COLLECTOR_ZIPKIN_HOST_PORT=:9411
- COLLECTOR_OTLP_ENABLED=true
restart: unless-stopped
Option B: Grafana Tempo (prod, scales):
# docker-compose.yml
version: '3.8'
services:
tempo:
image: grafana/tempo:2.3.0
command: ["-config.file=/etc/tempo.yaml"]
volumes:
- ./tempo.yaml:/etc/tempo.yaml
- tempo-data:/tmp/tempo
ports:
- "3200:3200" # Tempo HTTP
- "4317:4317" # OTLP gRPC
- "4318:4318" # OTLP HTTP
- "9411:9411" # Zipkin
restart: unless-stopped
volumes:
tempo-data:
Tempo config (tempo.yaml):
server:
http_listen_port: 3200
distributor:
receivers:
jaeger:
# ... (see EXAMPLES.md for complete configuration)
For prod with S3 storage:
storage:
trace:
backend: s3
s3:
bucket: tempo-traces
endpoint: s3.amazonaws.com
region: us-east-1
wal:
path: /tmp/tempo/wal
pool:
max_workers: 100
queue_depth: 10000
Got: Backend live. Ready for traces over OTLP. Jaeger UI or Grafana shows "no traces" first.
If fail:
- Ports in use?
netstat -tulpn | grep -E '(4317|16686|3200)' - Container logs:
docker logs jaegerordocker logs tempo - Test OTLP endpoint:
curl http://localhost:4318/v1/traces -v - For Tempo: check config syntax with
tempo -config.file=/etc/tempo.yaml -verify-config
Step 2: Instrument Apps (Auto)
Use OpenTelemetry auto-instrumentation. Common frameworks. Minimal code change.
Python with Flask:
pip install opentelemetry-distro opentelemetry-exporter-otlp
opentelemetry-bootstrap -a install
# app.py
from flask import Flask
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter
# ... (see EXAMPLES.md for complete configuration)
Go with Gin framework:
go get go.opentelemetry.io/otel
go get go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc
go get go.opentelemetry.io/otel/sdk/trace
go get go.opentelemetry.io/contrib/instrumentation/github.com/gin-gonic/gin/otelgin
package main
import (
"context"
"github.com/gin-gonic/gin"
"go.opentelemetry.io/otel"
# ... (see EXAMPLES.md for complete configuration)
Node.js with Express:
npm install @opentelemetry/api \
@opentelemetry/sdk-node \
@opentelemetry/auto-instrumentations-node \
@opentelemetry/exporter-trace-otlp-grpc
// tracing.js
const { NodeSDK } = require('@opentelemetry/sdk-node');
const { OTLPTraceExporter } = require('@opentelemetry/exporter-trace-otlp-grpc');
const { getNodeAutoInstrumentations } = require('@opentelemetry/auto-instrumentations-node');
const { Resource } = require('@opentelemetry/resources');
const { SemanticResourceAttributes } = require('@opentelemetry/semantic-conventions');
# ... (see EXAMPLES.md for complete configuration)
Got: Traces from instrumented services show in Jaeger UI or Grafana. HTTP requests auto-create spans.
If fail:
- Exporter endpoint reachable from app?
- Env vars set:
OTEL_EXPORTER_OTLP_ENDPOINT=http://tempo:4317 - Turn on debug logs:
OTEL_LOG_LEVEL=debug(Python),OTEL_LOG_LEVEL=DEBUG(Node.js) - Test with simple span — verify export pipe
- Check for version conflicts across OTel packages
Step 3: Add Manual Instrumentation
Custom spans for business logic, DB queries, external calls.
Python manual spans:
from opentelemetry import trace
tracer = trace.get_tracer(__name__)
def process_order(order_id):
# Create a span for the entire operation
# ... (see EXAMPLES.md for complete configuration)
Go manual spans:
import (
"context"
"go.opentelemetry.io/otel"
"go.opentelemetry.io/otel/attribute"
"go.opentelemetry.io/otel/codes"
"go.opentelemetry.io/otel/trace"
# ... (see EXAMPLES.md for complete configuration)
Span attrs best practice:
- Use semantic conventions:
http.method,http.status_code,db.system,db.statement - Business context:
user.id,order.id,product.category - Resource IDs:
instance.id,region,availability_zone - Record errors:
span.RecordError(err)+span.SetStatus(codes.Error, message) - Events for milestones:
span.AddEvent("cache_miss")
Got: Custom spans in trace view. Parent-child right. Attrs visible in span details. Errors highlighted.
If fail:
- Context propagation: parent span context passed to child?
- Span names descriptive, follow naming conventions?
- Spans ended? (
defer span.End()in Go,withblocks in Python) - Attr types: strings, ints, bools, floats only
- Semantic conventions: use standard attr names where applicable
Step 4: Wire Context Propagation
Trace context must flow cross service boundaries, async ops.
HTTP headers propagation (W3C Trace Context):
# Client side (Python with requests)
import requests
from opentelemetry import trace
from opentelemetry.propagate import inject
tracer = trace.get_tracer(__name__)
# ... (see EXAMPLES.md for complete configuration)
// Server side (Go with Gin)
import (
"go.opentelemetry.io/otel"
"go.opentelemetry.io/otel/propagation"
)
# ... (see EXAMPLES.md for complete configuration)
Message queue propagation (Kafka):
# Producer
from opentelemetry.propagate import inject
from kafka import KafkaProducer
producer = KafkaProducer(bootstrap_servers=['kafka:9092'])
# ... (see EXAMPLES.md for complete configuration)
# Consumer
from opentelemetry.propagate import extract
def process_message(msg):
# Extract trace context from Kafka headers
headers = {k: v.decode('utf-8') for k, v in msg.headers}
ctx = extract(headers)
# Continue the trace
with tracer.start_as_current_span("process_order_event", context=ctx):
order_id = json.loads(msg.value)['order_id']
handle_order(order_id)
Async ops (Python asyncio):
import asyncio
from opentelemetry import trace, context
async def async_operation():
# Capture current context
token = context.attach(context.get_current())
try:
with tracer.start_as_current_span("async_database_query"):
await asyncio.sleep(0.1) # Simulated async work
return "result"
finally:
context.detach(token)
Got: Traces span many services. Trace IDs consistent cross boundaries. Parent-child preserved.
If fail:
- W3C Trace Context propagator configured?
otel.propagation.set_global_textmap(TraceContextTextMapPropagator()) - Headers passed in HTTP requests?
- Kafka: headers supported by broker version (v0.11+)?
- Debug: log
traceparentheader value - Use trace viz to spot broken links
Step 5: Set Sampling Strategy
Sample to cut volume and cost. Keep visibility.
Sampling strategies:
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.sampling import (
ParentBased,
TraceIdRatioBased,
StaticSampler,
Decision
# ... (see EXAMPLES.md for complete configuration)
Tail-based sampling with Tempo:
In tempo.yaml:
overrides:
defaults:
metrics_generator:
processors: [service-graphs, span-metrics]
storage:
path: /tmp/tempo/generator/wal
remote_write:
- url: http://prometheus:9090/api/v1/write
send_exemplars: true
# Tail sampling (requires tempo-query)
ingestion_rate_limit_bytes: 5000000
ingestion_burst_size_bytes: 10000000
Use Grafana Tempo's TraceQL for dynamic sampling:
# Sample traces with errors
{ status = error }
# Sample slow traces (>1s)
{ duration > 1s }
# Sample specific services
{ resource.service.name = "checkout-service" }
Got: Trace volume cut to target percent. Error traces always sampled. Sampling decision in trace metadata.
If fail:
- Sampler applied before tracer provider init?
- Sampling decision attr in exported spans?
- Tail sampling: enough buffering? (
ingestion_burst_size_bytes) - Watch dropped traces:
otel_traces_dropped_totalmetric - Test with synthetic high-volume traffic to validate rate
Step 6: Tie Traces to Metrics and Logs
Link traces, metrics, logs. Unified observability.
Add trace IDs to logs (Python):
import logging
from opentelemetry import trace
# Custom log formatter with trace context
class TraceFormatter(logging.Formatter):
def format(self, record):
# ... (see EXAMPLES.md for complete configuration)
Generate metrics from traces (Tempo):
# tempo.yaml
metrics_generator:
registry:
external_labels:
cluster: production
storage:
# ... (see EXAMPLES.md for complete configuration)
Makes Prometheus metrics:
traces_service_graph_request_total- request count between servicestraces_span_metrics_duration_seconds- span duration histogramtraces_spanmetrics_calls_total- span call counts
Query traces from metrics (Grafana):
Add exemplar support to Prometheus datasource in Grafana:
datasources:
- name: Prometheus
type: prometheus
url: http://prometheus:9090
jsonData:
exemplarTraceIdDestinations:
- name: trace_id
datasourceName: Tempo
In Grafana dashboard, turn on exemplars:
{
"fieldConfig": {
"defaults": {
"custom": {
"showExemplars": true
}
}
}
}
Got: Click metric exemplar opens trace. Logs show trace IDs. Traces link to logs. Unified debugging cross signals.
If fail:
- Exemplar support on in Prometheus (v2.26+)?
- Trace ID format matches (32-char hex)?
- Metrics generator on in Tempo config?
- Remote write endpoint reachable from Tempo?
- Test exemplar queries:
histogram_quantile(0.95, rate(http_request_duration_seconds_bucket[5m])) and on() exemplar
Checks
- Backend receives spans from all instrumented services
- Traces show right parent-child cross services
- Span attrs include semantic conventions + business context
- Context propagates cross HTTP and message queues
- Sampling cuts volume to target percent
- Error traces always sampled (if error-aware sampling)
- Trace IDs in app logs, right format
- Grafana shows traces linked from metrics via exemplars
- Log panels have data links to trace viewer
- Trace retention matches storage policy
Pitfalls
- Context not propagated: Forgot to pass
contextdownstream → broken traces. Pass context explicit. - Spans never ended: Missing
defer span.End()(Go) orwithblocks (Python) → open spans, memory leaks. - Over-instrumentation: Span for every function bloats traces. Focus on service boundaries, DB calls, external APIs.
- Missing error recording: Skip
span.RecordError()→ lose debug info. Always record errors in spans. - High cardinality attrs: Unbounded values (user IDs, request bodies) as span attrs → storage pain. Sample or aggregate.
- Wrong span kind: CLIENT vs SERVER vs INTERNAL mixed up → wrong service graph. Follow semantic conventions.
- Sampling before context: Sampling must respect parent context. Use
ParentBasedsampler.
See Also
correlate-observability-signals- Unified debugging across metrics, logs, traces by trace IDsetup-prometheus-monitoring- Metrics from traces via Tempo generatorconfigure-log-aggregation- Trace IDs in logs for correlationbuild-grafana-dashboards- Viz trace-derived metrics and exemplars
GitHub 저장소
연관 스킬
evaluating-llms-harness
테스팅이 Claude Skill은 MMLU, GSM8K를 포함한 60개 이상의 표준화된 학술 과제에서 LLM 성능을 벤치마크하기 위해 lm-evaluation-harness를 실행합니다. 개발자들이 모델 품질을 비교하고, 학습 진행 상황을 추적하거나 학술 결과를 보고할 수 있도록 설계되었습니다. 이 도구는 HuggingFace와 vLLM 모델을 포함한 다양한 백엔드를 지원합니다.
cloudflare-cron-triggers
테스팅이 스킬은 cron 표현식을 사용하여 Worker를 스케줄링하기 위한 Cloudflare Cron Triggers 구현에 관한 포괄적인 지식을 제공합니다. 주기적 작업, 유지보수 작업, 자동화된 워크플로우 설정 방법을 다루며, 잘못된 cron 표현식이나 시간대 문제 같은 일반적인 이슈들을 해결하는 방법을 포함합니다. 개발자들은 이를 통해 스케줄된 핸들러 구성, cron 트리거 테스트, Workflows 및 Green Compute와의 연동 작업을 수행할 수 있습니다.
webapp-testing
테스팅이 Claude Skill은 Python 스크립트를 통해 로컬 웹 애플리케이션을 테스트하기 위한 Playwright 기반 툴킷을 제공합니다. 프론트엔드 검증, UI 디버깅, 스크린샷 캡처, 로그 확인 기능을 지원하며 서버 라이프사이클을 관리합니다. 브라우저 자동화 작업에 사용하되 컨텍스트 오염을 방지하기 위해 소스 코드를 읽지 않고 스크립트를 직접 실행하세요.
finishing-a-development-branch
테스팅이 스킬은 테스트 통과를 확인한 후 체계적인 통합 옵션을 제시하여 개발자가 완성된 작업을 마무리하도록 돕습니다. 구현이 완료된 후 머지, PR 생성, 브랜치 정리와 같은 워크플로우를 안내합니다. 코드가 준비되고 테스트가 완료되었을 때 개발 프로세스를 체계적으로 마무리하기 위해 사용하세요.
