Introductory MLOps Practices

NCA-ADS · 20 questions

• A public hospital district throughput desk notices the nightly RAPIDS pipeline runtime roughly doubles after a schema change. What monitoring action should they take first?
• An elections operations (non-partisan) data team sees overall pipeline time slip and stage timings show the feature-engineering step dominating. What should they do?
• A municipal fleet telematics desk has a flaky open-data pull that sometimes fails mid-run and risks corrupting nightly outputs. Which reliability pattern belongs in the monitored pipeline?
• A solid-waste route optimization analytics team schedules a citywide GPU retrain window and needs operational visibility during the job. What should they monitor?
• A county social-services case analytics team runs many district model trials and needs comparable history. What should they log in MLflow for each trial?
• A city planning zoning analytics team must pick a champion model before a public-health adjacent pilot. How should they compare candidates?
• A city open-data portal team is policy-blocked from SaaS experiment trackers but still needs comparable trial history. What should they do?
• A county public-health analytics team is asked to promote a production candidate whose only history is an undocumented notebook run. What should they require?
• A transit authority ridership desk has a winning cuML model and needs overnight batch scoring without retraining. What should they do?
• A water-utility sensor analytics team has a stored model artifact and must produce scores for next week's inspection list. What is the correct next step?
• An emergency-management fusion cell saves models for incident demand scoring and worries about silent overwrites. What persistence practice should they adopt?
• A housing-authority waitlist analytics team notices feature distributions shift after a new e-permitting form launches. What should production monitoring detect?
• A courts / clerk e-filing analytics team sees precision fall after a policy change shifts how labels are defined. What should they conclude?
• A library program attendance analytics team supporting flood-planning partners sees residual error rise on storm weeks for a deployed demand model. What should monitoring trigger?
• A parks and recreation demand desk wants reproducibility for a promoted attendance model. What should they store together as one artifact set?
• A school-district enrollment analytics team must ensure batch scoring uses the same columns the model saw at training. What should they pin with the artifact?
• A county assessor parcel desk is asked to promote a valuation model that lacks recorded hyperparameters and a data snapshot ID. What should they do?
• A 311/resident-service analytics team must decide whether CPU pandas or GPU cuDF better meets the weekly county extract SLA. What should guide the choice?
• A permitting and inspections bureau compared hosts and found transfer-plus-compute on GPU still beats end-to-end CPU time for their weekly feature job. What should they select?
• A fire/EMS response-time analytics team upgraded RAPIDS and is drafting a capacity request for more GPUs. What should they do first?