Deployment and Serving - Databricks ML Professional¶

Overview¶

This section covers model deployment and serving patterns, representing 15% of the exam. You need to understand serving endpoints, deployment strategies, inference tables, and production deployment patterns.

📖 Model Serving - Serving overview 📖 Inference Tables - Request logging

Key Topics¶

1. Model Serving Endpoints¶

📖 Serverless Serving - Endpoint management

Serving Options: | Type | Description | Best For | |------|-------------|----------| | Serverless | Fully managed, auto-scaling | Most production workloads | | Provisioned throughput | Dedicated compute, guaranteed performance | Latency-sensitive apps |

Key Concepts: - Endpoints auto-scale based on traffic (including scale-to-zero) - REST API for predictions: POST request with JSON payload - Authentication via API keys or OAuth tokens - Endpoint configuration specifies model version, compute size, and scaling - Multiple model versions can be served from a single endpoint

import requests

# Query serving endpoint
response = requests.post(
    f"https://{workspace_url}/serving-endpoints/{endpoint_name}/invocations",
    headers={"Authorization": f"Bearer {token}"},
    json={"dataframe_records": [{"feature1": 1.0, "feature2": "value"}]}
)
predictions = response.json()

2. Deployment Strategies¶

Key Concepts: - Blue/green: maintain two environments, switch DNS to route traffic - Canary deployment: start with 5-10% traffic to new model, increase if healthy - Shadow deployment: test new model with real traffic without affecting users - Feature flags enable/disable model versions without redeployment - Rollback strategy: revert to previous model version immediately on issues

3. Inference Tables¶

📖 Inference Tables - Request logging

Key Concepts: - Automatically log all requests and responses from serving endpoints - Stored as Delta tables for analysis and monitoring - Capture: inputs, outputs, timestamps, latency, endpoint metadata - Enable ground truth joining for model evaluation - Track prediction distributions for drift detection - Retention policies manage storage for high-volume endpoints

4. Batch Inference Deployment¶

# Batch inference with mlflow.pyfunc.spark_udf
predict_udf = mlflow.pyfunc.spark_udf(spark, model_uri="models:/my_model@champion")

# Score large dataset
predictions = (spark.table("features")
    .withColumn("prediction", predict_udf(struct("f1", "f2", "f3")))
    .write.format("delta").mode("overwrite")
    .saveAsTable("predictions"))

Key Concepts: - Use model aliases (@champion) to always score with the production model - Batch inference is scheduled as Databricks Workflows - Results stored in Delta tables for downstream consumption - Suitable for daily/hourly scoring of large datasets

5. Production Readiness¶

Key Concepts: - Health checks: verify endpoint is responding correctly - Circuit breakers: stop sending traffic to unhealthy endpoints - Load testing: validate throughput and latency before production - Latency budgets: define acceptable response time SLAs - Error handling: graceful degradation when model fails - Cost monitoring: track serving compute costs per endpoint

Exam Tips for This Domain¶

1. Deployment strategies - Know canary, blue/green, and shadow patterns
2. Inference tables - Automatic logging for monitoring and drift detection
3. Serverless vs provisioned - Tradeoffs between flexibility and guaranteed performance
4. Model aliases - Use @champion for production model references
5. Batch vs real-time - Know when each pattern is appropriate

Documentation Links Summary¶