← EY Interview Insights

EY·Data Scientist·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
May 2026

Summary

EY data scientist interview that went deep into data modeling territory. The whole session was basically one long architecture question broken into parts, which I wasn't fully expecting going in.

Questions Asked (5)

Q1

Design a logical data model for Accounts, Trades, Positions, Limits, and Customers. Walk through the entity relationships, primary/foreign keys, and cardinalities at an ERD level.

Data ModelingSystem DesignTechnical Trade-offs
Author's notes

I started drawing out the entities and immediately second-guessed whether Positions should be a derived entity or a first-class table.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business context and key requirements (e.g., trading frequency, limit types) to scope the model. Then, walk through each entity, its attributes, and relationships, using an ERD to visualize primary/foreign keys and cardinalities. Conclude by discussing trade-offs and how the model supports analytical needs.

Pro tip: Emphasize that the model should balance normalization for data integrity with denormalization for query performance, especially for time-series data like trades and positions. Mention that in financial services, auditability and temporal tracking are often critical, so consider adding valid-from/valid-to dates or versioning.

1. Clarify Requirements and Scope

Ask about the business domain: Is this for a bank, hedge fund, or retail brokerage? What are the key use cases (risk, reporting, trading)? This determines the level of detail and whether to include historical tracking.

2. Identify Core Entities and Attributes

List each entity (Customer, Account, Trade, Position, Limit) and its key attributes. For example, Customer has customer_id, name, type; Trade has trade_id, trade_date, quantity, price, side.

3. Define Relationships and Cardinalities

Determine how entities relate: A Customer can have multiple Accounts (1:N). An Account can have multiple Trades (1:N). A Trade results in a Position update (N:1 to Position? Actually, Position is derived from Trades, so it's a many-to-one from Trade to Position if Position is per instrument per account). Limits are associated with Accounts or Customers (1:N).

4. Specify Primary and Foreign Keys

Assign primary keys (e.g., customer_id, account_id, trade_id) and foreign keys (e.g., account.customer_id references customer.customer_id). Consider composite keys for junction tables if needed (e.g., for many-to-many between Limits and Accounts).

5. Discuss Trade-offs and Extensions

Talk about normalization vs. performance, handling historical data (slowly changing dimensions), and potential extensions like adding a Product or Instrument entity. Mention how the model supports analytical queries (e.g., aggregating trades to positions).

Key Points to Mention

  • Cardinalities: Customer 1:N Account, Account 1:N Trade, Account 1:N Position (or 1:1 if Position is current snapshot), Account 1:N Limit, Customer 1:N Limit (if limits apply at customer level).
  • Primary keys: Use surrogate keys (e.g., auto-increment) for simplicity, but natural keys like account_number may be used with caution.
  • Foreign keys: Ensure referential integrity; e.g., trade.account_id references account.account_id.
  • Temporal aspects: Trades are immutable events; Positions are derived and may need effective dating. Limits may have validity periods.
  • Normalization: Aim for 3NF to reduce redundancy, but consider denormalizing for performance in analytical workloads.
  • Trade-offs: Discuss whether to model Position as a separate entity or compute on-the-fly; consider partitioning for large trade tables.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

What normalization vs. denormalization tradeoffs would you make in this model, and why?

Data ModelingTechnical Trade-offs
Author's notes

Pretty comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business context and query patterns, then evaluate normalization vs. denormalization tradeoffs based on performance, storage, and maintainability. Recommend a balanced approach, such as a star schema with selective denormalization for critical analytics, and justify with concrete examples.

Pro tip: Always tie your recommendation to the specific use case and SLAs—interviewers at EY value client-focused pragmatism over dogmatic adherence to one approach. Mention that denormalization can reduce join complexity but may increase data redundancy and update anomalies, so it's a tradeoff.

1. Clarify Business Requirements

Ask about the intended use cases, query patterns, data volume, and performance expectations to ground your analysis.

2. Assess Normalization Benefits

Highlight how normalization reduces redundancy, ensures data integrity, and simplifies updates, which is crucial for transactional systems.

3. Evaluate Denormalization Benefits

Explain how denormalization improves read performance, simplifies queries, and is beneficial for analytical workloads with complex joins.

4. Propose a Hybrid Approach

Recommend a design that balances both, such as a star schema with dimension denormalization, and justify based on the tradeoffs.

5. Discuss Implementation Considerations

Mention maintenance overhead, data refresh strategies, and how to monitor performance to validate the chosen approach.

Key Points to Mention

  • Normalization reduces data redundancy and improves integrity but can lead to complex joins and slower reads.
  • Denormalization speeds up read queries and simplifies reporting but increases storage and risk of inconsistencies.
  • The choice depends on workload: OLTP favors normalization, OLAP often benefits from denormalization.
  • Star schema is a common denormalized model for analytics, while snowflake schema is more normalized.
  • Consider data volume, query complexity, and update frequency when deciding.
  • Document tradeoffs and involve stakeholders to align with business goals.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

How would you handle slowly changing dimensions for the Customers entity?

Data ModelingTechnical Trade-offs
Author's notes

SCD type 2 was my immediate answer and I walked through adding effective date and expiry date columns plus a current flag.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining slowly changing dimensions (SCDs) and their importance in data warehousing. Then, compare SCD types (1, 2, 3, 4, 6) with a focus on type 2 as the most common for customer data, and discuss trade-offs like storage, query complexity, and historical accuracy. Finally, recommend a specific approach based on business needs, such as using type 2 for tracking customer attribute changes over time.

Pro tip: Emphasize that the choice of SCD type should align with business requirements and analytics use cases, not just technical preference. Mention that in practice, a hybrid approach (e.g., type 6) often balances history and performance.

1. Define SCD and its relevance

Explain what slowly changing dimensions are and why they matter for maintaining historical accuracy in a data warehouse, especially for customer data.

2. Compare SCD types

Briefly describe SCD types 1, 2, 3, 4, and 6, highlighting their pros and cons in terms of history preservation, storage, and query complexity.

3. Recommend an approach

Propose a specific SCD type (e.g., type 2) for the Customers entity, justifying it with business needs like tracking customer attribute changes over time.

4. Discuss implementation and trade-offs

Outline how to implement the chosen SCD type, including ETL processes, surrogate keys, and handling late-arriving data, and discuss trade-offs such as increased storage and query complexity.

5. Align with business goals

Conclude by emphasizing that the choice should be driven by business requirements, such as regulatory compliance or customer analytics, and mention potential hybrid approaches.

Key Points to Mention

  • SCD Type 1: Overwrites old data, no history, simple but loses past information.
  • SCD Type 2: Adds new rows with effective dates and current flag, preserves full history, commonly used for customers.
  • SCD Type 3: Adds new columns for previous values, limited history, useful for specific comparisons.
  • SCD Type 4: Uses separate history table, keeps current dimension lean, but adds complexity.
  • SCD Type 6: Hybrid of types 1, 2, and 3, provides both current and historical views with a current flag and previous value columns.
  • Trade-offs: Storage cost, ETL complexity, query performance, and impact on downstream reports.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

How would you expose curated datasets for analytics consumers, and what are the tradeoffs between a star schema approach versus a data lakehouse architecture?

Data ModelingTechnical Trade-offsSystem Design
Author's notes

Star schema vs lakehouse is a debate I've had at actual jobs so this felt more natural.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business context and consumer needs, then describe how you would expose curated datasets through a semantic layer or governed views. Compare star schema and data lakehouse across dimensions like performance, flexibility, cost, and governance, and conclude with a recommendation tailored to the scenario.

Pro tip: Emphasize that the choice depends on use cases: star schema for BI and SQL analytics, lakehouse for ML and diverse data. Mention that a hybrid approach (lakehouse with star schemas on top) is often pragmatic in consulting.

1. Clarify Requirements

Ask about the consumers (BI analysts, data scientists), data volume, variety, latency, and governance needs. This ensures your answer is context-driven.

2. Exposure Strategy

Describe how you would expose curated datasets: e.g., through a semantic layer, governed views, or APIs. Highlight the importance of documentation, access control, and discoverability.

3. Compare Architectures

Contrast star schema and data lakehouse on key tradeoffs: performance, flexibility, cost, scalability, governance, and support for advanced analytics.

4. Recommendation

Provide a recommendation based on the context, possibly suggesting a hybrid approach where the lakehouse stores raw and curated data, and star schemas are built for BI.

Key Points to Mention

  • Star schema: optimized for BI, fast queries, simple for business users, but rigid and costly to scale for diverse data.
  • Data lakehouse: combines data lake flexibility with warehouse performance, supports ML and diverse data, but can be complex to govern and optimize.
  • Tradeoffs: performance vs. flexibility, cost, governance, and time-to-insight.
  • Exposure methods: semantic layer, views, APIs, and data catalogs for discoverability.
  • Governance: access control, data quality, and lineage are critical in both approaches.
  • Hybrid approach: use lakehouse for storage and processing, and star schemas for BI consumption.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

Describe the read path you'd design for three different consumers: business dashboards, regulatory data extracts, and ad-hoc analyst notebooks.

System DesignData ModelingProduct Analytics & Metrics
Author's notes

Structured this as three separate paths off a shared curated layer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the distinct requirements of each consumer—latency, data freshness, query complexity, and governance—then propose a tailored read path for each, such as pre-aggregated tables for dashboards, immutable snapshots for regulatory extracts, and a flexible query layer for analyst notebooks. Emphasize how these paths can coexist on a shared data platform without compromising performance or compliance.

Pro tip: Highlight the importance of data contracts and SLAs for each consumer to manage expectations and ensure reliability, and mention how you would monitor and evolve these paths as needs change.

1. Clarify consumer requirements

For each consumer, identify key needs: latency (real-time vs. batch), data freshness, query patterns (aggregations vs. raw scans), concurrency, and governance (audit, lineage, access control).

2. Design the read path for business dashboards

Propose a pre-aggregated, denormalized layer (e.g., star schema or materialized views) optimized for low-latency, high-concurrency queries, with caching and a BI tool-friendly interface.

3. Design the read path for regulatory extracts

Recommend immutable, versioned snapshots with full audit trails, strict access controls, and scheduled batch exports to ensure reproducibility and compliance.

4. Design the read path for ad-hoc analyst notebooks

Suggest a flexible, SQL-based interface over raw or lightly processed data (e.g., data lake or warehouse) with on-demand compute, allowing exploration without impacting other consumers.

5. Address integration and governance

Explain how these paths share a common data foundation (e.g., lakehouse architecture) and how you enforce consistent security, lineage, and monitoring across all consumers.

Key Points to Mention

  • Separation of concerns: different read paths for different latency, concurrency, and governance needs
  • Use of pre-aggregation and caching for dashboards to ensure sub-second response times
  • Immutable snapshots and versioning for regulatory extracts to support audits and reproducibility
  • On-demand compute and flexible schemas for analyst notebooks to enable exploration
  • Unified data platform (e.g., lakehouse) to avoid data silos and reduce duplication
  • Data contracts and SLAs to define expectations and ensure reliability for each consumer

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.