Private Cloud Deployment

1# AWS VPC with private subnets and NAT
2resource "aws_vpc" "atlas" {
3  cidr_block           = "10.0.0.0/16"
4  enable_dns_hostnames = true
5  enable_dns_support   = true
6}
7
8resource "aws_subnet" "private" {
9  count             = 3
10  vpc_id            = aws_vpc.atlas.id
11  cidr_block        = "10.0.${count.index + 1}.0/24"
12  availability_zone = data.aws_availability_zones.available.names[count.index]
13}
14
15resource "aws_security_group" "atlas" {
16  vpc_id = aws_vpc.atlas.id
17
18  ingress {
19    from_port   = 443
20    to_port     = 443
21    protocol    = "tcp"
22    cidr_blocks = ["10.0.0.0/16"]
23  }
24
25  egress {
26    from_port   = 0
27    to_port     = 0
28    protocol    = "-1"
29    cidr_blocks = ["0.0.0.0/0"]
30  }
31}

1# Azure Virtual Network with private endpoints
2resource "azurerm_virtual_network" "atlas" {
3  address_space = ["10.0.0.0/16"]
4}
5
6resource "azurerm_subnet" "private" {
7  address_prefixes                          = ["10.0.1.0/24"]
8  enforce_private_link_endpoint_network_policies = true
9}
10
11resource "azurerm_private_endpoint" "atlas" {
12  name                = "atlas-endpoint"
13  private_dns_zone_id = azurerm_private_dns_zone.atlas.id
14
15  private_service_connection {
16    name              = "atlas-connection"
17    is_manual_connection = false
18    subresource_names = ["Atlas"]
19  }
20}

1# GCP VPC without public IPs
2resource "google_compute_network" "atlas" {
3  auto_create_subnetworks = false
4}
5
6resource "google_compute_subnetwork" "private" {
7  ip_cidr_range = "10.0.1.0/24"
8  private_ip_google_access = true
9}
10
11resource "google_compute_router" "nat" {
12  name = "atlas-nat"
13}
14
15resource "google_compute_router_nat" "nat" {
16  name                               = "atlas-nat-config"
17  nat_ip_allocate_option            = "AUTO_ONLY"
18  source_subnetwork_ip_ranges_to_nat = "ALL_SUBNETWORKS_ALL_IP_RANGES"
19}

1# Calculate optimal reserved instance commitment
2import boto3
3
4ec2 = boto3.client('ec2')
5
6# Analyze on-demand usage patterns
7on_demand = ec2.describe_instances(Filters=[
8  {'Name': 'instance-state-name', 'Values': ['running']}
9])
10
11# Build RI recommendation based on:
12# - Consistent high-performance nodes (3-year RI)
13# - Variable inference nodes (1-year RI or on-demand)
14# - Burst capacity (spot instances)
15
16# Expected savings: 70% for 3-year commitment
17annual_compute_cost = 120000  # estimated
18ri_annual_cost = annual_compute_cost * 0.3
19spot_savings = 40000  # 50% more for burstable capacity
20
21print(f"Reserved: ${ri_annual_cost:,.0f}")
22print(f"Spot savings: ${spot_savings:,.0f}")
23print(f"Total annual: ${ri_annual_cost + spot_savings:,.0f}")

1# Monitor actual resource utilization
2kubectl top nodes --containers
3
4# Check pod CPU/memory usage
5kubectl top pods -A --sort-by=memory
6
7# Right-size recommendations:
8# GPU Utilization < 30% → Reduce instance type
9# Memory Utilization < 40% → Reduce memory allocation
10# CPU Bursting frequent → Use larger instances (better $/performance)
11
12# Example AWS Graviton instances:
13# - 60% cheaper than x86
14# - Excellent for Arabic NLP workloads
15# - Reduced egress with local ML acceleration