MLOps Engineer question bank
100 questions. Filter by level and skill, track what you've solved, and see how many people get each one right.
- What a model registry is for
- Why version data
- Why put a model in a container
- What experiment tracking records
- What data drift is
- When batch inference fits
- Same code, different results
- Canary deployment
- Data used to tune hyperparameters
- Recall from prediction counts
- What concept drift is
- Blue-green deployment
- A/B test versus canary
- What CI should check for a model
- Image versus container
- Traffic before the model has loaded
- What a pipeline DAG is
- Reading a cron schedule
- Where API keys belong
- What p99 latency means
- Throughput of four replicas
- Memory for 7B parameters in fp16
- Why quantise a model
- Why export a model to ONNX
- Logging a metric every epoch
- Making a training run repeatable
- What training-serving skew is
- What a model card is for
- Scaling a stateless model API
- Bad data reached training
- Protocol for internal inference calls
- Pod stuck in CrashLoopBackOff
- What a vector database does
- Cost of a GPU training job
- Shadow deployment
- Scheduling a GPU pod
- When to retrain
- Slow image builds
- Online and offline feature stores
- Low GPU use under load
- Safe retries in pipelines
- Labels arrive weeks later
- Computing PSI for two bins
- KV cache size for one long request
- Joining features to training labels
- Old start date with catchup on
- Pod killed while loading the model
- Autoscaling a GPU inference service
- Concurrent requests at peak
- Reading a KS drift statistic
- Serving many fine-tuned variants
- Smaller model with similar quality
- Switching embedding models in RAG
- Purpose of a multi-stage build
- Deploying by alias, not version
- Automatic gate before promotion
- Training on spot GPUs
- Speeding up training on 8 GPUs
- Why bf16 is easier than fp16
- Splitting requests into batches
- What to log for each LLM call
- Fleet-wide p99 from many pods
- Online features stopped updating
- CPU pods filling GPU nodes
- Silent wrong predictions from bad input
- Cost per 1,000 predictions
- Instructions hidden in a document
- p99 doubled right after a release
- p99 latency spikes during traffic bursts
- Rollback that made things worse
- Why LLM servers use continuous batching
- Same seed, different results
- Drift alerts firing all the time
- Hundreds of low-traffic models
- Model affects its own training data
- Loading a model from a public hub
- Where to use tensor parallelism
- KV cache with grouped-query attention
- Memory to fully fine-tune 7B with Adam
- Shadow model sent real emails
- Drift check with a classifier
- Canary that cannot catch quality drops
- Daily batch missing late events
- Pipeline reused yesterday output
- Offline and online scores disagree
- Batches of mixed text lengths
- Accuracy fell after int8 quantisation
- Same tag, different image
- Tail latency at low CPU usage
- More threads, no extra throughput
- State leaking between requests
- Streaming counts that run high
- Retries that make an outage worse
- Error budget for 99.9%
- Retrieval got worse as the index grew
- User asks for data deletion
- Semantic cache returns wrong answers
- GPU shows 100% but is underused
- High drift, stable performance
- Loss spike after resuming training