implementing-observability（实施可观察性）

Production Observability with OpenTelemetry

Purpose

Implement production-grade observability using OpenTelemetry as the 2025 industry standard. Covers the three pillars (metrics, logs, traces), LGTM stack deployment, and critical log-trace correlation patterns.

When to Use

Use when:

• Building production systems requiring visibility into performance and errors
• Debugging distributed systems with multiple services
• Setting up monitoring, logging, or tracing infrastructure
• Implementing structured logging with trace correlation
• Configuring alerting rules for production systems

Skip if:

• Building proof-of-concept without production deployment
• System has < 100 requests/day (console logging may suffice)

The OpenTelemetry Standard (2025)

OpenTelemetry is the CNCF graduated project unifying observability:

┌────────────────────────────────────────────────────────┐
│          OpenTelemetry: The Unified Standard           │
├────────────────────────────────────────────────────────┤
│                                                         │
│  ONE SDK for ALL signals:                              │
│  ├── Metrics (Prometheus-compatible)                   │
│  ├── Logs (structured, correlated)                     │
│  ├── Traces (distributed, standardized)                │
│  └── Context (propagates across services)              │
│                                                         │
│  Language SDKs:                                         │
│  ├── Python: opentelemetry-api, opentelemetry-sdk      │
│  ├── Rust: opentelemetry, tracing-opentelemetry        │
│  ├── Go: go.opentelemetry.io/otel                      │
│  └── TypeScript: @opentelemetry/api                    │
│                                                         │
│  Export to ANY backend:                                │
│  ├── LGTM Stack (Loki, Grafana, Tempo, Mimir)          │
│  ├── Prometheus + Jaeger                               │
│  ├── Datadog, New Relic, Honeycomb (SaaS)              │
│  └── Custom backends via OTLP protocol                 │
│                                                         │
└────────────────────────────────────────────────────────┘

Context7 Reference: /websites/opentelemetry_io (Trust: High, Snippets: 5,888, Score: 85.9)

The Three Pillars of Observability

1. Metrics (What is happening?)

Track system health and performance over time.

Metric Types: Counters (always increase), Gauges (up/down), Histograms (distributions), Summaries (percentiles).

Brief Example (Python):

from opentelemetry import metrics

meter = metrics.get_meter(__name__)
http_requests = meter.create_counter("http.server.requests")
http_requests.add(1, {"method": "GET", "status": 200})

2. Logs (What happened?)

Record discrete events with context.

CRITICAL: Always inject trace_id/span_id for log-trace correlation.

Brief Example (Python + structlog):

import structlog
from opentelemetry import trace

logger = structlog.get_logger()
span = trace.get_current_span()
ctx = span.get_span_context()

logger.info(
    "processing_request",
    trace_id=format(ctx.trace_id, '032x'),
    span_id=format(ctx.span_id, '016x'),
    user_id=user_id
)

See: references/structured-logging.md for complete configuration.

3. Traces (Where did time go?)

Track request flow across distributed services.

Key Concepts: Trace (end-to-end journey), Span (individual operation), Parent-Child (nested operations).

Brief Example (Python + FastAPI):

from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor

app = FastAPI()
FastAPIInstrumentor.instrument_app(app)  # Auto-traces all HTTP requests

See: references/opentelemetry-setup.md for SDK installation by language.

The LGTM Stack (Self-Hosted Observability)

LGTM = Loki (Logs) + Grafana (Visualization) + Tempo (Traces) + Mimir (Metrics)

┌────────────────────────────────────────────────────────┐
│                  LGTM Architecture                      │
├────────────────────────────────────────────────────────┤
│                                                         │
│  ┌──────────────────────────────────────────────┐      │
│  │           Grafana Dashboard (Port 3000)      │      │
│  │  Unified UI for Logs, Metrics, Traces       │      │
│  └──────┬──────────────┬─────────────┬─────────┘      │
│         │              │             │                 │
│         ▼              ▼             ▼                 │
│  ┌──────────┐   ┌──────────┐  ┌──────────┐            │
│  │   Loki   │   │  Tempo   │  │  Mimir   │            │
│  │  (Logs)  │   │ (Traces) │  │(Metrics) │            │
│  │Port 3100 │   │Port 3200 │  │Port 9009 │            │
│  └────▲─────┘   └────▲─────┘  └────▲─────┘            │
│       │              │             │                   │
│       └──────────────┴─────────────┘                   │
│                      │                                 │
│              ┌───────▼────────┐                        │
│              │ Grafana Alloy  │                        │
│              │  (Collector)   │                        │
│              │  Port 4317/8   │ ← OTLP gRPC/HTTP       │
│              └───────▲────────┘                        │
│                      │                                 │
│         OpenTelemetry Instrumented Apps                │
│                                                         │
└────────────────────────────────────────────────────────┘

Quick Start: Run examples/lgtm-docker-compose/docker-compose.yml for a complete LGTM stack.

See: references/lgtm-stack.md for production deployment guide.

Critical Pattern: Log-Trace Correlation

The Problem: Logs and traces live in separate systems. You see an error log but can't find the related trace.

The Solution: Inject trace_id and span_id into every log record.

Python (structlog)

import structlog
from opentelemetry import trace

logger = structlog.get_logger()
span = trace.get_current_span()
ctx = span.get_span_context()

logger.info(
    "request_processed",
    trace_id=format(ctx.trace_id, '032x'),  # 32-char hex
    span_id=format(ctx.span_id, '016x'),    # 16-char hex
    user_id=user_id
)

Rust (tracing)

use tracing::{info, instrument};

#[instrument(fields(user_id = %user_id))]
async fn process_request(user_id: u64) -> Result<Response> {
    // trace_id/span_id automatically included
    info!(user_id = user_id, "processing request");
    Ok(result)
}

See: references/trace-context.md for Go and TypeScript patterns.

Query in Grafana

{job="api-service"} |= "trace_id=4bf92f3577b34da6a3ce929d0e0e4736"

Quick Setup Guide

1. Choose Your Stack

Decision Tree:

• Greenfield: OpenTelemetry SDK + LGTM Stack (self-hosted) or Grafana Cloud (managed)
• Existing Prometheus: Add Loki (logs) + Tempo (traces)
• Kubernetes: LGTM via Helm, Alloy DaemonSet
• Zero-ops: Managed SaaS (Grafana Cloud, Datadog, New Relic)

2. Install OpenTelemetry SDK

Bootstrap Script:

python scripts/setup_otel.py --language python --framework fastapi

Manual (Python):

pip install opentelemetry-api opentelemetry-sdk \
    opentelemetry-instrumentation-fastapi \
    opentelemetry-exporter-otlp

See: references/opentelemetry-setup.md for Rust, Go, TypeScript installation.

3. Deploy LGTM Stack

Docker Compose (development):

cd examples/lgtm-docker-compose
docker-compose up -d
# Grafana: http://localhost:3000 (admin/admin)
# OTLP: localhost:4317 (gRPC), localhost:4318 (HTTP)

See: references/lgtm-stack.md for production Kubernetes deployment.

4. Configure Structured Logging

See: references/structured-logging.md for complete setup (Python, Rust, Go, TypeScript).

5. Set Up Alerting

See: references/alerting-rules.md for Prometheus and Loki alert patterns.

Auto-Instrumentation

OpenTelemetry auto-instruments popular frameworks:

from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor

app = FastAPI()
FastAPIInstrumentor.instrument_app(app)  # Auto-trace all HTTP requests

Supported: FastAPI, Flask, Django, Express, Gin, Echo, Nest.js

See: references/opentelemetry-setup.md for framework-specific setup.

Common Patterns

Custom Spans

from opentelemetry import trace

tracer = trace.get_tracer(__name__)

with tracer.start_as_current_span("fetch_user_details") as span:
    span.set_attribute("user_id", user_id)
    user = await db.fetch_user(user_id)
    span.set_attribute("user_found", user is not None)

Error Tracking

from opentelemetry.trace import Status, StatusCode

with tracer.start_as_current_span("process_payment") as span:
    try:
        result = process_payment(amount, card_token)
        span.set_status(Status(StatusCode.OK))
    except PaymentError as e:
        span.set_status(Status(StatusCode.ERROR, str(e)))
        span.record_exception(e)
        raise

See: references/trace-context.md for background job tracing and context propagation.

Validation and Testing

# Test log-trace correlation
# 1. Make request to your app
# 2. Copy trace_id from logs
# 3. Query in Grafana: {job="myapp"} |= "trace_id=<TRACE_ID>"

# Validate metrics
python scripts/validate_metrics.py

Integration with Other Skills

• Dashboards: Embed Grafana panels, query Prometheus metrics
• Feedback: Alert routing (Slack, PagerDuty), notification UI
• Data-Viz: Time-series charts, trace waterfall, latency heatmaps

See: examples/fastapi-otel/ for complete integration.

Progressive Disclosure

Setup Guides:

• references/opentelemetry-setup.md - SDK installation (Python, Rust, Go, TypeScript)
• references/structured-logging.md - structlog, tracing, slog, pino configuration
• references/lgtm-stack.md - LGTM deployment (Docker, Kubernetes)
• references/trace-context.md - Log-trace correlation patterns
• references/alerting-rules.md - Prometheus and Loki alert templates

Examples:

• examples/fastapi-otel/ - FastAPI + OpenTelemetry + LGTM
• examples/axum-tracing/ - Rust Axum + tracing + LGTM
• examples/lgtm-docker-compose/ - Production-ready LGTM stack

Scripts:

• scripts/setup_otel.py - Bootstrap OpenTelemetry SDK
• scripts/generate_dashboards.py - Generate Grafana dashboards
• scripts/validate_metrics.py - Validate metric naming

Key Principles

1. OpenTelemetry is THE standard - Use OTel SDK, not vendor-specific SDKs
2. Auto-instrumentation first - Prefer auto over manual spans
3. Always correlate logs and traces - Inject trace_id/span_id into every log
4. Use structured logging - JSON format, consistent field names
5. LGTM stack for self-hosting - Production-ready open-source stack

Common Pitfalls

Don't:

• Use vendor-specific SDKs (use OpenTelemetry)
• Log without trace_id/span_id context
• Manually instrument what auto-instrumentation covers
• Mix logging libraries (pick one: structlog, tracing, slog, pino)

Do:

• Start with auto-instrumentation
• Add manual spans only for business-critical operations
• Use semantic conventions for span attributes
• Export to OTLP (gRPC preferred over HTTP)
• Test locally with LGTM docker-compose before production

Success Metrics

1. 100% of logs include trace_id when in request context
2. Mean time to resolution (MTTR) decreases by >50%
3. Developers use Grafana as first debugging tool
4. 80%+ of telemetry from auto-instrumentation
5. Alert noise < 5% false positives