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

SeniorPrefer not to say

May 2026

Summary

TikTok system design round for a software engineering role. The whole thing was one big open-ended question about building a privacy detection pipeline over a massive database, and it went deep fast. Walked away feeling like I covered maybe 60% of what they were actually looking for.

Questions Asked (1)

Design an end-to-end system to automatically detect and classify privacy-sensitive or PII data across a very large database with billions of rows, multiple schemas, and unstructured content like logs and messages. Your design should cover requirements, architecture, detection methods (rule-based, ML, LLM/RAG), scalability, evaluation, feedback loops, and any security concerns in the pipeline itself.

System DesignTechnical Trade-offsData Modeling

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 constraints, then propose a layered architecture that separates data ingestion, detection, classification, and feedback. Emphasize a hybrid detection approach combining rule-based, ML, and LLM/RAG methods, and discuss scalability, evaluation, and security throughout.

Pro tip: Highlight the importance of a human-in-the-loop feedback mechanism to continuously improve detection accuracy and adapt to evolving PII patterns. Also, stress that the pipeline itself must be secure to prevent leaks of sensitive data during processing.

1. Clarify Requirements and Scope

Ask questions to understand data volume, variety, latency needs, compliance requirements, and accuracy targets. Define what constitutes PII and privacy-sensitive data for the organization.

2. Design High-Level Architecture

Outline a scalable pipeline: data ingestion (batch/stream), preprocessing, detection engines (rule-based, ML, LLM/RAG), classification, storage of results, and feedback loop. Consider distributed processing frameworks like Spark or Flink.

3. Detail Detection Methods

Explain how rule-based (regex, dictionaries), ML (NER, classification models), and LLM/RAG (for unstructured text) methods work together. Discuss trade-offs: rules are fast but brittle, ML is adaptable but needs training, LLMs are powerful but costly.

4. Address Scalability and Performance

Describe partitioning, parallelization, sampling strategies, and caching to handle billions of rows. Discuss how to optimize each detection method for scale, e.g., using approximate algorithms or model distillation.

5. Cover Evaluation, Feedback, and Security

Propose metrics (precision, recall, F1), A/B testing, and human review for continuous improvement. Emphasize securing the pipeline with encryption, access controls, and auditing to prevent data leaks.

Key Points to Mention

Hybrid detection approach combining rule-based, ML, and LLM/RAG methods for different data types and accuracy needs.
Scalability considerations: distributed processing, partitioning, sampling, and efficient model inference.
Evaluation metrics and human-in-the-loop feedback for continuous improvement and adaptation to new PII patterns.
Security measures within the pipeline: encryption, access control, auditing, and data anonymization.
Trade-offs between detection methods: cost, latency, accuracy, and maintainability.
Handling unstructured data like logs and messages with NLP and LLM techniques.

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