← EY Interview Insights

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

SeniorPrefer not to say
May 2026

Summary

EY data science interview that was basically one massive system design case study about building a cloud data platform for a retail bank. Dense question, lots of moving parts, felt like I was presenting a consulting engagement proposal rather than answering an interview question.

Questions Asked (5)

Q1

Design a cloud data platform for a mid-tier retail bank operating under PII and data-residency constraints. Walk through the full architecture: ingestion, storage, governance, transformation, and consumption layers.

System DesignData ModelingTechnical Trade-offs
Author's notes

I started with the residency constraints because ignoring those upfront would've made the whole architecture wrong.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the regulatory constraints (PII handling, data residency) and the bank's key data use cases (e.g., risk modeling, customer analytics). Then propose a layered architecture that enforces governance and security at every stage, emphasizing trade-offs between flexibility and control. Conclude by discussing how the design supports data science workflows while meeting compliance requirements.

Pro tip: Demonstrate awareness of data residency by proposing region-specific storage and processing, and mention the importance of data lineage and audit trails for regulatory audits. Also, highlight the need for a data catalog and metadata management to enable self-service analytics without compromising governance.

1. Clarify Requirements and Constraints

Ask questions to understand the bank's regulatory environment (e.g., GDPR, local banking laws), data types, volume, and key use cases. Identify which data is PII and where it must reside.

2. Design Ingestion Layer

Propose batch and streaming ingestion with data classification at source. Use tools like Kafka for streaming and AWS Glue or Azure Data Factory for batch, ensuring data is tagged and encrypted in transit.

3. Design Storage and Governance Layers

Recommend a data lakehouse (e.g., Databricks Delta Lake) with region-specific buckets, encryption at rest, and fine-grained access control. Implement a data catalog (e.g., Collibra) and policy engine for PII masking and residency enforcement.

4. Design Transformation and Consumption Layers

Use a medallion architecture (bronze, silver, gold) for transformation, with data quality checks and lineage tracking. Enable consumption via BI tools, SQL endpoints, and ML platforms, ensuring all access is governed and audited.

5. Address Trade-offs and Operational Concerns

Discuss trade-offs between centralized vs. federated governance, cost vs. performance, and flexibility vs. control. Mention monitoring, alerting, and disaster recovery for compliance.

Key Points to Mention

  • Data residency: store and process data in specific geographic regions, using region-locked cloud services.
  • PII protection: encryption, tokenization, dynamic data masking, and role-based access control.
  • Governance: data catalog, lineage, audit logs, and policy enforcement for compliance.
  • Scalable storage: data lakehouse with ACID transactions and schema enforcement.
  • Transformation: medallion architecture with data quality and validation.
  • Consumption: secure access for data scientists and analysts via governed APIs and notebooks.

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

Q2

How would you define and enforce data contracts and lineage tracking across this platform?

Data ModelingSystem DesignTechnical Trade-offs
Author's notes

Honestly not my strongest answer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining data contracts as formal agreements between data producers and consumers, specifying schema, semantics, quality, and SLAs. Then explain how to enforce them through automated validation at ingestion and transformation stages, and how to track lineage to trace data flow and impact. Emphasize the balance between strict enforcement and flexibility, and how this enables trust and compliance in a consulting environment like EY.

Pro tip: Highlight that data contracts should be versioned and treated as code, and that lineage tracking should be automated to reduce manual overhead. Mention that in consulting, demonstrating quick wins with a pilot on a critical data domain can drive adoption.

1. Define Data Contracts

Collaborate with stakeholders to specify schema, data types, quality rules, SLAs, and ownership for each data product. Document these contracts in a central repository, version them, and treat them as code.

2. Enforce Contracts

Implement automated validation at data ingestion and transformation points using tools like Great Expectations or custom checks. Integrate with CI/CD pipelines to fail builds on contract violations and alert producers.

3. Implement Lineage Tracking

Use metadata management tools (e.g., Apache Atlas, DataHub) to automatically capture lineage from source to consumption. Ensure lineage is granular (column-level) and integrated with the data catalog.

4. Monitor and Iterate

Set up dashboards to monitor contract compliance and lineage completeness. Use feedback loops to refine contracts and address gaps, and conduct regular audits to ensure adherence.

5. Govern and Scale

Establish a governance council to oversee contracts and lineage, and scale the approach across domains. Promote a data culture where contracts are mandatory for new data products.

Key Points to Mention

  • Schema evolution and backward compatibility
  • Data quality metrics and SLAs
  • Automated testing and CI/CD integration
  • Column-level lineage and impact analysis
  • Metadata management and data catalog
  • Governance and ownership models

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

Q3

What SLAs and SLOs would you set for this platform, and how would you handle partitioning, backfill scenarios, and late-arriving data?

System DesignProduct Analytics & MetricsTechnical Trade-offs
Author's notes

This is where I felt more comfortable.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the platform's purpose and data consumers, then propose SLAs/SLOs that balance business needs with technical feasibility. For partitioning, backfill, and late-arriving data, discuss trade-offs between cost, latency, and accuracy, and suggest practical solutions like tiered storage and idempotent pipelines.

Pro tip: Tie SLAs/SLOs to business impact (e.g., revenue, customer trust) and propose monitoring/alerting to ensure adherence. Show you understand that perfect data is impossible; focus on graceful degradation and clear communication with stakeholders.

1. Clarify Requirements and Constraints

Ask about the platform's users, data volume, velocity, and criticality. Understand what 'freshness' and 'accuracy' mean for different consumers (e.g., real-time dashboards vs. monthly reports).

2. Define SLAs and SLOs

Propose measurable SLAs (e.g., 99.9% uptime, <5 min data latency) and SLOs (e.g., 95% of queries <2s). Align them with business goals and cost constraints, and mention error budgets.

3. Design Partitioning Strategy

Recommend partitioning by time (e.g., daily) and possibly by key (e.g., customer ID) to optimize query performance and manage data lifecycle. Discuss trade-offs like partition size and maintenance overhead.

4. Handle Backfill Scenarios

Outline a backfill process: use idempotent writes, versioned data, and separate backfill pipelines to avoid impacting live traffic. Emphasize validation and reconciliation.

5. Manage Late-Arriving Data

Propose strategies like watermarks, allowed lateness windows, and reprocessing. Discuss how to handle updates (e.g., upserts) and maintain data consistency across layers.

Key Points to Mention

  • SLAs vs SLOs: SLAs are contractual, SLOs are internal targets; use error budgets to balance reliability and innovation.
  • Partitioning: time-based partitioning for efficient pruning; consider composite keys for even distribution.
  • Backfill: idempotent pipelines, versioning, and isolated resources to prevent disruption.
  • Late-arriving data: watermarks, allowed lateness, and reprocessing with upserts.
  • Monitoring and alerting: track SLO compliance, data freshness, and quality metrics.
  • Trade-offs: cost vs. latency vs. accuracy; choose based on business impact.

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

Q4

What cost guardrails would you put in place, and how would you structure a three-wave rollout with measurable success criteria?

Roadmap PrioritizationProduct StrategyCross-functional Alignment
Author's notes

The rollout structure came naturally: foundational infrastructure and governance first, then core data domains, then self-serve and advanced analytics.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing cost guardrails as a governance mechanism that balances innovation with financial discipline, then outline a three-wave rollout that progressively scales from pilot to full deployment, each with clear, quantifiable success criteria. Emphasize cross-functional alignment and iterative learning to ensure buy-in and adaptability.

Pro tip: Tie every guardrail and success metric to a business KPI (e.g., ROI, cost per prediction, adoption rate) to show you understand how data science drives value, not just model performance.

1. Define Cost Guardrails

Establish budget caps, resource utilization thresholds, and cost-per-unit metrics (e.g., cost per model training, inference cost) to prevent overspend. Include automated alerts and a review process for exceptions.

2. Design Three-Wave Rollout

Wave 1: Pilot with a small, controlled scope to validate assumptions and refine guardrails. Wave 2: Scale to a broader user base or use case, optimizing costs and performance. Wave 3: Full deployment with continuous monitoring and optimization.

3. Set Measurable Success Criteria

For each wave, define specific, quantifiable metrics (e.g., cost reduction %, model accuracy, adoption rate, time-to-insight) with target thresholds. Ensure criteria are tied to business outcomes and reviewed at gate points.

4. Ensure Cross-Functional Alignment

Engage finance, engineering, product, and business stakeholders to agree on guardrails and success metrics. Establish a steering committee or regular check-ins to maintain alignment and address issues.

5. Iterate and Communicate

Use learnings from each wave to adjust guardrails and rollout plans. Communicate progress and results transparently to maintain trust and secure ongoing support.

Key Points to Mention

  • Cost guardrails: budget caps, cost-per-prediction, resource quotas, automated monitoring
  • Three-wave rollout: pilot, scale, optimize—each with clear entry/exit criteria
  • Measurable success criteria: ROI, cost savings, model performance, adoption rate, time-to-value
  • Cross-functional collaboration: finance, engineering, product, business stakeholders
  • Iterative feedback loops and gate reviews to adjust based on data
  • Alignment with business KPIs and EY's client-focused, value-driven approach

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

Q5

What is the one capability you would deliberately defer in this rollout, and what's your reasoning?

Roadmap PrioritizationAdaptability & AmbiguityTechnical Trade-offs
Author's notes

Said real-time streaming ingestion.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Choose a capability that is valuable but not critical for the initial rollout, and justify deferring it by weighing impact, effort, and dependencies. Show that you prioritize delivering core value early while managing technical debt and stakeholder expectations.

Pro tip: Tie your deferral to a concrete trigger or milestone (e.g., 'once we have X data volume or Y user feedback') to show you're not just dropping it, but sequencing it deliberately.

1. Clarify rollout goals and constraints

Restate the primary objective of the rollout (e.g., MVP, pilot) and key constraints like time, budget, and data availability to frame your decision.

2. Identify candidate capabilities

List 2-3 capabilities under consideration (e.g., real-time scoring, advanced model interpretability, automated retraining) and briefly note their value and cost.

3. Select one to defer with rationale

Pick the capability with the lowest immediate impact-to-effort ratio or highest dependency risk, and explain why it's safe to postpone without compromising the rollout's core value.

4. Outline a deferral plan

Describe how and when you would revisit the capability, including any interim workarounds or monitoring to ensure it doesn't become a blocker.

5. Connect to business and technical trade-offs

Summarize how this decision balances short-term delivery with long-term scalability, and how you'd communicate it to stakeholders.

Key Points to Mention

  • Alignment with business priorities and MVP definition
  • Impact vs. effort analysis (e.g., RICE, MoSCoW)
  • Dependencies and technical debt considerations
  • Stakeholder communication and expectation management
  • Risk mitigation and fallback plans
  • Clear criteria for when to revisit the deferred capability

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