ML System DesignEasy
Production-readiness ML-systemsHow to check that a ML system is ready for production: what contracts, rollout, monitoring, rollback and quality gates are needed before launch?
Satel Generation
Назад к тренажеруНайти понятиеOpen full bank
Банк вопросов из реальных собеседований
Направления, темы и вопросы из записей интервью. Фильтры ниже сохраняются в ссылке.
ML System Design
274 questions from real interviews
Topics
Difficulty
Company
Language
QuestionCompanySourceAction
ML System DesignEasy
RAG in simple wordsHow to explain RAG in simple words: retrieval, augmentation and generation, and why does it look like two-stage recommendations?
Dubai Tech
ML System DesignEasy
Fast baseline ranking is better than randomThere are relevant search candidates, but the final order is random. What simple solution can be launched quickly?
Constructor
ML System DesignEasy
Multi-stage assembly of a Docker imageWhy divide the build and work image into several stages of Dockerfile and what does it change in the release of the service?
FlameTree
ML System DesignEasy
A cold start to a new user and productWhat is a cold start in a recommendation system and what practical solutions are suitable separately for a new user and a new product?
Dodo
ML System DesignMedium
Где искать latency, quality и cost деградацию LLM-сервисаLLM-сервис стал медленнее, хуже или дороже. Какие проверки делать?
Unimatch
ML System DesignMedium
Airflow Pipeline for Learning and InferenceHow to arrange an Airflow-pipeline for regular retraining and offline inference models? What components, model files and optimizations are needed?
GamerAM
ML System DesignMedium
ALS and matrix model for basket recommendationsHow to use an existing matrix model or ALS for recommendations to the current shopping cart without reducing the task to a common user profile?
GRII RED
ML System DesignMedium
API, queue, and scheduler for a GPU training platformHow should responsibilities be split between the API, queue, and scheduler in a distributed training platform?
Distributed AI infrastructure role, company not confirmed
ML System DesignMedium
BERT vs GPT and Transformer basicsExplain the difference between BERT and GPT in terms of Transformer architecture and training objective.
Undisclosed company
ML System DesignMedium
BERT pooling, sentence transformers and metric-learning analogyHow can you get a sentence embedding from BERT, how do sentence transformers differ, and why is this similar to metric learning for image pairs?
CorSearch
ML System DesignMedium
BERT: encoder, pretraining and attentionWhat is BERT, how does an encoder differ from a decoder, and what are the pretraining mechanisms of BERT?
Sber
ML System DesignMedium
Continuous batching in LLM inferenceWhat is continuous batching and why is it needed in the inference of large language models?
Mirai
ML System DesignMedium
Data drift and production model monitoringWhat is data drift, how can you detect it in data, and which signals should you monitor for an ML model in production?
inDrive
ML System DesignMedium
Deterministic orchestration instead of free agentWhen is a LLM assistant better than a free agent/tool calling?
TripleTen
ML System DesignMedium
Event stream instead of one summary from LLMHow to turn documents into features for forecasting: one summary, JSON-state or an event feed?
Mining / Commodities
ML System DesignMedium
Function calling and structured output in LLM applicationHow does function calling work and how does structured output work?
Waibee
ML System DesignMedium
Hidden student profile and prompt injectionHow to use the hidden profile of the student and prevent the user from pulling it out via prompt injection?
TripleTen
ML System DesignMedium
Kafka, S3 and ClickHouse: component rolesThe CTR panel has a stream of events about impressions and clicks. How to separate the roles of Kafka, S3 and ClickHouse?
Ads / Marketplace
ML System DesignMedium
Leakage from pretraining LLM on historical backtestWhy can the historical backtest of a LLM feature be unfair, even if the documents are submitted with the correct dates?
Mining / Commodities
ML System DesignMedium
LLM memory math and quantization for a 72B modelA 72B-parameter LLM is served on an A100 80GB. Estimate whether FP16 fits and explain what quantization changes.
Apriori
ML System DesignMedium
ML model deployment model files and rolloutYou trained and evaluated an ML model. What model files do you version, how do you package the service and how do you roll it out safely?
Apriori
ML System DesignMedium
Multimodal features in the RecSys pipelineHow to add text and visual features to a recommendation system without breaking the serving pipeline?
DROM
ML System DesignMedium
MVP without interaction historyWhat baseline should I launch for a new video feed if there are no clicks or purchases on it yet?
Undisclosed company