Data platform engineer databases interview questions

As asked

Users update their profile and immediately see old data because reads go to replicas. How would you fix this?

Sample answer outline

First confirm the symptom is replication lag by measuring primary commit time, replica replay lag, and the request path. Common fixes include read-your-writes routing to the primary for a short window after a write, session stickiness, or using a version token that waits until a replica has caught up. Do not route all reads permanently to the primary unless the capacity tradeoff is acceptable. Add lag-aware load balancing and alerts so unhealthy replicas stop serving freshness-sensitive reads. Strong answers define which reads require consistency and which can tolerate stale data.

Expect these follow-ups

• How would you expose read consistency requirements in an application API?
• What happens if replica lag is caused by one expensive analytical query?
• How does this differ in eventually consistent NoSQL stores?

As asked

You are standardising a data lake used by batch analytics, streaming ingestion, and ML training. How would you choose between Iceberg, Delta, and plain Parquet?

Sample answer outline

Plain Parquet is a file format, not a full table-management layer, so it leaves you to solve schema evolution, atomic commits, deletes, and snapshot isolation yourself. Iceberg and Delta add table metadata, partition evolution, time travel, and safer concurrent writes. The right choice depends on engine support, governance needs, cloud constraints, and whether the organisation can operate the catalogue cleanly. Discuss migration path and interoperability instead of treating the choice as a benchmark contest. A senior answer also covers compaction, small files, retention, and rollback.

Expect these follow-ups

• How do small files hurt query and training performance?
• What does partition evolution buy you?
• How would you migrate existing Parquet data without breaking consumers?