Appendix A: The Rosetta Stone - Cloud MLOps Service Mapping

A.1. The Compute Primitives

When lifting and shifting MLOps stacks, the most common error is assuming “VM equals VM.” The nuances of underlying hypervisors, networking, and accelerator attachment differ significantly.

A.1.1. General Purpose Compute

A.1.2. Accelerated Compute (GPUs/TPUs)

A.2. The Data & Storage Layer

A.2.1. Object Storage (The Data Lake)

A.2.2. Managed Databases for MLOps

A.3. The MLOps Platform Services

A.3.1. Training & Orchestration

A.3.2. Serving & Inference

A.4. The Security & Governance Plane

A.4.1. Identity & Access Management (IAM)

• AWS IAM:
  • Model: Role-based. Resources assume roles. Policies attached to identities or resources.
  • Complexity: High. “Principal”, “Action”, “Resource”, “Condition”.
  • MLOps Pattern: SageMakerExecutionRole determines what S3 buckets the training job can read.
• GCP IAM:
  • Model: Project-centric. Service Accounts.
  • Complexity: Medium. “Member” bound to “Role” on “Resource”.
  • MLOps Pattern: Workload Identity federation for GKE.
• Azure Entra ID (fka AD):
  • Model: Enterprise-centric. Users/Service Principals.
  • Complexity: High (Enterprise legacy).
  • MLOps Pattern: Managed Identities (System-assigned vs User-assigned) avoid credential rotation.

A.4.2. Network Security

• AWS: Security Groups (Stateful firewall) + NACLs (Stateless). PrivateLink for accessing services without public internet.
• GCP: VPC Service Controls (The “perimeter”). Global VPCs (subnets in different regions communicate via internal IP).
• Azure: VNet + Private Endpoints. NSGs (Network Security Groups).

A.5. Generative AI (LLM) Services Comparison (2025)

A.6. Equivalent CLI Cheatsheet

For the engineer moving between clouds.

A.6.1. Compute & Auth

A.6.2. Storage

A.7. Architectural Design Patterns Mapping

A.7.1. The “Hub and Spoke” Networking

• AWS: Transit Gateway (TGW) connecting multiple VPCs.
• GCP: Shared VPC (XPN). A Host Project shares subnets with Service Projects.
• Azure: VNet Peering to a Hub VNet (usually containing Azure Firewall).

A.7.2. Monitoring & Observability

• AWS: CloudWatch (Metrics + Logs) + X-Ray (Tracing).
• GCP: Cloud Operations Suite (formerly Stackdriver). Managed Prometheus.
• Azure: Azure Monitor + Application Insights.

A.7.3. Infrastructure as Code (IaC)

• AWS: CloudFormation (YAML), CDK (Python/TS).
• GCP: Deployment Manager (deprecated) -> Terraform (First class citizen).
• Azure: ARM Templates (JSON) -> Bicep (DSL).

A.8. Decision Framework: Which Cloud for MLOps?

No cloud is perfect. Choose based on your “Gravity.”

1. Choose GCP if:

   • You are deep int Kubernetes. GKE is unmatched.
   • You need TPUs for massive training runs (Trillion param).
   • You are a “Data Native” company using BigQuery.

2. Choose AWS if:

   • You want Control. EC2/EKS/Networking gives you knobs for everything.
   • You are heavily invested in the OSS ecosystem (Airflow, Ray) on primitives.
   • You need the broadest marketplace of 3rd party tools (Snowflake, Databricks run best here).

3. Choose Azure if:

   • You are a Microsoft Shop (Office 365, Active Directory).
   • You need OpenAI (GPT-4) exclusive access.
   • You want a pre-integrated “Enterprise” experience.

A.9. The “Hidden” Services Mapping

Documentation often skips the glue services that make MLOps work.

A.10. Deep Dive: The Networking “Plumbing”

The number one reason MLOps platforms fail in production is DNS, not CUDA.

A.10.1. Private Service Access (The “VPC Endpoint” War)

MLOps tools (SageMaker, Vertex) often run in the Cloud Provider’s VPC, not yours. You need a secure tunnel.

The “Split-Horizon” DNS Trap:

• The Problem: Your laptop resolves sagemaker.us-east-1.amazonaws.com to a Public IP (54.x.x.x). Your EC2 instance resolves it to a Private IP (10.x.x.x).
• The Bug: If you hardcode IPs, SSL breaks. If you check DNS, you might get the wrong one depending on where you run nslookup.
• The Rosetta Fix:
  • AWS: enableDnsHostnames + enableDnsSupport in VPC.
  • GCP: private.googleapis.com VIP.
  • Azure: Link the Private DNS Zone to the VNet.

A.10.2. Egress Filtering (The Firewall)

ML models love to pip install from the internet. Security teams hate it.

A.11. Infrastructure as Code: The Translation Layer

How to say “Bucket” in 3 languages.

A.11.1. The Storage Bucket

AWS (Terraform)

resource "aws_s3_bucket" "b" {
  bucket = "my-ml-data"
}
resource "aws_s3_bucket_server_side_encryption_configuration" "enc" {
  bucket = aws_s3_bucket.b.id
  rule {
    apply_server_side_encryption_by_default {
      sse_algorithm = "AES256"
    }
  }
}

GCP (Terraform)

resource "google_storage_bucket" "b" {
  name          = "my-ml-data"
  location      = "US"
  storage_class = "STANDARD"
  uniform_bucket_level_access = true
}

Azure (Terraform)

resource "azurerm_storage_account" "sa" {
  name                     = "mymlstorage"
  resource_group_name      = azurerm_resource_group.rg.name
  location                 = "East US"
  account_tier             = "Standard"
  account_replication_type = "LRS"
}
resource "azurerm_storage_container" "c" {
  name                  = "my-ml-data"
  storage_account_name  = azurerm_storage_account.sa.name
  container_access_type = "private"
}

A.11.2. The Managed Identity

AWS (IAM Role)

resource "aws_iam_role" "r" {
  name = "ml-exec-role"
  assume_role_policy = jsonencode({
    Version = "2012-10-17"
    Statement = [{
      Action = "sts:AssumeRole"
      Effect = "Allow"
      Principal = { Service = "sagemaker.amazonaws.com" }
    }]
  })
}

GCP (Service Account)

resource "google_service_account" "sa" {
  account_id   = "ml-exec-sa"
  display_name = "ML Execution Service Account"
}
resource "google_project_iam_member" "binding" {
  role   = "roles/storage.objectViewer"
  member = "serviceAccount:${google_service_account.sa.email}"
}

Azure (User Assigned Identity)

resource "azurerm_user_assigned_identity" "id" {
  location            = "East US"
  name                = "ml-exec-identity"
  resource_group_name = azurerm_resource_group.rg.name
}
resource "azurerm_role_assignment" "ra" {
  scope                = azurerm_storage_account.sa.id
  role_definition_name = "Storage Blob Data Reader"
  principal_id         = azurerm_user_assigned_identity.id.principal_id
}

A.12. Closing Thoughts: The “Lock-In” Myth

Engineers fear Vendor Lock-in. Managers fear “Lowest Common Denominator.”

• The Reality: If you use Kubernetes, Terraform, and Docker, you are 80% portable.
• The Trap: Avoiding AWS S3 presigned URLs because “GCP doesn’t do it exactly the same way” leads to building your own Auth Server. Don’t do it.
• The Strategy: Abstract at the Library level, not the Infrastructure level. Write a blob_storage.py wrapper that calls boto3 or google-cloud-storage based on an ENV var.

This Rosetta Stone is your passport. Use it to travel freely, but respect the local customs.