Data scientist domain knowledge interview questions

As asked

OpenAI publishes a significant amount of technical work: GPT papers, the o-series reasoning models, InstructGPT, RLHF, and alignment research. How closely do you follow that research, and can you walk me through one paper or technique from the past year that changed how you think about building AI systems?

Sample answer outline

The candidate should name a specific paper or technique, not a vague area. A strong answer explains the key contribution in plain terms, then connects it to a concrete implication for how they would build or evaluate systems. This is not about reciting papers but showing genuine intellectual engagement with the field.

Expect these follow-ups

• What is a claim in that paper you are not fully convinced by?
• How would you evaluate whether that technique actually improves a real product metric?

As asked

A customer wants to share a live dataset with a business partner who also uses Snowflake, without copying any data or setting up ETL pipelines. Explain exactly how Snowflake's secure data sharing works and what the provider and consumer need to do.

Sample answer outline

Secure Data Sharing works by the provider creating a Share object, granting privileges on specific databases, schemas, or tables to the Share, and then adding the consumer's Snowflake account to the Share. The consumer creates a database from the Share and can immediately query the data with their own warehouse, reading the provider's live S3 micro-partitions directly with no data copied. A strong answer mentions that the consumer pays for compute but the provider pays for storage, and discusses Reader Accounts for partners who do not have a Snowflake account.

Expect these follow-ups

• What are the security implications if the provider accidentally grants access to a table with PII?
• How does the Snowflake Marketplace differ from direct account-to-account sharing?

As asked

Your team wants to run sentiment analysis on millions of customer support tickets stored in a Snowflake table without moving data to an external service. How would you use Snowflake Cortex LLM functions to do this at scale, and what are the cost and latency tradeoffs compared to calling an external model API from a Snowpark job?

Sample answer outline

Snowflake Cortex provides SQL functions like SNOWFLAKE.CORTEX.SENTIMENT() and SNOWFLAKE.CORTEX.COMPLETE() that run hosted LLMs (Llama, Mistral, etc.) natively in-warehouse with no data leaving Snowflake. The tradeoff versus external APIs: Cortex keeps data governance intact and billing is in Snowflake credits per token, but model choice is limited and latency is higher than a dedicated GPU endpoint. An external call from Snowpark gives more model flexibility but requires network egress, API key management, and external function latency. At millions of rows, Cortex is often simpler operationally.

Expect these follow-ups

• How would you use Cortex Search to build a semantic search feature over internal documents?
• What governance controls would you put in place before running Cortex on data that may contain PII?

As asked

Your Snowflake bill has grown from $50K to $400K per month in one year and your CFO wants a plan to reduce it. Walk me through how you would analyze where the credits are going and what levers you would pull to reduce cost without hurting performance.

Sample answer outline

Start with ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY to find which warehouses consume the most credits, and QUERY_HISTORY to find the top credit-consuming queries. Common levers: right-size warehouses (many teams default to LARGE or X-LARGE when SMALL would do), enable auto-suspend at 60 seconds or less, check for always-on warehouses with no traffic overnight, identify long-running queries that could be optimized, audit automatic clustering costs on tables that do not need it, and check storage costs from excessive Time Travel retention settings. Finally, implement resource monitors with hard credit caps per warehouse.

Expect these follow-ups

• How would you implement a chargeback model so each team sees and is accountable for their own Snowflake spend?
• If the main driver of cost is a single ETL pipeline running every hour, what are the options?

As asked

A client's executive asks you to explain how IBM's AI Fairness 360 toolkit detects bias in a machine learning model, without using jargon. They also want to know what they should do if bias is detected. How do you explain this clearly and what practical steps do you recommend?

Sample answer outline

A strong answer uses an analogy (the model gives different outcomes for people who are otherwise similar, except for a protected characteristic like age or gender), explains that IBM AI Fairness 360 measures statistical fairness metrics like demographic parity or equal opportunity, and describes remediation at three stages: pre-processing the training data, in-processing by adjusting the model objective, or post-processing the predictions. It should be honest that no single fairness definition satisfies all criteria simultaneously.

Expect these follow-ups

• How do you decide which fairness metric is most appropriate for a credit-scoring use case?
• What do you do if making the model fairer reduces its accuracy?

As asked

You are setting up Snowflake for a 500-person company with engineering, analytics, finance, and executive roles, each with different data access needs. Walk me through how you would design the role hierarchy using Snowflake's RBAC model, what functional roles versus access roles means in practice, and how you would automate provisioning.

Sample answer outline

Snowflake recommends separating functional roles (ANALYST, ENGINEER, FINANCE_READER) that reflect job functions from access roles (DB_PROD_READ, SCHEMA_PII_WRITE) that grant specific object privileges. Functional roles are granted to users; access roles are granted to functional roles. This avoids privilege sprawl. Use SYSADMIN for object creation, SECURITYADMIN for role and grant management, and never use ACCOUNTADMIN for daily work. Automation via Terraform (the Snowflake provider) or a custom provisioning service that triggers on HRIS events is the right answer for scale.

Expect these follow-ups

• How would you audit who has access to a specific sensitive table right now?
• A user changes teams and their old role should be revoked. How do you ensure that happens reliably?

As asked

A team has 2 PB of Parquet files on S3 that they do not want to load into Snowflake storage. They also need to share this data with both Snowflake and Spark users. What are the options, what are the tradeoffs, and when would you recommend Snowflake-managed Iceberg tables over a standard external table?

Sample answer outline

Standard External Tables point Snowflake at S3 files and create a read-only table over them. They are slow because metadata must be refreshed (AUTO_REFRESH via SQS or manual), do not support DML, and do not prune well without partitioning. Snowflake-managed Iceberg tables write Iceberg metadata alongside the data in S3, enabling both Snowflake and Spark to read the same files with proper schema evolution, time travel via Iceberg snapshots, and partition pruning. The tradeoff is that Snowflake-managed Iceberg tables require Snowflake to be the writer. If external writers (Spark, Flink) must also write, use an Iceberg catalog (Glue, Nessie) and configure Snowflake as a reader via an external Iceberg table.

Expect these follow-ups

• How does Snowflake's Time Travel interact with Iceberg table snapshots?
• If the S3 files are in Avro format instead of Parquet, what changes?