azure（Azure 部署）

Azure Production Rules

Cost Traps

• Stopped VMs still pay for attached disks and public IPs — deallocate fully with az vm deallocate not just stop from portal
• Premium SSD default on VM creation — switch to Standard SSD for dev/test, saves 50%+
• Log Analytics workspace retention defaults to 30 days free, then charges per GB — set data retention policy and daily cap before production
• Bandwidth between regions is charged both ways — keep paired resources in same region, use Private Link for cross-region when needed
• Cosmos DB charges for provisioned RU/s even when idle — use serverless for bursty workloads or autoscale with minimum RU setting

Security Rules

• Resource Groups don't provide network isolation — NSGs and Private Endpoints do. RG is for management, not security boundary
• Managed Identity eliminates secrets for Azure-to-Azure auth — use System Assigned for single-resource, User Assigned for shared identity
• Key Vault soft-delete enabled by default (90 days) — can't reuse vault name until purged, plan naming accordingly
• Azure AD conditional access policies don't apply to service principals — use App Registrations with certificate auth, not client secrets
• Private Endpoints don't automatically update DNS — configure Private DNS Zone and link to VNet or resolution fails

Networking

• NSG rules evaluate by priority (lowest number first) — default rules at 65000+ always lose to custom rules
• Application Gateway v2 requires dedicated subnet — at least /24 recommended for autoscaling
• Azure Firewall premium SKU required for TLS inspection and IDPS — standard can't inspect encrypted traffic
• VNet peering is non-transitive — hub-and-spoke requires routes in each spoke, or use Azure Virtual WAN
• Service Endpoints expose entire service to VNet — Private Endpoints give private IP for specific resource instance

Performance

• Azure Functions consumption plan has cold start — Premium plan with minimum instances for latency-sensitive
• Cosmos DB partition key choice is permanent and determines scale — can't change without recreating container
• App Service plan density: P1v3 handles ~10 slots, more causes resource contention — monitor CPU/memory per slot
• Azure Cache for Redis Standard tier has no SLA for replication — use Premium for persistence and clustering
• Blob storage hot tier for frequent access — cool has 30-day minimum, archive has 180-day and hours-long rehydration

Monitoring

• Application Insights sampling kicks in at high volume — telemetry may miss intermittent errors, adjust MaxTelemetryItemsPerSecond
• Azure Monitor alert rules charge per metric tracked — consolidate metrics in Log Analytics for complex alerts
• Activity Log only shows control plane operations — diagnostic settings required for data plane (blob access, SQL queries)
• Alert action groups have rate limits — 1 SMS per 5 min, 1 voice call per 5 min, 100 emails per hour per group
• Log Analytics query timeout is 10 minutes — optimize queries with time filters first, then other predicates

Infrastructure as Code

• ARM templates fail silently on some property changes — use what-if deployment mode to preview changes
• Terraform azurerm provider state contains secrets in plaintext — use remote backend with encryption (Azure Storage + customer key)
• Bicep is ARM's replacement — transpiles to ARM, better tooling, use for new projects
• Resource locks prevent accidental deletion but block some operations — CanNotDelete lock still allows modifications
• Azure Policy evaluates on resource creation and updates — existing non-compliant resources need remediation task

Identity and Access

• RBAC role assignments take up to 30 minutes to propagate — pipeline may fail immediately after assignment
• Owner role can't manage role assignments if PIM requires approval — use separate User Access Administrator
• Service principal secret expiration defaults to 1 year — set calendar reminder or use certificate with longer validity
• Azure AD B2C is separate from Azure AD — different tenant, different APIs, different pricing