Study Databricks ML-PRO Automated Retraining: key concepts, common traps, and exam decision cues.
On this page
Retraining questions are not really about automation alone. They are about deciding when a retraining trigger is justified and how a replacement model earns promotion safely.
Retraining map
Requirement
Better first instinct
retrain when drift or degradation passes a real threshold
triggered retraining workflow
choose the best candidate from retraining output
explicit top-model selection strategy
avoid promoting a worse model automatically
promotion gate after retraining
What the exam is really testing
If the stem says…
Strong reading
“triggered by data drift detection or performance degradation”
automation should be signal-driven, not arbitrary
“selecting top-performing models”
retraining needs a comparison and promotion rule
“automated retraining”
automated does not mean ungated
Decision order that usually wins
Separate the trigger from the promotion decision.
Confirm that the signal crossing a threshold is meaningful enough to justify retraining.
Define how new candidates will be compared after retraining runs.
Keep an explicit gate between “new model produced” and “new model promoted.”
Prefer automation that is evidence-driven and reversible.
ML-PRO does not reward blind automation. The better answer usually makes retraining faster while keeping candidate evaluation and release discipline intact.
Scenario triage
Scenario
Better first move
drift or degradation crosses a meaningful threshold
trigger retraining workflow
retraining produced several candidates
apply explicit model-selection logic
organization wants fully automatic promotion
add gates before release, not just triggers
metric wiggles but business effect is unclear
avoid noisy retraining loops
Common traps
Trap
Better rule
retraining on every small metric wiggle
use meaningful thresholds
promoting the newest retrained model automatically
winning candidate selection still matters
confusing retraining trigger with release decision
a triggered pipeline can still produce a rejected candidate