Databricks project glossary

Use these meanings when reading the project walkthrough. The file/action column tells you where to make a decision or inspect its evidence in a generated project. Configuration names describe this checkout; use the matching Core release/build for your project revision.

Data and preprocessing

Term	Plain meaning	File or action
Record key	Stable columns identifying a source record and its published result. These are data identity, not a DataFrame's temporary row index.	Set record_key_columns in config/training.yml; verify uniqueness in the source.
Target / label	The actual outcome the estimator learns to predict, such as observed revenue or whether a customer churned. A prediction is not an actual label.	Set target_column in config/training.yml; maintain real outcomes upstream.
Pre-split	Fixed edits and eligibility filters applied before dividing labeled rows into training and holdout populations. They must not learn values from the complete dataset.	Select pre_split_recipe from config/pre_split.yml; decide target-filter reuse in config/scoring.yml before preview.
Fitted preprocessing	Transformations whose learned state, such as imputation means or scaling statistics, is fitted on training rows and reused for evaluation and prediction.	Define recipes in config/preprocessing.yml, select preprocessing_recipe in training.yml, then train a new version after edits.
Holdout	Rows reserved from model fitting for the final evaluation. They are distinct from the validation folds used during tuning.	Set split/window fields in config/training.yml; inspect the saved split membership and held-out metrics in training_report.
Cross-validation (CV)	Repeated train/validation splits within the training population, used to estimate or select model settings. It does not replace the final holdout.	Set cv_enabled, cv_type and related fields in config/training.yml; inspect saved fold reports.
Training window	The source time range eligible for a training run, before the final holdout boundary and other eligibility rules are applied.	Set training_window_mode, event/date parsing and window fields in config/training.yml.
Observation window	The prediction or request-time interval measured by monitoring. A completed performance window may wait for actual outcomes before assessment.	Set window_hours and label_delay_hours in monitoring_performance_policies; inspect monitoring reports. This differs from the training window and the job schedule.
Recipe	A named, ordered list of steps in config/pre_split.yml or config/preprocessing.yml. Only the selected recipe runs.	Select pre_split_recipe and preprocessing_recipe in config/training.yml. See the recipe catalog.
Preprocessing probe	An optional check that replays the saved preprocessing on up to 256 holdout rows, whole, repeated, in chunks and reversed. It does not refit or block promotion.	Set save_preprocessing_diagnostics: true; read preprocessing diagnostics.
Row, group, window and global context	What a step needs at prediction time: one row and saved state (row), all rows of a group (group), ordered earlier rows (window) or the whole batch (global).	Read what context means.
Outcome availability	When a real label became available for use. This can be later than the underlying event or prediction.	Configure result_available_at_column and cutoff rules for training; configure monitoring_result_available_at_column for monitoring labels.

Read the recipe catalog, training field reference and context-dependent preprocessing guide. For the target-reading filter error, see troubleshooting.

Model lifecycle and scoring

Term	Plain meaning	File or action
Candidate	A fitted, registered model version with saved evaluation evidence. Registration alone does not make it active.	Run the train job, then inspect training_report and its concrete version.
Champion	The active model version selected through the lifecycle's champion alias. The alias can move; the concrete version identifies one saved artifact.	Use the approved activation action and inspect its receipt. score_model_selection: champion resolves the active version for scoring.
Challenger	A candidate staged for comparison or an approval decision. Its existence does not route production requests or replace the champion.	Inspect the challenger alias and saved candidate report; use approve/reject according to policy.
Alias	A named pointer such as champion or challenger to a registered version. It is mutable, unlike a concrete version number.	Inspect registry aliases and lifecycle receipts. Use score_model_selection: pinned_version with model_version when a fixed score version is required.
Promotion / activation	A guarded decision to make a candidate active. A passing training task or deployment alone is insufficient.	Set promotion_policy, metric and quality thresholds in config/training.yml; inspect the activation receipt.
Competition	Several estimators compete for one target under shared data, split and selection rules. Only the activated winner becomes the production model.	Use training_layout: model_competition and candidate entries in config/training.yml.
Model set / multi-target	One coordinated release containing independent models for several targets, such as revenue and cost. It is not a competition for one winner.	Use training_layout: multi_target; define branches in config/training.yml, and set activation/publication policy in config/inference.yml. See model sets and its replacement rule.
Scoring eligibility	Rules determining whether an input receives a prediction. Excluded rows remain in the result with their status and reason.	Configure mode, target-step skipping and rules in config/scoring.yml before fitting the version.
Publication	Committing predictions and their metadata to the output Delta table, optionally exposing selected model-set views. This does not promote a registry version.	Set output and publication policy in config/inference.yml; inspect the score task and committed output.
Change Data Feed (CDF)	Delta's record of source inserts, updates and deletes over retained versions. Incremental scoring uses it to discover source changes after a saved checkpoint.	Enable CDF on the configured source upstream. Ordinary single/competition scoring accepts inserts; only model sets expose model_set.on_source_change in config/inference.yml for readable corrections. Use the documented recovery action if required history expires.
Full rebuild	Recompute the configured prediction population under the selected model rather than only processing changes since a checkpoint.	Review model/source change policy and the explicit recovery action in config/inference.yml; inspect the resulting output generation.
Expected champion and none	The champion version an operator action expects to replace. The action fails if the alias moved. none means no champion exists yet.	Copy it from the report's next actions. See operator actions.
previous_champion / previous_challenger	History aliases. Promotion moves the old champion to previous_champion; a new candidate moves the displaced challenger to previous_challenger.	Inspect registry aliases. See challenger history.
next_actions	Ready-to-copy parameters in the training report for the next approve, reject or rollback run.	Open training_report > Output. See operator actions.
lifecycle_action	The train job's Run now parameter: train, approve, reject or rollback. Only train fits a model.	Set it with Run now with different parameters. See operator actions.
Comparison digest	A SHA-256 hash of the reviewed candidate, data, metric policy and expected champion. Approval checks it instead of reevaluating.	Usually resolved from the candidate's saved evidence. See approve without training.
Promotion receipt	The saved record of a completed alias change. Rollback needs it to restore the prior champion.	Copy promotion_receipt_json from the report. See reject or roll back.
Prediction receipt	The saved record of a committed scoring write: source identity, versions and output. Recovery reads it before any retry.	Inspect the score task output. See scoring.
Watermark	The last committed source Delta version. The next score run reads changes after it.	Never edit it by hand. See how new rows are discovered.
Batch follow_champion	Whether a successful champion change starts the score job (true) or not (false). It is fixed at deployment.	Set it in config/inference.yml, then redeploy. See operator actions.
Serving follow_champion	Whether a successful champion change starts the serving endpoint job, in shared and personal development targets alike (true) or leaves it to a manual run (false). It is read from YAML at run time.	Set serving_endpoint.follow_champion in config/serving.yml and deploy the updated file; no job-parameter change is needed. See the serving job.
Handoff	A saved request from one job to another, such as activation requesting scoring or serving, or scoring requesting monitoring. It carries concrete evidence rather than leaving the child to guess the current alias.	Set the relevant follow_champion in config/inference.yml or config/serving.yml; follow the native child job in the run graph.

Read registry actions, scoring and recovery and training policy fields.

Features and online requests

Term	Plain meaning	File or action
Feature group	An optional upstream job unit that reads one source, runs the configured transform, and writes its own feature output table.	Add a group in config/features.yml and its transform under src/features/groups/; refresh the feature graph, deploy and run the feature job.
Merged table	The optional output obtained by joining the configured base table with feature-group outputs using declared keys and timestamp rules. It is a source for training/scoring only when you select it.	Configure the merge in config/features.yml, then point training_table and score_source_table to its output. Empty groups: {} creates neither a feature job nor a merged table.
Feature lookup	Retrieval of explicitly named features from a Unity Catalog feature table by entity keys and optional time rules. It is separate from the template's feature-group merge job.	Configure feature_lookup in config/training.yml; verify key/feature metadata and input columns.
Point-in-time lookup	A lookup that selects historical features according to each observation's timestamp and configured lookback, avoiding the use of later values.	Configure timestamp lookup settings in feature_lookup; inspect saved training lookup/snapshot evidence.
Online publication / synchronization	Copying eligible feature data into the configured online feature service so an endpoint can retrieve it. Publishing features does not train a model or deploy an endpoint.	Configure and run the optional publication/sync steps in the online feature guide and generated SERVING.md.
Online lookup	A live entity-key lookup made while handling a prediction request. Missing or stale required features can reject the request.	Configure the model's online_features required columns/freshness rules and the serving feature lookup contract. Inspect the endpoint response and logs.
Pinned endpoint	A serving endpoint tied to a concrete registered model version and inspected input/output contract.	Use the pinned serving helpers and generated SERVING.md; keep the endpoint, version and inference-log identity together.
Rollout	A separate, guarded process controlling serving traffic or stage progression. A champion alias update alone does not establish that traffic changed.	Configure the optional config/serving.yml rollout and inspect saved health/decision evidence.

Read feature groups and merged tables, online features and serving rollout.

Runtime, engine and operations

Term	Plain meaning	File or action
Engine	The DataFrame execution implementation selected for fitted Core preprocessing, such as pandas or Polars. It is not a package version or cluster size.	Set engine in config/training.yml; follow the supported engine/route combinations in the inference reference.
Runtime	The installed Python, Skyulf and dependency environment required to load and execute a saved artifact. A different engine or compute target does not make incompatible dependencies safe.	Inspect deployment/*requirements.txt, the saved artifact manifest and job environment. Install the matching release/build.
Inference mode	The route that applies a saved model, such as local scoring or the supported Spark UDF route. Spark inference does not turn training into Spark-native fitting.	Set batch_compute and Spark UDF options in config/inference.yml; verify the selected artifact's support boundary.
Admission / certified package	Spark workers score only a saved MLflow package with a safety certificate. Admission rechecks it and rejects unsupported steps.	Log a new certified version if admission fails. See distributed inference.
workflow_contract	A version marker on the generated job tasks. The runtime rejects a deployed graph and project that disagree.	Deploy the matching files together. See project upgrade.
Production-mode target	The shared test, syst and prod targets. Only they enroll models in the shared monitoring store; development targets can opt in to their own separate store.	Select the target with -t. See deployment.
Monitoring enrollment	The shared inventory entry binding a project/model/version to its prediction evidence, labels and policy. It is separate from a dashboard display.	Inspect activation register_monitor and the score-to-monitor handoff; configure monitoring target variables.
Dashboard refresh	Rerunning the dashboard's published datasets over saved monitoring evidence. It neither computes a new observation nor retrains the model.	Configure the native refresh task per target in deployment/monitoring.yml; inspect the monitoring job separately.
Drift	A measured change in feature distributions or declared schema checks. It is not automatically a loss in predictive quality.	Inspect the saved monitoring report and drift thresholds; enable guarded retraining separately if desired.
Performance degradation	A measured worsening against a matching baseline under an explicit metric, tolerance, window and label-coverage policy. Missing labels are unavailable evidence.	Configure monitoring_performance_policies; read the saved window and baseline evidence before interpreting its verdict.

Read inference artifacts and runtime checks, Spark execution boundaries and the monitoring overview.