Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/flyio-performance-tuning/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill flyio-performance-tuning. Copyright stays with the author (MIT).

Fly.io Performance Tuning

Overview

Optimize Fly.io performance: eliminate cold starts, right-size VMs, leverage multi-region for low latency, and tune concurrency settings.

Prerequisites

• A redacted baseline for latency, error rate, saturation, cost, and region-level health.
• A staging application, synthetic load, named change owner, and a tested rollback mechanism.

Output

Publish a tuning receipt with baseline/post-change aggregate metrics, VM/concurrency settings, regions affected, canary result, owner, and rollback outcome. Exclude runtime secrets, request bodies, and user data.

Error Handling

• Stop rollout on health, latency, saturation, or cost threshold breaches and revert the canary.
• Reduce concurrency or capacity changes before retrying a failed region promotion.
• Keep diagnostics redacted and route deployment incidents to the on-call owner.

Examples

Run a synthetic load test against a staging region, adjust one VM setting, and compare aggregate p95 latency and error rate. Simulate a failed health check and confirm the release rolls back before traffic expands.

Instructions

Step 1: Eliminate Cold Starts

# fly.toml — suspend instead of stop for faster resume (~100ms vs ~5s)
[http_service]
  auto_stop_machines = "suspend"   # Suspend to RAM, not full stop
  auto_start_machines = true
  min_machines_running = 1          # Always-warm in primary region

# For latency-critical: keep machines running in all regions
# min_machines_running applies globally

Step 2: Right-Size VMs

# Check current allocation
fly scale show -a my-app

# Start small, scale up based on metrics
fly scale vm shared-cpu-1x --memory 256    # Start here
fly scale vm shared-cpu-1x --memory 512    # If memory-constrained
fly scale vm shared-cpu-2x --memory 1024   # If CPU-bound
fly scale vm performance-2x --memory 4096  # For compute-heavy workloads

Step 3: Multi-Region Latency Optimization

# Deploy close to your users
fly scale count 1 --region iad    # US East
fly scale count 1 --region lhr    # Europe
fly scale count 1 --region nrt    # Asia Pacific

# Fly automatically routes to nearest region via Anycast
# Verify: curl with timing
curl -w "DNS: %{time_namelookup}s, Connect: %{time_connect}s, Total: %{time_total}s\n" \
  -o /dev/null -s https://my-app.fly.dev/health

Step 4: Connection Pooling for Postgres

// Use connection pooling for Fly Postgres
// PgBouncer runs on port 5433 (pooled) vs 5432 (direct)
const pooledUrl = databaseConfig.pooledEndpoint;

// Enable PgBouncer mode through the client's typed configuration.

Step 5: Tune Concurrency

[http_service.concurrency]
  type = "requests"       # or "connections"
  hard_limit = 250        # Max before rejecting
  soft_limit = 200        # Start scaling at this point

Resources

• Auto Stop/Start
• Machine Sizing
• Suspend/Resume

Next Steps

For cost optimization, see flyio-cost-tuning.