Datadog APM Skill
Analyze distributed traces and service performance in Datadog APM.
Tools
| Tool | Description |
|---|---|
search_traces | Search traces with query syntax and filters |
get_trace_details | Get detailed span information for a trace |
list_services | List all instrumented services |
get_service_summary | Get service health and performance summary |
Example Queries
Search for slow traces in the network service → search_traces(query="service:network-api @duration:>1s", time_range="1h") Find error traces → search_traces(query="status:error", time_range="30m") Get details of a specific trace → get_trace_details(trace_id="abc123def456") List all services → list_services() Get service performance summary → get_service_summary(service_name="network-api")
Workflows
Latency Analysis
- Search slow traces:
search_traces(query="@duration:>500ms") - Identify common patterns (resource, operation)
- Drill into trace waterfall:
get_trace_details(trace_id="...") - Find bottleneck spans
- Compare with baseline performance
Error Investigation
- Search error traces:
search_traces(query="status:error service:network") - Group by error type and root cause
- Review span-level errors in trace details
- Correlate with recent deployments or changes
Service Health Review
- List all services:
list_services() - Get summary for critical services:
get_service_summary(service_name="...") - Check error rate, latency percentiles, throughput
- Identify degraded or unhealthy services
- Review service dependencies map
Root Cause Analysis
- Start from incident or alert
- Find related traces:
search_traces(query="trace_id:abc123") - Walk span tree to identify failure point
- Check upstream/downstream service health
- Correlate with infrastructure metrics
Prerequisites
DD_API_KEYDatadog API keyDD_APP_KEYDatadog application keyDD_SITEDatadog site (optional, defaults to datadoghq.com)
Server
This skill uses the datadog-mcp server via remote MCP transport.
Notes
- APM requires instrumented services sending traces to Datadog
- Trace retention depends on your Datadog plan
- Large trace searches may be paginated
Failure Behavior
- On a tool error (timeout, unreachable host, malformed response), report the failure and its error message directly to the user rather than fabricating or guessing at results.
- For a confirmed read-only call, check connectivity and retry once if appropriate. For any call that changes state or sends a message, a timeout does not prove the action failed: inspect current state or delivery status before retrying, preserve the required approval/change gates, and do not repeat an action whose outcome is unknown.
