← Google Interview Insights

Google·Software Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
Jul 2026

Summary

Google system design round for a software engineer role. The whole question was basically one giant prompt about building an AI-powered anomaly detection service for OS snapshots, with a lot of moving pieces to cover in not enough time.

Questions Asked (1)

Q1

Design an AI-based anomaly detection service that ingests OS snapshots, classifies machines as normal or abnormal, triggers automated actions like shutdown or quarantine, sends warning emails, and lets clients query warning history. Cover data collection, model serving, decisioning, action orchestration, audit storage, access controls, and failure rollback.

System DesignTechnical Trade-offsAPI & Integrations
Author's notes

This is a beast of a question.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying requirements and scale, then design a high-level architecture that covers data ingestion, model serving, decisioning, action orchestration, and audit storage. Emphasize reliability, security, and failure handling, and discuss trade-offs between latency, accuracy, and cost.

Pro tip: Focus on the end-to-end flow and failure modes rather than deep ML details; Google values system thinking and operational excellence. Propose a phased rollout with human-in-the-loop for high-risk actions to build trust and safety.

1. Clarify Requirements and Scale

Ask about the number of machines, snapshot frequency, data size, latency requirements, and action criticality. Clarify what 'abnormal' means and the cost of false positives/negatives.

2. Design Data Collection and Processing

Outline how OS snapshots are collected (agent-based), transported (message queue), and processed (feature extraction, normalization). Consider batch vs. streaming and storage for raw and processed data.

3. Model Serving and Decisioning

Describe how models are trained, deployed, and served (e.g., TensorFlow Serving, Vertex AI). Explain how predictions are turned into decisions using thresholds, rules, or ensembles, and how to handle model updates.

4. Action Orchestration and Audit

Design a workflow engine to trigger actions (shutdown, quarantine, email) with idempotency, retries, and rollback. Store all decisions and actions in an audit log for querying and compliance.

5. Security, Access Control, and Failure Handling

Implement authentication/authorization (e.g., OAuth, IAM), encryption, and audit trails. Discuss failure scenarios (model errors, action failures) and mitigation (circuit breakers, dead-letter queues, manual approval).

Key Points to Mention

  • Scalable ingestion pipeline with message queue (e.g., Kafka, Pub/Sub) and stream processing (e.g., Dataflow, Flink).
  • Model serving infrastructure with versioning, A/B testing, and monitoring for drift.
  • Decision engine with configurable thresholds and human-in-the-loop for high-risk actions.
  • Action orchestration using workflow engine (e.g., Temporal, Airflow) with idempotency and rollback.
  • Audit storage with immutable logs and query API for warning history (e.g., BigQuery, Elasticsearch).
  • Access control using IAM, service accounts, and encryption at rest/in transit.

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