Hard MLOps Engineer questions
32 questions · 20 points each for a correct first answer.
- 1.p99 latency spikes during traffic burstsautoscaling, model-servinghard
- 2.Rollback that made things worsedeployment, feature-storehard
- 3.Why LLM servers use continuous batchingllm-serving, gpuhard
- 4.Same seed, different resultsreproducibility, gpuhard
- 5.Drift alerts firing all the timemonitoring, drifthard
- 6.Hundreds of low-traffic modelscost-optimisation, model-servinghard
- 7.Model affects its own training dataretraining, monitoringhard
- 8.Loading a model from a public hubsecurity, model-registryhard
- 9.Where to use tensor parallelismdistributed-training, llm-servinghard
- 10.KV cache with grouped-query attentionllm-serving, gpu-memoryhard
- 11.Memory to fully fine-tune 7B with Adamgpu-memory, distributed-traininghard
- 12.Shadow model sent real emailsdeployment-strategieshard
- 13.Drift check with a classifierdrift, model-monitoringhard
- 14.Canary that cannot catch quality dropsdeployment-strategies, experimentationhard
- 15.Daily batch missing late eventspipelines, data-qualityhard
- 16.Pipeline reused yesterday outputpipelines, reproducibilityhard
- 17.Offline and online scores disagreetraining-serving-skewhard
- 18.Batches of mixed text lengthsbatching, inference-optimisationhard
- 19.Accuracy fell after int8 quantisationquantisationhard
- 20.Same tag, different imagedocker, supply-chain-securityhard
- 21.Tail latency at low CPU usagekubernetes, latencyhard
- 22.More threads, no extra throughputpython, performancehard
- 23.State leaking between requestspython, model-servinghard
- 24.Streaming counts that run highstreaming, feature-storehard
- 25.Retries that make an outage worsereliabilityhard
- 26.Error budget for 99.9%slo, reliabilityhard
- 27.Retrieval got worse as the index grewvector-databases, raghard
- 28.User asks for data deletiongovernance, privacyhard
- 29.Semantic cache returns wrong answersllm-serving, llm-securityhard
- 30.GPU shows 100% but is underusedgpu-performance, monitoringhard
- 31.High drift, stable performancedrift, model-monitoringhard
- 32.Loss spike after resuming trainingdistributed-training, reproducibilityhard