XenorsDeploying ML Models on AWS SageMaker
SageMaker manages training machines, model hosting and scaling so you do not run servers yourself. You pay for what runs, so cleaning up matters as much as deploying.
SageMaker building blocks
A training job runs your script on managed instances and writes model.tar.gz to S3. A model points to that artifact and a container. An endpoint hosts the model behind an HTTPS API. Alternatives include batch transform for offline scoring, serverless inference for spiky traffic and asynchronous inference for long requests.
| Need | Option |
|---|---|
| Low latency, steady traffic | Real-time endpoint |
| Spiky or rare traffic | Serverless inference |
| Large offline scoring | Batch transform |
Training script
SageMaker passes data and output locations through environment variables.
# train.py
import argparse, os, joblib, pandas as pd
from sklearn.ensemble import RandomForestClassifier
if __name__ == "__main__":
p = argparse.ArgumentParser()
p.add_argument("--n-estimators", type=int, default=100)
a = p.parse_args()
df = pd.read_csv(os.path.join(os.environ["SM_CHANNEL_TRAIN"], "train.csv"))
X, y = df.drop(columns="target"), df["target"]
model = RandomForestClassifier(n_estimators=a.n_estimators, random_state=0).fit(X, y)
joblib.dump(model, os.path.join(os.environ["SM_MODEL_DIR"], "model.joblib"))
def model_fn(model_dir): # used at inference time
return joblib.load(os.path.join(model_dir, "model.joblib"))Train and deploy with the Python SDK
Run this from a SageMaker notebook or any machine with AWS credentials and an execution role. Check current framework versions in the AWS documentation.
import sagemaker
from sagemaker.sklearn.estimator import SKLearn
session = sagemaker.Session()
role = sagemaker.get_execution_role() # or an IAM role ARN
train_uri = session.upload_data("train.csv", key_prefix="demo/train")
est = SKLearn(entry_point="train.py", role=role, instance_type="ml.m5.large",
framework_version="1.2-1", hyperparameters={"n-estimators": 200})
est.fit({"train": train_uri})
predictor = est.deploy(initial_instance_count=1, instance_type="ml.m5.large")
print(predictor.predict([[5.1, 3.5, 1.4, 0.2]]))Invoke from an application
Applications call the endpoint through the runtime API, never the training SDK.
import boto3, json
rt = boto3.client("sagemaker-runtime")
resp = rt.invoke_endpoint(EndpointName=predictor.endpoint_name,
ContentType="application/json",
Body=json.dumps([[5.1, 3.5, 1.4, 0.2]]))
print(json.loads(resp["Body"].read()))Control cost and security
Endpoints bill while they run, so delete them when you finish a lab.
predictor.delete_endpoint()
predictor.delete_model()Hands-on practice
Run the examples in the downloadable Jupyter notebook. The extra checks below focus on mistakes that appear in real projects, not just syntax.
Keep inference payloads small
import json
payload = json.dumps({'features':[0.2,1.4,3.1]})
print(len(payload.encode()), payload)31 {"features": [0.2, 1.4, 3.1]}Large payloads raise latency and can hit endpoint limits.
Delete test endpoints after validation
# boto3 example
import boto3
sm=boto3.client('sagemaker')
# sm.delete_endpoint(EndpointName='demo-endpoint')
print('cleanup command prepared')cleanup command prepared
Cloud tutorial resources should have an explicit cleanup step to prevent idle cost.
Self-check
- Can you explain why the first practice result is trustworthy rather than just plausible?
- What would you monitor or test before using this AWS SageMaker workflow in production?
- Which assumption in the example would break first on a larger or messier dataset?
Frequently asked questions
Is SageMaker free?
There is a limited free tier for new accounts, but real-time endpoints and training instances are billed by usage. Check current AWS pricing.
Real-time endpoint or serverless inference?
Use real-time for steady low-latency traffic and serverless for infrequent or spiky requests where cold starts are acceptable.
How do I update a deployed model?
Create a new model version and update the endpoint configuration, ideally with a canary or blue-green rollout.
Primary references
Use these official sources when exact syntax, defaults or definitions matter beyond the examples here.
