Data scientist system design interview questions

As asked

We need to score transactions for fraud within 50ms. Features include user 30-day spend history, merchant risk score, and device fingerprint. Design a feature store that serves these features with low latency.

Sample answer outline

The candidate should separate the offline store (batch-computed features in a data warehouse or object store) from the online store (low-latency key-value retrieval via Redis or DynamoDB). A feature pipeline computes aggregates in batch (30-day spend) and writes to both stores; a point-in-time correct training dataset is assembled from the offline store using the event timestamp. They should discuss cache invalidation, TTLs for spend aggregations, and how to handle cold starts for new users.

Expect these follow-ups

• How do you ensure training-serving skew does not occur when the online store feature computation differs from the batch pipeline?

As asked

Your company runs 100 concurrent experiments across web and mobile. Design the system that handles experiment assignment, event logging, and statistical analysis at scale.

Sample answer outline

Key components: an assignment service that deterministically hashes user_id plus experiment_id to a bucket (consistent hashing ensures the same user always gets the same variant), an event collection pipeline (Kafka to a data warehouse like BigQuery or Snowflake), and an analysis layer that computes metrics on demand. The candidate should discuss experiment isolation (one user can be in multiple experiments only if orthogonal), holdout groups, and how to support sequential testing to allow peeking safely.

Expect these follow-ups

• How do you handle users who are not logged in, and what are the risks of cookie-based assignment?

As asked

Design an end-to-end recommendation system for a video platform with 50 million users and 5 million videos. It must serve recommendations within 100ms.

Sample answer outline

Two-stage architecture: (1) candidate retrieval using approximate nearest neighbor search (Faiss or ScaNN) on user and item embeddings generated by a two-tower model trained offline; (2) ranking with a point-wise or pairwise model that takes the top 500 candidates and scores them with richer features. Candidate embeddings are pre-computed and indexed. The candidate should discuss how to incorporate real-time signals (last 3 watched videos) via user embedding updates and cold-start handling for new videos.

Expect these follow-ups

• How would you handle the cold-start problem for a video uploaded 5 minutes ago?

As asked

We need to ingest 10TB of raw transaction logs daily from S3, transform them into analytics tables, and make them available to analysts by 6am. Design this pipeline with failure recovery in mind.

Sample answer outline

The candidate should propose a DAG-based orchestrator (Airflow, Prefect, or dbt with a scheduler), partitioned S3 ingestion using Spark or dbt with incremental models, and idempotent transformation steps so reruns are safe. They should discuss data quality checks (Great Expectations or dbt tests) as a gate before writing to production tables, SLA monitoring with alerting, and a backfill strategy for catching up after failures.

Expect these follow-ups

• How do you implement idempotency in a dbt incremental model to avoid double-counting on reruns?

As asked

Design a system that flags anomalous user sessions in real time, for example a user suddenly downloading 100x their normal number of files. It needs to alert within 2 minutes of the anomalous event.

Sample answer outline

Stream processing layer (Kafka plus Flink or Spark Streaming) consumes events and maintains rolling feature aggregates per user (count in last 5 minutes, 1 hour, 24 hours). Anomaly scoring runs in-stream using a pre-trained model loaded into Flink, or a simple statistical baseline (z-score versus a user's historical distribution). Alerts are published to a downstream notification service. The candidate should discuss state management for streaming aggregates and the tradeoff between false positive rate and latency.

Expect these follow-ups

• How would you handle a user who has only been active for one day and has no history to baseline against?

As asked

Your team trains 20 models per week across 5 product areas. Design a system to track experiments, manage model versions, and deploy models to production with canary releases.

Sample answer outline

The candidate should describe a model registry (MLflow or custom) that stores model artifacts, metrics, hyperparameters, and lineage to training data. Deployment pipeline: CI triggers on model promotion, builds a Docker container with the model, and deploys to a serving layer (BentoML, Seldon, or a custom FastAPI service) behind a feature flag system that controls traffic split for canary testing. Rollback is triggered by automated metric monitoring comparing canary to baseline.

Expect these follow-ups

• How do you version the training data alongside the model to ensure reproducibility?

As asked

Design a system that predicts user churn weekly, scores all 5 million users, and triggers personalized retention campaigns through the marketing platform.

Sample answer outline

Weekly batch job assembles features from the data warehouse (aggregated behavioral signals), runs inference via a batch scoring script or ML platform job, and writes scores to a scoring table. A downstream integration reads from the scoring table and pushes high-risk users into the CRM or marketing automation tool with a personalized campaign tag. The candidate should discuss score freshness, score threshold calibration against campaign capacity, and holdout groups to measure campaign lift.

Expect these follow-ups

• How do you measure whether the retention campaign actually caused retention, versus the model simply selecting users who would have stayed anyway?

As asked

Your company's analytics depend on 50 upstream tables updated daily. Design a system to detect and alert on data quality issues before they reach analysts.

Sample answer outline

The candidate should describe automated checks at the schema level (column presence, dtype), statistical level (row count, null rate, distribution of key metrics within expected bands), and business rule level (revenue must be non-negative). Tools: dbt tests, Great Expectations, or Monte Carlo for anomaly detection on historical distributions. Alerts should trigger in Slack or PagerDuty before the downstream tables are published, using a quality gate pattern in the DAG.

Expect these follow-ups

• How do you set thresholds for acceptable null rates without generating too many false-positive alerts?

As asked

Design a machine learning system to rank 10 million products for search queries at an e-commerce company. The system must update rankings daily based on user interactions.

Sample answer outline

Two-stage: retrieval (BM25 or dense retrieval with a bi-encoder) narrows candidates to 1,000, then a learning-to-rank model (LambdaMART or cross-encoder) scores them using query-product features (text similarity, sales velocity, click rate, price) plus user personalization signals. Training data is collected from logged click-through events with position bias correction. Daily retraining pipeline and online A/B test of ranking changes.

Expect these follow-ups

• How do you correct for position bias when using click logs as implicit relevance signals?

As asked

You are building an LLM-powered writing assistant. Design an evaluation pipeline to test new model versions before shipping, covering quality, safety, and latency.

Sample answer outline

The pipeline needs three tracks: (1) automated metrics on a fixed benchmark set (ROUGE, BERTScore, task-specific metrics); (2) LLM-as-judge scoring using a stronger model to evaluate quality dimensions; (3) human evaluation queue for a sample of outputs on safety, factuality, and tone. Results are compared to a baseline and a regression threshold blocks shipping. Latency is tested under load. The candidate should discuss benchmark contamination risk and the limits of LLM-as-judge for safety.

Expect these follow-ups

• How do you prevent benchmark contamination when your LLM was trained on data that may include your test set?

As asked

Adobe Analytics processes billions of web and app events per day. Design the pipeline that ingests raw click-stream events from client SDKs, deduplicates them, enriches them with user and session context, and makes them queryable within 30 seconds of ingestion.

Sample answer outline

The pipeline has four stages: ingestion (load balancers accepting UDP or HTTPS beacons, writing to Kafka partitioned by visitor ID), stream processing (Flink or Kafka Streams for deduplication by event ID within a time window, session stitching, geolocation enrichment), hot storage (Druid or ClickHouse for sub-second queries on recent data), and cold storage (Parquet on S3 for long-term reporting). The candidate should address out-of-order events (watermarks), exactly-once semantics in the Kafka consumer, and how to replay a failed enrichment step without re-ingesting from clients.

Expect these follow-ups

• A customer's SDK is sending duplicate events at 10x the normal rate. How do you protect the pipeline without silently losing their data?
• You need to add a new enrichment field to all events. How do you backfill historical data without blocking the live stream?

As asked

Airbnb needs to detect fraudulent hosts who create fake listings to collect deposits or scam guests. Design a system that flags suspicious listings and accounts in near real time. What signals would you use, and how would you act on them?

Sample answer outline

Strong answers identify signals at multiple levels: account (new account, unverified identity, no prior bookings), listing (copied photos from other sites, prices far below market, vague location), and behavioral (rapid listing creation, rapid message replies asking to pay off-platform). They propose a streaming pipeline (Kafka) feeding rule-based filters and a ML model, with a human review queue for borderline cases, and discuss the precision/recall tradeoff and the cost of false positives (blocking legitimate hosts).

Expect these follow-ups

• How do you avoid disparate impact where your fraud model flags hosts from certain regions at higher rates?
• How do you update the model as fraudsters adapt to your signals?

As asked

Design a search system for Amazon's product catalog. It needs to support full-text search, filtering by category and price, ranking by relevance and sales rank, and return results in under 200ms for hundreds of millions of queries per day.

Sample answer outline

A strong answer separates the indexing pipeline from the query path. The indexing side ingests product updates through a message queue, builds an inverted index (Elasticsearch or a custom Lucene-based system), and refreshes incrementally. The query path hits a distributed search cluster, applies ranking (a mix of BM25 textual relevance and real-time sales signals), and returns results through a CDN-cached response where possible. The candidate should discuss how to handle the long tail of rare queries versus hot queries, and how to update the index without re-indexing the entire catalog on every product change.

Expect these follow-ups

• How do you handle spell correction and query understanding?
• How do you personalize results per user without blowing your latency budget?

As asked

When a customer places an order, DoorDash needs to assign a Dasher. The naive approach is first-available, but that ignores Dasher proximity, multi-order batching, restaurant prep time, and Dasher acceptance rates. How would you design the dispatch algorithm, and what data inputs does it need?

Sample answer outline

Strong answers frame this as an optimization problem with multiple objectives (minimize delivery time, maximize Dasher utilization, reduce unassigned orders) and explain why a greedy heuristic often beats ML models in latency-constrained dispatch. Candidates should mention the signal inputs (Dasher GPS, restaurant historical prep time, traffic ETA from a maps API, Dasher historical acceptance rate), batching logic for nearby orders, and a fallback when no Dasher is within a threshold radius.

Expect these follow-ups

• How does your algorithm handle a sudden surge where demand outpaces supply by 3x?
• How would you evaluate whether a new dispatch model is better than the current one in production?

As asked

DoorDash regularly runs promotions like '50% off your first three orders.' Some users create multiple accounts to claim new-user discounts repeatedly. Design a system that detects and blocks this abuse in real time at the order checkout step, with low false-positive rates for legitimate new users.

Sample answer outline

Strong answers identify the abuse signals (same device fingerprint, same payment method, same delivery address, same phone number, overlapping IP address, account creation timing patterns) and design a risk-scoring service called at checkout that aggregates these signals from a feature store. The score triggers a block, a CAPTCHA challenge, or a hold for manual review depending on confidence. Candidates should discuss the asymmetry between false positives (blocking real new users is worse than letting one abuser through) and propose a feedback loop from fraud team decisions to retrain the model.

Expect these follow-ups

• An abuser figures out your device fingerprint signal and starts rotating devices. How do you adapt?
• How do you measure whether your fraud system has net-positive unit economics after accounting for false positives?

As asked

LinkedIn shows members who viewed their profile in the last 90 days. Design the data pipeline that captures view events, counts views per viewer, handles privacy settings (some viewers choose to be anonymous), and serves the 'Who viewed your profile' feature in near real time. What does your architecture look like?

Sample answer outline

A strong answer covers an event ingestion layer (Kafka), a stream processing job that aggregates view events per (viewed_member, viewer_member) with a 90-day rolling window, privacy filtering (anonymous viewers contribute to count but not to the list), and a serving layer (Redis sorted set or time-bucketed row store). The candidate should discuss exactly-once counting, windowing strategies (tumbling vs sliding), and storage estimates at 900 million members.

Expect these follow-ups

• How do you handle the privacy downgrade case: a viewer switches from named to anonymous mid-window?
• How would you make the count exactly correct rather than approximate?

As asked

Lyft faces fraud where bad actors create fake trips by colluding as both rider and driver to collect promotional credits or inflate driver earnings. Design a system to detect this pattern in near real time and flag suspicious trips before payouts process.

Sample answer outline

Strong answers identify behavioral signals (rider and driver accounts created at the same time, GPS traces that never deviate, payment method linked to the same IP or device, trips that always start and end at the same coordinates), describe a streaming fraud scoring pipeline, and explain how they balance precision and recall given the cost of false positives (blocking legitimate drivers) versus false negatives (paying out fraud).

Expect these follow-ups

• How would you prevent fraudsters from learning your detection rules and adapting around them?
• What is your appeal process when you incorrectly flag a legitimate driver?

As asked

Design a recommendation system that generates a personalized homepage for each of Netflix's 200 million subscribers. The page must load in under 200 milliseconds. Walk me through the offline training pipeline, the online serving layer, and how you handle cold-start for a new subscriber.

Sample answer outline

A strong answer separates the offline batch pipeline (collaborative filtering or two-tower neural models trained on viewing history) from the online serving layer (pre-computed embeddings cached in Redis, real-time context like time of day and device blended at request time). The candidate should cover feature stores, the ranking layer that merges multiple candidate generators, A/B experimentation on ranking, and a cold-start strategy using popularity signals and onboarding questions.

Expect these follow-ups

• How do you handle the explore-exploit tradeoff so you don't just keep recommending what someone already watched?
• A new title launches. How does it get recommended before it has viewing history?

As asked

Netflix needs to track every play, pause, seek, and quality switch event from 200 million devices in real time. Design a pipeline that ingests these events, makes them available for real-time dashboards within 30 seconds, and stores them for batch analysis. Walk me through the entire path from device to dashboard.

Sample answer outline

A strong answer covers client-side event batching and compression before sending (to reduce request overhead), a high-throughput Kafka ingest layer with partitioning by user or session, stream processors (Flink or Spark Streaming) for real-time aggregations written to a low-latency store (Druid or ClickHouse), and a cold path to S3 via Parquet for batch queries in Spark or Presto. The candidate should discuss event deduplication, out-of-order events, and schema evolution via Avro or Protobuf.

Expect these follow-ups

• How do you handle a client that was offline for 20 minutes and then flushes a backlog of events all at once?
• A dashboard shows a 40% drop in plays starting two minutes ago. How do you tell if it's a real content issue versus a pipeline lag?

As asked

Netflix runs hundreds of simultaneous A/B tests across its product, from recommendation algorithms to UI layout to encoding parameters. Design the infrastructure that assigns users to test cells, logs exposure events, computes statistical significance, and prevents users from seeing conflicting experiments.

Sample answer outline

A strong answer covers a hashing-based assignment service that deterministically maps user IDs to experiment cells without network round-trips, a namespace collision system that prevents the same user from being in two mutually exclusive tests, server-side and client-side exposure logging at the moment of fork rather than at page load, and a statistics layer that computes confidence intervals and guards against peeking by requiring fixed-horizon or sequential analysis. The candidate should discuss how engineering teams integrate the SDK without slowing request paths.

Expect these follow-ups

• How do you handle a user who clears cookies or switches devices mid-experiment?
• An experiment shows 2% improvement in engagement but 0.1% increase in video errors. How do you decide whether to ship it?

As asked

Suppose you are building a document retrieval system that uses OpenAI embeddings and approximate nearest-neighbor search. Walk me through the full design: how you generate and store embeddings, how you keep them fresh as documents update, how you choose an ANN index, and how you handle queries that the embedding model does not handle well.

Sample answer outline

A strong answer covers the embedding pipeline (batch vs on-write), the choice between Pinecone, pgvector, or Faiss and why (latency, update frequency, infrastructure simplicity), handling document updates without full re-indexing, hybrid search combining BM25 keyword retrieval with vector retrieval for robustness, and reranking as a latency-quality tradeoff. Edge cases: very short documents, multilingual queries, embedding model version changes.

Expect these follow-ups

• How do you handle a document that is 50,000 tokens long, well above the embedding model context window?
• What metrics would you use to evaluate the quality of this system?

As asked

You are building the ontology layer for a Gotham-style threat intelligence platform. Analysts need to link entities like persons, locations, vehicles, and events across dozens of ingested data sources, many with conflicting or incomplete records. How would you design the underlying data model, and what are the biggest trade-offs you would make?

Sample answer outline

A strong answer covers entity resolution and probabilistic merging, a graph-based or property-graph model for linking heterogeneous entities, versioning of entity state over time, and how schema-on-read differs from schema-on-write for messy ingested data. The candidate should explicitly call out trade-offs between query performance and flexibility, and how you handle conflicting attribute values from different sources.

Expect these follow-ups

• How would you handle an analyst who disputes an automated entity merge?
• What indexing strategy would you use to support sub-second graph traversals over billions of edges?

As asked

Palantir serves multiple government agencies from the same platform. An analyst at Agency A must never see data that belongs to Agency B, but a small set of data must be shared between them. Data is stored in a shared data lake on S3 backed by a query engine. How do you design the access control layer?

Sample answer outline

A strong answer covers row-level and column-level security, attribute-based access control policies tied to the data ontology, enforcement at the query layer versus the storage layer, and audit logging. The candidate should discuss the risk of policy misconfiguration, how you test that no data leaks across tenants, and the performance overhead of applying fine-grained policy on every query. Bonus for mentioning the tension between centralised policy and delegated administration.

Expect these follow-ups

• How do you verify that your access control is correct and that no cross-tenant leakage exists?
• What happens when a dataset is reclassified and existing queries or cached results must be invalidated?

As asked

A military client has 500,000 IoT sensors each emitting two events per second. You must ingest all events, store them durably, and make them queryable by analysts within 30 seconds of arrival. Walk me through your end-to-end architecture.

Sample answer outline

A strong answer covers a distributed log (Kafka or Pulsar) as the ingest backbone, partitioned by sensor ID for ordering guarantees, a stream processor (Flink or Spark Structured Streaming) for enrichment and transformation, and a time-series-friendly storage layer (ClickHouse, InfluxDB, or a columnar format in S3). The candidate must address back-pressure, schema evolution, exactly-once semantics, and hot spots when one sensor produces vastly more data than others.

Expect these follow-ups

• A burst of 10x traffic arrives when sensors all simultaneously detect the same event. How does your system cope?
• How do you handle sensors that arrive with timestamps 10 minutes in the past due to network delay?

As asked

You need to build a near-real-time change data capture pipeline that detects inserts, updates, and deletes on a large Snowflake source table and incrementally loads a downstream analytics table. How would you use Snowflake Streams and Tasks to do this, and what are the failure modes you would watch for?

Sample answer outline

A stream tracks a hidden offset into the table's change log and exposes changed rows via the METADATA$ACTION column (INSERT or DELETE) and the METADATA$ISUPDATE boolean column. Updates appear as a DELETE row for the old value and an INSERT row for the new value, both with METADATA$ISUPDATE = TRUE. A scheduled Task consumes the stream in a MERGE statement into the target table. Key failure modes include the stream going stale if not consumed within the table's data retention period, tasks silently failing without alerting, and the MERGE statement taking longer than the task schedule and overlapping. A strong answer also mentions Serverless Tasks vs warehouse-driven tasks for cost.

Expect these follow-ups

• How would you handle schema evolution on the source table without breaking the pipeline?
• What monitoring would you put in place to catch a stale stream before it causes data loss?

As asked

You are building a multi-tenant SaaS product where each customer's data must be fully isolated. How would you architect the data layer on Snowflake: one database per tenant, one schema per tenant, or row-level isolation with a tenant_id column? Walk through the tradeoffs.

Sample answer outline

A strong answer compares the three patterns: one database per tenant gives the strongest isolation and simplest permissioning but creates management overhead at thousands of tenants; one schema per tenant is a middle ground with easier cross-tenant analytics via views but still requires automating schema provisioning; a shared schema with tenant_id is operationally simplest but requires row-level security policies and careful query rewriting to prevent data leakage. The candidate should mention Snowflake's Row Access Policies for the shared model and discuss how warehouse sizing changes across patterns.

Expect these follow-ups

• How would you handle a tenant that is generating 100x the query load of the average tenant?
• If a tenant requests a full export of their data for GDPR compliance, which model makes that easiest?

As asked

You need to ingest 5 TB of JSON event data per day into Snowflake from thousands of IoT devices. Latency can be up to 15 minutes. Design the full ingestion architecture, naming the specific Snowflake features and complementary cloud services you would use, and explain how you would handle schema drift.

Sample answer outline

A strong answer describes devices publishing to a message bus (Kafka or AWS Kinesis), a processing layer (Kafka Connect Snowflake connector or a Flink job) writing Parquet files to S3/GCS, then Snowpipe automatically ingesting files via SQS event notifications. For schema drift, VARIANT columns absorb any JSON structure changes, and schema evolution can be handled by defining the final schema in Snowflake and using PARSE_JSON plus lateral flatten transforms at query time or in a transformation layer like dbt. The candidate should address error handling (dead-letter staging tables), monitoring (Snowpipe ingestion history views), and cost at scale.

Expect these follow-ups

• If latency needs to drop to under 1 minute, what changes to the architecture?
• How would you handle a burst of 10x normal volume without throttling or dropping events?

As asked

Design the end-to-end system behind a feature like Discover Weekly: a personalized playlist of 30 songs generated once a week for roughly 600 million users. Walk me through the data pipeline, the model serving layer, and how you would keep recommendations fresh and relevant without regenerating everything from scratch each week.

Sample answer outline

A strong answer separates offline training from online serving and explains the batch pipeline that computes user embeddings from listening history. It covers the candidate generation stage (collaborative filtering, content embeddings), the ranking stage, and how diversity constraints prevent filter-bubble playlists. It discusses the cold-start problem for new users and new tracks, and addresses how to store and serve 600 million playlists efficiently using object storage plus a CDN-like precompute strategy.

Expect these follow-ups

• How would you handle a user who listens only in one genre versus someone with very eclectic taste?
• Discover Weekly refreshes weekly. How would you A/B test a change to the recommendation model without users noticing?

As asked

Every song play, skip, pause, and seek generates an event at Spotify. At peak, this is hundreds of thousands of events per second. Walk me through how you would build the ingestion layer using Kafka, what guarantees you need around exactly-once delivery versus at-least-once, how you partition topics, and how downstream consumers like the recommendation engine and billing system would subscribe.

Sample answer outline

A strong answer explains partition strategy by user ID or track ID depending on downstream ordering needs, discusses consumer group isolation so billing and recommendations do not affect each other's offsets, and addresses the tradeoff between exactly-once semantics (expensive) and idempotent at-least-once (simpler with deduplication keys). It covers schema evolution with Avro or Protobuf in a schema registry, and lag monitoring so a slow consumer does not silently fall behind.

Expect these follow-ups

• If a consumer group falls 10 minutes behind during a traffic spike, how do you decide whether to scale horizontally or to shed load?
• How do you handle a malformed event that keeps crashing a consumer?

As asked

Stripe Radar scores every charge for fraud risk in under 100 ms before authorization. Describe how you would build the feature pipeline that takes a raw charge event and produces a fraud score. What features do you compute, how do you serve them at low latency, and how do you keep them fresh as merchant behavior evolves?

Sample answer outline

A strong answer distinguishes real-time features computed at inference time (card BIN, user agent, IP reputation) from precomputed features updated in batch (velocity counters per card, per email, per device). Serving at sub-100ms requires a low-latency feature store (Redis or Aerospike) accessed in parallel rather than sequentially. The candidate should address feature drift: how you detect that a feature no longer correlates with fraud and how you retrain or reweight without a full pipeline rewrite.

Expect these follow-ups

• How would you handle a new merchant with no transaction history for velocity features?
• What monitoring would you add to catch when your fraud score distribution shifts unexpectedly?

As asked

Tesla's Autopilot team collects billions of short video clips from the fleet every day. Design the end-to-end pipeline that ingests raw camera clips from vehicles, labels interesting edge cases automatically, routes ambiguous clips to human labelers, and delivers a clean training dataset to the model training cluster. Focus on throughput, cost, and labeling quality.

Sample answer outline

A strong answer covers: compressed upload from vehicle to cloud (cellular bandwidth constraints), cheap pre-filtering with a lightweight classifier to discard boring clips, automatic labeling via a teacher model with confidence scoring, active learning to route low-confidence examples to human labelers, deduplication to avoid overrepresenting common scenarios, versioned dataset snapshots, and feedback loops from model evaluation back into the labeling pipeline. The candidate should discuss throughput targets, storage cost (petabytes of video), and how to detect and correct systematic labeling errors.

Expect these follow-ups

• How do you avoid your training set becoming dominated by easy highway driving and underrepresenting rare but important edge cases?
• If a labeling error gets into training and causes a regression, how do you find and remove all affected examples?

As asked

Design a system that identifies the top 10 trending topics on X in near real-time, refreshing every 30 seconds, at a scale of 500,000 tweets per minute. Walk me through data ingestion, aggregation, and how you would handle the thundering-herd problem when a major event causes a sudden spike.

Sample answer outline

A strong answer sketches a Kafka ingestion layer, stream processors (Flink or Spark Streaming) using a Count-Min Sketch or a sliding-window heavy-hitter algorithm to approximate top-K, and a periodic snapshot to Redis Sorted Sets. The candidate should address locality (trending should be geo-scoped), deduplication of retweets, and how a load-shedding strategy prevents the aggregation tier from collapsing under a viral event.

Expect these follow-ups

• How would you distinguish a legitimate trending topic from a coordinated spam campaign artificially inflating a hashtag?
• If your stream processor falls 2 minutes behind, how does that affect the trending output and what would you do?

As asked

X's 'Who to Follow' feature suggests accounts to users. Describe how you would build a recommendation system that suggests users who share second-degree connections, common interests (inferred from tweets), or topic overlap. How would you scale this to run for all 300 million daily active users?

Sample answer outline

A strong answer covers graph traversal (BFS up to 2 hops in a sharded graph store like FlockDB or a custom adjacency list in Cassandra), collaborative filtering on follow patterns, topic modeling on tweet text (LDA or embedding similarity), and a candidate generation plus scoring pipeline that runs nightly for most users and near-real-time for new accounts. The candidate should quantify the candidate set size before and after ranking and explain why pre-computation is necessary at this scale.

Expect these follow-ups

• How do you prevent the recommendation engine from creating echo chambers by only recommending accounts the user is already similar to?
• A brand-new user has zero follows and zero tweets. How do you cold-start their recommendations?

As asked

X gets tens of thousands of new account registrations per hour, many of which are bots or spam accounts. Design a system that scores each new account within 30 seconds of registration and automatically restricts accounts above a threshold without requiring human review.

Sample answer outline

A strong answer covers signals available at registration (email domain, phone number carrier, IP reputation, device fingerprint, registration speed, username pattern), a lightweight online classifier (logistic regression or a small neural net) running in real time on an event stream, a staged action system (shadow-ban before hard-ban for FP reduction), and a human review queue for borderline cases. The candidate should address adversarial adaptation (bots evolve to defeat signals) and how to retrain the model without introducing label bias from your own shadow-banning.

Expect these follow-ups

• Your spam classifier has a 5% false positive rate. At 50,000 registrations per hour, how many legitimate users are incorrectly restricted per day and how do you build the appeal flow?
• A sophisticated bot farm uses residential proxies and real phone numbers. Which signals do you still have and which are you now blind to?

As asked

We need to score transactions for fraud within 50ms. Features include user 30-day spend history, merchant risk score, and device fingerprint. Design a feature store that serves these features with low latency.

Sample answer outline

The candidate should separate the offline store (batch-computed features in a data warehouse or object store) from the online store (low-latency key-value retrieval via Redis or DynamoDB). A feature pipeline computes aggregates in batch (30-day spend) and writes to both stores; a point-in-time correct training dataset is assembled from the offline store using the event timestamp. They should discuss cache invalidation, TTLs for spend aggregations, and how to handle cold starts for new users.

Expect these follow-ups

• How do you ensure training-serving skew does not occur when the online store feature computation differs from the batch pipeline?

As asked

Your company runs 100 concurrent experiments across web and mobile. Design the system that handles experiment assignment, event logging, and statistical analysis at scale.

Sample answer outline

Key components: an assignment service that deterministically hashes user_id plus experiment_id to a bucket (consistent hashing ensures the same user always gets the same variant), an event collection pipeline (Kafka to a data warehouse like BigQuery or Snowflake), and an analysis layer that computes metrics on demand. The candidate should discuss experiment isolation (one user can be in multiple experiments only if orthogonal), holdout groups, and how to support sequential testing to allow peeking safely.

Expect these follow-ups

• How do you handle users who are not logged in, and what are the risks of cookie-based assignment?

As asked

Design an end-to-end recommendation system for a video platform with 50 million users and 5 million videos. It must serve recommendations within 100ms.

Sample answer outline

Two-stage architecture: (1) candidate retrieval using approximate nearest neighbor search (Faiss or ScaNN) on user and item embeddings generated by a two-tower model trained offline; (2) ranking with a point-wise or pairwise model that takes the top 500 candidates and scores them with richer features. Candidate embeddings are pre-computed and indexed. The candidate should discuss how to incorporate real-time signals (last 3 watched videos) via user embedding updates and cold-start handling for new videos.

Expect these follow-ups

• How would you handle the cold-start problem for a video uploaded 5 minutes ago?

As asked

We need to ingest 10TB of raw transaction logs daily from S3, transform them into analytics tables, and make them available to analysts by 6am. Design this pipeline with failure recovery in mind.

Sample answer outline

The candidate should propose a DAG-based orchestrator (Airflow, Prefect, or dbt with a scheduler), partitioned S3 ingestion using Spark or dbt with incremental models, and idempotent transformation steps so reruns are safe. They should discuss data quality checks (Great Expectations or dbt tests) as a gate before writing to production tables, SLA monitoring with alerting, and a backfill strategy for catching up after failures.

Expect these follow-ups

• How do you implement idempotency in a dbt incremental model to avoid double-counting on reruns?

As asked

Design a system that flags anomalous user sessions in real time, for example a user suddenly downloading 100x their normal number of files. It needs to alert within 2 minutes of the anomalous event.

Sample answer outline

Stream processing layer (Kafka plus Flink or Spark Streaming) consumes events and maintains rolling feature aggregates per user (count in last 5 minutes, 1 hour, 24 hours). Anomaly scoring runs in-stream using a pre-trained model loaded into Flink, or a simple statistical baseline (z-score versus a user's historical distribution). Alerts are published to a downstream notification service. The candidate should discuss state management for streaming aggregates and the tradeoff between false positive rate and latency.

Expect these follow-ups

• How would you handle a user who has only been active for one day and has no history to baseline against?

As asked

Your team trains 20 models per week across 5 product areas. Design a system to track experiments, manage model versions, and deploy models to production with canary releases.

Sample answer outline

The candidate should describe a model registry (MLflow or custom) that stores model artifacts, metrics, hyperparameters, and lineage to training data. Deployment pipeline: CI triggers on model promotion, builds a Docker container with the model, and deploys to a serving layer (BentoML, Seldon, or a custom FastAPI service) behind a feature flag system that controls traffic split for canary testing. Rollback is triggered by automated metric monitoring comparing canary to baseline.

Expect these follow-ups

• How do you version the training data alongside the model to ensure reproducibility?

As asked

Design a system that predicts user churn weekly, scores all 5 million users, and triggers personalized retention campaigns through the marketing platform.

Sample answer outline

Weekly batch job assembles features from the data warehouse (aggregated behavioral signals), runs inference via a batch scoring script or ML platform job, and writes scores to a scoring table. A downstream integration reads from the scoring table and pushes high-risk users into the CRM or marketing automation tool with a personalized campaign tag. The candidate should discuss score freshness, score threshold calibration against campaign capacity, and holdout groups to measure campaign lift.

Expect these follow-ups

• How do you measure whether the retention campaign actually caused retention, versus the model simply selecting users who would have stayed anyway?

As asked

Your company's analytics depend on 50 upstream tables updated daily. Design a system to detect and alert on data quality issues before they reach analysts.

Sample answer outline

The candidate should describe automated checks at the schema level (column presence, dtype), statistical level (row count, null rate, distribution of key metrics within expected bands), and business rule level (revenue must be non-negative). Tools: dbt tests, Great Expectations, or Monte Carlo for anomaly detection on historical distributions. Alerts should trigger in Slack or PagerDuty before the downstream tables are published, using a quality gate pattern in the DAG.

Expect these follow-ups

• How do you set thresholds for acceptable null rates without generating too many false-positive alerts?

As asked

Design a machine learning system to rank 10 million products for search queries at an e-commerce company. The system must update rankings daily based on user interactions.

Sample answer outline

Two-stage: retrieval (BM25 or dense retrieval with a bi-encoder) narrows candidates to 1,000, then a learning-to-rank model (LambdaMART or cross-encoder) scores them using query-product features (text similarity, sales velocity, click rate, price) plus user personalization signals. Training data is collected from logged click-through events with position bias correction. Daily retraining pipeline and online A/B test of ranking changes.

Expect these follow-ups

• How do you correct for position bias when using click logs as implicit relevance signals?

As asked

You are building an LLM-powered writing assistant. Design an evaluation pipeline to test new model versions before shipping, covering quality, safety, and latency.

Sample answer outline

The pipeline needs three tracks: (1) automated metrics on a fixed benchmark set (ROUGE, BERTScore, task-specific metrics); (2) LLM-as-judge scoring using a stronger model to evaluate quality dimensions; (3) human evaluation queue for a sample of outputs on safety, factuality, and tone. Results are compared to a baseline and a regression threshold blocks shipping. Latency is tested under load. The candidate should discuss benchmark contamination risk and the limits of LLM-as-judge for safety.

Expect these follow-ups

• How do you prevent benchmark contamination when your LLM was trained on data that may include your test set?

As asked

Adobe Analytics processes billions of web and app events per day. Design the pipeline that ingests raw click-stream events from client SDKs, deduplicates them, enriches them with user and session context, and makes them queryable within 30 seconds of ingestion.

Sample answer outline

The pipeline has four stages: ingestion (load balancers accepting UDP or HTTPS beacons, writing to Kafka partitioned by visitor ID), stream processing (Flink or Kafka Streams for deduplication by event ID within a time window, session stitching, geolocation enrichment), hot storage (Druid or ClickHouse for sub-second queries on recent data), and cold storage (Parquet on S3 for long-term reporting). The candidate should address out-of-order events (watermarks), exactly-once semantics in the Kafka consumer, and how to replay a failed enrichment step without re-ingesting from clients.

Expect these follow-ups

• A customer's SDK is sending duplicate events at 10x the normal rate. How do you protect the pipeline without silently losing their data?
• You need to add a new enrichment field to all events. How do you backfill historical data without blocking the live stream?

As asked

Airbnb needs to detect fraudulent hosts who create fake listings to collect deposits or scam guests. Design a system that flags suspicious listings and accounts in near real time. What signals would you use, and how would you act on them?

Sample answer outline

Strong answers identify signals at multiple levels: account (new account, unverified identity, no prior bookings), listing (copied photos from other sites, prices far below market, vague location), and behavioral (rapid listing creation, rapid message replies asking to pay off-platform). They propose a streaming pipeline (Kafka) feeding rule-based filters and a ML model, with a human review queue for borderline cases, and discuss the precision/recall tradeoff and the cost of false positives (blocking legitimate hosts).

Expect these follow-ups

• How do you avoid disparate impact where your fraud model flags hosts from certain regions at higher rates?
• How do you update the model as fraudsters adapt to your signals?

As asked

Design a search system for Amazon's product catalog. It needs to support full-text search, filtering by category and price, ranking by relevance and sales rank, and return results in under 200ms for hundreds of millions of queries per day.

Sample answer outline

A strong answer separates the indexing pipeline from the query path. The indexing side ingests product updates through a message queue, builds an inverted index (Elasticsearch or a custom Lucene-based system), and refreshes incrementally. The query path hits a distributed search cluster, applies ranking (a mix of BM25 textual relevance and real-time sales signals), and returns results through a CDN-cached response where possible. The candidate should discuss how to handle the long tail of rare queries versus hot queries, and how to update the index without re-indexing the entire catalog on every product change.

Expect these follow-ups

• How do you handle spell correction and query understanding?
• How do you personalize results per user without blowing your latency budget?

As asked

When a customer places an order, DoorDash needs to assign a Dasher. The naive approach is first-available, but that ignores Dasher proximity, multi-order batching, restaurant prep time, and Dasher acceptance rates. How would you design the dispatch algorithm, and what data inputs does it need?

Sample answer outline

Strong answers frame this as an optimization problem with multiple objectives (minimize delivery time, maximize Dasher utilization, reduce unassigned orders) and explain why a greedy heuristic often beats ML models in latency-constrained dispatch. Candidates should mention the signal inputs (Dasher GPS, restaurant historical prep time, traffic ETA from a maps API, Dasher historical acceptance rate), batching logic for nearby orders, and a fallback when no Dasher is within a threshold radius.

Expect these follow-ups

• How does your algorithm handle a sudden surge where demand outpaces supply by 3x?
• How would you evaluate whether a new dispatch model is better than the current one in production?

As asked

DoorDash regularly runs promotions like '50% off your first three orders.' Some users create multiple accounts to claim new-user discounts repeatedly. Design a system that detects and blocks this abuse in real time at the order checkout step, with low false-positive rates for legitimate new users.

Sample answer outline

Strong answers identify the abuse signals (same device fingerprint, same payment method, same delivery address, same phone number, overlapping IP address, account creation timing patterns) and design a risk-scoring service called at checkout that aggregates these signals from a feature store. The score triggers a block, a CAPTCHA challenge, or a hold for manual review depending on confidence. Candidates should discuss the asymmetry between false positives (blocking real new users is worse than letting one abuser through) and propose a feedback loop from fraud team decisions to retrain the model.

Expect these follow-ups

• An abuser figures out your device fingerprint signal and starts rotating devices. How do you adapt?
• How do you measure whether your fraud system has net-positive unit economics after accounting for false positives?

As asked

LinkedIn shows members who viewed their profile in the last 90 days. Design the data pipeline that captures view events, counts views per viewer, handles privacy settings (some viewers choose to be anonymous), and serves the 'Who viewed your profile' feature in near real time. What does your architecture look like?

Sample answer outline

A strong answer covers an event ingestion layer (Kafka), a stream processing job that aggregates view events per (viewed_member, viewer_member) with a 90-day rolling window, privacy filtering (anonymous viewers contribute to count but not to the list), and a serving layer (Redis sorted set or time-bucketed row store). The candidate should discuss exactly-once counting, windowing strategies (tumbling vs sliding), and storage estimates at 900 million members.

Expect these follow-ups

• How do you handle the privacy downgrade case: a viewer switches from named to anonymous mid-window?
• How would you make the count exactly correct rather than approximate?

As asked

Lyft faces fraud where bad actors create fake trips by colluding as both rider and driver to collect promotional credits or inflate driver earnings. Design a system to detect this pattern in near real time and flag suspicious trips before payouts process.

Sample answer outline

Strong answers identify behavioral signals (rider and driver accounts created at the same time, GPS traces that never deviate, payment method linked to the same IP or device, trips that always start and end at the same coordinates), describe a streaming fraud scoring pipeline, and explain how they balance precision and recall given the cost of false positives (blocking legitimate drivers) versus false negatives (paying out fraud).

Expect these follow-ups

• How would you prevent fraudsters from learning your detection rules and adapting around them?
• What is your appeal process when you incorrectly flag a legitimate driver?

As asked

Design a recommendation system that generates a personalized homepage for each of Netflix's 200 million subscribers. The page must load in under 200 milliseconds. Walk me through the offline training pipeline, the online serving layer, and how you handle cold-start for a new subscriber.

Sample answer outline

A strong answer separates the offline batch pipeline (collaborative filtering or two-tower neural models trained on viewing history) from the online serving layer (pre-computed embeddings cached in Redis, real-time context like time of day and device blended at request time). The candidate should cover feature stores, the ranking layer that merges multiple candidate generators, A/B experimentation on ranking, and a cold-start strategy using popularity signals and onboarding questions.

Expect these follow-ups

• How do you handle the explore-exploit tradeoff so you don't just keep recommending what someone already watched?
• A new title launches. How does it get recommended before it has viewing history?

As asked

Netflix needs to track every play, pause, seek, and quality switch event from 200 million devices in real time. Design a pipeline that ingests these events, makes them available for real-time dashboards within 30 seconds, and stores them for batch analysis. Walk me through the entire path from device to dashboard.

Sample answer outline

A strong answer covers client-side event batching and compression before sending (to reduce request overhead), a high-throughput Kafka ingest layer with partitioning by user or session, stream processors (Flink or Spark Streaming) for real-time aggregations written to a low-latency store (Druid or ClickHouse), and a cold path to S3 via Parquet for batch queries in Spark or Presto. The candidate should discuss event deduplication, out-of-order events, and schema evolution via Avro or Protobuf.

Expect these follow-ups

• How do you handle a client that was offline for 20 minutes and then flushes a backlog of events all at once?
• A dashboard shows a 40% drop in plays starting two minutes ago. How do you tell if it's a real content issue versus a pipeline lag?

As asked

Netflix runs hundreds of simultaneous A/B tests across its product, from recommendation algorithms to UI layout to encoding parameters. Design the infrastructure that assigns users to test cells, logs exposure events, computes statistical significance, and prevents users from seeing conflicting experiments.

Sample answer outline

A strong answer covers a hashing-based assignment service that deterministically maps user IDs to experiment cells without network round-trips, a namespace collision system that prevents the same user from being in two mutually exclusive tests, server-side and client-side exposure logging at the moment of fork rather than at page load, and a statistics layer that computes confidence intervals and guards against peeking by requiring fixed-horizon or sequential analysis. The candidate should discuss how engineering teams integrate the SDK without slowing request paths.

Expect these follow-ups

• How do you handle a user who clears cookies or switches devices mid-experiment?
• An experiment shows 2% improvement in engagement but 0.1% increase in video errors. How do you decide whether to ship it?

As asked

Suppose you are building a document retrieval system that uses OpenAI embeddings and approximate nearest-neighbor search. Walk me through the full design: how you generate and store embeddings, how you keep them fresh as documents update, how you choose an ANN index, and how you handle queries that the embedding model does not handle well.

Sample answer outline

A strong answer covers the embedding pipeline (batch vs on-write), the choice between Pinecone, pgvector, or Faiss and why (latency, update frequency, infrastructure simplicity), handling document updates without full re-indexing, hybrid search combining BM25 keyword retrieval with vector retrieval for robustness, and reranking as a latency-quality tradeoff. Edge cases: very short documents, multilingual queries, embedding model version changes.

Expect these follow-ups

• How do you handle a document that is 50,000 tokens long, well above the embedding model context window?
• What metrics would you use to evaluate the quality of this system?

As asked

You are building the ontology layer for a Gotham-style threat intelligence platform. Analysts need to link entities like persons, locations, vehicles, and events across dozens of ingested data sources, many with conflicting or incomplete records. How would you design the underlying data model, and what are the biggest trade-offs you would make?

Sample answer outline

A strong answer covers entity resolution and probabilistic merging, a graph-based or property-graph model for linking heterogeneous entities, versioning of entity state over time, and how schema-on-read differs from schema-on-write for messy ingested data. The candidate should explicitly call out trade-offs between query performance and flexibility, and how you handle conflicting attribute values from different sources.

Expect these follow-ups

• How would you handle an analyst who disputes an automated entity merge?
• What indexing strategy would you use to support sub-second graph traversals over billions of edges?

As asked

Palantir serves multiple government agencies from the same platform. An analyst at Agency A must never see data that belongs to Agency B, but a small set of data must be shared between them. Data is stored in a shared data lake on S3 backed by a query engine. How do you design the access control layer?

Sample answer outline

A strong answer covers row-level and column-level security, attribute-based access control policies tied to the data ontology, enforcement at the query layer versus the storage layer, and audit logging. The candidate should discuss the risk of policy misconfiguration, how you test that no data leaks across tenants, and the performance overhead of applying fine-grained policy on every query. Bonus for mentioning the tension between centralised policy and delegated administration.

Expect these follow-ups

• How do you verify that your access control is correct and that no cross-tenant leakage exists?
• What happens when a dataset is reclassified and existing queries or cached results must be invalidated?

As asked

A military client has 500,000 IoT sensors each emitting two events per second. You must ingest all events, store them durably, and make them queryable by analysts within 30 seconds of arrival. Walk me through your end-to-end architecture.

Sample answer outline

A strong answer covers a distributed log (Kafka or Pulsar) as the ingest backbone, partitioned by sensor ID for ordering guarantees, a stream processor (Flink or Spark Structured Streaming) for enrichment and transformation, and a time-series-friendly storage layer (ClickHouse, InfluxDB, or a columnar format in S3). The candidate must address back-pressure, schema evolution, exactly-once semantics, and hot spots when one sensor produces vastly more data than others.

Expect these follow-ups

• A burst of 10x traffic arrives when sensors all simultaneously detect the same event. How does your system cope?
• How do you handle sensors that arrive with timestamps 10 minutes in the past due to network delay?

As asked

You need to build a near-real-time change data capture pipeline that detects inserts, updates, and deletes on a large Snowflake source table and incrementally loads a downstream analytics table. How would you use Snowflake Streams and Tasks to do this, and what are the failure modes you would watch for?

Sample answer outline

A stream tracks a hidden offset into the table's change log and exposes changed rows via the METADATA$ACTION column (INSERT or DELETE) and the METADATA$ISUPDATE boolean column. Updates appear as a DELETE row for the old value and an INSERT row for the new value, both with METADATA$ISUPDATE = TRUE. A scheduled Task consumes the stream in a MERGE statement into the target table. Key failure modes include the stream going stale if not consumed within the table's data retention period, tasks silently failing without alerting, and the MERGE statement taking longer than the task schedule and overlapping. A strong answer also mentions Serverless Tasks vs warehouse-driven tasks for cost.

Expect these follow-ups

• How would you handle schema evolution on the source table without breaking the pipeline?
• What monitoring would you put in place to catch a stale stream before it causes data loss?

As asked

You are building a multi-tenant SaaS product where each customer's data must be fully isolated. How would you architect the data layer on Snowflake: one database per tenant, one schema per tenant, or row-level isolation with a tenant_id column? Walk through the tradeoffs.

Sample answer outline

A strong answer compares the three patterns: one database per tenant gives the strongest isolation and simplest permissioning but creates management overhead at thousands of tenants; one schema per tenant is a middle ground with easier cross-tenant analytics via views but still requires automating schema provisioning; a shared schema with tenant_id is operationally simplest but requires row-level security policies and careful query rewriting to prevent data leakage. The candidate should mention Snowflake's Row Access Policies for the shared model and discuss how warehouse sizing changes across patterns.

Expect these follow-ups

• How would you handle a tenant that is generating 100x the query load of the average tenant?
• If a tenant requests a full export of their data for GDPR compliance, which model makes that easiest?

As asked

You need to ingest 5 TB of JSON event data per day into Snowflake from thousands of IoT devices. Latency can be up to 15 minutes. Design the full ingestion architecture, naming the specific Snowflake features and complementary cloud services you would use, and explain how you would handle schema drift.

Sample answer outline

A strong answer describes devices publishing to a message bus (Kafka or AWS Kinesis), a processing layer (Kafka Connect Snowflake connector or a Flink job) writing Parquet files to S3/GCS, then Snowpipe automatically ingesting files via SQS event notifications. For schema drift, VARIANT columns absorb any JSON structure changes, and schema evolution can be handled by defining the final schema in Snowflake and using PARSE_JSON plus lateral flatten transforms at query time or in a transformation layer like dbt. The candidate should address error handling (dead-letter staging tables), monitoring (Snowpipe ingestion history views), and cost at scale.

Expect these follow-ups

• If latency needs to drop to under 1 minute, what changes to the architecture?
• How would you handle a burst of 10x normal volume without throttling or dropping events?

As asked

Design the end-to-end system behind a feature like Discover Weekly: a personalized playlist of 30 songs generated once a week for roughly 600 million users. Walk me through the data pipeline, the model serving layer, and how you would keep recommendations fresh and relevant without regenerating everything from scratch each week.

Sample answer outline

A strong answer separates offline training from online serving and explains the batch pipeline that computes user embeddings from listening history. It covers the candidate generation stage (collaborative filtering, content embeddings), the ranking stage, and how diversity constraints prevent filter-bubble playlists. It discusses the cold-start problem for new users and new tracks, and addresses how to store and serve 600 million playlists efficiently using object storage plus a CDN-like precompute strategy.

Expect these follow-ups

• How would you handle a user who listens only in one genre versus someone with very eclectic taste?
• Discover Weekly refreshes weekly. How would you A/B test a change to the recommendation model without users noticing?

As asked

Every song play, skip, pause, and seek generates an event at Spotify. At peak, this is hundreds of thousands of events per second. Walk me through how you would build the ingestion layer using Kafka, what guarantees you need around exactly-once delivery versus at-least-once, how you partition topics, and how downstream consumers like the recommendation engine and billing system would subscribe.

Sample answer outline

A strong answer explains partition strategy by user ID or track ID depending on downstream ordering needs, discusses consumer group isolation so billing and recommendations do not affect each other's offsets, and addresses the tradeoff between exactly-once semantics (expensive) and idempotent at-least-once (simpler with deduplication keys). It covers schema evolution with Avro or Protobuf in a schema registry, and lag monitoring so a slow consumer does not silently fall behind.

Expect these follow-ups

• If a consumer group falls 10 minutes behind during a traffic spike, how do you decide whether to scale horizontally or to shed load?
• How do you handle a malformed event that keeps crashing a consumer?

As asked

Stripe Radar scores every charge for fraud risk in under 100 ms before authorization. Describe how you would build the feature pipeline that takes a raw charge event and produces a fraud score. What features do you compute, how do you serve them at low latency, and how do you keep them fresh as merchant behavior evolves?

Sample answer outline

A strong answer distinguishes real-time features computed at inference time (card BIN, user agent, IP reputation) from precomputed features updated in batch (velocity counters per card, per email, per device). Serving at sub-100ms requires a low-latency feature store (Redis or Aerospike) accessed in parallel rather than sequentially. The candidate should address feature drift: how you detect that a feature no longer correlates with fraud and how you retrain or reweight without a full pipeline rewrite.

Expect these follow-ups

• How would you handle a new merchant with no transaction history for velocity features?
• What monitoring would you add to catch when your fraud score distribution shifts unexpectedly?

As asked

Tesla's Autopilot team collects billions of short video clips from the fleet every day. Design the end-to-end pipeline that ingests raw camera clips from vehicles, labels interesting edge cases automatically, routes ambiguous clips to human labelers, and delivers a clean training dataset to the model training cluster. Focus on throughput, cost, and labeling quality.

Sample answer outline

A strong answer covers: compressed upload from vehicle to cloud (cellular bandwidth constraints), cheap pre-filtering with a lightweight classifier to discard boring clips, automatic labeling via a teacher model with confidence scoring, active learning to route low-confidence examples to human labelers, deduplication to avoid overrepresenting common scenarios, versioned dataset snapshots, and feedback loops from model evaluation back into the labeling pipeline. The candidate should discuss throughput targets, storage cost (petabytes of video), and how to detect and correct systematic labeling errors.

Expect these follow-ups

• How do you avoid your training set becoming dominated by easy highway driving and underrepresenting rare but important edge cases?
• If a labeling error gets into training and causes a regression, how do you find and remove all affected examples?

As asked

Design a system that identifies the top 10 trending topics on X in near real-time, refreshing every 30 seconds, at a scale of 500,000 tweets per minute. Walk me through data ingestion, aggregation, and how you would handle the thundering-herd problem when a major event causes a sudden spike.

Sample answer outline

A strong answer sketches a Kafka ingestion layer, stream processors (Flink or Spark Streaming) using a Count-Min Sketch or a sliding-window heavy-hitter algorithm to approximate top-K, and a periodic snapshot to Redis Sorted Sets. The candidate should address locality (trending should be geo-scoped), deduplication of retweets, and how a load-shedding strategy prevents the aggregation tier from collapsing under a viral event.

Expect these follow-ups

• How would you distinguish a legitimate trending topic from a coordinated spam campaign artificially inflating a hashtag?
• If your stream processor falls 2 minutes behind, how does that affect the trending output and what would you do?

As asked

X's 'Who to Follow' feature suggests accounts to users. Describe how you would build a recommendation system that suggests users who share second-degree connections, common interests (inferred from tweets), or topic overlap. How would you scale this to run for all 300 million daily active users?

Sample answer outline

A strong answer covers graph traversal (BFS up to 2 hops in a sharded graph store like FlockDB or a custom adjacency list in Cassandra), collaborative filtering on follow patterns, topic modeling on tweet text (LDA or embedding similarity), and a candidate generation plus scoring pipeline that runs nightly for most users and near-real-time for new accounts. The candidate should quantify the candidate set size before and after ranking and explain why pre-computation is necessary at this scale.

Expect these follow-ups

• How do you prevent the recommendation engine from creating echo chambers by only recommending accounts the user is already similar to?
• A brand-new user has zero follows and zero tweets. How do you cold-start their recommendations?

As asked

X gets tens of thousands of new account registrations per hour, many of which are bots or spam accounts. Design a system that scores each new account within 30 seconds of registration and automatically restricts accounts above a threshold without requiring human review.

Sample answer outline

A strong answer covers signals available at registration (email domain, phone number carrier, IP reputation, device fingerprint, registration speed, username pattern), a lightweight online classifier (logistic regression or a small neural net) running in real time on an event stream, a staged action system (shadow-ban before hard-ban for FP reduction), and a human review queue for borderline cases. The candidate should address adversarial adaptation (bots evolve to defeat signals) and how to retrain the model without introducing label bias from your own shadow-banning.

Expect these follow-ups

• Your spam classifier has a 5% false positive rate. At 50,000 registrations per hour, how many legitimate users are incorrectly restricted per day and how do you build the appeal flow?
• A sophisticated bot farm uses residential proxies and real phone numbers. Which signals do you still have and which are you now blind to?