NCP-ADS Practice Questions: 20 Exam-Style Scenarios Explained (2026)

A team ports a pandas ETL job to cuDF. Most steps get faster, but one step now dominates the runtime: a row-wise .apply() that calls a Python function to normalize product codes with string slicing and conditionals. The dataset is 80 million rows. What is the most effective fix?

A 200 GB Parquet dataset must be joined and aggregated, but the workstation has a single GPU with 80 GB of memory. The team wants to stay in the RAPIDS ecosystem. Which TWO approaches make this workload feasible? (Select TWO)

A analytics group has thousands of lines of existing pandas code and wants GPU speedups this quarter without a rewrite. Some of their code uses uncommon pandas corner cases. Which RAPIDS capability fits this constraint best?

During a cuDF merge of a 500-million-row fact table with a dimension table, GPU memory usage spikes far above the input sizes and the job dies. The join keys are 64-bit integers with only 40,000 distinct values, and several object-dtype string columns ride along. What change most directly reduces the memory spike?

A fraud model must respond to online transactions with p95 latency under 50 ms, and traffic arrives as single requests, one prediction at a time. The team serves it from a GPU. Which serving approach uses the GPU efficiently under this constraint?

Two data scientists train the same cuML model from the same notebook and get different metrics on different machines. The team needs GPU training runs that reproduce across laptops, CI, and the cluster. Which TWO practices most directly deliver that? (Select TWO)

A GPU-accelerated churn model has been in production for five months. Accuracy on the monthly holdout looks stable, but the business reports the model increasingly misses a new customer segment. Which monitoring addition would surface this class of problem earliest?

A model trains in a RAPIDS 25.02 container but is served from a hand-built Python environment on the inference hosts, where preprocessing occasionally produces different encodings than training did. Which practice eliminates this class of training/serving skew?

A 300-million-row cuDF DataFrame has a transaction_amount column with 4% missing values and a heavy right skew from a small number of very large transactions. The column feeds a linear model. What is the most defensible single imputation choice?

A features table includes merchant_id with 1.2 million distinct values, and the team one-hot encodes it before training a cuML gradient-boosted model, which promptly exhausts GPU memory. Traffic patterns per merchant are believed to be predictive. What should the team do instead?

Log text arrives with mixed casing, stray whitespace, and embedded device codes that must be extracted into their own column. The pipeline is cuDF end to end and processes 50 million rows hourly. Which TWO techniques keep this preparation step fast on the GPU? (Select TWO)

A feature engineering pipeline crashes with an out-of-memory error on a 40 GB GPU. Profiling shows the working DataFrame is 21 GB, all numeric columns are float64 by default, and precision beyond float32 is not required by the downstream model. What is the highest-leverage first change?

A workstation has four GPUs, and a Dask-cuDF workload should use all of them. Which cluster setup is correct?

A team retrains models nightly on cloud GPUs. The job takes about three hours, checkpoints every ten minutes, and can resume from the latest checkpoint. Finance wants the GPU bill cut without touching the schedule. Which change delivers the largest saving with acceptable risk?

A scikit-learn random forest takes six hours to run a hyperparameter search on CPU. The team moves the search to cuML on a single GPU. Beyond the drop-in speedup, which practice makes the sweep itself most efficient?

An XGBoost model trains on a 90-million-row engineered feature set that already lives in GPU memory as a cuDF DataFrame. Training on CPU takes 70 minutes. What is the correct way to move training onto the GPU?

A fraud dataset has 0.4% positive labels. A cuML classifier reports 99.6% accuracy, and the team is celebrating. Which TWO changes give an honest picture of model quality? (Select TWO)

An analyst needs exploratory statistics (grouped aggregations, correlations, quantiles) on a 150-million-row dataset that already sits in GPU memory. Their current habit is sampling 1% into pandas "so it is fast enough." What should they do instead?

A payments team models accounts as nodes and transfers as edges, roughly 2 billion edges, and wants influence scores to prioritize fraud investigations. Their NetworkX prototype works on a 1% sample but cannot scale. What is the RAPIDS-native path?

An analyst wants to visualize the geographic distribution of 120 million delivery points to spot density patterns. A matplotlib scatter plot of all points locks the notebook, and a 50,000-point sample hides the structure. What is the right approach?