Amazon SageMaker Canvas

SageMaker Canvas provides a drag-and-drop interface for building ML models. Business analysts can build classification, regression, and time-series forecasting models without writing code or understanding ML algorithms.

What Canvas Does

Business Analyst
  │
  ├── Upload CSV/Connect to S3/Redshift/Snowflake
  ├── Select target column (what to predict)
  ├── Canvas builds model automatically
  │     ├── Automatic data preparation
  │     ├── Algorithm selection
  │     ├── Hyperparameter tuning
  │     └── Model validation
  └── Generate predictions (batch or real-time)

Model Types

Type	Use Case	Example
Binary Classification	Yes/No prediction	Will customer churn?
Multi-class Classification	Category prediction	What product category?
Numeric Prediction (Regression)	Number prediction	How much will they spend?
Time Series Forecasting	Future values	Forecast demand for next 30 days
ML Models (Tabular)	Any tabular data	AutoML on your data

Getting Started

1. Open Canvas

# Create SageMaker Canvas app
aws sagemaker create-app \
  --domain-id domain-xxxxx \
  --user-profile-name analyst-01 \
  --app-name canvas \
  --app-type JupyterLab

Then open Canvas from SageMaker Studio.

2. Connect Data

Canvas supports:

Upload: CSV, XLSX files directly
S3: Browse and select S3 buckets
Redshift: Query warehouse data
Snowflake: Query Snowflake data
Databricks: Query Databricks data

3. Build Model

# Or via SageMaker Canvas API
import boto3
 
canvas = boto3.client('sagemaker-canvas', region_name='us-east-1')
 
# Create model
canvas.create_model(
    modelId='canvas-auto',
    modelName='customer-churn-predictor',
    dataset='s3://my-bucket/dataset.csv'
)
 
# Start build
canvas.start_model_training(
    modelName='customer-churn-predictor',
    objective='classification',
    targetAttribute='churned'
)

Canvas Build Options

Quick Build

Builds in 20-30 minutes
Uses smaller dataset sample
Less accurate but faster
Good for: exploration, POC

Standard Build

Builds in 2-4 hours
Uses full dataset
More accurate
Good for: production models

Evaluating Models

Canvas provides:

Accuracy score
F1 score (classification)
RMSE (regression)
Feature importance (which columns matter most)
Confusion matrix (classification)
Predictions vs actuals (regression)

# Get model metrics
metrics = canvas.get_model_metrics(modelName='customer-churn-predictor')
print(metrics['binary_classification_metrics'])
# {'accuracy': 0.92, 'f1_score': 0.89, 'precision': 0.91, 'recall': 0.87}

Generating Predictions

Batch Predictions

# Generate batch predictions
canvas.batch_transform(
    modelName='customer-churn-predictor',
    dataset='s3://my-bucket/new-customers.csv',
    outputUri='s3://my-bucket/predictions/'
)

Real-time Predictions

# Register for real-time
canvas.deploy(
    modelName='customer-churn-predictor',
    inferenceContainers=['default']
)
 
# Invoke endpoint
runtime = boto3.client('sagemaker-runtime')
response = runtime.invoke_endpoint(
    EndpointName='canvas-customer-churn-predictor',
    ContentType='text/csv',
    Body='age=35,income=75000,tenure=24,...'
)

Time Series Forecasting

Canvas auto-generates forecasts:

# Build time series model
canvas.start_model_training(
    modelName='demand-forecast',
    objective='forecasting',
    targetAttribute='sales',
    timeSeriesConfiguration={
        'timeColumn': 'date',
        'forecastFrequency': 'D',  # Daily
        'forecastHorizon': 30     # Forecast 30 days
    }
)

Forecast outputs:

Point predictions
Confidence intervals (80%, 95%)
Trend, seasonality, holiday effects

Pricing

Component	Cost
Canvas app (SageMaker Studio)	Included in SageMaker Studio cost
Quick build	$0.40/hour
Standard build	$1.20/hour
Batch predictions	Free (uses inference)
Real-time predictions	Standard SageMaker inference pricing

Limits

Resource	Limit
Dataset size	100K rows, 100 columns
Training time	48 hours
Forecast horizon	730 periods
Model storage	100 models

References

Homepage: https://aws.amazon.com/sagemaker/canvas/
Documentation: https://docs.aws.amazon.com/sagemaker-canvas/
Pricing: https://aws.amazon.com/sagemaker/canvas/pricing/

Nuggets & Gotchas

Canvas models are NOT exportable — you must use Canvas to predict: Canvas produces a proprietary model artifact. You can’t download the model and deploy elsewhere. For portable models, use SageMaker Autopilot or train with Python.
Canvas has a 100K row / 100 column limit — for larger datasets, sample before importing: If your dataset is larger, sample down before importing into Canvas. Or use Athena to pre-aggregate.
Canvas doesn’t support all data types — dates must be in a recognized format: If your date column isn’t recognized, format it as ISO 8601 (YYYY-MM-DD) before importing.
Canvas time series forecasting requires a UNIQUE time series identifier — if you have multiple products/stores, configure group columns: A single forecast model handles multiple series (e.g., forecast each product separately in one model). Configure group columns in Canvas UI.
Canvas Quick Build is NOT for production — use Standard Build for any decision-making: Quick build uses a sample of data and faster hyperparameters. The accuracy is significantly lower than Standard Build.

cloudnative wiki

Explorer

SageMaker Canvas