Infra aws data platform review

AWS Data Platform Review

When to Use

• Designing or reviewing an AWS-native data platform
• Auditing IAM permissions for data pipelines
• Choosing between AWS managed services vs self-managed
• Optimizing S3 data lake costs
• Setting up Lake Formation for fine-grained data access control

---

S3 Data Lake Architecture

# Terraform: 3-zone data lake (Bronze/Silver/Gold)
resource "aws_s3_bucket" "zones" {
  for_each = toset(["bronze", "silver", "gold"])
  bucket   = "${local.prefix}-data-lake-${each.key}"
  tags     = merge(local.common_tags, { Zone = each.key })
}

# Encryption with separate KMS keys per zone
resource "aws_s3_bucket_server_side_encryption_configuration" "zones" {
  for_each = aws_s3_bucket.zones
  bucket   = each.value.id
  rule {
    apply_server_side_encryption_by_default {
      sse_algorithm     = "aws:kms"
      kms_master_key_id = aws_kms_key.zones[each.key].arn
    }
    bucket_key_enabled = true   # reduces KMS API calls by 99%
  }
}

# Lifecycle: tiered storage by zone age
resource "aws_s3_bucket_lifecycle_configuration" "bronze" {
  bucket = aws_s3_bucket.zones["bronze"].id
  rule {
    id     = "raw-data-tiering"
    status = "Enabled"
    transition {
      days          = 30
      storage_class = "STANDARD_IA"
    }
    transition {
      days          = 90
      storage_class = "GLACIER_IR"
    }
    expiration { days = 365 }
  }
}

# Block all public access
resource "aws_s3_bucket_public_access_block" "all" {
  for_each                = aws_s3_bucket.zones
  bucket                  = each.value.id
  block_public_acls       = true
  block_public_policy     = true
  ignore_public_acls      = true
  restrict_public_buckets = true
}

---

IAM — IRSA (IAM Roles for Service Accounts)

# Airflow on EKS: worker pods assume an IAM role without static keys
module "irsa_airflow" {
  source  = "terraform-aws-modules/iam/aws//modules/iam-role-for-service-accounts-eks"
  version = "~> 5.0"

  role_name = "${local.prefix}-airflow-worker"

  oidc_providers = {
    main = {
      provider_arn               = module.eks.oidc_provider_arn
      namespace_service_accounts = ["airflow:airflow-worker"]
    }
  }

  role_policy_arns = {
    s3_access    = aws_iam_policy.airflow_s3.arn
    glue_access  = aws_iam_policy.airflow_glue.arn
  }
}

# Scoped S3 policy
resource "aws_iam_policy" "airflow_s3" {
  name   = "${local.prefix}-airflow-s3"
  policy = jsonencode({
    Version = "2012-10-17"
    Statement = [
      {
        Effect   = "Allow"
        Action   = ["s3:GetObject", "s3:PutObject", "s3:DeleteObject"]
        Resource = ["${aws_s3_bucket.zones["bronze"].arn}/*",
                    "${aws_s3_bucket.zones["silver"].arn}/*"]
      },
      {
        Effect   = "Allow"
        Action   = ["s3:ListBucket"]
        Resource = [aws_s3_bucket.zones["bronze"].arn,
                    aws_s3_bucket.zones["silver"].arn]
      }
    ]
  })
}

---

MSK Kafka Configuration

resource "aws_msk_cluster" "main" {
  cluster_name           = "${local.prefix}-kafka"
  kafka_version          = "3.5.1"
  number_of_broker_nodes = 3

  broker_node_group_info {
    instance_type   = "kafka.m5.xlarge"
    client_subnets  = module.vpc.private_subnets
    storage_info {
      ebs_storage_info {
        volume_size                = 1000   # GB per broker
        provisioned_throughput {
          enabled           = true
          volume_throughput = 250           # MiB/s
        }
      }
    }
    security_groups = [aws_security_group.msk.id]
  }

  encryption_info {
    encryption_in_transit {
      client_broker = "TLS"
      in_cluster    = true
    }
    at_rest_encryption_info {
      encryption_at_rest_kms_key_arn = aws_kms_key.msk.arn
    }
  }

  client_authentication {
    sasl {
      iam = true    # SASL/IAM auth — no credentials to manage
    }
  }

  configuration_info {
    arn      = aws_msk_configuration.main.arn
    revision = aws_msk_configuration.main.latest_revision
  }
}

resource "aws_msk_configuration" "main" {
  kafka_versions = ["3.5.1"]
  name           = "${local.prefix}-kafka-config"
  server_properties = <<-EOF
    auto.create.topics.enable=false
    log.retention.hours=168
    num.partitions=3
    default.replication.factor=3
    min.insync.replicas=2
    compression.type=lz4
  EOF
}

---

Lake Formation — Fine-Grained Access Control

# Register S3 location with Lake Formation
resource "aws_lakeformation_resource" "data_lake" {
  arn      = aws_s3_bucket.zones["gold"].arn
  role_arn = aws_iam_role.lakeformation.arn
}

# Grant table-level access to analytics team
resource "aws_lakeformation_permissions" "analytics_read" {
  principal   = aws_iam_role.analytics_team.arn
  permissions = ["SELECT"]
  permissions_with_grant_option = []

  table {
    database_name = "gold"
    name          = "fact_orders"
  }
}

# Column-level security: hide PII columns from non-privileged role
resource "aws_lakeformation_permissions" "pii_columns" {
  principal   = aws_iam_role.data_scientist.arn
  permissions = ["SELECT"]

  table_with_columns {
    database_name = "gold"
    name          = "customers"
    column_names  = ["customer_id", "order_count", "total_spend"]
    # excluded: email, phone, address
  }
}

---

MWAA vs Self-Managed Airflow on EKS

Aspect	MWAA	EKS Helm Chart
Setup effort	Low (minutes)	High (days)
Cost (2 schedulers)	~$300/mo + workers	~$150/mo + workers
Airflow version	Limited (2-3 supported)	Any version
Plugin support	Limited	Unlimited
KubernetesExecutor	No	Yes
Custom Python deps	S3 requirements.txt	Dockerfile
HA	Managed	Manual (2+ schedulers)
Scaling	Limited	Full autoscaling
Best for	Small teams, quick start	Large teams, custom needs

---

Glue vs EMR vs Spark-on-EKS

Aspect	AWS Glue	Amazon EMR	Spark on EKS
Management	Serverless	Semi-managed	Self-managed
Startup time	1-2 min	5-10 min	30-60s (reuse)
Cost model	Per DPU-hour	Per EC2-hour	Per pod resource
Spot support	Limited	Full	Full
Custom jars	Limited	Full	Full
Iceberg support	Via Glue catalog	Full	Full
Best for	Simple ETL, serverless	Large Spark jobs	Kubernetes platform

---

Cost Optimization

# S3 Intelligent-Tiering: auto-tier objects not accessed in 30 days
resource "aws_s3_bucket_intelligent_tiering_configuration" "bronze" {
  bucket = aws_s3_bucket.zones["bronze"].id
  name   = "auto-tiering"
  status = "Enabled"

  tiering {
    access_tier = "DEEP_ARCHIVE_ACCESS"
    days        = 180
  }
}

# EMR spot configuration
resource "aws_emr_instance_fleet" "task" {
  cluster_id = aws_emr_cluster.main.id
  name       = "task-fleet"
  target_spot_capacity = 20   # 20 units of spot (80% cheaper)
  target_on_demand_capacity = 2

  instance_type_configs {
    instance_type     = "m5.xlarge"
    bid_price_as_percentage_of_on_demand_price = 100
    ebs_config {
      size = 100
      type = "gp3"
    }
  }
}

---

VPC Networking for Data Platform

module "vpc" {
  source  = "terraform-aws-modules/vpc/aws"
  version = "~> 5.0"

  name = "${local.prefix}-data-platform"
  cidr = "10.0.0.0/16"

  azs              = ["us-east-1a", "us-east-1b", "us-east-1c"]
  private_subnets  = ["10.0.1.0/24", "10.0.2.0/24", "10.0.3.0/24"]
  public_subnets   = ["10.0.101.0/24", "10.0.102.0/24", "10.0.103.0/24"]

  enable_nat_gateway     = true
  single_nat_gateway     = false  # one per AZ for HA
  enable_vpn_gateway     = false

  # S3 VPC endpoint (free data transfer within VPC)
  enable_s3_endpoint     = true
  # Glue VPC endpoint
  enable_glue_endpoint   = true
}

---

Review Checklist

[ ] S3 buckets: encryption (KMS), versioning, public access blocked, lifecycle
[ ] IAM: IRSA for EKS workloads (no static keys), least-privilege policies
[ ] MSK: TLS + IAM auth, min.insync.replicas=2, encryption at rest
[ ] Lake Formation: column-level security for PII tables
[ ] VPC: private subnets for all data services, S3/Glue VPC endpoints
[ ] EKS: node groups in private subnets, IRSA enabled
[ ] CloudTrail: enabled for all regions, S3 bucket with lifecycle
[ ] Cost: S3 Intelligent-Tiering, Spot for EMR/EKS batch, Savings Plans
[ ] Monitoring: CloudWatch → Prometheus metrics via CW exporter
[ ] Backups: automated RDS snapshots (7+ day retention), S3 versioning

---

Anti-Patterns

1. Static AWS credentials in Airflow connections — rotate every 90 days or face locked pipelines; use IRSA on EKS or instance profiles on EC2.
2. S3 public access not blocked — one misconfigured bucket policy exposes entire data lake; always set restrict_public_buckets = true.
3. No S3 lifecycle rules on bronze zone — raw data accumulates indefinitely; raw zone grows 10x faster than gold zone.
4. Lake Formation disabled with broad S3 bucket policies — no row/column-level control; enable Lake Formation for all gold/silver tables.
5. No VPC endpoints for S3/Glue — data transfers via internet gateway incur NAT Gateway data processing costs; S3 VPC endpoint is free.

---

References

• AWS Lake Formation: docs.aws.amazon.com/lake-formation/latest/dg/
• MSK best practices: docs.aws.amazon.com/msk/latest/developerguide/bestpractices.html
• IRSA: docs.aws.amazon.com/eks/latest/userguide/iam-roles-for-service-accounts.html
• EMR on EKS: docs.aws.amazon.com/emr/latest/EMR-on-EKS-DevelopmentGuide/
• Related skills: [[terraform-data]], [[infra-kubernetes-security-audit]], [[de-cost-optimization]]