Use this glossary when SparkML, MLflow, feature-engineering, monitoring, and deployment terms start to blur together. Keep it beside the cheat sheet and resources, not in place of scenario practice.
High-yield terms
| Term |
Short meaning |
Why it matters on ML-PRO |
| SparkML |
Spark’s distributed ML library for pipelines, estimators, transformers, and scalable inference |
core model-development term |
| Nested run |
MLflow tracking pattern that groups child runs under a parent experiment context |
key advanced experimentation term |
| Alias |
Stable label pointing to a chosen registered model version |
key release-control term |
| Point-in-time correctness |
Feature lookup behavior that prevents leakage by using only information available at that moment |
one of the highest-value feature-engineering concepts |
| Online table |
Databricks feature-serving storage for low-latency applications |
key online-feature term |
| Lakehouse Monitoring |
Databricks monitoring surface for data and model-quality signals |
key drift and monitoring term |
| Drift metric |
statistical signal that tracks change in data or model behavior over time |
key monitoring decision term |
| Data parallelism |
split data across workers while training the same model structure |
key scaling strategy term |
| Model parallelism |
split model computation itself across resources |
key large-model scaling term |
| Optuna |
hyperparameter tuning framework used in Databricks workflows and often paired with MLflow logging |
key tuning term |
| Ray |
distributed compute framework often contrasted with Spark for ML workloads |
key scaling trade-off term |
| Databricks Asset Bundle |
packaging and deployment structure for Databricks assets and environment promotion |
key MLOps term |
| Blue-green deployment |
deployment strategy that shifts traffic between two environments with a clear cutover path |
key rollout term |
| Canary deployment |
rollout strategy that exposes a small portion of traffic first |
key blast-radius-control term |
| Custom PyFunc model |
MLflow model packaged through the pyfunc interface for custom serving logic |
key deployment-interface term |
| Deploy code strategy |
lifecycle approach where code and environment transitions manage how models move across stages |
key MLOps architecture term |
Commonly confused pairs
| Pair |
Keep this distinction clear |
| MLflow run vs registered model version |
experiment record versus release artifact |
| alias vs serving endpoint |
release pointer versus deployed inference interface |
| point-in-time correctness vs feature freshness |
leakage prevention versus recency of values |
| drift vs rollout regression |
gradual distribution or quality change versus bad release event |
| SparkML vs single-node model |
distributed pipeline fit versus local model path |
| Spark vs Ray |
different distributed-training ecosystems and trade-offs |
| retrain vs rollback |
create a new candidate versus restore a known good state |
If three terms blur together
| Cluster |
Fast separation |
| run / version / alias |
track the experiment, govern the releasable artifact, point the release control at the chosen version |
| drift / outage / rollout regression |
gradual change, service failure, or bad deployment event |
| SparkML / Ray / single-node training |
distributed Spark pipeline, alternative distributed framework, or local model path |
| blue-green / canary / rollback |
cutover strategy, partial rollout, or revert to a trusted prior state |
| point-in-time correctness / leakage / online features |
correct historical lookup, future-information contamination, or low-latency feature serving |
One-sentence memory hooks
- If the model scored well offline, ask whether the feature path and release path are still safe.
- If production gets worse, separate drift, feature bug, rollout regression, and serving failure before acting.
- If scaling is the issue, choose fit before size.
- If monitoring fires, decide whether the right action is retrain, rollback, block promotion, or fix upstream data.