Machine Learning Associate ยท Deploying and serving models
Streaming inference with Delta Live Tables
Applying a model to a stream inside a Delta Live Tables pipeline, with managed autoscaling, versus Structured Streaming jobs.
Key points
- Delta Live Tables (now branded Lakeflow Spark Declarative Pipelines): declare a streaming table over an incremental source and apply a model to its rows, e.g. via a Spark UDF from a registered model.
- The pipeline manages orchestration, retries, optional compute autoscaling.
- Each update processes only new data.
- Hand-written Structured Streaming: same micro-batch engine, but you own cluster sizing, checkpoints, scheduling.
Source: https://docs.databricks.com/aws/en/dev-tools/cli/reference/pipelines-commands
Practice questions
What does a Delta Live Tables pipeline do that a plain notebook with Structured Streaming does not do for you?
- It manages orchestration, retries and compute for the declared datasets
- It removes the need for a registered model
- It converts streams into synchronous REST calls
- It guarantees sub-10-millisecond latency
Show answer
Answer: A. It manages orchestration, retries and compute for the declared datasets
A pipeline is a managed, declarative runtime: you describe datasets and it handles orchestration, retries and compute management. It does not remove the need for a model or change what a stream is.
Which statements describe streaming inference in a declarative pipeline correctly?
- Each update processes new data incrementally
- A model can be applied to rows of a streaming table
- Callers get a synchronous HTTP reply per record
- The pipeline retrains the model on every update
- Predictions can only be written to local files
Show answer
Answer: A. Each update processes new data incrementally, B. A model can be applied to rows of a streaming table
Pipelines process new data incrementally and the model is applied to rows of a streaming dataset. They are not request/response services and do not retrain the model automatically.
True or false: streaming inference means every single record is sent to a REST endpoint and the caller waits for the reply.
Show answer
Answer: False
False. That describes real-time serving. Streaming inference applies the model continuously to an incoming flow of records inside a pipeline, with no caller waiting.
Practice 8 questions on this topic in Bitzu
Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.
