How would you implement automated data validation before training?

WHAT THIS TESTS: This question evaluates whether you treat data as a production dependency that requires automated contracts and circuit breakers. Interviewers want to see that you understand the difference between exploratory data cleaning and production-grade validation, and that you know how to prevent bad data from silently corrupting a model.

A GOOD ANSWER COVERS: First, schema validation. Define a strict contract for types, nullability, and feature ranges using a tool like TensorFlow Data Validation or Great Expectations. Second, statistical validation. Compare incoming distributions against a reference baseline to catch drift, skew, or anomalous volumes. Third, pipeline integration. Run these checks as a discrete step after ingestion but before training, so the training task cannot start until validation passes. Fourth, failure handling. On any violation, the pipeline should fail fast, write the rejected batch to a quarantine storage location, emit structured logs, and page or email the on-call with actionable diagnostics like which constraint failed and the severity.

COMMON WRONG ANSWERS: A red flag is suggesting a human-in-the-loop approval for every batch, which breaks automation at scale. Another is proposing to log warnings but continue training, which lets bad data poison the model. Some candidates focus only on schema and forget statistical checks, or they suggest fixing data inline instead of isolating the root cause upstream.

LIKELY FOLLOW-UPS: The interviewer might ask how you handle schema evolution without breaking old pipelines, how you version the validation rules alongside code, or how you distinguish between a critical blocking error and a non-blocking anomaly that should still be tracked.

ONE CONCRETE EXAMPLE: Suppose a new partner sends daily CSVs. You generate an initial schema from the first month of data and commit it to version control. The pipeline step runs TFDV against each new file. One morning the step detects a forty percent drop in row count and three new categorical values in the country column. Validation fails, the training DAG halts, the file moves to gs://project-quarantine/20240115/, and a PagerDuty alert fires with a link to the failing Dataflow job and the specific anomaly list.

Source: Google Cloud Architecture Center, MLOps: Continuous delivery and automation pipelines in machine learning