Medium MLOps Engineer questions
34 questions · 10 points each for a correct first answer.
- 1.Shadow deploymentdeployment, testingmedium
- 2.Scheduling a GPU podkubernetes, gpumedium
- 3.When to retrainretraining, ml-pipelinesmedium
- 4.Slow image buildsdocker, ci-cdmedium
- 5.Online and offline feature storesfeature-storemedium
- 6.Low GPU use under loadmodel-serving, gpumedium
- 7.Safe retries in pipelinesml-pipelines, orchestrationmedium
- 8.Labels arrive weeks latermonitoring, driftmedium
- 9.Computing PSI for two binsdrift, model-monitoringmedium
- 10.KV cache size for one long requestllm-serving, gpu-memorymedium
- 11.Joining features to training labelsfeature-store, data-leakagemedium
- 12.Old start date with catchup onorchestration, airflowmedium
- 13.Pod killed while loading the modelkubernetes, resource-managementmedium
- 14.Autoscaling a GPU inference serviceautoscaling, kubernetesmedium
- 15.Concurrent requests at peakcapacity-planning, latencymedium
- 16.Reading a KS drift statisticdrift, statisticsmedium
- 17.Serving many fine-tuned variantsllm-serving, loramedium
- 18.Smaller model with similar qualityinference-optimisationmedium
- 19.Switching embedding models in RAGrag, vector-databasesmedium
- 20.Purpose of a multi-stage builddockermedium
- 21.Deploying by alias, not versionmodel-registry, deploymentmedium
- 22.Automatic gate before promotionci-cd, model-evaluationmedium
- 23.Training on spot GPUsdistributed-training, cost-managementmedium
- 24.Speeding up training on 8 GPUsdistributed-trainingmedium
- 25.Why bf16 is easier than fp16mixed-precision, trainingmedium
- 26.Splitting requests into batchespython, batch-inferencemedium
- 27.What to log for each LLM callllm-observability, monitoringmedium
- 28.Fleet-wide p99 from many podsmonitoring, latencymedium
- 29.Online features stopped updatingfeature-store, monitoringmedium
- 30.CPU pods filling GPU nodeskubernetes, gpu-schedulingmedium
- 31.Silent wrong predictions from bad inputmodel-serving, data-validationmedium
- 32.Cost per 1,000 predictionscost-management, capacity-planningmedium
- 33.Instructions hidden in a documentllm-security, guardrailsmedium
- 34.p99 doubled right after a releaseincident-response, deploymentmedium