SageMaker Processing

Mental model

• Managed data processing jobs (preprocess, feature engineering, evaluation, dataset prep) on ephemeral compute.
• Similar to training jobs but meant for general processing containers.

Where it fits

• Data prep for training/evals, batch feature generation, bias checks, model eval pipelines.

Knobs that matter

• Instance type/count, volume size
• Network/VPC: only when needed; watch NAT costs
• Caching: use pipeline caching or reuse outputs in S3
• Container image: keep it small; dependency bloat hurts startup time

Pricing mental model

• Cost ≈ (#instances × hours) (like training), so right-size aggressively.

Terraform template (processing job skeleton)

resource "aws_sagemaker_processing_job" "proc" {
  processing_job_name = "${var.name}-proc"
  role_arn            = var.sm_role_arn

  app_specification {
    image_uri = var.image_uri
  }

  processing_resources {
    cluster_config {
      instance_type  = var.instance_type
      instance_count = var.instance_count
      volume_size_in_gb = 50
    }
  }

  processing_inputs {
    input_name = "input"
    s3_input {
      s3_uri       = var.input_s3_uri
      local_path   = "/opt/ml/processing/input"
      s3_data_type = "S3Prefix"
    }
  }

  processing_outputs {
    output_name = "output"
    s3_output {
      s3_uri      = var.output_s3_uri
      local_path  = "/opt/ml/processing/output"
    }
  }
}

variable "name" { type = string }
variable "sm_role_arn" { type = string }
variable "image_uri" { type = string }
variable "instance_type" { type = string default = "ml.m5.xlarge" }
variable "instance_count" { type = number default = 1 }
variable "input_s3_uri" { type = string }
variable "output_s3_uri" { type = string }

---