Knowledge check

Instruction and application

In Progress

Your analytics team must query 2 TB/day of clickstream logs to power daily feature generation. Queries scan only a subset of columns repeatedly. Which storage format best supports scalability and cost efficiency?

A) Parquet (columnar) with partitioning by date/hourB) CSV in cloud object storageC) JSON lines with gzip compressionD) Relational DB rows exported daily.

A real-time event stream encoded as JSON is causing rising inference latency due to parsing overhead and schema drift. What should you do first to improve scalability?

A) Increase instance size (vertical scaling).B) Add retries to the consumer service.C) Switch to a binary format like Protobuf/Avro and enforce a schema registry.D) Split the topic into multiple Kafka partitions only.

Your training data shows 25% missing values across key features and inconsistent timestamp formats. Build costs are spiking. What change most directly reduces resource usage during training?

A) Add more GPU workers to finish faster.B) Use a larger batch size in the trainer.C) Cache intermediate features after imputation.D) Enforce upstream schema and validation to reject/repair records before storage.

A daily training job slows from two to five hours after a marketing campaign introduces duplicate records. What's the main resource impact and best immediate response?

A) CPU/GPU underutilization — increase workersB) Higher storage and compute from redundant samples — add deduplication in the ETL layer.C) Network saturation — add more bandwidth.D) Feature drift — lower the learning rate.

Your product catalogue comes from a third-party API that occasionally changes field names without notice, breaking feature generation. What's the best mitigation?

A) Poll the API more frequently to detect changes.B) Build a nightly cron to rename fields after the fact.C) Define a data contract/Schema ID with the provider and enforce schema validation and fallback to cached data.D) Move the model to a faster GPU.

Your team depends on a single cloud region and one GPU family for both training and inference. A regional capacity shortage blocks deployments. What's the most effective resilience strategy?

A) Design for portability (containers/Kubernetes) and pre-test alternative instance types or regions to enable quick failover.B) Raise your on-demand spending limit in that region.C) Reduce model size to fit CPU only.D) Wait for spot capacity to return.

You are deploying an ML-based recommendation engine that must serve 2,000 requests per second during peak traffic, with each inference taking 80 ms on current hardware. Which additional specification would most strengthen your capacity planning?

A) Ensure the model is retrained every month.B) Convert all input data to CSV format.C) Include both current and forecasted traffic in the capacity plan.D) Adopt a container-based deployment strategy.

Your news summarisation API expects 3× traffic during breaking events with a realistic mix of request sizes. What validation should you run first to confirm your plan?

A) Chaos testing onlyB) Load test at 3× with realistic payloads and think timeC) Stress test at 10× to find maxD) Benchmark single-node CPU

Your inference service must handle 1,000 queries per second (QPS) with 80 ms average service time per request on a single CPU core. What is the minimum number of CPU cores needed (ignore overhead)?

A) 50 coresB) 100 coresC) 80 coresD) 125 cores

A training job completes in 12 hours on one GPU. You need it done in ~3 hours and assume near-linear scaling. How many GPUs should you provision (ignoring communication overhead)?

A) 2 GPUsB) 3 GPUsC) 4 GPUsD) 6 GPUs

Evaluate your understanding of this unit by completing the

Knowledge check.

Action item: Knowledge check

In progress