Databricks ML Professional flashcards
Databricks ML Professional · Build recall with a focused review session.
33 cards. Read the term and the definition here. The deck at the top of this page is optional.
- FeatureLookup
- A Feature Store specification that joins stored features into a training set by matching a lookup key column to the feature table's primary key.
- create_training_set
- Feature Store API that builds a training DataFrame by joining a label DataFrame with feature tables via specified FeatureLookup objects.
- Online feature store
- A low-latency store published from a Feature Store table, used at inference time to fetch feature values for individual requests.
- MLflow signature
- A recorded input and output schema for a model that enables validation of request payloads at serving time and documents expected columns.
- PythonModel load_context
- A method on a custom pyfunc model that loads artifacts once when the model is loaded, making them available to predict via the context object.
- SparkTrials
- An MLflow-integrated Hyperopt trials class that distributes objective evaluations across Spark workers and logs each trial to MLflow.
- MLflow autologging
- Flavor-specific instrumentation that automatically records parameters, metrics, and model artifacts for supported libraries without explicit log calls.
- Registry stage transition
- Moving a model version among None, Staging, Production, and Archived, optionally archiving other versions already in the target stage.
- Unity Catalog model URI
- A three-level reference of catalog, schema, and model name, optionally with a version number, used to load or serve a registered model.
- Hyperopt fmin objective
- A function returning a scalar loss to minimize; fmin explores the search space and returns the best parameters and loss found.
- Feature table primary key
- A set of columns that uniquely identify rows in a Feature Store table and are used as join keys during training set creation.
- Training set label DataFrame
- The DataFrame passed to create_training_set that supplies the target and lookup keys, to which feature values are joined.
- MLflow run
- A single execution within an experiment that records parameters, metrics, tags, and artifacts for one training or evaluation attempt.
- Metric step
- A numeric counter attached to a logged metric that orders values over time, so later values at the same step overwrite earlier ones.
- PendingRegistration state
- A transient model version state during registration; success moves it to Ready, and failure moves it to FailedRegistration, which cannot be served.
- Registry webhook
- A subscription to model registry events such as version creation or stage transition that sends an HTTP notification to a configured URL.
- Job cluster
- A cluster created for a single Databricks job run and terminated at completion, isolating cost and configuration to that run.
- Delta time travel
- Querying a Delta table as of an earlier version or timestamp, enabling reproducible reads of training data from a prior state.
- Unity Catalog lineage
- Automatically captured relationships linking a model version to its originating runs, notebooks, and tables for traceability and auditing.
- Slice metrics
- Per-subgroup evaluation values logged under distinct metric keys so performance differences across segments remain visible in run comparisons.
- Lakehouse Monitoring
- A Databricks feature that computes profile and drift metric tables comparing recent data distributions against a baseline profile.
- Continuous trigger
- A job trigger mode that keeps the job running and restarts it after each completion, used for always-on streaming pipelines.
- Experiment
- A named container in MLflow that groups related runs, providing a shared workspace for comparing metrics and artifacts.
- Archive existing versions
- A transition option that moves other versions out of the target stage so exactly one version occupies Production or Staging.
- Model version alias
- A mutable named pointer, such as champion, to a model version that clients and endpoints can reference instead of a fixed version number.
- Serving endpoint payload
- A JSON request body using dataframe_records or dataframe_split keys that the endpoint converts into the model's expected input format.
- Scale to zero
- An endpoint setting that removes all capacity when idle, reducing cost at the price of cold-start latency on the next request.
- Provisioned concurrency
- Pre-warmed endpoint capacity kept ready to serve requests immediately, avoiding cold starts for latency-sensitive inference.
- spark_udf inference
- An MLflow helper that wraps a loaded pyfunc model as a Spark UDF, enabling distributed batch scoring over a DataFrame.
- Endpoint permissions
- Access controls on a serving endpoint that determine which users or groups may query it, separate from model privileges.
- Served entity
- The model version or feature specification an endpoint hosts, including the version or alias that determines which artifact answers requests.
- Streaming inference sink
- The output destination of a Structured Streaming scoring job, such as a Delta table, written with checkpointing for exactly-once processing.
- Dependency pinning at log time
- Recording exact library versions in the model's environment file so the serving container rebuilds an identical runtime and avoids import failures.