traces

Distributed Tracing Analysis Skill

Analyze distributed traces for performance investigation and service dependency mapping.

When to Use

• Latency investigation
• Performance bottleneck detection
• Service dependency mapping
• Slow request analysis
• Cross-service debugging
• User asks "traces", "latency", "slow", "performance"

Tracing Analysis Approach

1. Discover Tracing Sources

Use capability-based discovery:

# Discover available integrations via capability detection

# Scan for tracing capabilities by analyzing server descriptions:
# - tracing capability: Dedicated distributed tracing systems
# - apm capability: APM tools that include distributed tracing
# - telemetry capability: Unified observability with tracing support

2. Query Trace Data

For each discovered source:

• Slow traces (>SLO threshold)
• Error traces
• Service operation latencies
• Request volume by endpoint
• Service-to-service call patterns

3. Analyze Latency Breakdown

For slow traces:

• Total request duration
• Time spent in each service
• Database query time
• External API call time
• Queue/async processing time
• Network/serialization overhead

4. Map Service Dependencies

Build dependency graph:

• Service call chains
• Synchronous vs asynchronous calls
• Fan-out patterns
• Critical path services

5. Identify Bottlenecks

See refs/bottlenecks-detection.md and refs/bottlenecks-resolution.md for detailed patterns.

Common bottleneck types:

• N+1 query patterns: Many sequential DB calls
• Slow external calls: Third-party API latency
• Database bottlenecks: Long query times
• Sequential processing: Parallelizable operations
• Resource contention: Lock waits, queue depth

6. Correlate with Code

Use wicked-garden:search to find:

• Database queries in slow services
• External API calls
• Synchronous operations that could be async
• Missing caching opportunities

Integration Discovery

Capability	What to Look For	Provides
tracing	Distributed tracing, span collection, trace analysis	Distributed traces, span details
apm	Performance monitoring with distributed tracing	Traces with performance context
telemetry	Unified observability with traces and metrics	Unified traces and metrics

Fallback: Analyze code for call patterns via wicked-garden:search (database calls, HTTP clients, async operations).

Output Format

Provide trace analysis with:

• Performance summary (slow traces, latency percentiles)
• Service dependency map (tree view with timing)
• Bottleneck analysis (latency, type, root cause)
• Optimization opportunities (quick wins vs long-term)

See refs/bottlenecks-detection.md and refs/bottlenecks-resolution.md for detailed output templates.

Common Performance Patterns

See refs/bottlenecks-detection.md and refs/bottlenecks-resolution.md for detailed analysis of:

N+1 Query Pattern

Sequential database queries in a loop. Fix with batch queries or JOIN.

Synchronous External Calls

Waiting for external APIs sequentially. Fix with Promise.all or async processing.

Missing Caching

Repeated identical queries. Fix with caching layer (Redis, in-memory).

Database Query Inefficiency

Slow database operations. Fix with indexes, query optimization, pagination.

Resource Contention

Lock waits or queue delays. Fix with reduced lock scope, optimistic locking.

Service Dependency Analysis

Critical Path Detection

Identify services on critical path (required for request completion). Focus optimization here.

Fan-Out Pattern Detection

Single request triggers many downstream calls. Risk of amplified load and cascading failures.

Circular Dependency Detection

Service A calls B calls A (cycle). Risk of infinite loops and complex debugging.

See refs/dependencies.md for detailed dependency analysis patterns.

Integration with engineering

When bottlenecks identified, engage wicked-garden:engineering:backend with trace context, code locations, and optimization recommendations.

Integration with wicked-crew

During build phase:

1. Capture baseline latency metrics
2. Monitor post-deployment latency
3. Alert on p95 increases >20%
4. Recommend rollback if critical path slows

Emit events:

• observe:trace:slow:warning
• observe:correlation:found:success

Notes

• Focus on p95/p99, not averages (outliers matter)
• Identify critical path first (highest impact)
• Consider parallelization opportunities
• External calls often biggest bottleneck
• Add instrumentation if spans missing
• Correlate slow traces with error traces