Lambda

When to use (ML/GenAI)

• Burst + spiky inference wrappers, lightweight tool APIs, async workers (SQS), glue code.
• Great for orchestration, pre/post-processing, fanout, not heavy GPU inference.

Knobs that matter

• Memory (also scales CPU), timeout, ephemeral storage, architecture (x86/ARM).
• Concurrency: reserved concurrency per function, account concurrency.
• Cold start: VPC attachment increases risk; use VPC endpoints; consider Provisioned Concurrency if p99 matters.
• Event source mappings: batch size, max batching window, DLQ, partial batch response (SQS).
• Streaming responses (HTTP response streaming) when you need token streaming semantics. ([Amazon Web Services, Inc.][1])

Pricing mental model

• Meter is requests + GB-seconds duration (compute time × memory). ([Amazon Web Services, Inc.][1])
• Mental model: “If I double memory, I roughly double $/sec but may halve runtime — optimize on cost per request not memory.”

Heuristics

• If runtime is > 30–60s, or needs heavy deps/GPU, move to ECS/EKS/EC2.
• If you’re paying big NAT bills for Lambda-in-VPC → add endpoints (S3/Dynamo/etc).

Terraform (basic Lambda + IAM + logs + reserved concurrency)

resource "aws_cloudwatch_log_group" "lg" {
  name              = "/aws/lambda/${var.name}"
  retention_in_days = 14
}

data "aws_iam_policy_document" "assume" {
  statement {
    effect = "Allow"
    principals { type = "Service", identifiers = ["lambda.amazonaws.com"] }
    actions = ["sts:AssumeRole"]
  }
}

resource "aws_iam_role" "role" {
  name               = "${var.name}-role"
  assume_role_policy = data.aws_iam_policy_document.assume.json
}

resource "aws_iam_role_policy_attachment" "basic" {
  role       = aws_iam_role.role.name
  policy_arn  = "arn:aws:iam::aws:policy/service-role/AWSLambdaBasicExecutionRole"
}

resource "aws_lambda_function" "fn" {
  function_name = var.name
  role          = aws_iam_role.role.arn
  runtime       = "python3.12"
  handler       = "app.handler"
  filename      = var.zip_path

  memory_size = 1024
  timeout     = 30

  environment { variables = var.env }

  # Uncomment if needed:
  # vpc_config { subnet_ids = var.private_subnet_ids, security_group_ids = [var.sg_id] }
}

resource "aws_lambda_function_concurrency" "reserved" {
  function_name                 = aws_lambda_function.fn.function_name
  reserved_concurrent_executions = 50
}

variable "name"    { type = string }
variable "zip_path"{ type = string }
variable "env"     { type = map(string) default = {} }

---