Section 5 – Optimizing Performance and Cost

5.1 – Application Performance Monitoring

Cloud Profiler

• Continuous CPU and memory profiling in production (low overhead ~0.5%)
• Languages: Go, Python, Java, Node.js, Ruby, PHP
• Identifies: hot functions, memory allocation hotspots
• No sampling configuration needed — automatic continuous profiling

# Python — add profiler to app
pip install google-cloud-profiler

import googlecloudprofiler
googlecloudprofiler.start(
    service='my-service',
    service_version='1.0.0',
    verbose=3
)

Active Assist

• AI-powered recommendations across GCP services
• Available for: cost, security, performance, reliability, manageability
• Access via: Cloud Console → Active Assist, or gcloud recommender

# View cost recommendations
gcloud recommender recommendations list \
  --recommender=google.compute.instance.MachineTypeRecommender \
  --location=us-central1 \
  --project=PROJECT

# View idle resource recommendations
gcloud recommender recommendations list \
  --recommender=google.compute.instance.IdleResourceRecommender \
  --location=global \
  --project=PROJECT

Cloud Run Performance Monitoring

# Monitor cold starts
gcloud logging read 'resource.type="cloud_run_revision"' \
  --filter='jsonPayload.message:"Cold start"' --limit=100

# Key metrics
run.googleapis.com/request_latencies  # Request latency histogram
run.googleapis.com/container/startup_latency  # Cold start duration
run.googleapis.com/request_count  # Request count by response code

---

5.2 – FinOps Practices

Observability Cost Management

Observability costs often grow unexpectedly. Key cost drivers:

Component	Cost driver	Optimization
Cloud Logging	Log ingestion volume	Exclusion filters, reduce sampling
Cloud Monitoring	Custom metric samples	Reduce cardinality, drop unused metrics
Cloud Trace	Trace samples	Adjust sampling rate (0.1 = 10%)
BigQuery	Log export + query	Partition tables, use log-based metrics instead

# Check metric ingestion volume (Metrics Management page)
gcloud monitoring metrics list --project=PROJECT | wc -l

# Exclude high-volume health check logs
gcloud logging sinks update _Default \
  --add-exclusion='name=healthchecks,filter=httpRequest.requestUrl="/health" AND httpRequest.status=200'

# Reduce trace sampling
# In OTel SDK:
from opentelemetry.sdk.trace.sampling import TraceIdRatioBased
sampler = TraceIdRatioBased(0.1)  # Sample 10% of traces

---

Compute Pricing Models

Sustained Use Discounts (SUD)

• Automatic — no purchase required
• Applied to Compute Engine VMs running > 25% of a billing month
• Incremental discounts: 25%→50%→75%→100% usage = 0%→10%→20%→30% discount
• Does NOT apply to: App Engine flexible, Dataflow, GPU accelerators (some), E2 machine series

Committed Use Discounts (CUD)

• Commit to 1 or 3 years of resource usage
• Two types:

Type	Description	Discount
Resource-based	Commit to specific vCPU/RAM in a region/machine family	Up to 57% (1yr) / 70% (3yr)
Spend-based (Flex CUD)	Commit to minimum hourly spend	28% (1yr) / 46% (3yr)

• Flex CUD covers: Compute Engine VMs, GKE Standard, GKE Autopilot, Cloud Run
• CUDs cannot be cancelled — commit only to your stable baseline

# Purchase a CUD
gcloud compute commitments create my-commitment \
  --plan=12-month \
  --region=us-central1 \
  --resources=vcpu=100,memory=400GB

Spot VMs

• Up to 91% cheaper than on-demand
• Can be preempted with 30-second notice when GCP needs capacity
• No maximum runtime (unlike old preemptible VMs which had 24h max)
• Good for: batch jobs, CI/CD workers, ML training, data processing

# Create Spot VM
gcloud compute instances create my-spot \
  --machine-type=n2-standard-4 \
  --provisioning-model=SPOT \
  --instance-termination-action=STOP

# Use Spot in GKE node pool
gcloud container node-pools create spot-pool \
  --cluster=my-cluster \
  --spot \
  --num-nodes=3

When to use which

Workload	Recommendation
Stable, long-running (prod)	Resource CUD (1yr)
Variable, mixed workloads	Flex CUD
Batch, CI/CD, ML training	Spot VMs
Short-lived variable (< monthly)	SUD (auto)
Dev/test	On-demand (no commitment)

---

Network Cost Optimization

Network Tiers

Tier	Description	Egress cost
Premium	Google’s global backbone — low latency	Higher
Standard	ISP routing — acceptable latency	~30% cheaper

# Set network tier for instance
gcloud compute instances create my-vm \
  --network-tier=STANDARD  # or PREMIUM (default)

Minimize Egress Costs

• Keep compute and storage in same region → free intra-region traffic
• Use CDN (Cloud CDN) for static content — reduce origin egress
• Cloud Interconnect for high-volume on-prem ↔ GCP: cheaper than internet egress
• Use VPC Flow Logs sampling < 100% — significant logging cost reduction

---

GKE Cost Optimization

Node Efficiency

# Right-size nodes with VPA
kubectl apply -f vertical-pod-autoscaler.yaml

# Use Autopilot for automatic right-sizing
gcloud container clusters create-auto my-cluster \
  --region=us-central1

# Enable cluster autoscaler with scale-down
gcloud container node-pools update pool \
  --cluster=my-cluster \
  --enable-autoscaling \
  --min-nodes=1 --max-nodes=20

# Spot node pool for non-critical workloads
gcloud container node-pools create spot-pool \
  --cluster=CLUSTER --spot

Namespace-Level Cost Visibility

• Use GKE Cost Allocation (built-in): allocates costs to namespaces/labels
• Enable in cluster settings → available in BigQuery export
• Use Kubecost or custom Prometheus + BigQuery pipeline for detailed breakdowns

# Enable GKE cost allocation
gcloud container clusters update CLUSTER \
  --enable-cost-allocation

---

Cloud Run Cost Optimization

# Set minimum instances (avoid cold start penalty at cost of always-on)
gcloud run services update myapp --min-instances=1

# Set concurrency high to maximize instance utilization
gcloud run services update myapp --concurrency=1000

# Use CPU always-on only if needed (default: CPU throttled when not serving)
gcloud run services update myapp --no-cpu-throttling  # Always-on CPU (more expensive)

---

GCP Recommenders (Active Assist)

Recommender	What it finds
MachineTypeRecommender	Over-provisioned VMs → suggest smaller
IdleResourceRecommender	VMs with <1% CPU usage — idle
DiskIdleRecommender	Unused persistent disks
AddressIdleRecommender	Unused static IPs ($7.20/month each)
SnapshotRecommender	Old snapshots costing money
GKENodeRecommender	Over/under-provisioned GKE node pools
IAMRecommender	Overly broad IAM bindings (security + cost)
FirewallInsightRecommender	Unused firewall rules

# Apply a recommendation
gcloud recommender recommendations mark-claimed RECOMMENDATION_ID \
  --recommender=google.compute.instance.MachineTypeRecommender \
  --location=us-central1-a \
  --project=PROJECT

# Mark applied
gcloud recommender recommendations mark-succeeded RECOMMENDATION_ID \
  --recommender=google.compute.instance.MachineTypeRecommender \
  --location=us-central1-a \
  --project=PROJECT

---

Cost Monitoring and Budgets

# Create budget alert
gcloud billing budgets create \
  --billing-account=BILLING_ACCOUNT_ID \
  --display-name="Prod Monthly Budget" \
  --budget-amount=5000USD \
  --threshold-rule=percent=50 \
  --threshold-rule=percent=80 \
  --threshold-rule=percent=100,basis=FORECASTED_SPEND \
  --filter-projects=projects/prod-project-id

Cost Allocation Labels

# Label resources for cost attribution
gcloud compute instances add-labels my-instance \
  --labels=team=payments,env=prod,cost-center=engineering

# Apply default labels via Org Policy
# constraints/compute.requireShieldedVm + labels via ConstraintCustom

---

Summary: Key Cost Levers

Priority	Action	Typical Savings
🔴 High	Apply CUDs to stable compute baseline	30-70%
🔴 High	Use Spot VMs for batch/CI workloads	60-91%
🟡 Medium	Right-size instances (use Recommender)	20-40%
🟡 Medium	Delete idle resources (IPs, disks, snapshots)	Variable
🟡 Medium	Reduce log ingestion (exclusions)	20-50% of logging bill
🟢 Low	Enable Cloud CDN for static assets	30-50% of egress
🟢 Low	Use Standard network tier where latency allows	~30% of egress

---

Exam Tips

• SUD = automatic, no commitment; CUD = manual purchase, ⅓ year commitment
• Spot VMs = no max runtime (unlike old preemptible 24h limit) + 30s termination notice
• Flex CUD = spend-based commitment; covers GKE Autopilot + Cloud Run (not just VMs)
• Recommenders are per-resource-type and per-location — check correct recommender ID
• GKE Cost Allocation = built-in namespace/label cost breakdown → exports to BigQuery
• Cloud Profiler = production profiling with <0.5% overhead (safe in prod)
• Active Assist = umbrella term for all GCP recommenders + insights
• Idle static IPs cost ~$7.20/month each — small but adds up; use AddressIdleRecommender