Mlops
265 bites tagged Mlops — interview questions with model answers, and 60-second explainers.
Epsilon in differential privacy and its trade-offs
Epsilon is the privacy budget bounding how much one record can change outputs; smaller epsilon means stronger privacy but more noise and lower accuracy. privacy-utility-cost balance.
Declarative vs imperative ML platform design
Declarative GitOps gives auditable, reproducible, reviewable desired-state config with strong governance but a steeper learning curve; imperative SDKs are flexible and fast for scientists but harder to… platform architecture tradeoffs.
Flask/Gunicorn vs Triton/TorchServe for serving
Flask is simple and flexible but lacks dynamic batching, GPU scheduling, and multi-model management; Triton/TorchServe add those plus metrics and versioning. choosing serving infrastructure.
Hybrid parallelism for large-model training
Split the model itself via tensor or pipeline parallelism so each replica is smaller, shrinking gradient all-reduce; combine with data parallelism in 2D/3D. scaling training past data-parallel limits.
Why design ML pipeline steps to be idempotent?
Re-running a step with the same input yields the same result and no duplicate side effects; enables safe retries and backfills. reliable, retryable pipeline design.
Concept drift vs data drift in production models
Data drift is a shift in input distribution P(X); concept drift is a shift in the relationship P(Y|X). model decay diagnosis. conflating them or claiming input monitoring alone detects concept drift, which needs labels.
What is a model registry and how does it enable CD?
A registry versions models with metadata, lineage, and stage tags; CD watches stage transitions to trigger deploys. model lifecycle governance. treating it as just blob storage with no versioning, stages, or lineage.
What is shadow deployment for ML models?
New model receives mirrored live traffic but its predictions are logged, not served; validates real-world behavior and latency without user risk. safe ML rollout strategy.
Automated rollback for a failed model deploy
Detect failure via health and metric checks, automatically route traffic back to the last known-good version (blue-green or canary), keep the registry entry but unpromote, and alert. safe deployment recovery.
Stages of an end-to-end ML pipeline
Ingest, validate, preprocess and feature engineer, train, evaluate against a baseline, then register and package for deployment, with gates between stages. the ML lifecycle as automation.
Detecting data drift on a continuous feature
Data drift is when serving feature distributions shift from training; detect with a Kolmogorov-Smirnov test comparing distributions; a small p-value signals drift to alert on. input distribution change in production.
CI/CD for microservice-based ML systems
Independent per-service pipelines, contract testing to protect interfaces and schemas, and incremental deploys (canary, blue-green); manage data and model contracts, not just code. ML CI/CD at service granularity.
Fairness and robustness gates in CI/CD
Sliced fairness metrics across subgroups, robustness checks via perturbation and adversarial sets, all compared to thresholds that fail the build. automated model quality gates.
Managing model-as-a-feature pipelines
An upstream embedding model becomes a versioned dependency, creating cascading retraining, version skew, latency stacking, and lineage complexity. dependency chains between models.
Versioning a 10TB dataset as code
Treat data like code via content-addressed pointers in Git while bytes live in object storage; dedupe by hashing so versions share unchanged files. reproducible data versioning at scale. copying the whole dataset per version.
Staging Environments for ML Pipelines
A staging environment mirrors production so models and pipelines are validated on production-like data and infrastructure before release. It catches drift, integration breaks, and serving regressions early, making promotion to production a safe, repeatable…
Fault-tolerant checkpointing for thousand-GPU pre-training
Checkpoint weights, optimizer state, RNG, and data position together; use asynchronous sharded writes and automated detect-restart-resume. Resilient large-scale training.
Data lineage and machine unlearning for a fine-tuned LLM
Version and fingerprint datasets, record transforms and which checkpoint saw what, and enable unlearning via retraining, data sharding, or approximate gradient methods. Auditable lineage plus practical unlearning.
Why version data and model artifacts, not just code
Code alone cannot reproduce a model; data and artifact versioning enable rollback, debugging, audit. reproducibility discipline. thinking Git on code suffices, or ignoring rollback and regulatory traceability.
ML CI/CD versus traditional software CI/CD
Validates code plus data plus the model, auto-trains and evaluates, adds continuous training and monitoring. MLOps pipeline understanding. describing only code CI/CD and ignoring data, model gates, and retraining.
What a feature store solves: skew and consistency
Central repository of computed features, one definition serving training and inference, reuse across models. feature-store value. calling it just a database or ignoring the offline-online consistency guarantee.
Batch prediction versus online real-time prediction
Batch is scheduled bulk scoring, online is low-latency per-request scoring; contrast latency, freshness, cost; give a use case each. serving pattern selection. confusing batch with retraining or claiming online is always best.
Deploy a saved model as a REST prediction service
Load the artifact, wrap it in a predict API, containerize, host with autoscaling, add monitoring. end-to-end deployment basics. jumping to model training or forgetting preprocessing parity and input validation.
Design drift detection and retraining for a CV model.
Monitor input and prediction distributions plus delayed labels, detect data and concept drift, trigger an automated retrain-validate-deploy pipeline with versioning and rollback. MLOps drift handling.
Get Mlops bites daily.
Five a day, five minutes, offline. With quizzes so it sticks.
Open testing — you’ll join as an early tester.