Q: 8
Your team has a model deployed to a Vertex Al endpoint You have created a Vertex Al pipeline that automates the model training process and is triggered by a Cloud Function. You need to prioritize keeping the model up-to-date, but also minimize retraining costs. How should you configure retraining'?
Options
Discussion
D . Feature drift is directly tied to data changes that impact model accuracy, so retraining only triggers when it's actually needed. Costs stay low since you're not retraining just on a set schedule or for random anomalies. Agree?
Option D not B. Only D triggers on real feature drift so you don't retrain for no reason.
D . Had something like this in a mock and feature drift was flagged as the main signal for retraining triggers, since it points to real shifts in input data distribution. That way, you avoid kicking off expensive retrains just for random anomalies. Pretty sure that's what Google wants here but open to debate.
Why D over C? Model monitoring with feature drift directly ties retraining to real data shifts, not just any anomaly. Isn't C kind of a trap since anomalies can be short-term or unrelated to concept drift?
C/D? Both are triggered by monitoring, but drift (D) is more about actual changes in input data distribution which directly affects model performance. C talks anomalies, but not all anomalies need retrain. I think D fits the balance between cost and model freshness best but open if anyone thinks otherwise.
Yeah, D is it for this one.
B for me. Setting up a Cloud Scheduler with your preferred frequency lets you fully control costs, since retrains happen only as often as you allow and fit your budget. Pub/Sub event triggers (like feature drift in D) could end up being unpredictable cost-wise if there's noisy data. Not totally sure if that's what Google expects, but B seems like the safe route for minimizing spend. Anyone else see it this way?
Yeah, D here. Feature drift is a more targeted trigger for retraining than just anomaly alerts.
C. not D
D or maybe C? Feature drift (D) directly targets when retraining is actually needed, not just any weird anomaly. C could be a trap because not all anomalies mean the model's stale. Anyone think C fits better if cost wasn't a factor?
Be respectful. No spam.