Profiling .NET in Production Without Killing Performance

Production profiling is scary until you have the right tools. Here’s how we investigate latency without restarting services.

Continuous Profiling with dotnet-monitor

dotnet-monitor collect \
  --process-id 1234 \
  --duration 00:01:00 \
  --profile cpu \
  --output /tmp/trace.nettrace

Overhead: <1% CPU for 60-second captures. Safe for production.

Analyzing Traces

dotnet-trace report trace.nettrace \
  --top-n 20 \
  --inclusive \
  --format json

Look for:

• GC pauses > 50ms
• Thread pool starvation (long Wait times)
• Sync-over-async (Task.Wait() in async code)

Our Alerting Stack

1. P99 latency breach → auto-trigger 60s trace
2. Trace uploaded to shared storage
3. Slack notification with flame graph link
4. On-call investigates within 15 minutes

Quick Wins We Found

Issue	Impact	Fix
Sync JSON serialization	+40ms P99	Source generators
Missing DB index	+200ms on one endpoint	Added composite index
Large object heap pressure	GC pauses every 30s	Object pooling for buffers