Batch Transform¶

Overview¶

Run inference on large datasets without deploying an endpoint.

When to Use¶

Large dataset inference
No real-time requirements
Periodic batch predictions
Cost optimization (no always-on endpoint)

Creating a Batch Transform Job¶

from sagemaker.transformer import Transformer

transformer = Transformer(
    model_name="my-model",
    instance_count=2,
    instance_type="ml.m5.xlarge",
    output_path="s3://bucket/batch-output/",
    strategy="SingleRecord",  # or "MultiRecord"
    assemble_with="Line",
    max_payload=6  # MB
)

transformer.transform(
    data="s3://bucket/batch-input/",
    content_type="text/csv",
    split_type="Line"
)

transformer.wait()

Configuration Options¶

Parameter	Options	Description
strategy	SingleRecord, MultiRecord	How to batch records
split_type	Line, RecordIO, None	How to split input
assemble_with	Line, None	How to assemble output
max_payload	0-100 MB	Max payload per request
max_concurrent_transforms	1-100	Parallel processing

Data Formats¶

Input¶

s3://bucket/input/
├── file1.csv
├── file2.csv
└── file3.csv

Output¶

s3://bucket/output/
├── file1.csv.out
├── file2.csv.out
└── file3.csv.out

Batch Transform vs Endpoint¶

Aspect	Batch Transform	Real-time Endpoint
Use Case	Large batch processing	Real-time predictions
Latency	Minutes to hours	Milliseconds
Cost	Pay per job	Pay per hour
Scaling	Automatic	Manual/Auto
Infrastructure	Transient	Persistent

Best Practices¶

!!! tip "Optimization" 1. Use MultiRecord strategy for throughput 2. Increase max_concurrent_transforms for parallelism 3. Use appropriate instance types for your workload 4. Partition input data for faster processing

Exam Tips¶

!!! warning "Key Points" - Batch Transform for offline, large-scale inference - No endpoint maintenance required - Cost-effective for infrequent predictions - Supports data distribution across instances